NEWS

rxode2 5.1.8

New features

rxode2() now uses jump event sensitivities by default for any model with sensitivities; before, they were used only when eventSens was given. Use eventSens = "fd" for the old behavior. calcJac is still forced on only when eventSens is given.
New event bus (rxEventListen(), rxEventEmit(), rxEventScope() and friends) so packages such as nlmixr2log can be told when a top-level simulation, fit or result is finished. Only the outermost operation emits: work inside rxEventScope() is silent, and the scope depth is inherited by worker processes started inside it. rxSolve() now emits solveComplete (never for rxControl(), nested solves or errors). With no listener registered, emitting costs a single length check.
New linCmtMicro() returns the micro-constant parameterization (k, k12, k21, k13, k31, v and ka) of each linCmt() call in a model, as R expressions of the model variables. It lets translators (like babelmixr2) write a linCmt() model with another program's closed-form linear compartment solutions.
New assertions assertRxUiTransform(), assertRxUiErrType() and assertRxUiAddProp() refuse a model whose residual transformation (like boxCox() or lnorm()), residual error type (like add() + pow()) or add() + prop()/add() + pow() combination (combined1 or combined2, including the rxode2.addProp default) is not in the allowed set; endpoints without a residual error (like pois() or ll()) are not checked. They let an estimation method that supports only some residual error models stop with a clear error instead of fitting a different one.
New assertions assertRxUiNoFixedResiduals() and assertRxUiNoFixedOmega() refuse a model that fixes a residual error parameter (add.sd <- fix(0.7)) or a between-subject variability (eta.ka ~ fix(0.6)), for estimation methods that would otherwise estimate them anyway.
rxPriorBuildSpec() and rxPriorLogDensity() now evaluate lotri's univariate continuous priors beyond normal and Cauchy (dlnorm(), dgamma(), dbeta(), studentT(), dexp(), dunif(), dweibull(), dlogis(), dchisq(), invGamma() and the rest of the catalog except wiener()), with the log density and its gradient, truncated to the parameter's own bounds (#1387).
The Stan-based linCmt() kernels and their gradients, .solComp2(), .solComp3() and the rxDerived() conversions moved to the new 'rxode2lincmt' package, which rxode2 now imports. rxode2 no longer builds against 'StanHeaders', 'RcppEigen' or 'RcppParallel', which shortens its installation; exported functions and compiled model code are unchanged.
Add splitInfusion(), splitInfusionBolus() and splitBolusInfusion() model directives to split or relocate doses at etTrans() translation time, mirroring splitBolus(). splitInfusion() splits infusion records (both data RATE/DUR and modeled rate()/dur() infusions, including their stop records). splitInfusionBolus() and splitBolusInfusion() split bolus and infusion records; a plain bolus dose is split so the FIRST target (splitInfusionBolus()) or the LAST target (splitBolusInfusion()) receives a modeled infusion start/stop pair (it must declare a modeled dur() or rate() property, otherwise its copies stay boluses) while every other target receives a bolus copy -- so one dose record can feed both an infusion and a bolus path (Monolix-style double absorption with mixed zero-/first-order routes). Unlike splitBolus(), these directives apply at translation time only, not to evid_() doses pushed while solving. Only one splitting directive (splitBolus(), splitInfusion(), splitInfusionBolus() or splitBolusInfusion()) is allowed per model.
The Stan-based linCmt() kernels and their gradients, .solComp2(), .solComp3() and the rxDerived() conversions moved to the new 'rxode2lincmt' package, which rxode2 now imports. rxode2 no longer builds against 'StanHeaders', 'RcppEigen' or 'RcppParallel', which shortens its installation; exported functions and compiled model code are unchanged.
An error-model argument can now be an expression, like cp ~ add(cp.sd * exp(eta.cp.sd)). It becomes a hidden modeled variable assigned before the endpoint (rx.cp.add ~ cp.sd * exp(eta.cp.sd), then cp ~ add(rx.cp.add)), the same way a number like add(3) becomes a fixed parameter. logitNorm() and probitNorm() now take a modeled variable or expression for their standard deviation; their bounds stay numbers.
The adaptive dosing functions (bolus(), infuse(), infuseDur(), replace(), multiply(), phantom(), evid_()) now take any expression, including one whose variables appear nowhere else in the model, like infuseDur(100 * bsa, 1, central) or bolus(DOSE * 30); those variables become model parameters or covariates instead of an "undeclared ... assign first" error.
Building a model ui, and therefore model() piping (which rebuilds it), is faster: 3-5x on large models. User ui functions are looked up once per function name instead of once per call, the per-line model rebuild is no longer quadratic, and mu-referenced covariate derivatives are computed in one symengine environment and cached between builds. The resulting ui is unchanged.
A user function a model function defines in its body, or encloses (like one made by a function that returns the model function), now lives in the model: it is found first, kept in the model's meta and written into ui$fun. Such a function is converted to C, with its derivatives, for that model only; it is never registered globally, so another model does not see it and a global rxFun() of the same name is left alone. A function the C translator cannot handle (or that reads variables it does not define) is called from R instead (#1416).
Building, compiling and solving a model no longer writes rxode2 namespace variables with assignInMyNamespace(); the mutable state lives in an internal environment. Each such write cost milliseconds once other packages (like nlmixr2) registered S3 methods on rxode2 generics, about 60% of an evaluation-only FOCEi fit (#1425).
New rxSolve(..., nonmem = TRUE) evaluates statements that do not depend on a state (like NONMEM's $PK) with time set to the time of the data record that ends the interval being integrated; d/dt() and state-dependent statements keep the continuous time. addl repeats, infusion ends, lagged and run-time doses do not end an interval, while mtime() times do. It also covers linCmt(), matExp() and calcJac/calcSens models, where rxS(pkTime = TRUE) keeps that time apart as rx_time_pk~t when symengine inlines it into d/dt(). With covsInterpolation = "nocb" and addlKeepsCov = FALSE this matches NONMEM 7.4 for TIME in $PK, a covariate changing between ADDL doses and an MTIME change point (#1429).
dnorm() is now supported in model syntax, with 1-3 arguments (dnorm(x, mean, sd)) and symbolic derivatives. In model functions (ini()/model()), dnorm(), pnorm() and qnorm() also take R's arguments by name, including log= for dnorm() and lower.tail=/log.p= for pnorm() and qnorm().

Bug fixes

Event (eventSens = "jump") sensitivities with a modeled alag() now use the exact jump f(x-) - f(x+) at the dose time instead of its linearization J*(x- - x+), matching the fix in cOde. The old jump was wrong whenever the model is nonlinear in the dosed state, for example 3x too large for a Michaelis-Menten model. The second-order lag jumps were also corrected for these models.
Event sensitivities of replace (evid=5) and multiply (evid=6) records now include a modeled f() on the replaced value or factor, and their second-order jumps now include the lag and f() terms. The second-order lag jumps of rxSensMatExp() models now include the dependence of the rate constants on the parameter.
rxode2(..., calcSens = , indLin = TRUE) no longer expands the sensitivities twice; with calcSens2 it used to abort with free(): invalid next size.
calcJac = TRUE with a dotted state or parameter name (e.g. a sensitivity wrt eta.cl) no longer generates C that fails to compile.
method = "lsoda" now reads a model's analytic Jacobian (calcJac = TRUE or df()/dy() lines) untransposed; it used to get the transpose, so stiff sensitivity models could fail with too many steps.
Jump event sensitivities no longer treat the alag()/f() that a delay() model puts on its sensitivity compartments as event compartments; a 2nd-order delay() model with a dose used to abort.
eventSens = "jump" now falls back to "fd" for a model whose alag() depends on the states; the lag is evaluated at the dose's own time, where the jump runtime has no sensitivities.
An x(0) statement in a model no longer moves alag() and the splitBolus()-family directives onto the wrong compartment. A steady-state dose with alag() then gave 0 before the lag time instead of the steady state (#1452).
The parser now zeroes its tables when a model grows past their first block, and tracks the variable and string-variable table sizes separately; a model with more than 5000 compartments or string variables could read uninitialized memory or write past the end of a table.
Comparing a string variable assigned in the model (like tRACE <- "A") with a literal (tRACE == "B") now uses the numbering of the variable's assigned strings. Before, the literals were numbered separately, so the comparison could give the wrong answer unless both appeared in the same order (#1456).
Steady-state dosing in a model using delay() now brings the delay()-driven states to steady state; the steady-state iterations record the delay history and the converged dosing interval becomes the history before the dose (added to the existing history for ss=2). Before, delay() returned the initial condition during and after the steady-state solve (#1447).
rxFixRes() no longer fails with subscript out of bounds on a model with a literal residual value, like a named ordinal Y ~ c(p0 = 0, 1) or cp ~ add(0.5); the literal is already in the model, so its generated fixed parameter is left alone. This crashed nlmixr2 fits of these models (#1450). A fixed residual parameter is now assigned before the first model line that uses it, not just before its endpoint.
rxode2 now requires dparser (>= 1.3.1-14). Earlier dparser versions shared a first reduction path in reduce_one() across threads, which could segfault rxode2's parallel expression optimization (rxOptExpr()) on clang builds (nlmixr2/dparser-R#33; CRAN's clang checks of babelmixr2).
delay(x, tau) before the start of integration now returns the evaluated x(0) when x(0) is set from a parameter or a computed value, instead of 0; only a literal x(0) worked before (#1441).
rxControl() and rxSolve() now reject a zero atol, atolSens, ssAtol or ssAtolSens. A zero absolute tolerance gave a failed or silently wrong solve, since the states start at zero (#1440).
A time-varying covariate resampled with rxSolve(..., resample = TRUE) is now read from the sampled subject at the right time. Functional alag(), rate(), dur() and mtime() read outside the subject's data, and the ODE could see the sampled subject's covariate frozen at one value when that subject had not been solved yet (#1439).
Time-varying covariates other than the first are no longer read from the wrong data after a dose is pushed while solving (bolus(), evid_(), ...). A pushed record interpolates its covariates, and last() of a covariate skips pushed records.
A linearly interpolated covariate past its last non-missing value (a trailing NA, or past a pushed record) is now carried forward instead of becoming NaN.
lag(time) (or lag(t), diff(time), lead(time), ...) no longer aborts R while the model is compiled; it now returns the time of the subject's previous (or next) record. lag() of another reserved name like amt is a syntax error, and lag(x, 0) no longer normalizes to an empty expression (#1434).
When a variable read by lag() is assigned more than once, the symengine form of the model (rxS(), and the models built from it) now computes an lhs that reads it (directly or through diff()) with the value at that point, not the final value (#1435).
A d/dt() written between two assignments of a variable read by lag() now reads the value at that point in the symengine form of the model (rxS()), not the final value, matching how rxSolve() evaluates it (#1445).
A time-varying covariate (including DV when the model refers to it, as in res <- DV - central) no longer corrupts doses with modeled lag times and durations in more than one compartment. Interpolating the covariate looked up neighbouring records and overwrote the compartment of the dose being given, so a later infusion could be sent to the wrong compartment and the amounts go negative (nlmixr2/nonmem2rx#263).
When a lagged first dose sorted after a time-zero record, that record read its time-varying covariates (including the CMT that selects a multiple-endpoint prediction) from the dose record instead of its own, so a time-zero observation could get the wrong endpoint's IPRED (nlmixr2/nlmixr2est#1182). A missing (NA) covariate on the first sorted record is now filled like on any other record instead of staying NA.
The ar() estimation lines (FOCEi and the log-likelihood methods) now take the previous residual and time gap from the previous observation, skipping dose and evid=2 records, so the first observation gets the marginal likelihood again and a dose's (filled) DV no longer enters the AR(1) residual (#1453). The new internal model variable rx__isObs__ is 1 on an observation record and 0 otherwise. rxOptExpr() no longer hoists a lag()-family call into a shared temporary, which could move a self-referencing x <- f(lag(x, 1)) reference above the assignment of x.
qnorm() was documented as the normal pdf; it is the normal quantile function. qnorm(p, mean, sd) with more than one argument can now be translated to symengine (it errored before).
rxFromSE() now keeps the grouping of a substituted derivative that is a sum or starts with a minus, so a derivative like -(x)*dnorm(x) is no longer spliced into a product as --(...) or with the wrong precedence.
rxOptExpr() and the symengine translation no longer segfault with more than one thread on clang builds (CRAN's r-devel clang checks of nlmixr2est). They parse in parallel, but dparser before 1.3.2 shares one vector across concurrent parses; they now parse serially unless the loaded dparser is at least 1.3.2 (#1427).
ui$fun of a model that had been solved printed a stray NULL line in place of the cached simulation model.
$ on an rxUi no longer adds rxode2's own call frames (like rxPriorLogDensity()'s) to the environments searched for user-defined functions; it records the nearest frames outside rxode2 instead. What those frames held (like a large theta) is now freed when they return, instead of staying in memory until later calls pushed them out of the search list.
Parsing a model no longer reads freed memory when a garbage collection happens while the built-in function table is claimed; this showed up as intermittent "INTEGER() can only be applied to a 'integer'" errors or crashes (introduced in 5.1.7).
The per-thread LSODA work-array and context pools are no longer reported by valgrind as definitely lost at R exit (seen in nonmem2rx's checks).
ui$fun of a model that had been solved printed a stray NULL line in place of the cached simulation model.
$ on an rxUi no longer adds rxode2's own call frames (like rxPriorLogDensity()'s) to the environments searched for user-defined functions; it records the nearest frames outside rxode2 instead. What those frames held (like a large theta) is now freed when they return, instead of staying in memory until later calls pushed them out of the search list.
rigeom() and ripois() returned 0 for every subject, because model code called them through a pointer typed to return int while they return double; they now return their draws.
Function pointers between rxode2 and compiled models now match the functions they call (__assignFuns, user-defined functions and simeta()/simeps()), so solving is no longer flagged by UBSAN's -fsanitize=function (#1420).
The model parser now keeps its own copies of the model md5 and matrix-exponential code, which the code generator reads in a later call; it used to hold pointers into R strings that could be garbage collected in between (the model prefix global is copied too, so it never dangles). Generated C symbols are now keyed by the parsed md5 handed to the code generator, and its unused timeId argument was dropped (#1421).
rxFun() of a function whose body has no braces (like function(x) x + 1) translated each part of the expression as its own statement, so the C function returned 1 and its derivative was 0.
A covariate column named value or dose is now passed through when the data has an amt column. Without amt, evid, mdv or method columns it is read as the amt alias and now stops with an informative error instead of turning every record into a dose (#1386).
A column named exactly cmt is now used as the compartment even when the data also carries another compartment spelling (like an unused CMT, YTYPE, state or var column); those used to stop with "can only specify either 'cmt', 'ytype', 'state' or 'var'" (#1410).
linToOde() of a linCmt() ~ ... endpoint named the translated prediction rxLinCmt, which rxode2 reads back as a linCmt() model, so the ODE model failed with "'depot', 'central' are required for linCmt() but defined in ODE too". The prediction is now named rxLinCmtOde.
linToOde() of a model with linCmt() and other ODEs now keeps the compartment numbers of the linCmt() model (depot and central first), so data using compartment numbers dose the same compartments.
linToOde() now translates linCmt() that is part of an expression (like cp <- 1e6 * linCmt(), or inside an ODE); before, the model was returned with linCmt() unchanged.
linToOde() now translates a linCmt(ka, cl, v) ~ ... endpoint with arguments, and writes the ODEs once when several endpoints use the same linCmt() (like linCmt() ~ add(a) | phase1 and linCmt() ~ add(b) | phase2), which before ran past the translated linear compartment calls. The prediction of a linCmt() ~ endpoint gets a name the model does not already use.
A model with a numeric residual error like cp ~ add(3) can now be piped and rebuilt from its own function; the generated fixed parameter was added a second time and the model then failed to parse.
linMod() no longer drops the model's between subject variability: the iniDf it returned kept the thetas twice and no etas, so an eta like eta.cl became a covariate and lost its mu-reference.
Piping a line that calls a ui user function whose parser form differs from what is written (like model(p <- plogis(tp + eta.p))) no longer fails with a syntax error.
With options(rxode2.verbose.pipe = FALSE), promoting a covariate to a parameter with ini() (like ini(covwt = 0.5)) now removes it from $covariates and re-runs the mu-reference analysis, so the model is the same as with the default verbose setting; only the messages depended on it.
Promoting a covariate to a between subject variability with ini() (like ini(etav ~ 0.1)) now re-runs the mu-reference analysis, so the new eta is in $eta and mu-referenced as in a model written with it from the start.
A model echoed after a syntax error that is only found once the whole model is parsed (like an undeclared bolus() argument) no longer drops its first character (/dt(depot) instead of d/dt(depot)).
A model whose ui user function requests the model variables (like linModM()) now restores the ui parsing state after it is built.
A model subtracting a negated term (a - -b) or adding a positive one (a + +b) now compiles; the generated C code read the two signs as the --/++ operator. a - Inf also compiles now instead of generating aR_NegInf (#1399).
A mixest or mixunif data column can now be combined with other covariates; previously any other covariate made etTrans() fail with "mixest is time-varying but must be constant within an individual" or an out-of-bounds error, depending on the column order.
rxSymInvCholCreate() can treat a positive-definite Omega whose off-diagonal zeros do not split it into contiguous blocks as fully parameterized instead of erroring with "theta has to have N elements"; such zeros are not zeros of the Cholesky factor, so they become free parameters (#1365). It is off unless a callback registered with .rxSymInvBlockZeroFreeCallback() asks for it: the released nlmixr2est infers its own omega parameter positions from the zero pattern of the matrix it passes in, so accepting that matrix rather than its filled replacement makes est="vae" stop with a position-count mismatch. It becomes unconditional once a nlmixr2est that does not infer positions that way is released.
A solve with exactly one observation now simulates its residual error; previously the one-row draw was discarded and each eps came back as a fixed element of the sigma covariance matrix, regardless of the seed (#1364).
splitInfusion(), splitInfusionBolus() and splitBolusInfusion() are now kept in the simulation model of a model with a residual-error endpoint; before, only splitBolus() was carried over, so solving such a model silently skipped the dose split (#1381).
Second derivatives of registered functions now convert. A multi-variable Derivative(f(...), v1, v2) -- what symengine emits during a second-order sensitivity expansion -- aborted with "'Derivative' conversion only takes one function and one argument" unless every intermediate derivative happened to be a bare registered call, which most are not. Four defects contributed: the multi-variable loop searched the previous derivative's operands for the next variable instead of differentiating the expression; digamma and trigamma were missing from the arity table although registered derivatives emit them, so they were rejected as "requires 0 arguments (supplied 1)"; symengine's tuple-form simultaneous Subs(..., (a, b), (e1, e2)) is not parsable R and failed before the converter saw it; and the xi renaming was not idempotent, producing rxrx_xi_1 when text made a second pass, which leaked into the generated model as a free parameter.
A second derivative through linCmtB() is refused again rather than falling through to a numeric difference. linCmtB() is itself the first-order parameter sensitivity of the solved form, which carries no second-order state sensitivity, so the difference is not the second derivative of the prediction. Returning one made nlmixr2est's .foceiAddHdEta2() succeed where it is meant to fail, so a fast=TRUE FOCEi fit of a linCmt() model built an inner Hessian with a fabricated curvature term instead of falling back to finite differences; an ll() + linCmt() fit converged to -22.4 rather than 118.5. Second derivatives that do not involve linCmtB() still convert.

The failure was silent further up: a model needing these had no analytic second-order sensitivities, so the FOCEi-family analytic outer gradient fell back to finite differences without saying so.
The Boost special functions no longer promote double to long double internally. The promotion is Boost's default and costs several times the runtime for accuracy the surrounding solve cannot use.
Simulations no longer depend on the number of threads. With rxSetSeed() in force the seed sequence advanced by the thread count, the eta draws made before an ODE solve and rxRmvn(ncores=) split their normal draws by thread, and in-model draws such as rxnorm() continued each thread's stream across subjects. Parameter draws now give the single-threaded result at any cores/ncores; in-model draws are seeded per subject, so their values change from earlier versions (#1376).
Fixed installation with clang/LLVM OpenMP (CRAN r-devel-linux-x86_64-debian-clang and r-devel-linux-x86_64-fedora-clang). R_NO_REMAP is now defined on the compile command line (-DR_NO_REMAP in src/Makevars.in), which is the only way to have it in effect before any R header is included, as "Writing R Extensions" requires. Without it R's match macro expanded inside the declare variant match(...) pragma in LLVM's omp.h and the compile failed. CRAN released 5.1.7.1 as an interim fix for the same problem; see that version below.
rxSolve() of a function-style (ini()/model()) model no longer converts its ODEs to linCmt() unless asked. rxSolve.rxUi() still defaulted to useLinCmt=TRUE and ignored options(rxode2.useLinCmt=); it now uses a useLinCmt set in the model's meta block, else the option (default FALSE), like every other model type (#1389).
rxProgress() and rxProgressStop() now give an error for a zero-length argument such as rxProgress(NULL) instead of crashing R, and rxProgressAbort() falls back to its default message when error is empty (#1377).
rxSolve(omegaSeparation="tnpri") no longer draws the omega entries the thetaMat gives no variance from the dfSub inverse Wishart; they stay at their estimates, as documented. sigmaSeparation="tnpri" likewise ignores dfObs. The "thetaMat has too many items" message now says when an ignored column could be drawn with "tnpri" (#1388).
Parsing a model (rxModelVars(), rxode2(), rxNorm()) no longer takes time quadratic in the number of ;-terminated statements; a 1600-state chain went from about 19 s to 0.2 s. A lone ; is now a statement only as an empty if/while body; semicolons elsewhere parse as before (#1398).
Unloading rxode2 now releases the call frames it keeps for finding user defined functions before its final gc(), so what they hold is freed while rxode2 is still loaded; before, it stayed in memory after unloadNamespace("rxode2") (#1408).
Two fixed-rate infusions into the same compartment at the same rate are now paired with their own stop records when they overlap. dose() reported the wrong amount (60 instead of 100 for a 100 mg infusion with a 50 mg one nested inside it), the dosing records of addDosing = TRUE output carried the same wrong amount, and with f() each infusion's duration was scaled from the other's stop time, so the solved amounts were wrong too. etTrans() now records the start/stop pairing (the rxInfPair attribute) for the subjects where the solver could not recover it, so its output changed for those subjects only; the etTrans() golden snapshots were updated for the 48 nmtest cases that gain the attribute, which are otherwise unchanged. An infusion start written with a classic internal evid and no stop record of its own is now reported as having no duration instead of borrowing the stop of another infusion at the same rate (#1348).
A prior spec from rxPriorBuildSpec() (which nlmixr2est fits with priors keep) is now freed by R with its external pointer instead of by a finalizer in rxode2's DLL, so freeing it after unloadNamespace("rxode2") no longer crashes R where the DLL is actually unloaded, as on Windows (#1406).
The environment used to look up R user functions is now set only for the length of the rxode2 call that sets it (rxode2(), rxSolve(), $ on a ui, rxToSE(), ...). It used to be the first caller of the session, whose frame was then never freed and whose same-named function a later model could use instead of its own. A model now also finds its user function when an earlier model used a different one (#1409).

rxode2 5.1.7.1 (2026-09-29)

Published on CRAN by Prof Brian Ripley on 2026-09-29, not released from this repository.

Prof Brian Ripley fixed the rxode2 build so that rxode2, and the packages that depend on it, would install on the r-devel clang flavours again. src/rxomp.h undefines R's match macro when R's headers have already been included, before omp.h is reached, where it otherwise expanded inside LLVM's #pragma omp begin declare variant match(...). With thanks.

The source is kept in this repository on the cran-5.1.7.1 branch, tagged v5.1.7.1; upstream it is cran/rxode2@435808f390db2092b31364075e43e4881d9cb9a9. rxode2 5.1.8 addresses the same problem differently, by defining R_NO_REMAP on the compile command line.

rxode2 5.1.7 (2026-09-14)

New features

rxUiEtaDistAnchors() reports which model variable holds each argument of a declared distribution. rxEtaDistExpand() hoists every family argument onto its own line, rxEdA.<eta>.<role>, so the compiled model computes them per observation; this says where those values land, so an estimator can read them from the solve instead of evaluating the same argument expressions again in its own code. NA marks an argument the family emits no line for. Pass the declarations for an EXPANDED model: rxEtaDistExpand() removes etaDist from the iniDf, so they cannot be read back off the expanded ui.
The mu-reference scan now tells the two declared-distribution routes apart when a covariate appears in a dist() argument. A covariate written into a rxEdA.<eta>.<role> anchor reads locally like theta + coefficient*covariate and was claimed as a mu2 covariate on both routes. That is right on the cdf route, where the decoder reads the anchor and it is genuinely in the observation path, and wrong on the direct route, where the anchor feeds the prior alone. Measured on a subject-constant covariate arm with a true coefficient of 0.75 from a start of 0.35: the cdf route needs the claim (0.5526 with it, 0.1841 without), and the direct route needs it skipped (0.7441).
The declared-distribution expansion now hoists each family argument onto its own named line, rxEdA.<eta>.<role>, and the decoder refers to that name:
```
rxEdA.eta.cl.shape <- 1/exp(lclrv)
rxEdA.eta.cl.rate  <- 1/(exp(lclrv) * exp(lclm))
eta.cl <- gammapInv(rxEdA.eta.cl.shape, phiU(rxN.eta.cl))/(rxEdA.eta.cl.rate)
```
The line is keyed by lotri's ROLE rather than the family's argument name, so scale means the same thing across families and is a usable group key. The point is that a covariate on a declaration's rate becomes a term added to one named line -- an ordinary edit every downstream consumer already handles -- instead of a substitution buried inside a quantile call.

Both spellings of dist() emit the same anchors (the model({}) form carries them through the user-function before hook), so the two remain byte-identical, and only arguments the family's quantile template actually uses get an anchor -- a normal-based family collapses to its latent and never references some of its arguments.

Fits are unchanged: this is a refactor of the generated model text. Measured on a declared-gamma covariate model, the estimates and objective function value match the pre-anchor run to every printed digit.

rxEtaDistMuRef() knows about the new prefix. Its scan for generated lines was ^rx[NTLSUuc][.], so without this a declaration parameter that had moved onto an anchor came out silently non-mu-referenced -- the same failure its own comment warns about for rxT..
rxEtaDistMuRef() mu-references the parameters of a declared random-effect distribution. rxEtaDistExpand() writes them as bare thetas inside an inverse-CDF call, so none of them is in theta + eta form and none of them is mu-referenced -- the case saem and the FOCEi family handle worst. This carries each of them (and the copula correlation) on its own random effect with a small FIXED variance, which is the structure NONMEM gets from MU_5 = THETA(5) with $OMEGA (0.0 FIXED).

The helper variance must not be ~0, which is where the NONMEM idiom stops porting: nlmixr2's mu-theta M-step is weighted by omega^-1, so a ~0 variance pins the parameter at its ini() value while still reporting it as estimated. Values below 1e-6 are refused with that explanation.

The result is a different model -- the helper variance is real between-subject variability -- so it is a way to travel, not a way to finish; use it as stage one and refit the intended model from its estimates.
Random effects can now be declared non-Gaussian. lotri's dist() line says what a random effect's distribution is, and rxEtaDistExpand() turns it into an ordinary model:

ini({
  lclm  <- log(5)
  lclrv <- log(0.09)
  dist(eta.cl) ~ dgamma(shape=1/exp(lclrv), rate=1/(exp(lclrv)*exp(lclm)))
})
model({
  cl <- eta.cl   # a gamma distributed clearance
  ...
})

No variance is written for a declared random effect -- the declaration fixes its marginal and the latent scale is standard normal, so dist() implies ~ 1. A block is written out only for a correlation between two declared random effects.

The technique is Bauer's (NONMEM 7.5.1): keep the latent random effect standard normal and change the CDF -- z ~ N(0, 1), u = phiU(z), eta = Q(u) -- with correlation induced on the latent scale as a Gaussian copula. The expansion happens on the UI, so rxSolve() simulates from the declared distribution with no special case of its own, and it rewrites the declared correlation block into unconstrained rxCor.* thetas plus a fixed identity omega, which is a form every downstream estimation method already handles.

New model functions for the inverse CDFs this needs: phiU() (the normal CDF bounded away from 0 and 1, so a quantile cannot return an infinity), ibeta()/ibetaDer()/ibetaInv() (pbeta/dbeta/qbeta) and studentTCdf()/studentTDen()/studentTInv().
The event table and the runtime dose-pushing statements (evid_(), bolus(), infuse(), infuseDur(), replace(), multiply(), phantom(), reset()) now share one implementation of the NONMEM event semantics, so the same regimen written either way produces the same internal records. A test drives both over the same events and compares them record for record, so the two can no longer drift apart.

Bug fixes

A model's declared-distribution expansion is now derived once rather than on every solve. rxSolve() asks rxEtaDistExpand() whether the model declares a distribution, and for a solve-ready object answering that meant rebuilding the ui -- re-parsing a model that had already been parsed -- and, for a model that does declare one, building a NEW expanded model each time. Models live in the ODE model pool, so that re-set up the problem on every call instead of reusing the one already set up. The expansion is a pure function of the model, so the result is now cached on the object under the model md5 rxModelVars() already carries; a changed model is re-derived rather than answered from a stale memo. Measured on a three-eta linCmt() model that declares nothing: repeat calls 0.0331 s to 0.00010 s, and rxSolve() of twenty subjects 0.0781 s to 0.0398 s.
Solving a model no longer re-emits the warnings its parse already gave. rxSolve() asks rxEtaDistExpand() whether the model declares a distribution, and on a solve-ready object that lookup had to rebuild the ui, re-parsing the model and repeating every parse-time diagnostic. A model with no declaration at all could warn from inside rxSolve().
gammap()'s derivative with respect to its shape argument, and every derivative of gammapInv()/gammaqInv(), are now in the derivative table. They were absent, and a missing rule does not error: rxode2's symbolic differentiation silently substituted a one sided finite difference instead. Any model using an inverse incomplete gamma was therefore differentiated numerically without saying so -- worst exactly in the tails, where the inverse CDF is steepest. The new gammapDera(), ibetaDera(), ibetaDerb() and studentTCdfDnu() supply the shape derivatives that have no elementary closed form.
The DENSITIES now differentiate too: gammapDer(), ibetaDer() and studentTDen() have derivative rules, in closed form. This is the second derivative of an inverse-CDF chain -- d(q)/d(p) is one over the density at the quantile, so anything needing a second derivative of that (a Laplace inner Hessian, an analytic outer gradient) has to differentiate the density in turn. Without a rule that silently became a one sided finite difference taken of a function that is itself a finite difference. Any model using gammapDer() was affected, not only a declared random effect distribution.

Event translation

A steady-state dose into a compartment with a modeled alag() pushed from inside a model now expands the way the event table expands it, so the two spellings of the regimen agree. A steady-state dose has to be solved unlagged while the dose the subject receives is lagged, and only the event table split the record into that pair; the pushed form solved to something else (in the reported case the two differed by 9.05, and now agree to 9e-07) (#1349).
addl no longer repeats a pushed observation, "other" (evid=2) or reset (evid=3) record. A pushed evid_(t, 3, ..., addl = 2) reset the system three times; the event table warns and ignores addl for those records, and the push path now does the same.
A pushed phantom dose (evid=7) keeps a modeled rate or duration instead of silently becoming a bolus, and a pushed evid=2 record naming a compartment turns that compartment back on -- both matching what the same row does in the event table.
evid_() now accepts a negative compartment, by name or number (evid_(t, 2, 0, -depot, 0, 0, 0, 0)), to turn that compartment off -- the same signal the event table takes from a negative CMT column.
A steady-state constant infusion (ss=1, ii=0, amt=0) written as a hand-encoded classic internal evid (>= 100) carrying a duration now errors in the event table too. The runtime push path already refused it (#1350); the event table accepted it and steady-stated the compartment to zero.
A split bolus now splits every record a dose translates to rather than only the first, which a lagged steady-state bolus needs.
A steady-state constant infusion (ss=1, ii=0, amt=0) pushed from inside a model with a duration -- modeled (rate=-2, e.g. via evid_()), fixed (infuseDur()), or a hand-encoded classic internal evid (>= 100) -- now errors instead of silently steady-stating the compartment to zero. That combination never had a usable rate (a constant infusion never turns off, so there is nothing for the duration to measure), and the event-table path already refused it; the runtime push path did not check for it (#1350).
Piping an omega block into a model with ten or more etas no longer permutes the etas that were not piped over. They were renumbered with factor(paste(neta1)), which sorts the numbers as TEXT -- "10" before "2" -- so the survivors came back in an arbitrary order. That splits a correlated block across the matrix, and it can renumber a repeated (same()) block ahead of the block it repeats, which has no representation at all since the linkage is a relative offset backwards ($omega then errored with "must refer to an earlier parameter").
rxRename() now follows a repeated (same()) block's marker to the new name. The block a repetition mirrors is recorded BY NAME in the condition column -- which is what lets the marker survive renumbering -- so a rename has to be followed too; left alone it pointed at a name that no longer existed and $omega refused to assemble ("refers to '', which is not in this block").
Nested (inter-occasion) simulation gave every random effect the wrong variance whenever a level carried more than one parameter. The omega a level draws from is laid out occasion-major, with the parameters inside each stamp, but the expansion indexed it parameter-major, so the two were transposed: with lotri(a ~ 0.01, b ~ 1, cc ~ 100) | occ every parameter in occasion 1 drew variance 0.01, every one in occasion 2 drew 1, and every one in occasion 3 drew 100. A single parameter per level is unaffected, which is why this went unnoticed. Any covariance specified within an occasion was likewise placed between occasions of one parameter rather than between the parameters of one occasion (#1345).

New features

Correlated inter-occasion variability is now supported. A | occ block may carry off-diagonal elements, and they are simulated at the level they were declared at. assertRxUiIovNoCor() no longer reads a repeated (same()) block as a separate level of variability.
rxSymInvCholCreate() gains a same= argument. Blocks that repeat an earlier one -- NONMEM's $OMEGA BLOCK(n) SAME, written same() in a lotri/ini block -- share that block's parameters rather than being estimated separately. For a block diagonal omega whose blocks are identical the derivative with respect to a shared parameter is just the block diagonal sum of each copy's contribution, so this needs no change to the C++ inner problem.
$omegaSameMap reports, for each eta, which earlier eta it repeats, which is what rxSymInvCholCreate(same=) takes.
The same model variable can now be used by more than one residual-error endpoint, as long as each endpoint is named, as in cp ~ add(add.sd1) | phase1 and cp ~ add(add.sd2) | phase2. rxode2 gives each of those endpoints its own hidden alias (rx.cp.phase1, rx.cp.phase2, recorded in $endpointAlias) when the simulation or estimation model is assembled, so they get separate residual parameters, separate simulated residuals and separate ar() state, while the model({}) block still prints, extracts and pipes as it was written. Previously this needed hand-written aliases (cp.phase1 <- cp, ...) and otherwise failed to solve with "The simulated residual errors do not match the model specification". Two endpoints on the same linCmt() are supported the same way.
Model piping selects one of several endpoints by its condition, as in model(cp ~ prop(prop.sd) | phase2), and drops one with model(-phase2). Piping an endpoint that shares its variable without naming a condition now says which conditions are available instead of reporting the lhs as duplicated.
Two endpoints that share a condition (cp ~ add(a) written twice) now give a clear error at parse time asking for a | <name>, rather than failing later with a confusing message about the additive standard deviation.
linCmt() models can now report a PER-COMPARTMENT dose-time sensitivity. linCmtB(which1 = -3) differentiates with respect to one delay shared by every dose feeding the linear system, so a regimen that doses a lagged depot alongside an unlagged central -- the paired IV/oral design bioavailability is routinely estimated from -- had to be refused. Two new modes read a per-origin decomposition of the linCmt() amounts instead: which1 = -9 is the derivative with respect to a delay on one compartment's doses alone, and which1 = -10 returns the amounts that arrived through one compartment, which chain-rules to bioavailability as A^(q)/F_q. which2 packs the origin compartment and the wanted output (q*8 + out, out = 7 for the reported concentration). Both match central finite differences across one to three compartments, IV and oral, bolus, infusion and steady-state bolus regimens, including models that lag their linCmt() compartments differently. The decomposition is maintained only for a model that declares a modeled alag()/f() on a linCmt() compartment (nlmixr2/rxode2#1119).
linCmt()'s sensitivity state now belongs to the INDIVIDUAL rather than to whichever thread happens to be running it. The theta-keyed window of hoisted closed-form constants and the last-row value memo are the two pieces whose value across calls is the whole point of having them, and holding them per thread meant both had to be keyed on the subject id and were discarded whenever another individual landed on the same slot. They now live in the solving structure for the individual, allocated on first touch inside the solve and released with the subject alongside the delay history and the rate history; the subject-id keys are gone, and a solve being independent of how subjects were handed to threads is now structural rather than something the transition-matrix path has to restart per subject to enforce. Pure scratch -- the Eigen work matrices and the kernel object -- deliberately stays per thread, since it is overwritten on every call and a per-individual copy would be memory holding nothing. Results are bitwise identical, including across thread counts. Measured on an idle machine by alternating two cleanly built installations, a FOCEi fit runs 1.15x (two compartment) to 1.19x (one compartment) faster at an identical objective and an identical evaluation count. A plain rxSolve() shows almost nothing, and that is the expected shape: every subject of one solve shares a theta, so the window was already valid as a thread walked from subject to subject -- only a fit, where each individual carries its own random effects, has anything here to recover.
The linCmt() per-row path no longer allocates. getVc(), adjustF() and getJacCp() took a const Eigen::Matrix& while every caller passes an Eigen::Map over the parameter buffer, so each call materialized a heap-allocated vector and copied into it; they take an Eigen::Ref now, which binds the map directly. adjustF() is on the value path, so that temporary was being built for every row, every output and every inner pass of a fit.
New internal rxSetIdLvlFactors() C callable lets a host package (e.g. nlmixr2est during focei/saem estimation) populate the global subject-id factor table directly, so aggregated solver warnings can be attributed to the real subject id even when the data passed to rxSolve_() is not a classed rxEtTran table. It takes a character vector of ids, coerces an integer/real/logical one, and clears the levels for any other type. When the id still cannot be resolved, the aggregated-warning flush prints the 1-based internal solve index (e.g. internal #1) instead of a bare Unknown; a subject whose id is literally Unknown is still printed as itself. In a multiple-simulation solve the subjects past the first simulation are labelled by subject and simulation (e.g. 2 (sim 2)) rather than falling back.
rxIndLinExpStats() reports the per-thread matrix-exponential cache used by matExp()/method="indLin" solving: computed, reused, noSlot (exponentials taken by a thread with no cache slot of its own, so the pool was never sized or is narrower than the thread team) and slots. The same computed/reused counts already rode out of an rxSolve() as $counts$dadt/$counts$jac, but an estimation package drives the solver directly and never builds that data frame, so from inside a fit the cache was unobservable. The counters survive the pool free that runs at the start of the next solve, so a measurement can span the solves a fit is made of. Measured with the accessor on an optimized build, the cache it reports is worth 1.60x on a matExp() FOCEi fit (93.8% of the exponentials reused) and 1.15x on a 200-subject matExp() solve; RXODE2_INDLIN_NO_EXP_CACHE=1 is the A/B switch, and bench/indlin_expcache_ab.R is the harness.
A linCmt() sensitivity row's state-transition matrix and its parameter derivatives are now assembled from their CLOSED FORM in the constants the theta-keyed window already holds -- the eigenvalues and spectral matrices, ka, and each of their tangents -- rather than by probing the kernel with unit-basis prior states. The probe needs one kernel evaluation per direction per column, so it could only pay for itself on an interval that demonstrably recurs, and it had to exclude rate-bearing rows; the closed form costs about one kernel evaluation altogether, so it is assembled for any interval -- irregular designs, first occurrences and infusions included -- and cached where the interval repeats, which is where the reuse is worth more than any build. This is the same exact closed form summed in a different order (the matrix assembled, then applied), the order the transition-matrix path already shipped, so it can differ from the row-by-row order in the last few digits: measured against the row tail the disagreement is a few units in the last place, and against reverse mode it is the same spread the row tail itself has. Where the probe engages -- and its entries are exact by construction -- the two matrices agree to 1e-14. Measured on an optimized build against the previous default, the closed-form route is never meaningfully slower and is up to 2.2x faster on the designs the probe could not serve. Which designs those are is the whole of it: where the sampling is regular the probe already served nearly every row and nothing changes, and the gain is on schedules whose intervals do not repeat. It carries into a fit -- a 40 subject by 100 observation two compartment oral FOCEi fit runs 1.37x faster on an irregular schedule (53.4s to 38.9s) and unchanged on a uniform one, with the objective function identical to 1e-11. The probe-built route stays available as linCmtSensPhi=1, and FALSE still evaluates row by row in the order earlier versions used. linCmtSeqStats() reports the rows served as phiAnalyticRows.
The linCmt() sensitivity delta memo now keeps serving a row's own repeat executions after its give-up guard has disarmed. The guard stops the window building exponentials for gaps that never recur, which is what a schedule with no repeated interval produces; but one row is looked up several times under one set of parameters -- once per linCmtB() call the model generates, and again on every inner re-walk of a fit -- and disarming was discarding those repeats too, so an irregular design rebuilt the exponentials, and with linCmtSensPhi=2 reassembled the transition matrix, once per EXECUTION rather than once per row. A slot outside the round-robin now holds the current row's gap while the guard is disarmed; it is never read as evidence that an interval recurs, so the guard's reading of the design and the probe-built route's engage rule are unchanged. Results are bitwise identical either way (the memo is exact caching, tested with it forced on and off). Measured on a 40 subject by 100 observation two compartment oral FOCEi fit, an irregular sampling schedule runs 1.14x faster and a uniform one is unchanged, which brings the cost of an irregular schedule relative to a uniform one from 1.31x to 1.08x. linCmtSeqStats() reports the builds served this way as expSolo.
linCmtSensType="ADm" differentiates the linCmt() closed form with every requested direction carried through a SINGLE forward-mode pass, where "AD" repeats the whole evaluation once per direction with one tangent each. The solution itself -- each exponential, each division of the depot transfer, the eigen-decomposition -- is therefore evaluated once per row rather than once per direction. Results match "AD": the multi-direction scalar reproduces the operation order of every forward-mode rule it replaces, and carries the same Eigen cost traits so the small matrix products unroll the same way (validated over 1/2/3 compartments x IV/oral x every trans x every steady-state form x infusion x every requested-direction mask). On the amortized row path that is bitwise on every platform tested. The two are different template instantiations, though, so a compiler may contract a multiply-add into an FMA in one and not the other; through the full evaluator that is what clang on arm64 does, and there the two agree to round-off rather than to the bit. Unlike the amortized ordinary-row path this also shares the work on steady-state rows, which have no constants/tail factorization and so were paying a full evaluation per direction. linCmtSeqStats() reports the rows served as dualRows.
Model parsing is no longer superlinear in model size. statement in the dparser grammar could derive the empty string (its last alternative was a bare end_statement, which is (';')*), so statement_list : (statement)+ admitted any number of empty statements at every position. dparser resolved that ambiguity by greediness, re-walking the whole accumulated parse tree at each position, and every parse in the package paid for it -- rxode2(), rxNorm(), rxModelVars(), rxS() and rxOptExpr() alike. Requiring a semicolon in that alternative makes the grammar unambiguous. On a 301-line model rxNorm() drops from 3.18s to 0.07s and rxOptExpr() from 5.61s to 0.43s; per-line parse cost, which grew tenfold between a 25-line and a 301-line model, is now flat. Model text that is only whitespace and comments is now treated as a blank model, as an empty string always was.

One degenerate form is no longer accepted as a consequence: an if or while with no body at all and no braces (if (a > 1), while (a > 1), if (a > 1) else b <- 1). These used to parse because a statement could be empty, which is the ambiguity being removed -- an empty body makes if (a > 1) b <- 1 ambiguous, since b <- 1 could be the body or the next statement, so the old behavior cannot be kept alongside the fix. Write an empty body as {} or ; (if (a > 1) {}), both of which parse as before. Bodies that only look empty, such as a block holding nothing but a comment, are unaffected. R rejects all three of these too, so the grammar now agrees with R where it used to be more permissive, and they are reported as an ordinary model syntax error naming what to write instead rather than as a bare parser error.
The common-subexpression search in rxOptExpr() runs in C. It counted subexpressions in a named R list and looked them up with [[text]], a linear scan, so the search was quadratic in the number of distinct subexpressions -- which on a second-order sensitivity model is enormous. A new dparser grammar (inst/rxCse.g) and C pass (src/rxCse.c) count into a hash instead, one statement at a time across threads, and the counts are merged by the earliest position each subexpression was seen so the result does not depend on the thread count. On a 286-line second-order model the search drops from ~143s to ~1.5s. Output is byte-identical; anything the C pass will not reproduce exactly -- an unsupported left-hand side, past(), a numeric literal it cannot render the way as.character() would -- makes it decline so the R implementation runs. Set options(rxode2.optExprC = FALSE) to force the R implementation.
rxOptExpr() decides whether to chunk on the model's SIZE rather than its line count. Chunking amortizes the parse, and parse cost tracks characters, so a model with many short lines was being chunked when chunking made it slower. The threshold is options(rxode2.optExprChunkChars = ), default 512 KB; chunkLines= still caps the chunk size and chunkLines = 0 still forces a single whole-model pass. A model between 40 lines and the size threshold is now optimized whole, so it gets better sharing and no rx_expr_c<i>_ temporaries.
options(rxode2.compile.O=) now reaches the compiler. The level was written into the generated model's PKG_CFLAGS, and R CMD SHLIB puts PKG_CFLAGS ahead of R's own CFLAGS in ALL_CFLAGS, so R's -O2 came last and won -- the option had no effect and every model was built at -O2 whatever it was set to. The level is now applied through a temporary user Makevars for the duration of the build, so the documented default (-O3) is what models are actually compiled at. Only the -O is changed; R's other flags, and any CFLAGS the user sets in their own Makevars, are left alone.
rxNorm() parses a model once rather than twice. With no condition set it asked rxCondition() whether one was, and rxCondition()'s lookup key is a digest of the normalized model -- so the model was normalized to build the key, the text discarded, and then normalized again to be returned. It is now normalized once and the same text used for both. rxOptExpr() also dropped an unused rxModelVars() call, so one rxOptExpr() now parses once instead of three times. On a 301-line model rxNorm() drops a further 0.074s -> 0.039s and rxOptExpr() 0.43s -> 0.34s.
Symbolic derivative setup (rxS(), .rxJacobian(), .rxSens(), and so the nlmixr2est model builds that use them) is 4-5x faster. The cost was never symengine itself -- symengine::D() is under 1% of the total -- but the R text translation around it: rxFromSE() re-parsed every symengine string with R's parse(), emitted through nested paste0() and a nine-deep sub() regex chain, and saved and restored the whole options() list on every leaf. Two changes address it. The .rxSEcnt constant renderings are now computed once when the package is built instead of by eval(parse()) on each numeric leaf. A new dparser grammar (inst/seFromSE.g) and C emitter (src/seFromSE.c) then translate symengine output directly; expressions the emitter does not reproduce exactly fall back to the R walker, so output is unchanged. .rxSens() on a fifteen-state model drops from 1.27s to 0.19s. Set options(rxode2.symengineC = FALSE) to force the R walker. The symengine expressions of a jacobian or sensitivity build are now translated as one batch rather than one call per line, and the batch is spread across threads once it is large enough to pay for them.
linCmtSensType="auto" stays forward-mode AD ("AD"). An intermediate development version made reverse-mode AD ("ADr") the default on the strength of timings that had been taken through devtools::load_all(), which compiles at -O0; on an optimized build forward mode is at least as fast as reverse for every number of requested sensitivity directions on every configuration, and reverse is 2-4x slower on a three compartment oral model, because its per-row tape build costs more than the extra forward passes. "auto" therefore resolves to "AD" on every solve path (rxSolve(), threaded solves, ind_solve() and linCmtModelDouble()), and "ADr" remains an explicit option.
Reverse-mode AD (linCmtSensType="ADr") now solves across threads: the Stan tape is per thread under STAN_THREADS, so it is no longer forced onto one core. Results match forward mode to round-off.
The linCmt() forward-mode sensitivity evaluation is amortized across rows: the theta-only constants of the closed form (the elimination constant or the 2/3-compartment eigen-decomposition, and ka) and their derivatives are taken once per theta-keyed window, and each ordinary row costs one allocation-free forward pass per requested direction through the dt-dependent tail of the solution (steady-state rows and the finite-difference families keep the full evaluator). Together with the removal of per-row heap allocation from the sensitivity hot path this makes the sequential sensitivity solve 1.6x (two-compartment) to 2.1x (three-compartment) faster on an optimized build, with results identical to round-off. linCmtSeqStats() reports the window refills and how many rows took the amortized tail. The per-subject hybrid strategy this window machinery was first built for (rxSolve(linCmtSensStrategy=) and the linCmtHybrid* thresholds, introduced on this development branch and never released) was measured to win nowhere once the sequential path itself was amortized, and has been removed.
A last-row value memo removes the repeated work of the generated model executing the same linCmtB() value call many times per row (measured: about fifteen executions per row -- five full computations and ten restores). A repeat with an identical key returns the cached value with the Jacobian left standing for the sensitivity reads; sentinel calls and model reshapes invalidate the memo. Measured on an optimized build this makes the sequential sensitivity solve a further 1.5x (two-compartment) to 2.3x (three-compartment, sparse) faster; results are identical. linCmtSeqStats() now also reports the value-execution classes and the memo hits.
A thin value path consolidates the two per-row visits a fit makes to a pure linCmtB() row (the solver's state fill and the left-hand-side walk): a value re-execution of an already-solved row now returns the saved amounts and concentration scaling only, skipping the sensitivity setup, the rate cache, the Jacobian restore and the concentration gradient recompute. The Jacobian is restored lazily if a sentinel or read call for that row follows, so carry models are unaffected (tested against reverse mode). linCmtSeqStats() reports the served executions as valueLite; in a FOCEi posthoc evaluation the left-hand-side walk is served entirely by this path.
A delta-keyed memo caches the tail's dt-dependent exponentials (and their derivative in every requested direction) per distinct row gap under the theta window, so designs with repeated observation spacing evaluate their sensitivity rows without recomputing any exponential -- a uniform sampling schedule needs one exponential build per window and every other row is multiply-only. The memo is exact caching (bitwise identical results, tested), sized at four gaps from measured designs, and can be disabled with RX_LINCMT_DELTA_MEMO=off; a design with no gap reuse stops building speculatively after eight consecutive misses so it pays essentially nothing; a stretch whose gap repeats the previous row's -- what a regular sampling schedule produces and an irregular one does not -- re-arms it, so an irregular stretch no longer disables the memo for the regular rows that follow it under the same parameters. linCmtSeqStats() reports the builds and hits.
Where a linCmt() sensitivity row's interval repeats -- as it does under regular sampling, and across the dosing intervals of a repeated regimen -- that interval's state-transition matrix is now assembled once and the later rows of the same width propagate through it, instead of evaluating the closed form again in every requested direction. Measured at 1.1 to 1.6 times faster on those designs, most at three compartments and many directions. The matrix is built only when a different row is seen to share an interval width -- the one thing that shows the interval really does recur in the design -- so a design whose intervals never repeat builds none and is unaffected, in a fit as in a single solve. (A row re-queried while a fit's inner problem re-walks a subject is the same interval asked about twice, not a recurrence, and does not count.) Each subject starts from a blank interval state, so a solve is unchanged by the number of threads it runs on. Rate-bearing rows of an infusion and steady-state rows keep the previous route.

This is the same exact closed-form solution evaluated in a different order: the interval's matrix is summed first and then applied, where the previous route accumulated the same products as it went. Floating-point addition is not associative, so the two can differ in the last few digits -- neither is an approximation of the other and neither is the more correct. Measured over every compartment count, parameterization, regimen and direction mask, the largest disagreement was a few units in the last place of the values involved (against an independently integrated reference the new order was in fact the closer of the two slightly more often). rxSolve(linCmtSensPhi=FALSE) restores the previous order for anyone who needs to reproduce earlier results digit for digit; linCmtSeqStats() reports how many matrices were built and how many rows used one.
linCmtB() derivatives are now emitted as direct reads of the sensitivity state columns: the parser registers a derivative slot when a rx__sens_<cmt>_BY_<slot> state is referenced as a bare symbol, and the derivative table writes the concentration gradient as arithmetic over those states (matching the internal per-trans scaling exactly, so results are bitwise identical; a trans without a covered scaling keeps the call form). A FOCEi inner model drops from about eleven linCmtB() executions per row to the single value call. The measured effect is modest (the read calls already short-circuited cheaply): about 1.0-1.1x on the sensitivity solve and 1.04x on a dense FOCEi fit, with the identical objective.
rxPriorLogDensity(ui, theta, omega) evaluates a model's ini({}) priors as a Bayesian penalty at the current parameter values -- the value and gradient kernel an estimation method's objective function needs, as opposed to rxSolve()'s use of the same priors to simulate study-level variability around the initial estimate. Supports dnorm()/stdNormal(), dcauchy() (bounds-truncated using the parameter's own lower/upper), the joint theta+omega multiNormal() block, and invWishart() degrees of freedom on an omega block, with three methods: "general" (a textbook Bayesian log density), "nwpri" (NONMEM's own $PRIOR NWPRI omega parameterization, NONMEM7 Technical Guide eq. 1.157/1.159/1.170, which is not the same density as the textbook one), and "tnpri" (the assumption Monolix's Bayesian estimation and NONMEM's own estimation make -- all estimated parameters, including omega, are jointly normal; an om.<eta> member is on the same raw omega scale as "general", since neither Monolix nor NONMEM place the prior on a Cholesky factor). Neither "nwpri" nor "tnpri" has a Cauchy analogue (nlmixr2/nlmixr2est#929).

The value/gradient math itself (rxPriorLogDensityEval()) is pure C++ with no R/Rcpp call of any kind, added to rxode2's C function-pointer table (inst/include/rxode2prior.h) so a downstream package's own C++ objective function can call it directly from inside an OpenMP-parallel region -- rxPriorLogDensity() is a thin R convenience wrapper over it. rxPriorBuildSpec(ui, method) builds the reusable spec the evaluator needs (an R-only, one-time, main-thread step).

rxPriorOmegaToCholOmegaInvGrad(omega, gradOmega) chain-rules a raw-omega-scale gradient (from rxPriorLogDensity(), or from anywhere else) into the equivalent gradient with respect to the upper-triangular free elements of chol(Omega^-1) -- the parameterization FOCEI's own op_focei.cholOmegaInv varies internally (nlmixr2est/src/inner.cpp). A method that estimates omega directly (SAEM) does not need this conversion. Like the density kernel itself, the underlying rxPriorOmegaToCholOmegaInvGrad() C++ function is pure, with no R/Rcpp call, and is exposed through the same function-pointer table.
rxPriorLogDensity()/rxPriorBuildSpec() now evaluate a marginal dnorm()/dcauchy() prior placed directly on ONE omega covariance (off-diagonal) element, not just a diagonal (variance) one -- prior(eta.cl, eta.v) ~ dnorm(0, 0.1) in a model whose eta.cl/eta.v covary. This is distinct from a whole-block distribution (invWishart()/multiNormal()), which still requires naming the model's entire connected block; a marginal prior only needs its two names to covary with each other, so it also works on a two-name subset of a larger correlated block. Previously refused outright with "a prior on an off-diagonal omega element is not supported" -- unreachable through real ini() syntax until lotri's own prior(a, b) ~ dnorm(...) parsing was relaxed to allow it (see lotri's own NEWS). rxSolve(..., usePrior=TRUE)'s simulation path does not yet support this (it draws a deviation added to the model's own initial estimate, which this needs its own machinery for) and gives a clear error rather than the confusing one this would otherwise have produced.
rxSetActiveParLoader() / rxClearActiveParLoader() activate a registered parameter loader for solves that do not go through rxSolve.rxUi() -- an estimation method's internal solves, for example. Previously the only way in was to .Call() rxode2's compiled entry points by name from another package, which is not a supported interface (and R CMD check flags it).
rxRegisterUiAssembled() / rxRemoveUiAssembled() register a hook called with a freshly assembled rxUi before it is compressed. This is the point at which a package can attach parse-time state to a model: a user-defined function (rxUdfUi()) can only return code, and once the ui is compressed rxUiDecompress() yields a fresh environment on every call, so an in-place assignment made any later is invisible to the caller. The hook receives the ui environment and may assign into it; pair that with sticky so the slot survives piping and saveRDS().
A parameter carried in rxForcedPars() now counts as supplied when solve parameters are resolved. Previously the forced values were written after resolution had already rejected the solve, so a parameter that existed only as a forced parameter failed with "the following parameter(s) are required for solving" and needed a placeholder data column purely to get past the check. This lets a model own parameters that never appear in the data.
A solve-time ui preparation hook registered with rxRegisterUiPrep() may now take (ui, solveModel) rather than only (ui). solveModel is the model whose parameter order the gpars layout actually uses, which a hook that resolves parameter positions by name needs. Existing one-argument hooks are unaffected.
rxSolve() simulates parameter uncertainty from the prior distributions the model's ini({}) block specifies, which is what NONMEM does with $PRIOR NWPRI and $PRIOR TNPRI. Writing a prior in the ini({}) block needs lotri 1.0.5 or newer; with an older lotri the block cannot express one and prior simulation simply does not engage. omegaSeparation="tnpri" below works with any lotri. A model that carries priors uses them whenever variability is simulated, so rxSolve(model, ev, nStud=100) is all that is needed.

Each omega block is drawn from an inverse Wishart with its own degrees of freedom (prior(eta.cl, eta.v) ~ invWishart(20)), which the single dfSub argument cannot express; a block with no prior is left at its point estimate. A normal prior on a population parameter (tka ~ 0.01) gives the thetaMat, and a block may name omega elements as well (tcl + om.eta.cl ~ c(...)) for one joint variance over the thetas and the omega values.

A prior mean has to be what the model already says the entry is, since the draw is added to that value; a prior centered elsewhere is an error rather than a simulation that quietly differs from the model.

Because a jointly drawn omega is not guaranteed positive definite, such a draw is retried up to priorPdRetry times (10 by default); if none is, the nearest positive definite matrix of the kept draws is used with a warning, since the projection is biased toward singularity.

usePrior=FALSE ignores the priors. Priors take precedence over a thetaMat/dfSub carried in the model's meta block, with a warning; one given at the call site wins over the priors instead. Chunked solves (#1252) are a clear error rather than a solve that silently drops the prior.

Nested/occasion models are supported (#1253): each prior's degrees of freedom go on the nesting level holding its block, and a level with no prior stays at its estimate. Because a level is drawn as a whole, a prior covering only part of a level is an error -- drawing it would redraw the rest of the level and correlate blocks the model declared independent.
rxSolve(omegaSeparation="tnpri") (and sigmaSeparation="tnpri") draws the omega/sigma entries carried in a thetaMat jointly with the thetas, rather than redrawing their correlations with a separation strategy. This is the general form of the TNPRI above, for a thetaMat that did not come from an ini({}) block prior.

A covariance step already gives one: nonmem2rx emits a thetaMat with columns like IIVCL, omega1.2, IIVV1, ... and a nlmixr2 fit's $cov uses om.<eta>/cov.<eta1>.<eta2>. Both spellings are recognized, as are the sigma equivalents. Until now the off-diagonal entries were dropped as "too many items" and the correlations were redrawn from LKJ, discarding what the covariance step measured; drawing them jointly also keeps the covariance between a theta and an omega entry, which no separation strategy can carry.

It is opt-in because an eta-named thetaMat column already means that eta's variance under the existing strategy, so the same column cannot silently change meaning. omega has to be a matrix, since the draws are added to it.
rxUiPriors() returns the priors a model specifies, with the parameter name, the prior, its neta1/neta2 (NA for a population parameter) and the parameter's lower/upper bounds. The predicates testRxUiPriors(), testRxUiNormalPriors(), testRxUiOmegaDf() and testRxUiOmegaNormalPriors() report what kind of priors are present, so a method that implements priors can branch instead of asserting. testRxUiOmegaDf() and testRxUiOmegaNormalPriors() are mutually exclusive, since an omega prior is either degrees of freedom (NWPRI) or a normal prior (TNPRI) and lotri rejects a model that gives both.
assertRxUiNoOmegaDf() and assertRxUiNoOmegaNormalPriors() reject the omega prior forms a method cannot use, the NWPRI and TNPRI ones respectively.
assertRxUiNoPriors() and assertRxUiNormalPriors() let an estimation method declare which prior distributions it can use. A prior specified in the ini({}) block must never be silently ignored -- that would make the fit do something other than what the model says -- so a method that cannot use priors calls assertRxUiNoPriors() and one that only handles normal priors calls assertRxUiNormalPriors(). Both are no-ops when the installed lotri has no prior support, since then there are no priors to reject.
rxSolve(zeroVarParamHandle=) says what happens when params supplies a value for an omega/sigma item whose variance is zero (say eta.base ~ fix(0)). Such an item is dropped from the matrix that is simulated from and given to the model as a literal zero instead, which discards the supplied value: "warn" (the default) does that and says so, "ignore" does it silently, and "keep" uses the supplied value.
rxSolve(safeLog=2) floors log(0) at log(.Machine$double.eps) the way safeLog=TRUE does, but treats a negative argument as a domain error and returns NaN. safeLog=TRUE (the default) and safeLog=FALSE are unchanged. This is for a hand-written likelihood taking log() of a parameter that must stay positive: under safeLog=TRUE an invalid negative value returns a large finite number, which -log(sigma) turns into a reward of roughly +36 per observation instead of a rejection.
A function that produces models can now name them. rxModelName() is an s3 generic dispatched on the name of the function that was called, so a rxModelName.readModelDb() method names every model rxode2(readModelDb("PK_1cmt")) builds (here, "PK_1cmt") instead of leaving it named after the text of the call. The method is given the call and its (unevaluated) arguments, matched to the argument names of the function being called; a call with no method keeps the default name.
rxModelVars(m)$indLin$wIndLin now reports the states whose indLin(<state>) <- <expr> forcing references a compartment, rather than always being empty. It is worked out by replaying the parsed assignments and forcings in source order, so hand-written matExp() models are covered as well as converted ones, and a forcing that reaches a compartment only through an assigned variable (cp <- central/20) counts too; a forcing built only from parameters or covariates (e.g. indLin(Gc) <- Gprod) stays unflagged, as does one whose variables were reassigned to something state free before it reads them. A forcing inside an if/while may not run, so it adds to what the forcings before it established rather than replacing them. In a model that also has a linCmt(), every forcing is flagged: a solved concentration moves within the step, so such a forcing cannot be treated as constant over the interval the way a locf covariate can. The entries are the 0-indexed positions in $state, named with those states.
rxModelNameLhs() registers the name an assignment is making, for assignment operators like nlmixr2save's := (fit := nlmixr2(...)). It names the model when the model expression itself names nothing -- an anonymous model function, or a call with no rxModelName() method -- so the model is built with that name rather than none. rxModelNameFromExpr() exposes the whole naming sequence for packages that capture a model expression with substitute().
rxMemoryEstimate() now accounts for what method="indLin" allocates, as two new components: indLinExpCache (the per-thread matrix-exponential cache) and indLinWork (the per-thread solver scratch). Both depend on which driver the model runs -- a pure matExp() model holds one rate matrix, while true inductive linearization iterates and carries a Jacobian, P(h) and its inverse as well -- and both scale with cores rather than with subjects, so rxSolve() reaches the same out-of-memory decision for method="indLin" that it already reached for every other solver, and rxSolveChunked() sizes its chunks without charging per-thread buffers to each subject.
rxSolve(method="indLin") now solves subjects in parallel and honors cores. Inductive linearization was held to a single core because it was listed with the Fortran COMMON-block solvers (lsoda, lsode, bdf); it is not one of them, and its matrix-exponential and scheme caches were already per thread. It now goes through the same thread-safety switch as liblsoda, so a model whose functions are not thread safe still drops to one core with the usual warning. The answer, and the rxIndLinSteps() step counts, are unchanged from the single-core solve.
Solve-time hooks let a package change what a solve sees from outside the model text. rxForcedPars(ui) <- c(cl = 1.2) sets parameter values that override params/data and the initial estimates on every solve; they are stored on the ui (hidden from the printed model, registered sticky), so they survive piping and travel into a nlmixr2 fit built from that model, which keeps a fit carrying externally-owned values (e.g. trained weights) self-contained. For a block that is computed rather than fixed, a package registers a C par-loader (rxRegisterParLoader()) that runs once per solve, single-threaded, after the global parameter matrix is laid out and before integration; rxInjectedPars() reports what it changed, and those values are saved on the solved object so re-solving reproduces them in a session where the injecting package's buffer is gone. A loader registered with rxRegisterParLoaderNamed("<pkg>:<fn>") runs only for a model that flags that name with rxParLoader(ui) <- "<pkg>:<fn>", so an injector cannot reach an unrelated model; an unnamed loader keeps running on every solve. rxRegisterDydtForce() adds a term to a state derivative at the end of the generated model's RHS, so it is integrated like any other -- it runs inside the parallel per-subject solve, so such a callback must be thread safe and must check neq[0] before writing a dydt slot. rxRegisterUiPrep(name, fn) calls fn(ui) at the start of every ui solve, before parameters are loaded, so a package can rebuild C-side state that a saved-and-reloaded ui no longer has; resolve positions by name rather than a stored index, and keep it cheap and a no-op for models it does not own, since it runs on every ui solve. A failing prep hook is downgraded to a warning. Pair each registration with rxRemoveParLoader() / rxRemoveDydtForce() / rxRemoveUiPrep() in .onUnload(). See the solve-time hooks article.
Building a symengine environment with rxS() no longer rebuilds its function symbols on every call (#1283). The opaque symbols it loads (linCmtA(), delay(), lag(), the derivative helpers, ...) were made by splicing each name into a fresh function body, so R created -- and byte-compiled -- about 250 new closures per call. They now share one body, are built when the package itself is built, and are reused by every symengine environment; a user function registered at run time with rxFun() or rxD() is built the same way on first use. This removes a fixed per-call cost for every consumer that loads models into symengine repeatedly, such as an nlmixr2 fit; the saving is largest for small models and in a pkgload::load_all() session, where the discarded closures were byte-compiled again on each call.
Translating a model no longer grows process memory without bound (#1289). Three leaks, none of them visible to gc(), rxUnloadAll() or rxClean(), and all of them paid again every time an already-translated model was translated again:
- .rxModelVarsCharacter() derived the parse prefix from tempfile(), so it differed on every call. rxTrans.character() is memoised and memoise keys on its arguments, so each call missed the cache and added an entry to it -- about 0.4 MB per rxModelVars()/rxNorm() call. The prefix is now derived from the model, which is the only thing it has to distinguish. Translating also leaves the parsed model in the C parser's state, which the code generator reads, so the two callers that translate for that state (rather than for the returned model variables) now ask for a real translation -- otherwise a cache hit left the parser holding some other model and rxDelete() followed by $compile() could not regenerate its code.
- reset() (src/tran.c) allocated tb.lho beside tb.lh, but parseFree() freed only tb.lh, leaking MXSYM * sizeof(int) (~200 KB) per parse.
- .udfAddToSearch() appended the calling environment to a list that was never pruned, pinning a call frame per $ on a rxUi, per rxSolve() and per rxode2(); its index also had to mint a new name per environment, and R interns names for the life of the session. The list is now bounded by options(rxode2.udfSearchLimit = ) (default 20, oldest forgotten first) and membership is tested by identity.
Repeating the reprex from the issue -- 500 rxNorm() calls on one model -- grew process memory by ~190 MB before and does not measurably grow it now.
ind_solve() now solves the subject it is asked for, whatever order the solve loop is in (nlmixr2/nlmixr2est#1020). Its cid argument is a subject id -- ind_solve() itself indexes rx->subjects[cid] with it -- but most of the per-individual drivers it dispatches to mapped it through rx->ordId first, i.e. read it as a position in the run-time-ordered solve sequence. The two readings agree only while rx->ordId is the identity, and sortIds() deliberately reorders subjects most-expensive-first once there are at least throttle times more threads than subjects. From then on an external per-individual driver -- such as nlmixr2est's FOCEi, which solves one subject at a time through ind_solve() -- had its subject id reinterpreted as a position, so the wrong individual was integrated while the caller attributed the result to the subject it asked for. A fit's objective function and estimates therefore depended on the solve order, and since that order comes from wall-clock timing, the same fit on the same data could give different answers from run to run. Every driver now reads the argument as a subject id and the position -> id mapping happens in the par_*() loops that walk positions, so the load-balancing order is unchanged and rxSolve() results are unaffected.
method="lsoda", "lsode", "bdf" and "indLin" now draw the same random numbers for a subject as every other solver does. Each par_*() loop seeds the per-subject stream immediately before solving that subject; these four seeded seed0 + solveid - 1 where the others seed seed0 + id, so a model containing rxnorm() or similar gave a different simulation under lsoda than under liblsoda from the same seed=, and the first subject was seeded outside the block setRxSeedFinal() reserves, which could repeat a seed used by an earlier solve. Simulated values from these four methods therefore change; the other methods are unaffected.
Solving twice in one session with an explicit seed= no longer re-uses seeds the first solve already spent. Each par_*() loop claims a block of per-subject seeds with getRxSeed1(cores) but consumes one per subject, so it has to close the block with setRxSeedFinal(seed0 + nsolve); 97 solvers never did, leaving the global seed advanced by cores rather than by nsolve. par_cvodesadj() additionally never called setSeedEng1() at all, so its subjects inherited whatever stream happened to be current instead of a per-subject one. Simulated values from the affected methods change.
A matExp() sensitivity system is now exponentiated as k independent blocks rather than as one n(1+k) square. rxSensMatExp() emits a rate matrix that is block lower triangular with the same diagonal block A in every block row, so its exponential is k Frechet blocks of size 2n sharing one exp(A*dt). The structure is read off the operand -- the emitted diagonal blocks are assignments of the same constant, so they agree bitwise -- and the output accumulator and any infusion or forcing columns are carried along with it, so the assembled answer is the same matrix in a different summation order and agrees with the one it replaces to 1e-15. The split is taken only where it is cheaper than the exponential it replaces, and never changes which solver path a model takes. Measured on an optimized build, 40 subjects with distinct parameters over an irregular 100 point schedule, single thread: on the exponential path the solve is 1.0x to 4.1x faster, growing with the number of sensitivity parameters and with the compartment count (1.5x at two compartments with three parameters, 2.1x at three compartments with four, 4.1x at three compartments with seven). Models that still reach the iterative path are unchanged (1.03x), since the exponentials are a small share of their time. RXODE2_INDLIN_NO_BLOCK_EXP=1 forces the split off, and rxIndLinSteps() reports how many exponentials took it as blockExp.
A prior distribution can now be set by piping, not only written in the ini({}) block (#1254):

mod |> ini(prior(tka) ~ dnorm(0, 10))
mod |> ini(prior(eta.cl, eta.v) ~ invWishart(4))

Piping a prior replaces whatever was on that parameter, the way piping a label or an estimate does. The line is validated by lotri in the context of the real parameters rather than by a second implementation here, so a piped prior is checked exactly like one written in the block.

Note the normal prior shorthand keeps its piping meaning: mod |> ini(tka ~ 4) still changes the initial estimate, as it always has. Use the explicit prior() form to set a prior by piping.

Bug fixes

sortIds()'s run-time solve ordering is reachable again. The throttle is documented (and was originally written) to SUPPRESS the sort when nsubject * throttle <= nthreads; a refactor flattened the suppress-branch into the sort-branch without negating the comparison, so the sort was taken only when threads outnumbered subjects -- the one regime the throttle exists to exclude. At the default throttle of 2 a 131-subject fit needed 262 cores before it would reorder anything, so rx->ordId stayed the identity on any ordinary machine and the ordering was dead code. The comparison is now nall * throttle > cores, evaluated in 64 bits because throttle is user-settable and the product overflows 32. .rxSortIdsWanted() exposes the gate so the direction is asserted by a test rather than by a comment.
sortIds() now sorts in C++ instead of calling back into R's .order1(). The sort runs once per solve pass of an estimation, so with the gate reachable again the data.table round trip (~300us for a few hundred subjects, ~1s per fit) cost more than the ordering it computes saves: correcting the gate alone measured ~8% SLOWER on a 131-subject SAEM fit on 4 threads, and correcting it with the C++ sort measured ~4% faster. Sorting in C++ also removes a latent truncation: with forderForceBase(TRUE), or with data.table absent, .order1() drops NAs and returned fewer than nall positions, which left the tail of rx->ordId holding stale entries.

On that fit the ~4% is not the load balancing the sort is named for. Sorting by ascending cost, and a permutation carrying no cost information at all, are just as fast; a rotation, which reorders without scattering, is not. What pays is that the subjects a thread team works on concurrently stop being neighbours in rx->subjects. Scheduling the same identity order in coarser chunks does not reproduce it, so it is not simple adjacent-subject false sharing -- the per-subject slices of the gsolve slab are worth a look on their own. The cost ordering itself is worth at most ~0.5% here, which is the whole makespan schedule(dynamic,1) leaves on the table at ~98 subjects per thread.

Compilation

The Fortran sources are no longer compiled with a C-only diagnostic flag on check configurations whose FFLAGS carries one. flang accepts -Wall but reports it unusable once per Fortran file, which R CMD check collects as a significant installation warning. The flag comes from the configuration rather than from rxode2, which sets no PKG_FFLAGS, so configure drops it from the Fortran compile line only when the Fortran compiler is flang and flang itself reports it unusable. Every other toolchain, gfortran included, compiles exactly as before.
getSolvingOptionsInd() now walks the subject array at the stride it was allocated with rather than at its own translation unit's sizeof(rx_solving_options_ind). src/rx2api.c is the package's ABI surface and is compiled separately from the code that owns the array, and R's default make rules track no header dependencies, so an object file whose own source did not change is linked in unchanged after a change to the struct. The two views then differed by exactly the appended bytes and every subject but the first was read from the wrong address -- silent heap corruption for any caller that walks subjects, reported downstream as an absurd allocation size or a segfault (nlmixr2/nlmixr2est#1039). Fields keep their offsets by the append-only convention the struct already documents, so the stride was the whole of the disagreement.
The per-thread linCmtB() object is a tagged struct rather than an anonymous one named by its typedef, which is what recent clang warns about (-Wnon-c-typedef-for-linkage) since its members are C++ (Eigen matrices, a Stan object). The warning was the macOS R CMD check failure.
A model that fails to build now shows the compiler's own error lines (and only those -- warnings and progress chatter are dropped, and the list is capped by options(rxode2.compileErrLines=)), followed by how to get the rest (rxode2::rxLastCompile("stderr") for the full compiler output, rxode2::rxLastCompile("c") for the generated C code). The Rtools/C-compiler advice is only given when the failure actually looks like a toolchain problem: a diagnostic naming a source file and line is about the code that was compiled, so the message says the generated C code is at fault and points at the issue tracker, while a driver, linker or loader that fails without reaching the source still gets the setup advice. Previously every failure blamed the toolchain, which sent users off validating Rtools when the compiler had already named the generated-code defect (#1197).
A model that compiles but will not load reports what the loader said rather than the loader's error replacing the diagnosis, and a build failure found without recompiling (the model's dll was already present) no longer errors with could not find function ".badBuild" or reports a previous model's compiler output.
Re-compiling a linCmt() model from its own rxModelVars() no longer fails with implicit declaration of function 'linCmt'. Whether to expand linCmt() was read from a parser global that only an actual parse refreshes, and rxGetModel() returns model variables it is handed without re-parsing them, so the expansion was skipped -- or run on a model that had no linCmt() -- depending on what happened to be parsed last. The model itself is asked now, so rxode2(rxModelVars(ui)) builds and the result no longer depends on build order (#1227).
rxLastCompile() now prints its section rules -- cli::rule() was called but its result was never messaged -- and takes what= to choose which sections are messaged (rxLastCompile("stderr") for the compiler error alone). The returned list is unchanged.
The statement form of ifelse() -- ifelse(cond, stmt, stmt), where each branch is a statement rather than a value -- now compiles anywhere in a model. Its handler appended if ( to the code buffers without first clearing the preceding statement's text, so the generated C ran the two together (kin=3if (t<2) {) and only a model whose first statement was an ifelse() compiled. The construct now emits and normalizes exactly like the equivalent if (...) {...} else {...}, so it round-trips through rxNorm() and translates for symengine derivatives (sensitivities, FOCEi) the same way (#1211).
rxCompile() now re-parses the model it is handed whenever the parser's current model is a different one. Code generation reads the parser's global model state, and the old guard only checked whether some model was loaded, so a re-compile requested while an unrelated model was parsed wrote that other model's C under this model's name and handed back its model variables. Building a model with rxode2() never hit this (it parses, then compiles immediately), but re-loading one whose .so is gone did -- as when a saved fit is restored in a new session, since its DLL lived in the original session's tempdir(). Such a fit came back solving a different model, e.g. a restored SAEM fit failing with "The following parameter(s) are required for solving: eta.v, eta.cl".
Event ("jump") sensitivities now compile when a dosing modifier (dur(), f(), alag(), ...) depends on more than one estimated parameter. Each such parameter contributes its own assignment line to the same generated buffer, but the rewrite of nlmixr2's indexed THETA[n]/ETA[n] to the codegen locals _THETA_n_/_ETA_n_ only collected the indices used by the first line, so an index appearing only in a later line survived as raw symengine array syntax and the model failed with "'ETA' undeclared". This hit any model with, say, a food-effect duration built from two etas, whether or not the parameters were mu-referenced (#1196).
Building a model no longer hangs forever on a lock left behind by an interrupted session, and two processes no longer build the same model at once. rxTempDir() exports itself with Sys.setenv(), so every subprocess (a testthat parallel worker, for one) inherits ONE shared build directory -- but the build lock was acquired as file.exists() followed by sink(), a check-then-create pair that does not exclude, so two processes could both see no lock and both compile the same artifact into that directory. The losing write surfaced as "error building model", "cannot open the connection" or "cannot change working directory", depending on which step it lost. Separately, nothing ever removed a lock whose owner had been killed, and the wait for it was unbounded, so a single interrupted compile wedged that model for the life of the cache. The lock is now taken with dir.create() -- the portable atomic test-and-set -- and the wait for another builder is bounded (options(rxode2.buildLockTimeout=), 300s by default) before the abandoned lock is reclaimed.

Mixture models

A mix() model whose call has been expanded by symengine -- the form every estimation method's prediction model is built from -- is now still recognized as a mixture model. The expansion emits one reserved rx_mixsel_<k>_<n>_ selector per component, spelling out the component count the dropped mix() call carried, so mixnum reports it and a per-individual mixest supplied in the data or in iCov reaches the solve. The total is spelled out rather than inferred from the largest selector present, because a component whose expression folds to zero -- any sensitivity with respect to an eta only one component uses -- drops its selector out of the expression entirely. Previously such a model parsed with no mixture at all: the mixest column was discarded, ind->mixest stayed 0, and every mix()-derived variable solved as 0 -- which silently corrupted the predictions in the fit table (nlmixr2/nlmixr2est#1041).
ind->mixest is now set when the value is read from the data, not only inside _mix(). A model that reads mixest without calling mix() never ran _mix(), so the supplied assignment never reached it.
A mixest or mixunif column in iCov now splits a homogeneous event group when the model is a mixture model. Subjects that share an event table are solved as one group, and the group was only split on iCov columns that are model parameters; mixest is a reserved variable, so the whole group took the first subject's component.
rx_mixsel_<k>_<n>_ is now a reserved variable name, like mixest, mixnum and mixunif. A model cannot use it for anything else, and selectors that disagree with each other, or with a literal mix() in the same model, about the number of components are a syntax error.
Note that the expansion drops the mixture PROBABILITIES along with the mix() call, so an expanded model can be told which component a subject belongs to (mixest) but cannot sample one from a supplied mixunif. Simulation from mixunif needs the mix() call itself, which every hand-written model keeps.
An iCov column that a homogeneous solve group is split on no longer drops the subject when its value is NA. The split key came from interaction(), which is NA for a row with any NA, and the split() it feeds discarded that row -- so the subject vanished from the solve output instead of being rejected.

Model piping

Model piping no longer promotes a reserved rxode2 variable to a population parameter. Appending or prepending a line that used t, time, tlast, newind, rxFlag, one of the M_ constants or pi/NA/NaN/Inf added it to the ini({}) block, and the resulting model then failed to parse with "the following parameter(s) were in the ini block but not in the model block". Reserved names are now retained as-is, and the list comes from the parser itself rather than a second copy in R.
rxRename() now refuses to rename a parameter to a reserved rxode2 variable. rxRename(t = tcl), rxRename(lhs = tcl) and rxRename(cmt = tcl) produced the same unparseable model, and rxRename(pi = tcl) or rxRename(E = tcl) produced a model that parsed but silently ignored the renamed parameter, since the name reads back as the constant.
Piping a model's ini() into another model no longer silently leaves shared random effects behind. Three cases dropped an eta with no error and no message, leaving the destination model on its own initial estimate: when the two models shared exactly one eta, when a shared eta was fix()ed (it came across unfixed), and when the source model had random effects at more than one level (| occ), which dropped every eta.
ini() piping of a random effect now honors unfix() and a | condition. mod |> ini(eta.ka ~ unfix(0.7)) changed the estimate but left the eta fixed; mod |> ini(eta.occ ~ 0.2 | occ) stopped with incorrect number of dimensions; and a correlated block, mod |> ini(eta.a + eta.b ~ c(0.6, 0.01, 0.3) | occ), stopped with argument is of length zero.
A piped correlated block now records its covariance at the level of the etas it links rather than always at id, so a | occ block no longer lands split across two omegas.
ini() piping now says when a piped | condition does not match the level the random effect already sits at. Piping an estimate does not restructure the model, so the eta keeps its own level, which used to happen silently.
A piped random effect keeps its label the way a piped population parameter always has; subsetting the omega used to drop the labels with the estimates.
ini() piping no longer adds a covariance between two random effects that sit at different levels. The resulting $omega could not be assembled at all, so the model stopped as soon as anything asked for it.

Estimation / symengine translation

A model variable named after one of symengine's constants (e, I, Catalan, GoldenRatio or EulerGamma) is no longer shadowed by that constant in the symbolic layer. rxS() bound the constants into the model environment, so reading the name back gave the constant rather than the model variable and differentiating by it failed with "Input is not a SYMBOL"; the remaining symengine::D() call sites in the Jacobian, adjoint, delay-differential, event-sensitivity and mu-referencing code also passed the model-side name straight to symengine, which for the event-sensitivity code silently dropped the term instead of erroring. A matExp()/indLin() model likewise emitted k_p_q=exp(1) for a rate constant that was the parameter e, lag(e, 1) lagged Euler's number, and a delay whose duration depended on e lost its breaking-point correction terms and then computed its jump amplitude from M_E (#1359). E is the one exception: it is the symengine spelling of the model language's M_E, so the environment's E still means Euler's number and a model variable of that name is reached through its internal name instead.
A model using the modulo operator %% can now be estimated. %% was missing from the infix operator tables of the if/else rewriter (rxPrune()) and of rxOptExpr(), so both emitted it as the prefix call %%(a, b), which is not parsable rxode2. Since every nlmixr2 estimation method runs those two stages, a model that solved fine failed to fit with a syntax error -- blocking %% as the way to write a square-wave or circadian time-dependent parameter. Operands that are not a plain name or number are parenthesized, as the grammar requires (#1229).
floor(), ceil(), round(), trunc(), sign(), fround(), fprec() and fsign() can now be used with the nlmixr2 estimation methods. They parsed and solved, but symengine's Math group generic has no method for them, so loading such a model raised non-numeric argument to binary operator and no estimation method could run it -- which ruled out floor(time/24), the natural way to write a circadian or square-wave switch. They are now loaded as opaque function symbols (like rxMod()) and are locally constant, so their derivative is 0 at every order. fsign(x, y) transfers the sign of y onto abs(x), so it gets a real derivative instead: sign(x)*fsign(1, y) in x and 0 in y (#1230).
Every other parser-known function symengine has no method for now loads too, rather than silently corrupting the model. This covers the special functions (bessel_i(), bessel_j(), bessel_k(), bessel_y(), logspace_add(), logspace_sub(), fmax2(), fmin2(), gammaq(), gammapDer(), gammapInv(), gammapInva(), gammaqInv(), gammaqInva()) and the derivative helpers rxode2 itself emits (llikNormDmean(), dSELU(), d4GELU(), d2PReLU(), dSwish(), ...). The failed assignment used to be stored as the variable's value and written into the model as <var>=.expr, which failed later with no hint of where it came from -- or not at all, when nothing read the variable. The set is now a deny list of the functions symengine differentiates itself, so a function added to the parser is loadable by default, and an assignment that still cannot be loaded says which variable and why instead of continuing.
ftrunc(x) builds. Its arity was recorded as two arguments while C's Rf_ftrunc() takes one, so ftrunc(x) was rejected by the parser and ftrunc(x, digits) failed to compile -- the function could not be used at all.
dSwish() can be used with the estimation methods. Its symengine expansion was missing a closing parenthesis, so the text could not be parsed back and the model failed to load.
The parser no longer accepts a function it cannot generate compilable C for. abs0() and polygamma() exist only between rxToSE() and rxFromSE() (abs0(x) is written abs(x)/fabs(x), and polygamma(n, x) is psigamma(x, n)), and d2PReLU() had no implementation anywhere -- PReLU() is piecewise linear, so its second x derivative is the literal 0 rxode2parseD() already returns. Writing any of the three built C with an undeclared function, which rxode2 reported as a code-generation bug and asked the user to file; they now fail at the model text with the usual unsupported function message. Both symengine directions still convert them.
The description of fsign() in rxSyntaxFunctions said abs(x)*sign(y), which is wrong when y is 0: the function carries the sign of y onto abs(x) and treats 0 as positive, so it returns abs(x) there rather than 0.
rxDfdy() (and rxModelVars()$dfdy) report an ETA derivative as df(A)/dy(ETA[1]) instead of leaking the internal name df(A)/dy(_ETA_1_). Both THETA[n] and ETA[n] are translated back from their internal spellings for display, but the ETA translation was written into the buffer and then unconditionally overwritten, so only THETA[n] survived.
A model comparing a string covariate against a literal (e.g. `cl <- exp(tcl
- eta.cl + cllow * (LowID == "Yes"))) no longer translates to an undefined parameter. A character literal reaches symengine as the symbol rxQ____rxQ, and while the R translator decodes it back to a quoted string, the C translator added in 5.1.7 emitted the symbol verbatim -- so the generated model carried LowID==rxQ__Yes__rxQand solving it failed with "the following parameter(s) are required for solving: rxQ__Yes__rxQ". Only symengine's own underscore naming reached the C path (the bracket form declines and falls back to R), which is why it surfaced in the sensitivity models anlmixr2estFOCEi fit builds rather than in a plainrxode2()call. The C translator now decodes the literal, and hands the expression back to the R translator for any byte whosedeparse1()` spelling it cannot reproduce exactly.
Symbolic translation now simplifies constant arithmetic instead of emitting it. A fully constant expression folds to its value (1/gamma(2) is 1, not 1/1), extending the fold that was already applied to the right-hand operand of every binary operator, and the arithmetic identities are applied: x/1, x*1, 1*x, x+0, 0+x and x-0 all reduce to x, the same identity as the x^1 rule that was already there. Only the right-hand operand is folded, so 0-x and 1/x are correctly left alone. The named constants still win, so pi*2 remains M_2PI rather than becoming 6.28.... This shows up most in generated sensitivity code, where differentiating leaves a great many *1 and +0 terms behind; the emitted values are unchanged.

Event translation

A steady-state dose into a compartment with a modeled alag() pushed from inside a model now expands the way the event table expands it, so the two spellings of the regimen agree. A steady-state dose has to be solved unlagged while the dose the subject receives is lagged, and only the event table split the record into that pair; the pushed form solved to something else (in the reported case the two differed by 9.05, and now agree to 9e-07) (#1349).
addl no longer repeats a pushed observation, "other" (evid=2) or reset (evid=3) record. A pushed evid_(t, 3, ..., addl = 2) reset the system three times; the event table warns and ignores addl for those records, and the push path now does the same.
A pushed phantom dose (evid=7) keeps a modeled rate or duration instead of silently becoming a bolus, and a pushed evid=2 record naming a compartment turns that compartment back on -- both matching what the same row does in the event table.
evid_() now accepts a negative compartment, by name or number (evid_(t, 2, 0, -depot, 0, 0, 0, 0)), to turn that compartment off -- the same signal the event table takes from a negative CMT column.
A steady-state constant infusion (ss=1, ii=0, amt=0) written as a hand-encoded classic internal evid (>= 100) carrying a duration now errors in the event table too. The runtime push path already refused it (#1350); the event table accepted it and steady-stated the compartment to zero.
A split bolus now splits every record a dose translates to rather than only the first, which a lagged steady-state bolus needs.
A steady-state constant infusion (ss=1, ii=0, amt=0) pushed from inside a model with a duration -- modeled (rate=-2, e.g. via evid_()), fixed (infuseDur()), or a hand-encoded classic internal evid (>= 100) -- now errors instead of silently steady-stating the compartment to zero. That combination never had a usable rate (a constant infusion never turns off, so there is nothing for the duration to measure), and the event-table path already refused it; the runtime push path did not check for it (#1350).
Piping an omega block into a model with ten or more etas no longer permutes the etas that were not piped over. They were renumbered with factor(paste(neta1)), which sorts the numbers as TEXT -- "10" before "2" -- so the survivors came back in an arbitrary order. That splits a correlated block across the matrix, and it can renumber a repeated (same()) block ahead of the block it repeats, which has no representation at all since the linkage is a relative offset backwards ($omega then errored with "must refer to an earlier parameter").
rxRename() now follows a repeated (same()) block's marker to the new name. The block a repetition mirrors is recorded BY NAME in the condition column -- which is what lets the marker survive renumbering -- so a rename has to be followed too; left alone it pointed at a name that no longer existed and $omega refused to assemble ("refers to '', which is not in this block").
Nested (inter-occasion) simulation gave every random effect the wrong variance whenever a level carried more than one parameter. The omega a level draws from is laid out occasion-major, with the parameters inside each stamp, but the expansion indexed it parameter-major, so the two were transposed: with lotri(a ~ 0.01, b ~ 1, cc ~ 100) | occ every parameter in occasion 1 drew variance 0.01, every one in occasion 2 drew 1, and every one in occasion 3 drew 100. A single parameter per level is unaffected, which is why this went unnoticed. Any covariance specified within an occasion was likewise placed between occasions of one parameter rather than between the parameters of one occasion (#1345).
Translating an event table no longer slows down with the number of subjects alone. Whether an id had been seen was tested once per input row against vectors holding one entry per subject, as std::find() linear scans, so the cost was O(rows * subjects): with the row count held fixed at 120000, going from 100 to 12000 subjects cost 12.6x. The membership tests now use a set, which is flat -- 15x faster at 12000 subjects -- and the vectors are kept for their size and for the order the "IDs without observations" warning lists them in. A dosing-only id was also matched with a linear scan once per row when dropping those rows.
An infusion pushed from inside the model with evid_() now turns back off. evid=4 (reset + dose) used both slots of the translated event for the reset and the infusion start, so the stop record was dropped and the infusion ran for the rest of the solve; a modeled rate=-1/rate=-2 dose was pushed without its companion "off" record at all, so the solve failed outright with data error 997/886 instead of scheduling the infusion. The translator emits up to three records now, and a pushed infusion matches the same regimen written into the event table for fixed rate, fixed duration, modeled rate, modeled duration, evid=4, addl, ss=1, ss=2 and a split bolus. A steady-state dose pushed into a compartment that also carries a modeled alag() is still not expanded the way the event table expands it, and a steady-state constant infusion pushed with a duration rather than a rate is still not rejected the way the event table rejects it; both remain known gaps.
The last-record guard for a modeled rate()/dur() infusion start was off by one: handleTurnOnModeledRate()/handleTurnOnModeledDuration() rejected only idx >= n_all_times and then read (and, through updateRate()/updateDur(), wrote) record idx + 1. The only way to reach it was the lone modeled start the push path used to emit, so with that fixed the guard is defensive rather than a user-visible fix. Separately, _rxPushDose()'s event-array growth under-reserved when a bolus is split across compartments -- it counted the translated events rather than the records they expand into, which the idose growth beside it already did -- and now allocates the guard slot its own comment promises for ix and timeThread too.
updateRate() no longer leaves ind->idx pointing at the dose record when a modeled rate() evaluates to zero or less. Both of its error returns skipped the trailing restore of the saved index, so the corrupted value stayed live solver state until the error was picked up after the integration step. The restore now happens before the checks, as it already did in updateDur().
dose() and tad() no longer read the wrong infusion when two infusions run at the same rate into different compartments. The internal _getDur() scan that recovers an infusion's duration paired records by amount alone, so an infusion of +rate matched the first -rate it found, which may belong to another compartment's infusion (or, in the backward direction, be a bolus of the same amount). A fixed rate/duration infusion emits its stop record with the same internal event id as its start, so the scans now compare that too -- the pairing handleInfusionGetEndOfInfusionIndex() already performed. An overlapping 100 mg and 50 mg infusion both run at 10 mg/hr reported dose() = 60 for the first; it now reports 100. A steady state infusion with a modeled alag() reported dose() = 0 for the same reason and now reports the whole dose, and before the lagged dose lands dose(), tad() and tlast() are NA -- what a plain lagged infusion has always reported. Separately, an orphaned infusion end sitting at dose index 0 reports the missing start instead of falling through to the forward scan and returning a negated duration (nlmixr2/rxode2#1322).
Translating an event table no longer re-wraps its own output columns once per row. The row loop in etTrans() took each output column as a fresh Rcpp vector on every row -- a no-op cast, since the columns were allocated as the right type, but one that still paid Rcpp's preserve/release bookkeeping at every one of roughly nine sites per row. The pointers are taken once instead, which is 2.7x on etTrans() alone and 2.8-3.8x on a solve at 160000-320000 rows.
etTrans(allTimeVar = TRUE) no longer errors when an iCov covariate is supplied. Under allTimeVar every covariate is emitted as a per-row column, but the column for an iCov covariate was never allocated, so taking it failed instead of returning the covariate.
tad(), tafd(), tlast(), dosenum() and dose() no longer skip an infusion when the subject's dosing record starts with a bolus. The dose history asked for the infusion duration with the solver's running dose counter (ind->ixds), which the output pass never advances, so the lookup either failed -- the infusion was then dropped from the history entirely, leaving dosenum() un-incremented and tad() counting from the earlier dose -- or silently returned a different infusion's duration, which made dose() report the wrong amount. The duration is now looked up from the record being handled. Solving itself was never affected (#1316).
The dose history no longer measures an EXTRA dose's infusion against an unrelated dose record. The steady-state and modeled-lag infusion paths append extra doses whose amount and time live in ind->extraDose* rather than in ind->idose, so the duration lookup -- an index into idose -- had no entry to find and fell back to the solver's running dose counter, reading whichever regular record it happened to point at. For a steady-state infusion with a modeled alag() that produced a duration of 0, entering the infusion into the history with an amount of 0; another arrangement could find no matching off-record at all and drop the dose from the history. An extra dose's duration is now taken from the matching off-record in the extra-dose arrays, paired from the end of the pool so that an infusion longer than the inter-dose interval -- which overlaps itself, leaving more off records than on records -- is measured against its own off rather than against the one closing an earlier overlapping infusion (#1321).

Simulation

Residual error (sigma) is now simulated for every subject when the subjects come from the event table's id column (et(id = )) rather than from nSub=. Identical subjects are translated once and shared, so the residual draw was sized from that one representative: subject 1 got the only draws and every other observation reused the last of them, making the simulated residual nearly constant and any prediction interval built from it far too narrow. The counts that size the draw are now expanded by the shared group the way the rest of the solve setup expands them. omega was never affected (#1341).
A chunked solve (rxSolve(file=, chunkSize=)) with a sigma now reproduces the unchunked solve. rxSimThetaOmega() draws study by study, and inside one study it draws that study's etas and THEN that study's residuals, so a pre-draw that left the sigma out was a study short of the unchunked stream from study 2 onward -- every eta after study 1 was a different (still valid) draw -- and the residuals themselves were redrawn per chunk on top of that. The parent now draws the residuals for the whole solve and hands each chunk the slice its subjects own (#1339). The parent therefore holds one residual per observation, per study, for the whole solve.
confint() on a solved object now says whether the thetaMat the solve was given was actually drawn from. A thetaMat is ignored unless the variability is being simulated (nStud > 1, or simVariability=TRUE), so the message makes it clear whether the reported interval carries parameter uncertainty. Nothing is said when the solve had no thetaMat (#1308).
confint() on a solved object again uses the study dimension to build the confidence bands around the simulated percentiles when nStud > 1. When the event table holds a single subject, rxode2 numbers the nStud * nSub simulations in sim.id and emits no id column, and confint() read that sim.id as the individual identifier; it therefore ignored nStud, said "you need at least 2500 simulations", and returned plain pooled percentiles. It now recovers the study/individual split, so a nStud > 1 simulation run from a one-subject event table gives the same answer as the same simulation run from an event table that lists the subjects explicitly (#1308).
confint(mean="binom", ciMethod=) now reaches binomProbs(). The option was read out of an undocumented method argument, so the documented spelling was silently ignored and the interval always came back from binomProbs()'s own default. method= keeps working when it names a ciMethod, and is left alone otherwise (#1308).
confint() counts the individuals in the solved data rather than reading nSub back off the solve arguments, so a data set that carries its own subjects reaches the 2500 individual threshold that puts confidence bands around the percentiles. A solve of 2500 or more subjects supplied as data now returns the banded summary instead of the pooled percentiles (#1308).
A multi-subject rxSolve() with nsim/nStud > 1 no longer sizes the per-individual solve pool as nsub times the number of individual solves it needs. The over-allocation grew with the square of the number of subjects, so a large study either ran out of memory or overflowed the size to a negative number and stopped with nothing to solve -- which is what made nlmixr2est::addNpde() and vpcSim() fail on a large fit (nlmixr2/nlmixr2#412). Results are unchanged.
A chunked solve (rxSolve(file=/chunkSize=)) with nStud > 1 now simulates the omega uncertainty it was asked for. It previously returned a plausible looking result drawn entirely from the point estimate omega, with the between study variability silently gone (#1252).

The draw is made once in the parent, so every chunk shares the same per study omegas -- drawing per chunk would put subjects in different chunks into different studies. $omegaList/$sigmaList are reported on a chunked solve as they are on a plain one.

This changes existing chunked results with nStud > 1; they were wrong before. Note that reproducing a chunked solve exactly needs both seeds pinned, set.seed() as well as rxSetSeed(), because the omega draw runs on R's RNG while the etas run on rxode2's.
A chunked solve (rxSolve(file=/chunkSize=)) with a thetaMat now solves instead of erroring out with "when specifying 'thetaMat' the parameters cannot be a 'data.frame'/'matrix'" (#1263). The chunked solve hands each chunk a parameter data frame, but thetaMat was still forwarded alongside it -- a combination rxSolve() refuses -- so the solve died at any nStud, including nStud = 1.

The thetas are now drawn once in the parent, in the same draw the omega already uses, and the thetaMat/thetaDf/thetaLower/thetaUpper/ thetaIsChol arguments are stripped from what the chunks are forwarded -- so every chunk shares one draw, as it must for subjects in different chunks to belong to the same study. A thetaMat given without an omega is drawn too, and $thetaMat is reported on a chunked solve as it is on a plain one.

A joint (TNPRI) draw is the one case still refused: the omega/sigma entries a thetaMat carries under omegaSeparation="tnpri" are drawn with the thetas, which the one draw the chunks share cannot express, so a chunked solve asking for it is now a clear error rather than a result drawn from the point estimate omega. Prior simulation from ini({}) stays refused under a chunked solve for the same reason.
A chunked solve given an omega/thetaMat and a per-subject parameter data.frame now says so, rather than dying inside the draw with "Not compatible with requested type: [type=list; target=double]". The draw every chunk shares is made from a named parameter vector.
A chunked solve with dfObs > 0 no longer simulates sigma uncertainty per chunk. sigma is forwarded to each chunk (the residual draw is per observation, so it is not something a chunk's slice of the parameter table can carry), so every chunk drew its own per study sigma and subjects in different chunks ended up with different residual covariance inside the same study -- which $sigmaList did not report either. It is now a clear error rather than a wrong answer; a fixed sigma (dfObs = 0) is unaffected.

A fixed sigma still parts the two solves' random streams, which is worth knowing rather than fixed here: rxSimThetaOmega() interleaves each study's residual draw with that study's eta draw, and the shared pre-draw carries no residual draw, so from study 2 on a chunked solve draws different etas than the unchunked one. Each study's etas still come from that study's omega -- the simulation is right, it is simply not the same draw -- and this is unchanged from before, but it reaches any model with an error term (#1339).
A chunked solve is no longer refused outright for a prior written in ini({}). A prior on the population parameters is a thetaMat, which the shared pre-draw now covers. A prior on an omega block rides on arguments that draw cannot take and is still refused, now with a message that says so.
A chunked solve with nSub greater than the number of subjects the event table has is now an error rather than a solve of one subject. Chunks are cut by the ids the event table carries, so the nSub replication of a single-subject table never happened.
The lkj/separation omega strategy no longer hangs on a simulated standard deviation it cannot use (#1255). cvPost() retried a non-finite draw with no bound, but the failure is often not random: with the default omegaXform = "variance" the transform is a sqrt(), so a negative simulated standard deviation gives NaN on every attempt and the solve span at 100% CPU with no error, no warning and no way to tell a hang from a slow solve. The attempts are now bounded and the error names the cause and points at thetaLower = 0.

Initial conditions data frame

The iniDf now tolerates the prior column that lotri 1.0.5 adds for prior distributions (#1248). testIniDf()/assertIniDf() used to reject every model built with such a lotri, and the ini rows that are constructed by hand internally (adding a covariance between two etas, promoting a parameter, linMod()) hard-coded the column list and so failed to rbind() with "numbers of columns of arguments do not match". These now match whatever columns the iniDf actually has, so an iniDf without the column still works.

Parsing

Two or more subject-level random effects in one mu-referenced expression now name the clashing parameters and both fixes -- declaring one at its own level (etaVcOcc ~ 0.1 | OCC) or splitting the line -- rather than reporting currently do not theta + eta1 + eta2. The guard keeping occasion-level etas out of $nonMuEtas now reads env$info$level, where the level names live.
A variable that is used only as an argument to an adaptive dosing call (evid_(), bolus(), infuse(), infuseDur(), replace(), multiply(), phantom(), obs()) is now a parse-time error instead of an uncompilable model. These statements consume their arguments as text, so such a variable was never registered and the generated C referenced an undeclared identifier; the failure only showed up as a compiler error that looked like a broken toolchain. The message now names the variable, the argument and the function, and points at the fix (assign it to a model variable first):
```
undeclared 'DOSE' in 'amt' of 'infuseDur()'; assign first: 'amtVal <- DOSE'
```
The check runs once the whole model is parsed, so a variable assigned below the dosing statement still counts as declared (#1231).
A wrong-arity linCmtA()/linCmtB() call (e.g. linCmtA(a,1,1,0), one argument short) now reports a parse-time syntax error instead of segfaulting R. handleFunctionLinCmt() indexed its fixed argument positions without checking how many arguments were actually given, so a short call dereferenced a NULL parse node (#1266).
A param()/params() statement, and the interpolation statements (locf(), linear(), nocb(), midpoint()), no longer splice the preceding line into their own normalized text. They build that text by appending to the normalizing buffer and, unlike an assignment, did not reset it when the statement started, so "y=z*a;param(a,c);" normalized to "paramy=z*a(a,c);" -- text that no longer parses -- whenever such a statement followed an assignment or a d/dt() line (#1279).
Repeated param() statements in one model now normalize to the single merged declaration that rxModelVars()$params already reports, instead of being kept as separate statements. A generated model that appends a param() statement to an already-built model text (for example to add DV to a general-likelihood prediction model) previously left a normalized model whose first param() statement did not name every parameter, so code that read or edited that statement silently missed the later declarations. The merged statement takes the place of the first param() statement and spans the parameter vector up to the last declared parameter, so re-parsing the normalized text gives back the same parameter order (#1279).
A variable that is both declared in param() and assigned in the model now gets its interpolation recorded. Such a dual lhs/parameter takes a slot in the parameter vector like any other parameter, but the slot in the (uninitialized) interpolation vector was never written, so rxModelVars()$interp held a garbage code for it and printing it could fail with malformed factor.
Parsing a model no longer corrupts the caller's PROTECT stack. The translation table and _goodFuns were claimed on the protect stack by one function and released by another, with the whole parse in between; an Rf_error raised in that window (a model syntax error, say) unwound the stack while the outstanding count stayed set, so the next parse released entries it no longer owned and popped the caller's own protections. Callers holding a protect index across the parse then failed -- on macOS this surfaced as R_Reprotect: only 137 protected items, can't reprotect index 143 thrown out of vapply(), which made rxOptExpr() silently abandon chunking and fall back to optimizing the whole model. Those objects now use R_PreserveObject(), which an unwind does not undo.
loggamma() no longer fails to compile. It is symengine's name for lgamma() and the parser accepted it, but code generation emits the rxode2 name verbatim as the C name and there is no loggamma() in C, so a model using that spelling parsed and then failed at the compiler. (gammafn() and lgammafn() in the same table work only because they happen to coincide with Rmath.h.) Its derivative was never affected: symengine differentiates loggamma natively to polygamma(0, x), so a model differentiating it already got the exact digamma().
ceiling() is now a supported function. rxode2 knew C's ceil() but not the name R users actually write, and because ceiling was absent from the function table it fell through to the user-defined-R-function path, found base R's primitive ceiling, and reported "user function 'ceiling' requires 0 arguments (supplied 1)" -- formals() of a primitive is empty. floor() had worked the whole time. ceiling() now parses, compiles to ceil(), takes one argument, and is locally constant like ceil(), floor() and round(), so its derivative is 0 and it can be used in models that take sensitivities.
psigamma(), log1pmx() and polygamma() now check how many arguments they were given. All three guarded the count with length(x == n) instead of length(x) == n; x is a call, so x == n compares its elements and length() of that is always at least one, leaving the guard permanently true and the error below it unreachable. Too few arguments failed with "subscript out of bounds" from the missing element, and extra arguments were silently dropped -- psigamma(a,b,c) translated as psigamma(a,b) and polygamma(0,x,y) as polygamma(0,x). Correct calls are unaffected.

Model interface

ui$modelName is now always a single character string, as it was always documented to be. It came from as.character() of the substituted model expression, which returns one element per part of a call, so rxode2(readModelDb("PK_1cmt")) gave c("readModelDb", "PK_1cmt") and an anonymous model function gave a four-element vector including the deparsed body. The name is the tidied first deparsed line of the expression instead: a symbol keeps its name and a call becomes its own text (readModelDb("PK_1cmt")), unless a rxModelName() method or rxModelNameLhs() names it better. Names wider than 60 characters are truncated. An anonymous model function names nothing, so its modelName is NULL rather than a piece of its body. Values assigned by other packages (or read from models saved by earlier versions) are also collapsed to a single string on access (#1019).
A trailing # comment on an ini({}) line may now contain a double quote or a backslash. Such a comment is promoted to a label() call when the model is parsed with its source refs intact, and while the label text was escaped correctly it was then interpolated into the replacement argument of sub(), which parses backslashes and strips one level. The generated label("fixed to a "small value"") did not parse, so the model failed with a bare syntax error pointing into regenerated text rather than at the offending source line. Because the promotion only runs when source refs are kept, the same model resolved fine without them -- so a package build could be green while a test suite run with keep.source = TRUE was red on the identical file (#1195).
A trailing # comment on an ini({}) line keeps its label() when the comment itself contains a #, including the common ## comment form. The code portion of the line was matched greedily, so on a line with two # it ran on to the last one and left the first sitting in the generated code, where it commented out the label() that had just been appended. The label was dropped silently -- the model still parsed and built, it simply lost the label (#1205).
A # comment inside an ini({}) statement that spans more than one line no longer breaks the model. The comment was promoted by appending ; label("...") whether or not the statement on that line had finished, so a comment between an opening ( and its ) -- or on a line ending in an operator such as + or ~ -- put the ; in the middle of the statement and the model failed with a bare unexpected ';' pointing into regenerated text rather than at the offending source line. A comment is now promoted only where it trails a complete statement; one inside an unfinished statement stays a comment and is dropped. A comment on the line that closes the statement still becomes that parameter's label. As with #1195 the promotion only runs when source refs are kept, so the same model built fine from an installed package while failing under keep.source = TRUE (#1318).
A comment-only ini({}) line indented with a tab is no longer turned into the label of the preceding parameter. The comment-only test allowed leading spaces only, so a tab-indented comment fell through to the label branch, whose code portion then captured just the tab. The bare ; label("...") that produced parses -- a leading ; is legal -- so there was no error and the comment silently became the label of the parameter above it (#1318).
The "cannot find additive standard deviation" error tested a $predDf column that does not exist, so its multiple-endpoint hint was appended even for single-endpoint models.
A modeled ar() correlation on an endpoint written with a condition (cp ~ add(add.sd) + ar(corv) | phase1) is now found; the endpoint was matched against the left-hand side, which never carries the condition.
An endpoint's condition is no longer treated as a residual parameter, so model(cp ~ add(add.sd) | assay1) names the endpoint instead of failing with "the following parameter(s) were in the ini block but not in the model block: assay1".

Solving

The automatic ODE-to-linCmt() conversion (rxSolve(..., useLinCmt=TRUE), the default) no longer drops a right-hand side term that is not proportional to a compartment. transit() absorption, a zero-order or endogenous production rate and a dose carried in a covariate column were all parsed and then discarded by both the topology detector and the emitted linCmt() call, so the model solved was not the model written -- either identically zero, or non-zero and plausible but wrong (a transit chain silently became plain first-order absorption, reported here as nlmixr2/rxode2#1370). linCmt() is driven entirely by the event table's dosing records and has no parameter that can carry such a term, so a model containing one now keeps its explicit ODEs, and odeToLin() names the term it declined to convert.
The same conversion no longer substitutes a different rate constant for the one that was written. The emitted linCmt() call passes parameter NAMES only, so anything else in a rate coefficient or in the concentration line was discarded: - 2 * kel * central solved as if it eliminated at kel, cp <- central / (vc * 1000) reported central / vc (a thousandfold error), and a covariate factor written into the ODE ((cl / vc) * cms * central) was dropped. Detection now compares the system's own rate constants and reported volume against rxDerived() -- the same parameterization inference linCmt() itself uses, so the two cannot drift apart -- and keeps the explicit ODEs unless they agree. Folding such a factor into the parameter (cl <- exp(lcl) * cms) converts as before.
The same conversion no longer renumbers a model's compartments out from under its event data. linCmt() orders its compartments depot, central and keeps no state for a peripheral, so a model that declares d/dt(central) before d/dt(depot) numbers them the other way round: a record addressing compartment 1 by index (which includes an event table with no cmt column at all) dosed central before conversion and depot after, turning an IV profile into a plausible oral one. Such a solve now keeps the explicit ODEs. Addressing a compartment by name, and NONMEM-style data observing a one compartment model in cmt = 2, both still convert.
Every implicit method (ros4, iem, ros43, ...) and every AutoSwitch composite is much faster, because the analytic Jacobian model is no longer regenerated on every rxSolve(). The augmented model's text was cached but rxode2() was then re-run on it every call -- a full parse of a model carrying one df()/dy() line per Jacobian entry, so quadratic in the number of states -- followed by a second pass through rxSolve.default(). The compiled model is cached instead. On a single subject with seven doses over 0-168 h, "dop853+ros4" drops from 0.0199 s to 0.0132 s on a 1-cmt oral model and from 0.592 s to 0.0234 s on an 11-state PBPK model; ros4 on its own drops by the same amount. This was the whole of the composite slowdown reported in nlmixr2/rxode2#1307, and it applies equally to delay differential equations (whose default method is the composite) and to FOCEi, which solves once per iteration.
AutoSwitch composites now actually switch. dop853's stiffness detector is gated on its accepted-step count, which restarts on every dop853() call, and rxode2 calls it once per interval -- so reporting stiffness needed about 64 accepted steps inside one observation interval, which a PK interval never takes. In practice a composite only switched when dop853 failed outright, after exhausting maxsteps; "dop853+ros4" on a stiff TMDD model returned bit-identical output to plain dop853. The detector's estimate and verdict counters now persist across the intervals of a subject solve, and are evaluated on every accepted step. The same TMDD solve is now 1.7x faster than its own primary because it switches, while genuinely non-stiff models still do not switch at all.
A composite that switches no longer integrates the interval twice. The stiff secondary continues from the last step the primary completed instead of restarting from the interval start. In the dense path this also fixes an inconsistency: the fallback rewound the state and the delay history but not the observation cursor, so a segment that was supposedly re-solved kept the discarded attempt's observation values.
A composite whose primary is not dop853 -- "dop5+ros4", "bs+ros4", "f78+ros4" and the rest -- now switches on the main timeline. Those primaries' drivers ignored the stiff secondary entirely, so the composite ran as the plain primary there while its steady-state intervals did switch. The stiff Robertson problem, which neither dop5 nor bs can solve alone, now solves through both composites. dense = TRUE for a composite is documented and enforced as "dop853+ros4" only, which is what was ever implemented; asking for it on another composite used to silently run the plain primary's own dense stepper.
The documented AutoSwitch controls do something again. autoSwitchNonstifftol sets the ratio at which a step is called stiff (the same eigen_est * h / stability_size test Julia's AutoSwitch uses), autoSwitchStifftol sets the same ratio for the optimistic re-probe after a subject has already switched, so lowering it makes stiff mode stickier, autoSwitchStiffFirst starts the solve on the stiff secondary, and autoSwitchSwitchMax guards the switch back. All five had been left without a consumer when the composite became reactive. autoSwitchDtfac has nothing to mean in a reactive scheme and is now documented as accepted and inert rather than removed. The wait before the primary is probed again now doubles on each probe that fails, so a persistently stiff subject stops paying for probes.
dop853's stiffness estimate used the signed step size, so on a reverse-time integration it was negative and the detector was silently dead. Steady-state intervals solved with dop853 used scalar tolerances while the rest of the same solve used the per-compartment ones, so a sensitivity model -- whose tolerances are scaled per equation -- integrated its steady-state intervals to different tolerances than everything else.
rxSolve() on a model function's rxUi no longer loses what the model's meta block carries -- most visibly a sigma, whose residual variables the solve then rejected as unsupplied parameters. rxSolve.rxUi() is not a registered S3 method, so a call from user code lands on rxSolve.default(), which hands the model back to rxSolve() with the whole rxControl() expanded into named arguments; the meta block is only read for options the caller did not name, so naming all of them hid it. The entries meta supplies that are still at their default are now left unnamed on the way back. (A call from inside the package, including from test_check(), found rxSolve.rxUi() directly and was never affected.)
rxSolve(method="indLin") on a model function's rxUi now solves. The matExp() conversion ran before that same hand-back, so it replaced the ui with a plain model built from the ui's equations and the ini() values were never supplied -- the solve stopped asking for the population parameters. A function or rxUi is now converted on re-entry, when the simulation model and its parameters are both in hand.
A subject's sticky ODE tolerance now stays with that subject. The per-individual tolFactor (and the loosening nlmixr2est applies through atolRtolFactor_()) was folded into the tolerance arrays a thread already held, so it survived into whatever subject that thread solved next and compounded once per subject; every subject after a loosened one was solved at the wrong tolerance, with the result depending on how the subjects happened to be distributed over the threads. Each subject's tolerances are now derived from the solve's base atol/rtol/ssAtol/ssRtol and its own factor. The factor is also recorded on the subject itself rather than on the thread's copy, so it is no longer lost, and it is bounded as a multiplier instead of being clamped to maxAtolRtolFactor -- which had made a request to loosen tolerances tighten them on the next solve. A loosening requested when no subject is being solved no longer attaches itself to whichever subject was solved last.
Modeled duration (rate = -2) and modeled rate (rate = -1) doses that fall at exactly the same time now solve. Each such dose is expanded into a start/stop pair sharing one time and the solver pairs the two positionally, but the event sort keys on the compartment-bearing evid, so tied doses interleaved (start2 start1 stop2 stop1) and the solve failed with data errors 686/886 (or 797/997 for a modeled rate) -- even for doses into different compartments, which is a legal data set. etTrans() now re-pairs each start with its own stop after the sort, matching on compartment and infusion type; the pass only runs when the data set has a modeled rate/duration dose and it leaves already-correct records in place. The four data-error messages now say what the problem is instead of only naming a number (#1218).
Fixed an out-of-bounds read of the extra-dose pool while advancing to the first extra dose at or after the current step. The index was bounds checked before it was incremented rather than after, so a subject whose extra doses all precede the step -- reachable with tied modeled duration steady state doses -- read one element past the end and then dereferenced it as a record index, corrupting the heap.
A parallel chunked solve (rxSolve(file=, chunkSize=, parallel=)) no longer fails outright when the mirai daemons load a different rxode2 than the parent is running -- a source checkout, or a library updated underneath a long-lived pool. The whole control list is forwarded to each daemon by name, and rxSolve() rejects an argument it has no formal for, so a parent one version ahead lost every chunk to unused argument. A control the daemons cannot take is now dropped, with a warning naming it and the version they loaded, rather than losing the solve over a setting that version had no notion of. What they can take is asked of the daemon itself, so a matching pool drops nothing.
rxSolve()'s thread-safety dispatch no longer has a code path that could silently substitute liblsodaR for the solving method a user explicitly requested. The path was reachable only once a currently-disabled model classification was re-enabled, so it never fired in a release, but every par_* solver already reseeds its RNG per subject as a pure function of that subject's position, so the swap was never buying the reproducibility it was meant to protect (#1240).
An event pushed by the model with evid_() (and the bolus(), infuse(), replace(), multiply(), reset(), phantom() and obs() helpers) now gives the same solution as the identical event written in the data, on every solving method. The ODE methods fired the model body from dydt() at the start of the next integration interval: the time value was right, but the event was inserted only after the solver had been asked to integrate past it, so liblsoda, dop853 and cvode applied the jump one observation late and lsoda dropped it altogether. evid_() now fires from a single shared point at the record itself -- once per distinct record time, with the pushed event landing in the slot immediately after that record -- so ODE, linCmt() and indLin() models agree with each other and with the explicit event. A model that pushes an event but defines no lhs variable also compiled to an empty calc_lhs() and never pushed anything; its body is now emitted. A pushed event that extends the timeline past its original last record is no longer truncated by the dense dop853 driver, and dense=TRUE is now dropped (with a warning) for a model that pushes: a dense segment integrates across every observation between two key events at once, which cannot honour an event the model decides on at one of those observations. A model that combines delay() with a pushed event is now an error rather than silently returning one of two wrong answers: delay() requires the dense output that a pushed event rules out.
An adaptive dosing helper guarded by t == <mtime> no longer pushes its dose twice when that mtime() names a time the event table already contains. The same model written as a function (ini({})/model({})) and as an rxode2({}) block disagreed, because rxSolve() defaults to useLinCmt=TRUE for a function model: that one was auto-converted to a linCmt() model, and the linCmt() driver fired evid_() from both its own internal model evaluation and a second pass for the same-time observation. Both forms now push once, and the doubled dose (silent except in the state at the next time point) is gone.
rxSolve() no longer returns silently wrong, run-to-run varying results when a multi-row params data.frame (one parameter set per id) is combined with omega = NA or sigma = NA. c() on a data.frame drops the data.frame class and yields a ragged list -- the per-id columns keep their length while the appended zeros have length one -- which was then read out of bounds while solving, so the random effects that omega = NA fixes at zero were filled from unrelated memory instead. With eight or more subjects this changed the solved values on every solve of identical input, occasionally to non-finite ones. A multi-row params matrix hit the same problem from the other side: c() dropped its dim, so omega = NA/sigma = NA failed outright with "The following parameter(s) are required for solving". omega = NA on a model with no between subject variability (which failed with "invalid 'times' argument") and sigma = NA on a model with no residual error are now the no-ops they should be.
An omega/sigma entry whose variance is zero (say eta.base ~ fix(0)) is now supplied to the model as a literal zero when params is a matrix, as it already was for a data.frame or a named numeric vector. Such an entry is dropped from the matrix that is simulated from, so a matrix params reached the solver without it and rxSolve() failed with "The following parameter(s) are required for solving". A matrix that did supply the item kept its value where a data.frame had it replaced by zero; both replace it now, and zeroVarParamHandle= chooses (see New features).
A params matrix that supplies a random effect is now recognized as supplying it, so that effect is no longer simulated on top of the supplied value. rxSolve() decided whether params already had a random effect with names(params), which is NULL for a matrix -- its names are the column names -- so the answer was always "no". The supplied column was silently ignored and a random draw used in its place: with eta.base = 100 supplied for every subject, a data.frame gave 101 102 103 ... and a matrix gave 0.92 1.84 2.67 .... There was no warning, and the values look reasonable unless you know what they should be.
Supplying a value for one random effect no longer stops the others from being simulated. A supplied effect is dropped from the omega before solving, but the subset that drops it took a single remaining effect down to a scalar, whose dim is NULL, and all(NULL == c(0L, 0L)) is TRUE -- so the whole omega was dropped and the remaining effect was neither simulated nor supplied (The following parameter(s) are required for solving: eta.b). The matching sigma code already guarded this.
?rxSolve no longer states that method="dop853" cannot solve in parallel. Only the non-reentrant Fortran COMMON block solvers (lsoda, lsode, bdf) are excluded from threading by solveMethodThreadSafe(); dop853 has been thread-safe and has honored cores. Documentation only, no behavior change (#1305).
rxMemoryEstimate() no longer double-counts the ODE state output matrix. gsolve_n0 is a piece of gsolve, not a sibling of it -- rxFillMemLayout() adds n0 into gsolve_total -- but total summed every reported element, so it counted the single largest allocation of an ordinary solve twice and could approach double the real figure. gsolve_n0 is still reported (and still printed indented under gsolve), it is just no longer added to total. The out-of-memory guard in rxSolve() and the chunk sizing in .rxOomChunkSize()/rxSolveChunked() both act on total, so a solve that fits is no longer refused and chunks are no longer about half the size they should be.
rxMemoryEstimate() now counts the per-individual event and solve arrays. When op$indOwnAlloc is set -- which rxSolve() defaults to the model's evid_ parser flag, so any dose-pushing model (bolus(), obs()) gets it, as does anyone passing rxSolve(..., indOwnAlloc = TRUE) -- rxAllocInd() gives every individual its own dose/ii/all_times/timeThread/evid/ ix/idose/solve arrays, and gsolve is still allocated at its full size regardless, so those arrays are memory on top of it. The estimate ignored them entirely, which understated total -- the direction that makes an out-of-memory guard useless. They are now reported as an indOwnAlloc component, computed by rxFillIndAllocTotal() in inst/include/rxMemoryCalc.h alongside the rest of the layout.
rxMemoryEstimate() now scales the event-indexed buffers with nSub, nStud and nsim. The replicated subject count was applied to the subject total but never to the event total, so rxControl(nSub = 100) on a one-subject table reported the one-subject figure for gsolve_n0, gall_times, gevid and gpars -- an undercount of the largest allocation by the full replicate factor, and again in the direction that makes the out-of-memory guard useless. nsub and nsim now mean to the estimate what they mean in rxData.cpp: nsub is the subjects of ONE simulation and nsim is how many times that block is replicated, so nStud lands in nsim (where it also correctly pays for the extra-simulation copies in gall_timesS) and nSub grows the events of a single simulation. effectiveSubs still reports the total individual count.
rxMemoryEstimate() sizes ordId by individuals rather than events. rx$ordId is the solve ORDER over individuals -- nsub * nsim ints -- but the estimate charged one int per event, overstating it by the number of events per subject.
rxMemoryEstimate() now reports gEtaPre, the pre-generated eta draws. rxPreGenEta() mallocs nsim * nsub * neta doubles before the parallel solve loop whenever the model has etas and a nonzero omega, and none of it was counted.
rxMemoryEstimate() now reports gSampleCov, allocated when rxControl(resample=) asks for covariate resampling, and counts the per-thread pointer table that accompanies the gInfusionRate buffers.
rxMemoryEstimate() now charges the two per-individual history buffers: the delay() dense history (ind$delayHist) and the linCmtB() output-time rate history (ind$linCmtRateHist), neither of which was counted at all. Both grow by doubling inside the solve rather than being sized up front, so unlike every other component these are a documented BOUND rather than a mirror of a calloc: the capacity the doubling reaches at roughly one stored step per event, floored at the initial allocation. They are zero for a model that uses neither.
rxMemoryEstimate() no longer returns NA for very large event counts. The per-subject event totals were summed in integer arithmetic, so a solve past 2^31 events -- exactly the size this estimate exists to judge -- overflowed and then failed with "missing value where TRUE/FALSE needed".
An integer or logical covariate column no longer makes rxSolve() quadratic in the number of rows. While building the solving data set each covariate column was coerced to double once per output row; for a column that is already double that coercion is free, but for an integer or logical column it allocated and converted a copy of the whole column on every row. The result was correct but progressively slower -- roughly 14x a double column at 20000 rows and 90x at 160000 -- and the only hint was the timing, since a logical covariate arises naturally from ordinary R code such as flag = x == "value". Each covariate column is now coerced once, so every storage mode solves at the speed a double column always did.
A model that reads CMT as a covariate no longer slows down with the number of subjects. CMT is carried as an integer column, and the copy of each covariate into the solving buffer -- done once per subject -- coerced the whole column to double every time, costing about 3x an otherwise identical double covariate at 20000 subjects. The columns are now coerced once, as above.

Sensitivities

A model reading only some linCmtB() sensitivity directions (a FOCEi inner model with fewer etas than linCmt() parameters) solved every row as NA after a steady-state dose: the Jacobian columns nobody requested were carried into the next row's state reconstruction as the NA the buffer starts with, instead of zero.
A 3-compartment oral linCmt() model whose depot amount goes negative (a negative dose larger than what is left in the depot) no longer drops the depot from that interval's solution. The depot branch was skipped for a negative amount, which returned the wrong amounts and sensitivities under forward-mode AD and crashed under reverse-mode AD (linCmtSensType="ADr") (#1275).
A 3-compartment oral model's linCmtB(which1 = -2, which2 = 6) read (the d/d(ka) sensitivity column) is now registered by the parser; it was rejected as an unknown read, which left that column unfilled.
linCmtB() gained internal per-subject sensitivity-carry sentinels (which1 = -4 to -8) that nlmixr2est uses to keep a linCmt() eta gradient exact when a time-varying covariate makes a parameter differ between rows (-8 pins a subject's pass to the full transition advance so an event-modifier jump fed to -7 is propagated); they are not reached by a model that does not request them.
linCmtB(which1 = -3) -- the dose-time (moving boundary) sensitivity a modeled alag() on a linCmt() compartment needs -- no longer reports a biased value for an individual whose regimen does not carry one shared delay. The -dA/dt identity it rests on assumes every dose feeding the linear system is delayed together; which compartments the model lags is known when the model is built, but which ones an individual actually doses is data, so it is now decided while solving. An individual dosing only unlagged linCmt() compartments gets an exact 0 (its amounts do not depend on the delay), one mixing lagged and unlagged doses gets NA instead of the single-delay answer, and one dosing only lagged compartments is unchanged. A paired IV/oral design -- the case that most often estimates a modeled alag()/f() -- previously took a silently wrong gradient on its IV arm. The same rule fixes an individual with no linCmt() dose at all whose compartments were started from an initial condition: those amounts decay in time, so -dA/dt reported a nonzero sensitivity, but an initial condition is not delayed by alag() and the answer is 0. A plain amt = 0 dose is not counted, so it cannot refuse a regimen it contributes nothing to, and neither is a dose into a mixed model's ODE compartment -- including a steady-state infusion there, which used to cost the linCmt() compartments an answer that was exact (#1119, #1237).
linCmtSensH (the fixed finite-difference step used by the forwardH/ centralH/forward3H/endpoint5H linCmtSensType options) is now read from its own control slot instead of linCmtSensType's. rx->sensH was populated from the linCmtSensType control index a second time, so a fixed- step linCmt() sensitivity used the integer sensType code itself as its step size (e.g. 10.0 for forwardH) instead of the intended default of 1e-4, wildly distorting those finite-difference sensitivities (harmless for the AD sensitivity types, which never read sensH) (#1276).
rxode2() now refuses to build a model that calls linCmtB(which1 = -3) (the dose-time sensitivity) while its linCmt() compartments carry more than one distinct modeled alag() -- e.g. alag(depot) driven by one parameter and alag(central) by another. which1 = -3 is the derivative wrt ONE delay applied to every dose feeding the linear system; such a model previously solved and silently returned that single-delay answer instead of the true per-compartment sensitivity, which the entry point cannot compute (#1237).
linCmtB() gained a dose-time (moving boundary) sensitivity, which1 = -3: the derivative of a linCmt() model with respect to a delay applied to every dose feeding it, which is what a modeled alag() on its dosed compartment produces. which2 = -3 gives it for the reported concentration, which2 >= 0 for the amount in that compartment; chain-rule it with d(alag)/dp for the sensitivity wrt a model parameter. The system is linear and its whole input is delayed together, so the derivative is exactly -dA/dt -- it matches a finite difference to round-off for bolus, infusion, and steady-state-bolus regimens across one to three compartments, IV and oral. It reports NA for a steady-state infusion (its amounts are established analytically, so the infusion rate afterward is not well defined at that index -- #1236) and requires that every dose reaching the linear system share the same alag() (#1119).
A model that mixes linCmt() with d/dt() now expands its sensitivities once. The linCmt() call has to be resolved before the sensitivity expansion, and the model was re-parsed with calcSens= afterwards, which differentiated the already-expanded model a second (and, with eventSens=, a third) time. The result carried rx__sens_rx__sens_<state>_BY_<p>___BY_<p>__ compartments nobody asked for and an interleaved compartment layout, which the event-sensitivity map then read as a second-order Hessian block. The linCmt() text is now built first and the sensitivities expanded once, from that text. As a consequence summary() of such a model prints the linCmt() model as written, without the generated rx__sens_* equations after it (#1119).
.rxLinCmt() no longer invents a peripheral1 compartment for a one compartment oral linCmt() (nor a peripheral2 for a two compartment oral one): the compartment count it decodes includes the depot, and it was read as the number of disposition compartments. An ODE state named like the invented compartment was dropped from rxStateOde(), so it never got a sensitivity expansion -- its rx__sens_<state>_BY_<param>__ compartment did not exist at all -- and it also raised a bogus "share a name with linCmt() reserved compartments" warning (#1119).
eventSens="jump" now applies to the ODE compartments of a model that also has a linCmt(). Every linCmt() model was downgraded to finite differences because the moving-boundary jump for a modeled alag()/f() on a linCmt() compartment is not implemented; the ODE compartments of such a model carry ordinary solved sensitivity compartments and are unaffected by that. The downgrade is now limited to the models that need it: a pure linCmt() model, a reserved-name collision, or a modeled alag()/f()/rate()/dur() on a linCmt() compartment itself (#1119).
The event-sensitivity jump map is now checked against the true compartment indices rather than assuming them. The runtime injection addresses the sensitivity compartment of (state k, parameter p) as nState + p*nState + k; a model whose compartments do not lie that way falls back to finite differences instead of having jumps written into the wrong compartment (#1119).

Delay differential equations

rxOptExpr() no longer fails on a past(state, tau) whose delay duration is an expression rather than a name or a number (past(G, exp(lT)), past(G, tau*2)), which raised unsupported lhs in optimize expression and printed the duration into the middle of the progress bar. This made optExpression=TRUE unusable for such a delay differential equation; it now optimizes, and the duration follows the same common subexpression its delay() terms do, so the history stays matched to them.
A generated delay differential equation model (rxode2(..., calcJac=TRUE), calcSens=, or an nlmixr2 estimation model) now resolves the past() delay duration the same way it resolves the history itself. A duration written as an intermediate (T <- exp(lT)) was emitted verbatim while every delay() had its duration inlined, so the generated model named a duration no delay() used any more and rxSolve() rejected it with duration 'T' does not match any delay(...). This also covers a duration or a history written with THETA[n]/ETA[n], as every mu-referenced model is: they were left unresolved, and an unresolved history additionally emitted no per-parameter sensitivity pre-history at all.

Matrix exponential / inductive linearization

meOnly()/indLin() no longer write past the end of their buffers when a downstream package sets a per-individual effective state count (setIndNeqOverride()). Those buffers were sized by the effective count while the model-generated ME()/IndF()/calc_jac() bodies always index by the compiled state count, so a shortened count overran them -- quadratically for ME(). The generated code is now called through a full-size buffer and the leading effective block copied back; with no override, which is every path rxode2 itself takes, the calls and the numerics are unchanged.
rxToIndLin() -- and therefore rxSolve(method="indLin") -- now converts a model that mixes linCmt() with d/dt(). It walked $state, which counts the linCmt() pseudo-compartments (depot, central, peripheral*); those have no d/dt() behind them, so it emitted cmt()/indLin() lines for derivatives that do not exist -- one of them the literal R variable name .tmp -- and the generated model did not parse. Only the d/dt() block is converted now; the solved compartments stay with the analytic solver and are copied back after each step. A term reading a linCmt() goes to the indLin() forcing rather than into a rate constant, since a solved compartment moves within the matrix-exponential step, and such a forcing takes the iterating path so the driver refines it.
A df(<state>)/dy(<state>) Jacobian entry may now reference linCmt(). A linCmt() call retyped the whole statement, so the entry lost its Jacobian routing and was emitted into dydt(), where __PDStateVar__ does not exist: the model failed to compile. For the same reason a matExp() rate constant or indLin() forcing built from a linCmt() concentration now reaches the ME()/IndF() functions instead of reading a stale value.
rxSensMatExp(calcSens3=) now carries the indLin() forcing at third order, as calcSens and calcSens2 already did. Only the rate-matrix cross terms were generated, so third-order sensitivities of a nonlinear model were short every term the forcing contributes; the warning that said so is gone.
The Al-Mohy matrix exponential evaluated the wrong Pade numerator below degree 13. The coefficients depend on the degree, and the routine read a fixed table -- the degree-13 row -- and truncated it, which is not the degree-p numerator. The answer stayed convergent but only to a few 1e-12 where every other backend reaches machine precision, and only against an exact solution is that visible. The row is now built for whichever degree was selected.
The Al-Mohy matrix exponential returned a wrong answer for a very large matrix norm. The squaring count was returned as the factor 2^s in an int and clamped so it could not overflow, but clamping caps the scaling while leaving the norm untouched, so degree-13 Pade ran far outside its range and produced a plausible finite number: a one-compartment model with a rate constant of 1e20 returned 5.1e-08 for a quantity that underflows to zero. The squaring count is now carried as a count.
rxSolve(indLinMatExpType=) now defaults to "Al-Mohy" rather than "expokit". With the degree bug below fixed, all four backends agree to solver tolerance and take the same steps on every problem tested, and "Al-Mohy" is the cheapest per exponential: about 4-5% on a Michaelis-Menten population and 34% on a stiff van der Pol one, where an exponential-Rosenbrock step rebuilds its operator every step and the exponential cache cannot help. On a linear model the difference is unmeasurable, the cache serving nearly every call. Results move in the last digits, as any change of exponential kernel does; pass indLinMatExpType="expokit" to keep the previous one.
rxSolve(indLinMatExpType="Al-Mohy") chose its Pade degree and its scaling inconsistently, which could return a silently wrong answer or a solve that never finished. The scaling came from the Al-Mohy-Higham threshold table, whose entries each belong to one specific degree, while the degree itself came from indLinMatExpOrder (default 6) -- so any matrix with a 1-norm up to the table's largest threshold, 5.37, was evaluated at degree 6 with no scaling at all where the table calls for degree 13. Both are now taken together from the norm, as the taylor backend already did. A two-compartment linear model returned 1.8e-06 against about 1e-11 for every other backend, and one van der Pol subject at mu = 95.7866 under an exponential Rosenbrock step ran for over 390 seconds -- a bad exponential can make the error estimate unsatisfiable, so the step controller shrinks the step without limit instead of failing -- where the other backends took 0.03 s. Both now agree with the other backends, and on a 50-subject stiff population Al-Mohy goes from not completing in 418 s to 0.92 s, the fastest of the four. Consequently indLinMatExpOrder no longer applies to Al-Mohy; it still applies to expokit.
rxSolve(<function or rxUi model>, method="indLin") failed with "Can only parse scalar data". With the default useLinCmt=TRUE the ODE was first rewritten into linCmt(), leaving a model with linCmt() pseudo-compartments and no d/dt() for the matrix-exponential conversion to work from. That rewrite is now skipped when method="indLin" is requested, so such a model integrates its own rate matrix rather than being replaced by the analytic solution.
A steady-state infusion (ss=1 or ss=2 with a rate) gave a diverging solve under method="indLin". Its solver was the only one that never drained the pending-dose queue, which is where the infusion's off record is held, so the steady state itself was found correctly and the infusion was then left running for the rest of the timeline. Steady-state boluses and ordinary (non-steady-state) infusions were unaffected.
method="indLin" is substantially faster. The ODE-to-matExp() conversion ran on every rxSolve() call although it is a pure function of the model, and cost several times the solve it was preparing for; it is now done once per model (options(rxode2.indLinConvCache=FALSE) restores the old behaviour). The matrix exponential itself was recomputed on every fixed-point pass even though the rate matrix cannot change between them, and identical exponentials are now reused (RXODE2_INDLIN_NO_EXP_CACHE disables this). Together these are several times faster on a nonlinear model and more on a linear one; no result changes.
indLinRichardson extrapolated indLinIteration="exprb32" with the factors for a second-order base step, which exprb32 is not -- it is third order, so each level took its leading term down by a constant instead of removing it, and the step was sized from an estimate a whole order off. Asking for a level therefore made the answer worse: on a Michaelis-Menten model at 1e-8, indLinRichardson="always" delivered 3.7e-6 where "never" delivered 1.0e-7. The tableau now takes both the base order and how far a level advances it from the scheme, so "always4" is 1.2e-8 for a ninth of the steps "never" needs. Only exprb32 is affected: it is neither the default nor reachable from "auto", which never raised its level.
rxSolve(indLinForcing=) chooses how method="indLin" carries the indLin() forcing across one relinearization step. It was folded into an augmented column exactly as a constant infusion rate is, so it was frozen for the whole step. "ramp" (the new default) evaluates it at both ends of the step and integrates the line between them exactly -- the phi2 term -- with the rate matrix taken at the step midpoint; "constant" is the previous scheme, which reaches the same second order by averaging a start-linearized and an end-linearized answer. It applies to the "picard" and "newton" schemes; the exponential Rosenbrock ones never freeze the forcing.

Only the endpoint value moves with the iterate, so the rest of the step is built once and a pass costs a forcing evaluation and a matrix-vector product rather than a matrix exponential. The converged ramp step is symmetric, so its error expands in even powers of the step alone and indLinRichardson now removes two orders per level instead of one -- third order becomes fourth, fourth becomes sixth, fifth becomes eighth. That is where the difference shows up: under the default indLinRichardson="auto" a nonlinear model is several times to a hundred times more accurate at the same tolerance for the same or fewer steps, while with no extrapolation the two are a wash, both being second order there.
rxSolve(indLinJac=) chooses where the forcing Jacobian comes from when method="indLin" needs one, which is only under "newton", "exprb" and "exprb32" -- Picard needs none, so a non-stiff model under the default scheme never forms one. "symbolic" uses the model's own analytic Jacobian, which the matExp() conversion already emits as df()/dy() lines, and costs no extra forcing evaluations; "fd" central-differences the forcing at 2n evaluations. "auto" (the default) takes the symbolic one when the model carries it and falls back to finite differences otherwise, which is what happens above getOption("rxode2.indLinJacMaxStates") states where the emission is skipped.

On cost the two are a wash at compartmental sizes -- within about 25% of each other either way from 3 to 16 states, with no consistent ordering, and the symbolic emission adds a fraction of a second once at model build. The reason "auto" prefers symbolic anyway is exactness rather than speed: an exponential Rosenbrock step's order conditions assume the Jacobian is exact, and on a stiff van der Pol the symbolic one delivered a smaller error for the same work.
rxSolve(indLinIteration="exprb32") adds the Luan-Ostermann third-order exponential Rosenbrock pair. Its embedded second-order member is "exprb" itself, so the two differ by a computable quantity and it sizes its step from that rather than from the extrapolation column -- which is what "exprb" has to use and why "exprb" is held at fourth order. It is NOT the default and is not selected by "auto": measured at matched delivered accuracy it wins only on a stiff problem at a loose tolerance, by about 1.2 to 1.7 times, and loses elsewhere, badly so on a non-stiff population. The reason is the cost of the third phi function, which needs an augmented matrix three rows wider than the plain step; at the small dense systems compartmental models produce, widening the exponential costs more than the extra order saves.
rxSolve(indLinIteration=) chooses how method="indLin" solves each relinearization step: "picard" (the previous and only behaviour), "newton", or "exprb", an exponential Rosenbrock step that does not iterate at all. Which is cheapest depends entirely on the problem -- on a non-stiff model the iteration never limits the step and Picard is cheapest, while on a stiff one it is the only thing limiting it -- so "auto" (the default) starts on Picard and switches only once steps are actually being cut for non-convergence. A model that never needs a Jacobian therefore never forms one. On a van der Pol oscillator integrated over a full relaxation period at matched accuracy this is about 39 times faster than Picard at mu = 100 (593 relinearizations against 45,913) and about 426 times faster at mu = 1000 (581 against 1,001,968), which takes a full cycle at that stiffness from impractical to routine; a Michaelis-Menten model is left on Picard and unchanged. With both schemes given their best extrapolation level, that division holds: Picard is ahead on a non-stiff model at working tolerances and the exponential Rosenbrock step is ahead on a stiff one, and at a delivered error of 1e-8 on a non-stiff model. "exprb" runs at fourth order or above, since its error estimate comes from the extrapolation column and the third-order one is not reliable enough to size a step from.
method="indLin" extrapolates further when it pays. Each relinearization step could previously be raised from second to third order by running it also at half length; it can now use a Romberg column of up to four entries (h, h/2, h/4, h/8) for up to fifth order, at 3, 7 and 15 fixed-point solves per step. indLinRichardson="auto" (the default) raises the level as the step the controller settles on crosses each break-even. "always4" and "always5" force the new levels. On a 200-subject Michaelis-Menten solve this halves the time at atol=rtol=1e-8, and on a single subject at 1e-12 it is over seven times faster than the third-order step.
indLinRichardson="auto" keeps the extrapolation level it has earned for the rest of the subject, instead of dropping back to second order at every observation and re-earning it. A step that only needs a few relinearizations never reached the break-even, so a model observed at a dozen times ran most of its profile at second order however low the break-even was set: on a 200-subject Michaelis-Menten solve the default took 0.626 s to reach a delivered error of 1e-4 where forcing the fourth-order column took 0.098 s. It is now 0.112 s, and 0.341 s rather than 0.646 s at 1e-6. The break-evens themselves are also measured rather than derived, and differ between the fixed-point and exponential-Rosenbrock steps, whose costs per level differ. A model whose forcing reads no state is unaffected: the matrix exponential is already exact for it and there is nothing to extrapolate.

Two consequences for anyone reading step counts. A loose tolerance now does use extrapolation -- it turns out to pay there too, taking fewer steps than the second-order path rather than the same number -- and the delivered error at a loose tolerance is much smaller than before, so a ratio of errors across a tolerance sweep is no longer a way to read off the order of the default path. Use indLinRichardson="never" for that.
rxSolve(indLinMatExpType="taylor") adds a Taylor scaling-and-squaring matrix exponential, which needs no linear solve; its degree is chosen per call from the norm. It is as accurate as the default "expokit" on every problem tested, including a linear system where "Al-Mohy" at its default order is six orders of magnitude worse. The default is unchanged: profiling puts all of LAPACK at roughly 3% of a solve, so avoiding the linear solve does not pay on nonlinear problems.
$counts$dadt and $counts$jac report the matrix exponentials computed and reused for a method="indLin" solve. Both counters were previously unused on this path.
method="indLin" no longer uses the R API from inside the parallel solve. The Al-Mohy matrix-exponential backend took its workspace from R_alloc, and the default expokit backend warned through RWarn on a singular Pade denominator; neither is safe from a worker thread. The singular case also used to continue with an unfinished matrix, and now reports and returns zeros.
method="indLin" no longer throws from inside the parallel solve. Two code paths in the inductive-linearization solver raised an R-level error from a worker thread, which crashes the session rather than reporting an error; both now report through the usual per-subject error flag.
An indLin(<state>) <- <expr> forcing that references a compartment is now evaluated at that compartment's current value. The generated forcing function took no state vector, so the compartment kept its NA_REAL declaration and any state-dependent forcing (e.g. Michaelis-Menten elimination, indLin(central) <- -vmax*central/(km+central)) solved to NA under method="indLin". A forcing that references no state is unchanged.
method="indLin" iterates again, so it is inductive linearization rather than one relinearization per hmax substep. Within each substep the matrix and the forcing are rebuilt at the latest iterate while propagation restarts from the substep's starting state, until the states reported by rxModelVars(m)$indLin$wIndLin stop moving to within atol/rtol. Plain Picard iteration only barely contracts once the substep is comparable to the forcing's own time scale -- a Michaelis-Menten forcing with no linear elimination sits right at the stability boundary, oscillating for ~1e5 passes -- so each step is relaxed by a secant estimate of the iteration's contraction ratio. Relaxation does not move the fixed point, so the converged answer is the undamped one. Models with no forcing, or with a forcing that reads no state, keep the single-pass path and are unchanged.
Converting an ODE model to matExp() form (rxToIndLin(), and therefore rxode2(..., indLin = TRUE) and method="indLin"'s auto-conversion) now puts the nonlinear part of a right-hand side into an indLin() forcing instead of into a rate constant. A rate constant that reads a compartment is not a rate constant -- the matrix exponential assumes the rate matrix is constant in the states -- and burying the nonlinearity there meant the solver could not iterate it. Michaelis-Menten elimination now converts to indLin(central) <- -vmax*central/(km + central) with an empty rate matrix rather than to k_central_output = vmax/(km + central). This also removes a rate constant that was singular when the compartment was empty. Because these models now reach the iterating solver path, they are far more accurate: a one-compartment Michaelis-Menten solve that was about 70% off at the default settings is now within about 0.01%. rxIndLinStrategy() and rxIndLinState() no longer affect the conversion, since no way of factoring a multi-state product yields a state-free rate constant; both are kept so existing code keeps working.
rxSensMatExp() (rxToIndLin(calcSens=)) splits the system the same way. It used to take its rate matrix from the full Jacobian, so for a nonlinear model every rate constant read a compartment and the propagated primal was A(X).X rather than f(X). The nonlinear part now rides in an indLin() forcing, and each sensitivity compartment gets its own forcing d(f)/dp + (df/dy).S^p, at first and second order. A state-free input term (d/dt(x) = k0 - ke*x), which the Jacobian never saw, is carried too instead of being dropped. Michaelis-Menten forward sensitivities now match the ordinary ODE calcSens path, and since the rate matrix is constant the matrix exponential is cached across substeps. A rate constant that reads a compartment is a parse error for a sensitivity model as well now. Third-order sensitivities do not yet get a forcing contribution.
eventSens="jump" gets the right event-time (alag) jump sensitivity on a matExp() model that has an indLin() forcing. The replace()/multiply() jump rows need the right-hand side at the pre-event state, which was taken from the model Jacobian dotted with the state -- correct only while the whole right-hand side is the rate matrix times the state. With a forcing it is short by the forcing's own contribution, which on a Michaelis-Menten model put those sensitivities about 3.6% out. It now comes from the rate matrix and the forcing function directly. This also affects hand-written indLin() models, not only the ones rxSensMatExp() generates.
rxSolve(indLinRichardson=) Richardson-extrapolates each method="indLin" relinearization step, raising it from second to third order: the step is run once whole and twice at half length, and since a second-order step has a quarter the error at half the length, the two answers together cancel it. That costs three fixed-point solves per step instead of one, so it only pays once the tolerance is tight enough that taking far fewer steps outweighs it. "auto" (the default) decides per interval: after the first accepted step it compares the step the controller settled on against what is left of the interval, and switches when finishing at that step would take more than 27 of them -- the break-even point, since a second-order step needing N steps becomes a third-order one needing about N^(2/3) at three times the cost each. "always"/TRUE and "never"/FALSE force the choice. On a Michaelis-Menten model the switch-over lands at about atol=rtol=1e-5; at 1e-8 "auto" takes 544 steps where the second-order step takes 12,865.
rxSolve(indLinStepSearch=) and rxSolve(indLinMaxIter=) control the fixed-point iteration method="indLin" runs inside each relinearization step. indLinStepSearch="secant" (the default) estimates the iteration's contraction ratio from the last two residuals and relaxes by it, which costs nothing extra and is what makes an oscillating iteration converge at all; "exact" spends one more matrix exponential per iteration to locate the residual-minimizing factor in closed form; "none" is plain, undamped Picard. All three converge to the same answer -- relaxation does not move the fixed point -- so the choice trades iterations against work per iteration; on a Michaelis-Menten model the default is about five times faster than "none". indLinMaxIter (default 20) caps the iterations per step; running out is not an error, since the iteration contracts in proportion to the step and the solver reads it as a step that is too long.
A matExp() rate constant that depends on a compartment is now a parse error rather than a silently invalid model. The matrix exponential is only correct when the rate matrix is constant over the step, so a k_from_to that reads a state breaks the method's central assumption; the error names the constant and the compartment it reaches and points at indLin(), which is where a state-dependent term belongs and where the solver can iterate it. This applies to sensitivity models built by rxSensMatExp() as well.
method="indLin" chooses its own relinearization step for models with a state-dependent indLin() forcing, instead of subdividing each interval evenly by hmax. The forward answer (matrix built at the step's starting state) and the converged backward answer bracket the truth symmetrically, so their difference is a local error estimate that costs nothing extra; the step is then chosen from it the same way the other adaptive solvers choose theirs. atol/rtol and maxsteps now control the accuracy of these models and hmax only bounds the step. An iteration that will not converge is treated as a step that is too long and is retried shorter rather than reported, so stiff forcings that previously failed outright now solve; non-convergence is reported only once the step or the step budget runs out. $counts$slvr reports the relinearization steps actually taken, where it used to read zero. One consequence worth knowing: as with every adaptive method, the solution is now a piecewise function of the parameters, which adds a little noise to finite-difference gradients taken through it.

Each step also advances on the average of the two answers the error estimate is built from, whose leading errors are equal and opposite, so what is propagated is second order where either alone is first. This costs nothing -- both are already in hand -- and it is what brings the step count down: the error now falls roughly in proportion to atol/rtol rather than to their square root, so the work needed for a given accuracy grows like 1/sqrt(error) instead of 1/error. On the Michaelis-Menten model above, matching the accuracy the old scheme delivered at its default now takes about a twentieth of the steps, and the gap widens the more accuracy is asked for.

The forward answer is evaluated at the step's starting time as well as its starting state, so that it and the converged answer really are the two ends of one quadrature. Evaluating both at the step end cancels the state error but leaves the explicit-time error, which silently dropped any forcing that reads t back to first order: on a Michaelis-Menten model with an exp(-t) input the error at atol=rtol=1e-9 falls from 4.6e-03 to 1.1e-07.
rxSensMatExp() no longer emits an indLin() forcing that is algebraically zero, which had been demoting every sensitivity model with two or more compartments to the fixed-point iteration. The generator splits the system term wise as dX/dt = A.X + F(X) and keeps whatever rhs - A.X leaves as the forcing; symengine holds A_ij * X_j as a product of a sum and a symbol and does not distribute it, so from two compartments up the subtraction left a residual that prints as non-zero and is zero. A structurally non-zero forcing is what classifies a model as state dependent, so the whole solve took the inductive-linearization driver -- Picard/Newton substepping with error control and several exponentials per substep -- instead of one cached matrix exponential per interval. One compartment emitted no forcing at all, which is why it was the only configuration where matExp() was competitive. The residual is now cancelled before it is tested, and the same cancellation collapses the un-simplified k_<cmt>_output constants the split produced (-q/v-(-q/v-cl/v) is now cl/v), which the generated model re-evaluated on every ME() call. An expansion that takes a forcing from reading a compartment to reading none is kept however long it gets, since that is the difference between the two drivers; everywhere else it buys no reclassification and is kept only where it does not lengthen what it replaced, so a genuinely nonlinear forcing is untouched. Measured on an optimized build, 40 subjects over an irregular 200 point schedule with three sensitivity parameters, single thread, the solve is 9.6x faster at two compartments and 12.5x at three, with the solved values unchanged to 3e-14 -- the dropped terms contributed nothing but cost. bench/indlin_zero_forcing_ab.R is the harness.

Serialization

A saved solver state now round-trips the indLin() convergence set (op->indLin, from rxModelVars()$indLin$wIndLin). Only its length was written, so restoring a state for a model with an indLin() forcing left the set itself empty and the relinearization iteration indexed a null pointer.
A saved solver state now round-trips the initial-condition and scale vectors it claims to. Their lengths were taken from the distance to the next pointer in the gsolve slab rather than from the vectors themselves, so they spanned the intervening lhs and tolerance blocks and no state could be restored at all: rxLoadState() failed with a size mismatch for every model. The two lengths now travel with the state, which is what the format version is bumped to 3 for; a state written by an earlier version is rejected with a message asking for it to be re-saved.

Installation / linking

rxode2 again installs and works with lotri 1.0.4 (the requirement was relaxed from 1.0.5). Priors (prior(x) ~ ...) and repeated blocks (same()) still need lotri >= 1.0.5, and their tests are skipped with an older lotri.
On Windows, STAN_THREADS and the TBB link are kept when building against RcppParallel >= 6.2.0, which ships tbb.dll/tbbmalloc.dll with the package again. configure now decides whether to strip the TBB flags by looking for the TBB library in RcppParallel's lib directory rather than by the shape of the -L flags it emits, so the TBB-less build introduced in 5.1.6 is used only with RcppParallel 6.0.0--6.1.1, which shipped no TBB library on Windows. (The 5.1.6 release notes had this backwards: RcppParallel 6.2.0 restored the TBB library on Windows rather than dropping it.)

rxode2 5.1.6 (2026-08-04)

New features

Added the event ("jump") sensitivity shape to rxode2's linked function-pointer API, so a downstream package can install a model's shape from C++ without an R round trip: rxode2EventSensLoadFull() (all six dims, where the older rxode2EventSensLoad() omitted nParam3/useCalcJac), rxode2EventSensGetDims()/rxode2EventSensSetDims(), rxode2EventSensSetActive(), rxode2EventSensDeactivate(), and rxode2EventSensShapeSize()/rxode2EventSensShapeSave()/ rxode2EventSensShapeRestore(), which snapshot and reinstall a whole shape (dims plus the model's dosing-derivative function pointers) through a caller-owned opaque buffer. This lets several peer models with different shapes be solved through one shared solve pool, installing each batch's shape and restoring the previous one afterwards.
Added setIndCmt() to the function-pointer API, the writer counterpart of getIndCmt(), so a downstream package can re-base the per-observation CMT covariate without reaching into op->cmtCov/ind->cov_ptr by field. getIndCmt() reports a missing CMT as NA_INTEGER, distinct from the 1 it returns for a model with no CMT covariate at all (where every observation really is compartment 1), so a caller re-basing the column can leave missing rows alone.

Bug fixes

Compilation cache

The compiled-model cache key now includes eventSensCode, so two builds of one model whose generated C differs -- event sensitivities on vs off -- no longer share a .c/.so path in the rxode2 cache directory. Previously the second build overwrote the first while any model object created earlier kept resolving its entry points by name, so it silently began executing the other variant: the declared lhs width was unchanged but most slots were never written, and rxSolve() returned whatever was left in the buffer. A model with no event-sensitivity code keeps exactly the prefix it had, so no existing cache entry is invalidated (#1171).

Dependencies

The suggested xgxr is now required to be >= 1.1.6. Its xgx_scale_x_log10()/xgx_scale_y_log10() return the ggplot2 scale itself from that version on, rather than a length-one list wrapping it.

Installation / linking

Fixed the Windows build against RcppParallel 6.2.0, which no longer ships the TBB library there (6.0.0 still built). configure already dropped the -ltbb/-ltbbmalloc link flags when they are unavailable, but still compiled with -DSTAN_THREADS and -DRCPP_PARALLEL_USE_TBB=1, which pulls stan::math's ad_tape_observer (a tbb::task_scheduler_observer) into the objects and left undefined references to tbb::detail::r1::observe at link time. Those defines are now dropped together with the link flags, stan-math's init_chainablestack.hpp is kept out of the build, and the main thread's AD tape -- which that observer would otherwise have created -- is constructed in src/linCmt.cpp instead (by Jeroen Ooms).

Solving

Fixed heap corruption when simeta()/simeps() resample inside a solve. Both go through simvar(), which reseeded its threefry stream with getRxSeed1(); unless rxSetSeed() had been called that draws from R's own random number generator, which allocates R objects and can trigger a garbage collection. Doing so from an OpenMP worker thread corrupted R's heap, and the session then failed later in an unrelated place (cannot get data pointer of 'NULL' objects, 'rho' must be an environment, corrupted double-linked list, or a segfault). The in-solve resample now draws from a per-thread engine seeded on the main thread and touches no R API.
The simeta()/simeps() resample no longer replays the simulated parameters. Its engines are keyed off a threefry draw rather than off the runif()-derived seed handed out at solve setup, which rxSolve() goes on to reuse for the simulated omega/sigma deviates; a resampled eta could therefore come out exactly equal to another subject's simulated eta.

rxode2 5.1.5 (2026-07-28)

New features

library(rxode2) is faster. .onLoad() no longer calls requireNamespace() on the suggested packages (pillar, tibble, arrow, dplyr, nlme, units, digest) before registering their S3 methods. The registration helper already installs an onLoad hook when the other namespace is not loaded yet, so the eager loads only added startup cost; the methods are still registered at the same point in time from the user's perspective.
rxControl(sigdig=) now derives the ODE solver tolerances with one solver-independent formula -- the same for stiff, non-stiff and auto-switching solvers. The rtol exponent IS sigdig and atol sits three orders below it: rtol = 10^(-sigdig) and atol = 10^(-sigdig-3). The sensitivity tolerances match the main solve (rtolSens = rtol, atolSens = atol), since gradients and covariances are built from them, and the steady-state tolerances run one order looser (ssRtol = ssRtolSens = 10*rtol, ssAtol = ssAtolSens = 10*atol). This matches how nlmixr2est derives solver tolerances from its optimization sigdig, so the same sigdig means the same thing whether it is used for estimation or for a plain rxSolve().

sigdig remains NULL by default and continues to have no effect unless you pass it, so solves that do not name sigdig are unchanged.

Two notes for callers who do pass it. First, the mapping is keyed to sigdig as a request for that many significant digits, which for small sigdig is looser than what the previous symmetric atol = rtol = 0.5*10^(-sigdig-2) gave: at sigdig = 4, rtol moves from 5e-7 to 1e-4, which is also looser than the 1e-6 default rtol (atol moves the other way, from 5e-7 to 1e-7). If you were using sigdig to tighten a solve, raise it or set atol/rtol directly. Second, each tolerance is resolved independently and only when you did not supply it, so an explicit atol/rtol overrides the main solve but does not propagate to the sensitivity or steady-state tolerances -- set those directly if they should change too.
The SUNDIALS public headers are now vendored into the package (src/sundials_inc/) alongside the already-vendored SUNDIALS C sources, and the LinkingTo: sundialr dependency has been dropped. This guarantees the vendored sources always compile against headers from the same SUNDIALS release, instead of silently drifting when sundialr updates its bundled SUNDIALS (#1155). The vendored include is injected via PKG_CPPFLAGS so it precedes the LinkingTo include flags; otherwise the older SUNDIALS copy bundled inside StanHeaders would shadow it.
rxSerialize() now writes the base R types only ("xz", "bzip2", "base"); qs2 is no longer a write format. rxDeserialize() still reads qs2/qdata-serialized data and base91-encoded strings, so objects stored by earlier versions remain readable. Test data was converted from .qs2 to .rds.

Bug fixes

Serialization

qs2 moved to Imports. rxDeserialize() used it without declaring it anywhere, and because the call sites named the package as a string, the dependency was invisible to R CMD check as well. Environments that build their library from the declared dependency graph could therefore end up without qs2 and fail to read objects stored while qs2 was an allowed rxode2.serialize.type -- for example the origData slot of fits saved by earlier versions. qs2 is now only ever read, never written.

Error models / transformations

The first and second derivatives of the Yeo-Johnson transform (rxTBSd() and rxTBSd2()) had the wrong sign for negative values when lambda was exactly 2. There yj(x) = -log(1 - x), so the derivatives are 1/(1 - x) and 1/(1 - x)^2, both positive; the special case returned them negated, which also contradicted the general formula's limit and made the derivative discontinuous in lambda at 2. Since Yeo-Johnson is monotone increasing, its first derivative must be positive everywhere.
Review of the fix above found further errors in the same transform family (all pre-existing, none introduced by that fix):
- rxTBSd2() returned a wrong second derivative for the logit transform (an algebra error in the closed form) and for the composed logit + yeoJohnson / probit + yeoJohnson transforms, where the chain rule used the first Yeo-Johnson derivative in place of the second and dropped the inner-transform curvature term.
- rxTBSi() did not invert the composed logit + yeoJohnson / probit + yeoJohnson transforms: it applied the forward Yeo-Johnson transform (or skipped it entirely) instead of the inverse, so rxTBSi(rxTBS(x)) did not return x for lambda != 1. This affected simulation back-transforms of those error models.
- The lambda gradient of the transform log-Jacobian (powerDL, used by estimation routines) was wrong on the negative Yeo-Johnson branch (-log1p(x) instead of -log1p(-x), NaN for x < -1), returned a spurious 0 at exactly lambda == 1 for boxCox/yeoJohnson, was missing the probit + yeoJohnson case (returned NA), and returned a spurious log(x) (instead of 0) for the lambda-free lnorm transform. The log-Jacobian itself (powerL) clamped the wrong term in its logit guard, giving an unprotected log(0) at the upper bound.
- For boxCox/lnorm, rxTBSd() and rxTBSd2() returned the clamp constant sqrt(.Machine$double.eps) itself for x at or below the clamp instead of clamping x and evaluating the derivative formula, making the derivatives discontinuous (and ~15 orders of magnitude too small) at the boundary. The clamp now feeds the usual formula, matching how every other transform in the family handles the guard.

Estimation / symengine translation

The symbolic derivatives of the relational operators (>, <, >=, <=) are now centered on the discontinuity a == b: the atanh(2*tol - 1) shift that placed the smoothed nascent-delta bump at a - b ~ +/-0.46 was removed. Since the forward pass evaluates relationals as hard booleans, the shifted bump gave sensitivity/exact-gradient consumers (e.g. FOCEI's analytic gradient paths) a spurious derivative in a band next to the threshold; the centered rule makes the derivative consistent with the forward value. This also makes the first derivatives of abs(), min(), and max() exact away from the boundary (#1159).

Solving

Fixed heap corruption when OMP_NUM_THREADS is set below the number of cores a solve asks for -- as it is on CRAN check machines, which set OMP_NUM_THREADS=2. The extra-dosing pools were sized once when the package loaded, from omp_get_max_threads() (which honors OMP_NUM_THREADS), but they are indexed by the solving thread id, which is bounded by op$cores; rxSolve(cores=) overrides OMP_NUM_THREADS through OpenMP's num_threads clause. Every thread past the first OMP_NUM_THREADS therefore wrote off the end of those arrays, corrupting the heap and crashing the session later in an unrelated allocation. The pools now grow to cover op$cores at solve setup, like the other per-thread pools.
Fixed an out-of-bounds thread index that could segfault a solve. The internal thread id used to slice the per-thread solving buffers was not bounded by the number of threads those buffers were allocated for (op$cores). A larger id read past the end of gInfusionRate[] -- an array of pointers -- and the resulting garbage pointer crashed iniSubject(); the flat per-thread arrays were silently overrun in the same way. The id is now clamped to the last valid slot, matching what rx_get_thread() already did.
Fixed a cross-subject leak in batched multi-subject linCmt() solves: the per-thread inter-event amount buffer was never cleared between subjects, so with cores < nSub every subject after the first on a thread could start from the previous subject's compartment amounts (surfaced by a modeled alag()) (#1153; by Hidde van de Beek).
delay()/past() models containing an if/else block failed to solve with unexpected 'else': the DDE helpers parsed the rxNorm() text directly, which puts } and else on separate top-level lines; the normalized text is now parsed wrapped in a { } block. In addition, a past() history inside an if/else branch is now rejected with a clear error (it was invisible to validation), and delay-duration root-variable resolution now sees assignments made inside if/else branches (#1151).

Event tables

ev$id on an event table now returns the per-row id column (matching as.data.frame(ev)$id) instead of the unique subject ids, so idiomatic subsets like ev[ev$id == 3, ] and per-subject assignments like ev$wt <- 50 + 20 * ev$id no longer silently recycle a short vector; the unique ids remain available via ev$env$ids. [.rxEt now errors on a logical row index whose length matches neither 1 nor the number of rows, and columns assigned with ev$col <- value (new covariates as well as previously hidden canonical columns such as cmt) now round-trip through as.data.frame(ev) (#1154).
Columns assigned explicitly on an event table (ev$wt <- 70) are now shown in the tibble printed by print(ev), in ev$get.EventTable(), and in the compressed preview printed for ev$get.dosing()/ev$get.sampling(), matching as.data.frame(ev). They were kept and used when solving, but never displayed, so they looked like they had disappeared (#1154).
ev$get.dosing() and ev$get.sampling() now print the same columns print(ev) does regardless of how the event table is stored internally. Previously an un-grouped table printed every internal column, including hidden ones such as low/high/dur and covariates that only rode along with an imported data frame, while a compressed one printed only the shown columns. Every column is still present on the returned data frame for programmatic access, a column added or renamed on the returned frame still prints, and dplyr verbs turn it back into a plain data frame the way they already did for rxEt -- including the column verbs (select(), relocate()), which subset with [ rather than going through dplyr_reconstruct().
Explicitly assigned columns now survive a round trip through a data frame. as.data.frame() tags them in a rxEtExtraCols attribute that et(), as.et() and $import.EventTable()/$importEventTable() read back, so et(as.data.frame(ev)) keeps showing wt instead of demoting it to a hidden imported covariate. A data frame built by hand carries no tag, so its covariate columns stay hidden as before.
as.data.frame() on an event table still hides covariate columns that simply rode along with an imported data frame (et(data)), while showing columns assigned explicitly on the event table (ev$wt <- 70, #1154). The covariate is still used when solving. Showing every non-canonical column broke code that imports events and then joins its own covariates back onto as.data.frame(ev), since the join produced wt.x/wt.y and the model parameter disappeared.

Model compilation

The parsed_md5 of a model no longer depends on how many models were built before it in the session. linCmtSens was folded into the hash but only assigned after the model was parsed, so the first build of a session hashed with an unset value and every later build hashed with the previous call's value. Because the compiled DLL is named from parsed_md5, the same model could get two different cache keys (and hence a redundant recompile) depending on build order. It is now set before the parse.

Installation / linking

RcppParallel is now a runtime import (added to Imports with an importFrom), so its shared library is loaded into the process before rxode2's. rxode2 links against RcppParallel (-lRcppParallel); with RcppParallel only in LinkingTo its DLL was not guaranteed to be loaded first, so on Windows library(rxode2) could fail with LoadLibrary failure: The specified module could not be found. This surfaced with RcppParallel 6.0.0, which statically links TBB and no longer ships the tbb.dll stub that previously happened to pull the library in.
On Windows with RcppParallel >= 6.0.0 (which statically links TBB through Rtools and no longer loads tbb.dll), the stale -ltbb/-ltbbmalloc flags and the -L path to RcppParallel's old dynamic TBB directory that StanHeaders::LdFlags() still emits are stripped at configure time, so the rxode2 DLL no longer records an unresolvable runtime dependency on tbb.dll. The strip is keyed to that stale -L<RcppParallel/lib> signature, so a future StanHeaders that emits corrected flags -- or a user-supplied TBB via TBB_LINK_LIB/TBB_LIB -- is left untouched (#1161).
The vendored SUNDIALS *NewEmpty constructors now allocate with calloc instead of malloc, so any struct fields added by a newer SUNDIALS release are NULL (and safely ignored) rather than uninitialized (#1155).

rxode2 5.1.4 (2026-07-20)

Bug fixes

Model piping

Model piping no longer shares the meta environment by reference between the original and the piped model. .newModelAdjust() assigned the previous model's meta env directly (to retain sticky items), so both models shared one env -- including the cached simulation model ($meta$.simModelBase). Whichever model was solved first cached its simulation model for both, so a piped model could silently drop an appended compartment/state (e.g. a nonmem2rx import: mod %>% model(d/dt(AUC) <- f, append=TRUE)) or the original model could silently gain the piped model's states/estimates. The meta env is now copied via .copyEnv() (which drops .simModelBase), so each model keeps its own cache.

Compilation

Silenced the CRAN -Wlto-type-mismatch warnings seen with LTO/gcc builds. The rxSolveWarnPush() forward declaration in src/init.c was missing the variadic ... of its definition, and the ODEPACK /DLS001/ common block was declared with two inconsistent (but memory-equivalent) layouts across the LSODE/LSODA step routines. Both are now declared consistently; the fixes are layout-preserving and the Fortran solvers produce identical results.

rxode2 5.1.3 (2026-07-19)

New features

rxOptExpr() gains chunkLines and parallel, to optimize a large machine-generated model (a sensitivity- or Jacobian-augmented model) in contiguous cost-balanced chunks rather than in a single pass.
Delay differential equations: delay(state, T) evaluates an ODE state at t - T (Monolix semantics), with past(state, T) <- expr defining the pre-history. Delayed states are interpolated from the solver's dense output; delay models default to the "dop853+ros4" composite and cap the step size to the smallest delay. The dense-output/history machinery is adapted from the 'dde' package (Rich FitzJohn, Wes Hinsley, Imperial College), whose authors are added as contributors.
Forward sensitivities for delay models, so delay() models estimate with gradient-based methods such as FOCEi. Parameter-dependent delays are supported at first order (rxDelayD()); second/third order are generated for constant delays (rxDelayD2()/rxDelayD3()) and rejected for parameter-dependent delays (use a numeric or Gauss-Newton Hessian there).
Many new ODE solver methods: a large suite of explicit Runge-Kutta tableaus (orders 3-14), stiff Rosenbrock and implicit Runge-Kutta methods ("ros43", "ros6", "radauiia5", "gauss6", "sdirk43", "backwardEuler", ...), symplectic steppers, SUNDIALS CVODE ("cvode", linear solver selectable via cvodeLinSolver=), and LSODE/BDF. Implicit methods auto-generate an analytic Jacobian. New helpers rxIsStiff(), rxIsNonStiff(), rxIsImplicit(), rxIsDense(), and rxIsAutoSwitch() classify methods; see the new "ODE solvers" article.
AutoSwitch composite methods written "primary+secondary" (e.g. "dop853+ros4"): a non-stiff primary with reactive fallback to a stiff secondary, in both the standard and dense-output paths.
Adjoint sensitivity solving: rxSolveAdjoint() and rxSolveAdjointRk4() return the same rx__sens_<state>_BY_<param>__ output as forward sensitivities via a backward sweep. Exact discrete adjoints exist for the one-step methods ("s" suffix, e.g. "dop853s"), "liblsodaadj", and "cvodesadj", including event jumps (dose/reset/replace/multiply), modeled alag/rate/dur, and steady state. Stiff adjoint and forward-sensitivity solvers integrate the augmented system with its analytic Jacobian.
Jump sensitivities for dosing events (based on https://github.com/dkaschek/EventSensitivities), replacing finite differences as the default (rxode2.eventSens option: "jump", "fd", "both"). Hybrid jump sensitivities are used for matrix exponential and linCmt() models (up to 3rd order for the ODE/matrix exponential cases).
Automatic conversion of linear ODE models to linCmt() at solve time (rxSolve(..., useLinCmt=TRUE), the default), passing detected PK parameters explicitly. Handles a central sub-system with an output-only peripheral observable; systems linCmt() cannot represent stay explicit ODEs, and a conversion that will not compile falls back to the ODE (only rxode2).
Adaptive dosing helpers (bolus(), infuse(), replace(), etc.) now work inside linCmt() and mixed linCmt()+ODE models, with Jacobian handling of the dosing events; odeToLin() preserves and renames them when converting.
linCmt() sensitivity (linCmtB) solves now run in parallel across subjects on the default forward-mode AD Jacobian path (linCmtSensType="AD"), which is stack-local with no shared Stan arena. The reverse-mode AD ("ADr") and finite-difference paths remain single threaded.
Inductive linearization and matrix exponentials rewritten with a more NONMEM-like interface (automatic ODE->syntax translation retained) and symbolic-differentiation gradients.
Added a forward automatic-derivative linear compartment model.
ar(cor) residual term simulating continuous-time AR(1) residuals for normal, t, and cauchy error models, addable per endpoint alongside any transform; cor is in [0, 1) and the lag correlation decays as cor^(time gap) (Karlsson, Beal and Sheiner 1995). Estimation is supported in nlmixr2est (nlm and focei families).
lag0()/lead0()/diff0(): like lag()/lead()/diff() but return 0 instead of NA when there is no prior/following record. A calculated variable may now reference itself through lag()/lag0()/diff() (a first-order recurrence); a non-lag self-reference is still a required input parameter.
rxOmegaVarCovDeriv(): non-Cholesky Omega path returning Omega^{-1}, log|Omega|, and their first/second derivatives with respect to each free variance-covariance element.
rxExpandSens3_() generates analytic third-order forward sensitivity equations; .rxSens() gained a vars3 argument.
For downstream packages: rxSetSolveAtolRtol()/rxGetSolveAtolRtol() in the C function-pointer API, and setRxThreadId() so a package can drive the per-subject solve from its own OpenMP team.
rxTest() test blocks now muffle stray progress messages (e.g. "calculate sensitivities"); set options(rxode2.test.verbose = TRUE) to see them. Messages asserted with expect_message() are unaffected.
coef() methods for rxUi models (and model functions). By default coef() returns the fixed-effect (theta) estimates; coef(model, level = "omega") returns the random-effect variability matrix and coef(model, level = "all") returns both. nlme::fixef() continues to return the fixed effects.

Bug fixes

The C accessors exposed through the function-pointer API (getRxNsub(), getSolvingOptions(), getSolvingOptionsInd(), and the other rx_solve* accessors) no longer segfault when handed a NULL or uninitialized solve. They fall back to the global solve; a scalar counter/flag accessor (nsub, nall, nobs, npars, ...) simply reports zero before any solve, exactly as before, so downstream code that probes those counts at load time keeps working (for example babelmixr2's PopED integration, which queries them from .onLoad). An accessor that must dereference a per-subject record (getSolvingOptionsInd()) instead raises a normal catchable R error stating that the solving environment is not set up, rather than dereferencing a NULL pointer and crashing the R process. This hardens downstream packages that call an accessor before their solve pointer has been populated (for example a cold first nls/nlm fit in nlmixr2est).
A Jacobian entry df(state)/dy(THETA[n]) or df(state)/dy(ETA[n]) (a bracketed parameter reference, which the grammar accepts) no longer segfaults. The synthetic _THETA_n_/_ETA_n_ symbol was never registered, so its index stayed -2 and the model validator read tb.ss.line[-2] out of bounds. This crashed nlm/FOCEi fits that re-parse their generated calcJac model (whose parameters are THETA[n]) in the residual/table step -- notably for a delay-differential-equation model whose delay parameter appears in a product of delayed states.
past(state, tau) on a state with no d/dt(state) now reports that cleanly instead of corrupting the heap. The error path appended nothing to the message buffer and then trimmed a trailing ', that was never written, moving the write offset before the start of the buffer; the damage surfaced as a double free or corruption abort on a later, unrelated parse rather than at the offending model. The message now names the property ('past(G)' present, but d/dt(G) not defined), and a property with no message branch can no longer underflow the buffer.
rxOptExpr() no longer fails on a model that uses past(state, tau) and is long enough to be optimized in chunks. A past() line only parses in a chunk that also holds the matching d/dt(), and sensitivity augmentation appends past() after every d/dt() -- so it reliably landed in a chunk of its own. It is now disguised for the duration of the optimization like any other compartment-scoped left-hand side, and restored byte-exactly afterwards. Together with the fix above this unblocks estimating a non-constant-history DDE (e.g. the rheumatoid arthritis model of Koch et al. 2014, J Pharmacokinet Pharmacodyn 41:291-318, Example 6).
rxAppendModel() now warns (instead of erroring) when the appended models have no variables in common, so the combined model is still returned; use common=FALSE to suppress the warning (#520).
rxFixPop() no longer tries to literally substitute a fixed mixture proportion (mix()). A mixture proportion must stay a named model-block variable, so substituting its value made the re-parse throw from mix() ("the probabilities in a mixture must be in the model block ..."); a downstream caller wrapping rxFixPop() in try() leaked that error to the console during otherwise-successful mixture fits. Fixed mixture proportions are now excluded from the substitution.
Tests that use datasets from the suggested nlmixr2data package (theo_sd, warfarin, nmtest) now guard their use with skip_if_not_installed("nlmixr2data"), so the test suite runs cleanly when nlmixr2data is not installed (#95).

Estimation / symengine translation (`rxFromSE()`)

Convert raw R comparison/logical operators (>, ==, &, ...), not only their rxGt()/rxEq() symengine forms; fixes "user function '>' requires 0 arguments" in FOCEi models with inter-occasion variability (nlmixr2/nlmixr2#390).
Recognize bare relationals on the second conversion pass of a Subs() over a Derivative(); unblocks FOCEi IOV models that also have a between-subject eta on a parameter without IOV.
The numeric-constant canonicalization now evaluates operands in baseenv() only and guards zero-length results, fixing an "argument is of length zero" error and silent substitution of user-workspace variables (#1109).
A trig function (sin/cos/tan) whose argument is a compound expression divided by something (for example sin(2 * 3.14 * (time - mtime1) / period)) no longer drops its whole argument. The division branch fell through without returning when the numerator was not a single token, so the argument became NULL and the emitted C code was sin() -- which failed to compile with "too few arguments to function 'sin'". Such models (for example an enterohepatic gallbladder model with a sinusoidal release) now build and fit (nlmixr2/nlmixr2est#513).

Model parsing / mu-referencing

Summing two or more population parameters in an expression that has no random effect (for example a combined residual error W <- sqrt(sigma.1. + sigma.2.)) is no longer misreported as "2+ single population parameters in a single mu-referenced expression". That check now fires only for a genuine mu-referenced expression (one that also contains an eta), and the message names the parameters that were actually summed instead of the first parameters in the model (#471).

Delay models

calcJac=TRUE rewriting (also used by the stiff ros4/dop853+ros4 path) no longer breaks delay models declaring literal THETA_n_/ETA_n_ parameters: constant ~ intermediates stay bound, the literal names are restored, and past() history lines are re-emitted.
A state read by delay() is always kept as a real ODE, so delayed states named like sensitivities (rx__sens_*) keep their defining d/dt() and can use the stiff/dense composite directly.
Delay models whose analytic Jacobian cannot be generated now fall back to dop853 (dense) instead of liblsoda, which recorded no delay history and silently returned pre-history values.
An lhs reading delay() is now reported correctly in the output data frame (#1140). The dense delay history was freed at the end of each subject's solve, so the post-solve lhs recalculation returned the constant pre-history (0) at every record even though the delayed value drove the ODE. The history is now kept until rxSolveFree() releases the subject, which also plugs a leak on the discrete-adjoint (rk4s) path where it was never freed.

`linCmt()` models

Fixed a compartment-indexing bug where a model with both an error model and an in-equation compartment reference (e.g. Cp <- peripheral1 / vp) read an unwritten slot.
tad(<state>)/tlast(<state>) no longer return NA or the wrong value when the model also declares an extra cmt() for an algebraic observable (nlmixr2est#685).
The automatic linCmt() conversion no longer fires on a nonlinear model whose nonlinearity is written through a state-derived observable (e.g. Michaelis-Menten via Cc <- central / vc).
The automatic linCmt() conversion no longer changes results when the event data addresses a compartment (in a dose or an observation record) by the name of an ODE compartment the conversion renames (e.g. an ODE centre compartment addressed as CMT = "centre", which the conversion renames to central). Such a solve now falls back to the original ODE model instead of routing the record nowhere and returning all-zero predictions.
Fixed the automatic linCmt() conversion cache reusing the first model's initial estimates for a later model that shares the same model({}) equations but has a different ini({}) block, which made structurally identical models with different parameters return identical predictions.
Fixed the string form of the compartment argument in the adaptive dosing helpers (e.g. bolus(50, cmt = "depot")).

Solving

Zero the LSODA solver work memory on allocation (alloc_mem, calloc instead of malloc). The shared work block (Nordsieck history yh, Jacobian workspace wm, acor/savf, ...) was left uninitialised and parts are read before the integrator writes them on some paths (e.g. a first stiff/BDF step at an extreme point), making a solve non-deterministic. Surfaced by valgrind as reads of uninitialised LSODA memory inside FOCEi/impmap inner solves, and downstream as an occasional blown-up importance-sampling fit run after a prior (parallel) fit. Solving is otherwise unchanged.
lag()/diff() (and first()/last()) previously returned a constant instead of the prior record's value for calculated variables and time-varying covariates; they now read the prior record (NA on each individual's first record) and work through the estimation/symengine path. Only lag(x, 1)/diff(x, 1) are supported for calculated variables.
Bug fix for mix() models and iCov models.
The rxMemoryEstimate() RAM detection no longer calls the defunct utils::memory.limit() (which warned on every Windows solve); total RAM is now queried natively in C (GlobalMemoryStatusEx on Windows, sysctl on macOS, sysconf on Linux) and available memory reuses the allocator preflight estimate (rxAvailableMemoryBytes()). This also drops the memuse suggested dependency and the shell-command fallbacks.
Fixed out-of-bounds heap reads (AddressSanitizer-confirmed; results unchanged): rxSolve() parameter setup when subjects share one event table in an nsim > 1 sorted solve; syncIdx() dose-index lookup; cvPost() with a 1x1 omega; linCmt.h linCmtStan2ssInf8; etTran() combineDvid; rxDerived() derived1.

rxode2 5.1.2 (2026-06-02)

geom_cens() / stat_cens() no longer emit "Ignoring unknown aesthetics" warnings when censoring aesthetics are mapped. Documentation corrected to describe the two supported lowercase forms: lower/upper (both required) or cens (with optional limit). The two forms cannot be mixed, lower and upper are now required together, and limit without cens is rejected rather than silently ignored.
Checks for is.loaded() before loading a rxode2 model. This helps fix the m1 ODR issue shown in nlmixr2est.
Moved dim.rxEt() here instead of in nlmixr2est

rxode2 5.1.1 (2026-05-28)

Various low level fixes to allow nlmixr2est to have parallelized focei.
Parallelized the rxode2 data.frame creation.
Added parallel solving mirai for clusters and HPC support.
Added out of memory solve using arrow/duckdb. These out-of-memory (rxSolveOom) solves behave like a standard solved object: it prints the $params and $inits (mirroring the rxSolve console output), supports $, head(), nrow(), ncol()/dim() and the usual as.data.frame()/as_tibble()/ as.data.table() coercions, and exposes the per-subject parameter table and initial conditions that are now persisted alongside the chunked data. A DuckDB query layer over the parquet chunks is used for lazy access (head(), single-column extraction, schema) when available. The chunks can also be queried lazily with dplyr (via as.arrow() or arrow::to_duckdb()) so that filtering and aggregation are pushed down to the on-disk chunks and a possibly out-of-memory result never has to be fully materialized. The storage/query engine can be pinned with the rxode2.oom.backend option ("auto", "duckdb", "arrow" or "rds"); the option is also forwarded to parallel (mirai) workers.
Use ALTREP for id, sim.id, repeated simulation event columns (evid, cmt, ss, amt, rate, dur, ii, time), covariates and kept variables when blocks are identical across simulations; falls back to filled out columns when runtime event mutation is detected (evid_() push growth / per-individual event reallocation). Also factors cannot currently be represented by altrep, so they are forced to be fully represented.
Change compile flags and compiler directives for rxode2 models to speed up how they run.
Have a pre-allocated context pool for lsoda in both liblsoda and lsoda (faster because memory doesn't need to allocated and deallocated so often)
Change OMP scheduling to dynamic to try to help load-balance the ode solving per subject.
Simulation normal random numbers before integrating them into your solve.
Add evid_() function to allow arbitrary doses and observations in a rxode2 model.
Add splitBolus() function to split or relocate doses in the final output. This is done at translation time (but is respected by evid_()) so in general is a bit faster then arbitrary doses in an estimation step for nlmixr2
Add %% operator to valid rxode2 syntax
Create per-individual ODE solving tolerances for use in focei.
Fix potential security and memory-management issues that could lead to crashes or undefined behavior including integer overflow
Change dop853 to allow per state tolerances and parallel solving like liblsoda.
Change dop853 to be able to use dense=TRUE for the 8th order dense polynomial interpolation between dosing events.
Now dop853 can be parallelized per thread.
Change mtime state-based dosing to use less memory.
Add plogis() translation inside rxode2 to it's c-based expit() functions
Refactored et() to be mostly in R, fixing many issues (#722 , #725, #858, #732, #723, #721, and #724) and allowing dosing/sampling windows to use ii, addl and until (realized immediately)
Add linToOde() convert linCmt() models to ODEs.
Fix IOV simulation issue observed in #982.
Fix sticky variable calculation (#1013, #1025)
More easily identify initial conditions (#948)
Fix sensitivities in the linCmt() that did not match the ODE (#1018, #1012)
Added in-solve addition of observations (obs()), bolus doses bolus(), infusion doses infuse() or infuseDur(), system resets reset(), compartment replacement replace(), multiplicaton events multiply(), and phantom/transit compartment events phantom(). For more granular control you can also use evid_().
Refactor string comparison in rxode2 so that it is actually doing an integer comparison when running the ODE solving routine (simulation and estimation) instead of using a string comparison. It makes using strings like (sex == "male") run faster.
Add rxMemoryEstimate() and rxMemSummary() to estimate the amount of memory that is required for a rxode2 solve.
Add tolFactor, a per individual change of the tolerances to be used in solving. This is used have individualized tolerances from nlmixr2est.
Add serializeFile as an option to save the rxode2 C fitting data and then restore as needed.
Add out of memory solve capabilities

rxode2 5.0.2 (2026-03-20)

Allow state-dependent dur(), rate(), alag(), mtime() now allow states to modify their behavior. The state value at the time of the event is used to calculate any changes.
Fix: all six ODE solve loops now use precomputed timeThread values for event times instead of recomputing via getTime_() with ypNA, preventing NA propagation for any state-dependent lag scenario.
Export the internal .rxGetSeed() and .rxSetSeed() for use in the nlmixr2save package.
Bug fix for .copyUi() with the new format (5.0+) of rxode2 ui models
With new versions of R, getOption() is no longer a bottleneck, so syncing to local variables is no longer done internally
Allow transforms to return NA.
Drop magrittr and use |> instead of %>% in the examples (requires R 4.1)
Change default model serialization to bzip2 and move binary code generation inside of C.
Fix where getting seed saves/modifies the RNG scope, as well as a bug fix for restoring the random seed state

rxode2 5.0.1 (2025-12-09)

Change random number generation to always return doubles internally as well as no longer take a rxode2 individual structure, this is inferred by the thread number.
Change string representation of model variables to internal binary C code (to avoid macOS M1 sanitizer issues with strings).
Allow user to change the internal serialization type with options("rxode2.serialize.type"); Currently can be one of "qs2", "qdata", "base", "bzip2" and "xz". This option must be set before rxode2 is loaded, once loaded it keeps the option initially set. This is set to xz which is from base R, but could be sped up with either "qs2" (more future proof) or "qdata" (a bit faster).
Removed lsoda CDIR$ IVDEP directive, as requested by CRAN.

rxode2 5.0.0 (2025-11-28)

Better error for tad(depot) when linCmt() doesn't include a depot compartment.
Remove qs dependency; For rxode2 ui objects, use lists instead of serialized objects. The internal C++ code still generates qs2 sterilization objects (#950)
Fixed translation for censoring/limit to account for a possible CMT variable before the CENS / LIMIT column (#951, #952)
Added dmexpit() for getting the diagonal Jacobian.
Added special handling of mixest and mixunif.

rxode2 4.1.1 (2025-10-08)

Stacking for multiple-endpoint ipredSim now matches multiple-endpoint sim; Issue #929
Fix occasional $props that threw an error with empty properties (when using properties like tad0()); Issue #924
Allow mixture models mix() to be loaded with rxS() as a step to support mixtures in nlmixr2's focei; Issue #933.
Identify the correct transformation type for iov variables (#936)
Fix multiple compartment simulation edge cases where simulations were not being performed (#939)
When referencing cmt in models, the variable is forced to be CMT (related to #939)
Added ability to use mixest or mixunif to preserve the selected mixture estimates when performing a table step for a nlmixr2 mixture model (#942)

rxode2 4.1.0 (2025-08-29)

Change rxui $ evaluation when completing in rstudio, fixes strange calculations popping up in rstudio (#909)
Add orphan rxode2 model unloading when using rxUnloadAll(), and change the return type to always be a boolean.
Add assertRxUiIovNoCor to assert IOVs have no correlations in them.
Handle the levels for inter-occasion variability in the ui better (#614)
Create a new function mix() that will allow mixture models to be simulated in preparation of mixture support in nlmixr2. This allows mixture models to be specified as: v = mix(v1, p1, v2, p2, v3) where the probability of having v=v1 is modeled by p1, v=v2 is modeled by p2, and v=v3 is modeled by probability 1-p1-p2.
Created new functions mlogit() and mexpit() to convert probabilities used in mixture models to log-scaled values. mlogit() converts the probabilities to log-scaled values (using root-finding) and mexpit() converts the log values into probabilities. The equation for the conversion of log to probabilities is $p_i = \frac{exp(x_i)}{1+\sum_{j=1}^{N-1}exp(x_j)}$
Added new assertion assertRxUiNoMix which throws an error when a mixture model is present (ie mix())
Fix for label processing when calling rxode2(uiModel)

rxode2 4.0.3 (2025-07-24)

For CRAN's m1 ASAN checks of nlmixr2est, loading and unloading the same dll or by deleting the dll and recreating the exact same code, and then loading the dll will cause the ASAN check to flag an odr violation. Because of this, a mechanism to not unload dlls has been added. This allows the next version of nlmixr2est to not have issues with Mac m1 san checks.

rxode2 4.0.2 (2025-07-21)

At the request of CRAN, be a bit more careful so that names are not duplicated. Now include the md5 hash, a global counter and random 4 digit and number combination. In addition add the name of the original function so it will be easier to debug in the future.
Fall back to data.frame rbind when rbind.rxSolve() fails

rxode2 4.0.1 (2025-07-17)

Add the ability to use rbind for solved rxode2 frames.
Fix LTO issue for _rxode2_calcDerived

rxode2 4.0.0 (2025-07-16)

Add more information errors about NAs during solving.
Fix rxDerived() for mixed vector and non-vector input.
Fix model variables for alag(cmt) when they are defined before d/dt() or linCmt()
Just in time use of state.ignore in the model variables, fixes negative length error observed in #857.
Fix steady state bug with time-varying covariates. Now the covariates are inferred at the time of the steady state (instead of searching through the subject based on the projected time).
Rework the linear solved systems to use the wnl solutions, and threaded linear systems solve (for non-gradient solutions). This new method closes a variety of linear compartment model bugs (#261, #272, #441, #504, #564, #717, #728, #827, and #855)
Added new types of bounds for event tables:
- 3 point bounds et(list(c(low, mid, high))) when specified this way, they will not change. Perfect for use with babelmixr2's PopED (#862, #863, #854)
- Intervals simulated by normal values instead of uniform. In this case the first seen interval will be 3 elements with NA at the end et(list(c(mean, sd, NA), c(mean, sd))), and the other elements can simply be 2 declaring the c(mean, sd)
- Of course the uniform windows of et(list(c(low, high))) still work
- Currently these different types of windows cannot be mixed.
Add ability to pipe a list or named numeric as an eta with %>% ini(~eta)
Added a fix for event tables where expanding IDs in non-sequential order. In particular if the first ID is not the minimum ID when expanding the first event table, the smallest ID was not in the output table. Now the smallest ID is in the event table. (Fixes #878, #869, #870)
Added ability to pipe ini() or lotri(), or any other expression that can be converted to an ini with as.ini(). Also allows ini() expressions to be converted to lotri with as.lotri(). Fixes #871
Added new type of variability expression for simulation and estimation with focei and likelihood related methods: +var(). This changes standard deviation parameters to variance parameters.
Added new type of endpoint expression for focei estimation +dv(). This only transforms the data and not the predictions. I can only see it being useful in model linearization.
Bug fix for parameters that are in both input ($params) and output ($lhs) that respects the order of the $lhs declaration (Fixes #876)
Add rxFixRes to literally fix the residual estimates in a model (#889)
Now modeled duration of 0 is treated as a bolus dose (#892)

rxode2 3.0.4 (2025-02-14)

Add stable hashes for rxUi objects (#838, #689)
Fix for iov simulation (#842)
Fix for rxnbinom() called directly from R (#847) and expand it to match more close with R's rnbinom() including allowing named mu= calls. In rxode2 ui, these are also now allowed.

rxode2 3.0.3 (2024-12-15)

Add logit/expit named expressions, that is logit(x, high=20) becomes logit(x, 0, 20) in ui models.
Updated random ui models like rxnorm(sd=10) to accept complex numeric expressions like rxnorm(sd=10+1).
Updated random ui models to accept complex non-numeric expressions like rxnorm(sd=a+b)
Rework the tad() and related functions so they use the same interface as compartments (this way they do not depend on the order of compartments); See #815. For mu-referencing, Also allow dummy variables to ignore state requirements (ie podo(depot) in a single line will not error when parsing mu-referenced equations).
Add getRxNpars to api. This allows the development version of babelmixr2 to better check what model is loaded and unload/reload as necessary.
Add rxUdfUiControl() to rxode2 user function to get control information from something like nlmixr2
Bug fix for tracking time after dose when dosing to 2 compartments occur at the exact same time (#804, #819)
Change transit() model so that it uses tad0(), podo0() and related functions for a bit more stable simulation and estimation
Fix compile flags to work with BH 1.87 (#826)

rxode2 3.0.2 (2024-10-30)

Bug fix for api, the censoring function pointer has been updated (#801).
Query rxode2.verbose.pipe at run time instead of requiring it to be set before loading rxode2.
Have correct values at boundaries for logit, expit, probit, and probitInv (instead of NA). For most cases this does not break anything.
Add a new style of user function that modifies the ui while parsing or just before using the function (in the presence of data).
Used the new user function interface to allow all random functions in rxode2 ui functions to be named. For example, you can use rxnorm(sd=3) instead of having to use rxnorm(0, 3), although rxnorm() still works.

rxode2 3.0.1 (2024-09-22)

Explicitly initialize the order vector to stop valgrind warning (requested from CRAN)

rxode2 3.0.0 (2024-09-18)

Breaking Changes

The model properties was moved from $params to $props so it does not conflict with the low level rxode2 model $params
Error when specifying wd without modName
With Linear and midpoint of a time between two points, how rxode2 handles missing values has changed. When the missing value is lower than the requested time, it will look backward until it finds the first non-missing value (or if all are missing start looking forward). When the missing value is higher than the requested time, the algorithm will look forward until it finds the first non-missing value (or if all are missing, start looking backward).
The order of ODEs is now only determined by the order of cmt() and d/dt(). Compartment properties, tad() and other compartment related variables no no longer affect compartment sorting. The option rxode2.syntax.require.ode.first no longer does anything.
The handling of zeros "safely" has changed (see #775)
- when safeZero=TRUE and the denominator of a division expression is zero, use the Machine's small number/eps (you can see this value with .Machine$double.eps)
- when saveLog=TRUE and the x in the log(x) is less than or equal to zero, change this to log(eps)
- when safePow=TRUE and the expression x^y has a zero for x and a negative number for y replace x with eps.
Since the protection for divide by zero has changed, the results will also change. This is a more conservative protection mechanism than was applied previously.
Random numbers from rxode2 are different when using dop853, lsoda or indLin methods. These now seed the random numbers in the same way as liblsoda, so the random number provided will be the same with different solving methods.
The arguments saved in the rxSolve for items like thetaMat will be the reduced matrices used in solving, not the full matrices (this will likely not break very many items)

Possible breaking changes (though unlikely)

iCov is no longer merged to the event dataset. This makes solving with iCov slightly faster (#743)

New features

You can remove covariances for every omega by piping with %>% ini(diag()) you can be a bit more granular by removing all covariances that have either eta.ka or eta.cl by: %>% ini(diag(eta.ka, eta.cl)) or anything with correlations with eta.cl with %>% ini(diag(eta.cl))
You can also remove individual covariances by %>% ini(-cov(a, b)) or %>% ini(-cor(a,b)).
You can specify the type of interpolation applied for added dosing records (or other added records) for columns that are kept with the keep= option in rxSolve(). This new option is keepInterpolation and can be locf for last observation carried forward, nocb which is the next observation carried backward, as well as NA which puts a NA in all imputed data rows. See #756.
- Note: when interpolation is linear/midpoint for factors/characters it changes to locf with a warning (#759)
- Also note, that the default keep interpolation is na
Now you can specify the interpolation method per covariate in the model:
- linear(var1, var2) says both var1 and var2 would use linear interpolation when they are a time-varying covariate. You could also use linear(var1)
- locf() declares variables using last observation carried forward
- nocb() declares variables using next observation carried backward
- midpoint() declares variables using midpoint interpolation
linear(), locf(), locb(), midpoint(), params(), cmt() and dvid() declarations are now ignored when loading a rxode2 model with rxS()
Strings can be assigned to variables in rxode2.
Strings can now be enclosed with a single quote as well as a double quote. This limitation was only in the rxode2 using string since the R-parser changes single quotes to double quotes. (This has no impact with rxode2({}) and ui/function form).
More robust string encoding for symengine (adapted from utils::URLencode() and utils::URLdecode())
Empty arguments to rxRename() give a warning (#688)
Promoting from covariates to parameters with model piping (via ini()) now allows setting bounds (#692)
Added assertCompartmentName(), assertCompartmentExists(), assertCompartmentNew(), testCompartmentExists(), assertVariableExists() testVariableExists(), assertVariableNew(), assertVariableName(), and assertParameterValue() to verify that a value is a valid nlmixr2 compartment name, nlmixr2 compartment/variable exists in the model, variable name, or parameter value (#726; #733)
Added assertRxUnbounded(), testRxUnbounded(), warnRxBounded() to allow nlmixr2 warn about methods that ignore boundaries #760
Added functions tad0(), tafd0(), tlast0() and tfirst0() that will give 0 instead of NA when the dose has not been administered yet. This is useful for use in ODEs since NAs will break the solving (so can be used a bit more robustly with models like Weibull absorption).
rxode2 is has no more binary link to lotri, which means that changes in the lotri package will not require rxode2 to be recompiled (in most cases) and will not crash the system.
rxode2 also has no more binary linkage to PreciseSums
The binary linkage for dparser is reduced to C structures only, making changes in dparser less likely to cause segmentation faults in rxode2 if it wasn't recompiled.
A new model property has been added to $props$cmtProp and $statePropDf. Both are data-frames showing which compartment has properties (currently ini, f, alag, rate and dur) in the rxode2 ui model. This comes from the lower level model variable $stateProp which has this information encoded in integers for each state.
A new generic method rxUiDeparse can be used to deparse meta information into more readable expressions; This currently by default supports lower triangular matrices by lotri, but can be extended to support other types of objects like 'nlmixr2's foceiControl() for instance.

Bug fixes

Fix ui$props$endpoint when the ui endpoint is defined in terms of the ode instead of lhs. See #754
Fix ui$props when the ui is a linear compartment model without ka defined.
Model extraction modelExtract() will now extract model properties. Note that the model property of alag(cmt) and lag(cmt) will give the same value. See #745
When assigning reserved variables, the parser will error. See #744
Linear interpolation will now adjust the times as well as the values when NA values are observed.
Fix when keeping data has NA values that it will not crash R; Also fixed some incorrect NA interpolations. See #756
When using cmt() sometimes the next statement would be corrupted in the normalized syntax (like for instance locf); This bug was fixed (#763)
keep will now error when trying to keep items that are in the rxode2 output data-frame and will be calculated (#764)

Big change

At the request of CRAN, combine rxode2parse, rxode2random, and rxode2et into this package; The changes in each of the packages are now placed here:

rxode2et (no changes before merge)

rxode2et 2.0.13

Fix import of data where there are NA times

rxode2et 2.0.12

Fix formatting issues identified by m1mac, as requested by CRAN

rxode2et 2.0.11

Make the stacking more flexible to help rxode2 have more types of plots
Add toTrialDuration by Omar Elashkar to convert event data to trial duration data
Fix Issue #23 and prefer variable values over NSE values

rxode2et 2.0.10

Fix dollar sign accessing of objects (like data frames), as pointed out by @frbrz (issue #16)
Use rxode2parse functions for internal event table creation (where they were moved to).
Dropped C++14 and let the system decide.

rxode2et 2.0.9

Split off et(), eventTable() and related functions.
Also split off rxStack() and rxCbindStudyIndividual() in this package.
Added a NEWS.md file to track changes to the package.

rxode2random (before merge)

Fix a bug when simulating nested variables (#25)

rxode2random 2.1.0

Breaking Change changed distributions from the standard C++ <random> to boost::random. Since this is not dependent on the compiler, it makes the random numbers generated from Mac, Windows and Linux the same for every distribution. Unfortunately with a new random number transformation, the simulation results will likely be different than they were before. The exception to this is the uniform number, which was always the same between platforms.

rxode2random 2.0.13

Fixed formatting issues (as requested by CRAN and identified on m1mac)

rxode2random 2.0.12

Added function dfWishart which gives (by simulation) an approximation of the degrees of freedom of a Wishart to match a rse value.
Added function swapMatListWithCube which swaps omegaList with omegaCube values
Ensure that the outputs are integers (instead of long integers) as requested by CRAN for some checking functions.

rxode2random 2.0.11

Fix qassert LTO

rxode2random 2.0.10

Moved fast factor to rxode2parse to allow etTrans to be moved there

rxode2random 2.0.9

Initial release of rxode2random, which separates the parallel safe, random number generation from 'rxode2' into a separate package to reduce 'rxode2' compilation time. This should make CRAN maintenance a bit easier.
Added a NEWS.md file to track changes to the package.

rxode2parse (fixed before merging)

As requested by CRAN remove the C code SET_TYPEOF which is no longer part of the C R API.

rxode2parse 2.0.19

Added a evid suffix of 60 for cases where evid=2 adds an on event (fixes tad() calculation in certain edge cases)
Initialize all variables to NA

rxode2parse 2.0.18

Removed linear compartment solutions with gradients from rxode2parse (and rxode2) when compiled with intel c++ compiler (since it crashes while compiling).
Fixed m1mac string issues as requested by CRAN

rxode2parse 2.0.17

Added ability to query R user functions in a rxode2 model (will force single threaded solve)
Moved core rxFunParse and rxRmFunParse here so that C and R user function clashes can be handled
Model variables now tracks which compartments have a lag-time defined
For compartment with steady state doses (NONMEM equivalent SS=1, SS=2), an additional tracking time-point is added at to track the time when the lagged dose is given. As an upshot, the lagged dose will start at the steady state concentration shifted by + ii - lag in rxode2 (currently for ode systems only)
This release calculates non bio-availability adjusted duration for all rates instead of trying to figure the rate duration during solving.
Make double assignment an error, ie a <- b <-
NA times are ignored (with warning)
Steady state bolus doses with addl are treated as non steady state events (like what is observed in NONMEM)
Timsort was upgraded; drop radix support in rxode2 structure
etTrans now supports keeping logical vectors (with the appropriate version of rxode2).
Security fixes were applied as requested by CRAN

rxode2parse 2.0.16

Import data.table explicitly in the R code (before was imported only in C/C++ code)

rxode2parse 2.0.15

Updates the make flags to support CXX17.

rxode2parse 2.0.14

'linCmt()' translations of 'alpha', 'beta', 'gamma', 'k21', 'k31', 'vc' now error instead of ignoring 'gamma' and 'k31' to give 2 cmt solution
transit compartment internal code now changes dose to 0.0 when no dose has been administered to the depot compartment. This way dosing to the central compartment (without dosing to the transit compartment) will not give a NA for the depot compartment (and consequently for the central compartment)
Moved rxDerived here and added tests for it here as well.
Moved etTransParse here and added tests for it here as well (makes up most of etTrans). In addition the following changes were made to etTransParse()/etTrans():
- The internal translation (etTrans()) will not drop times when infusions stop. Before, if the infusion stopped after the last observation the time when the infusion stopped would be dropped. This interferes with linCmt() models.
- Breaking change/bug fix evid=2 are considered observations when translating data to internal rxode2 event structure
- Fix edge case to find infusion duration when it is the first item of the dosing record at time 0.
Fixed a bug for certain infusions where the rate, ii and/or ss data items were dropped from the output when addDosing=TRUE
Also have internal functions to convert between classic NONMEM events and rxode2 events
Have an internal function that gives information on the linear compartmental model translation type, which could be useful for babelmixr2
'time' in model is now case insensitive
Use function declaration in rxode2parseGetTranslation() to determine thread safety of functions available to rxode2
Add check for correct number of function arguments to parser.
Like R, known functions can be assigned as a variable and the function can still be called (while not changing the variable value). For example you can have a variable gamma as well as a function gamma().
Fix garbled error messages that occur with certain messages.
Fixed errors that occurred when using capitalized AMT variables in the model.

rxode2parse 2.0.13

Version bump for dparser (so binaries will be built correctly)

rxode2parse 2.0.12

Bug fix for strict prototypes
Removed sprintf as noted by CRAN
Made rxode2parse dll binary independent of rxode2()

rxode2parse 2.0.11

Bug fix for strict aliasing as requested by CRAN

rxode2parse 2.0.10

Use strict aliasing as requested by CRAN

rxode2parse 2.0.9

Initial release to split of rxode2parse from rxode2 to reduce compilation time of 'rxode2'

rxode2 2.1.3 (2024-05-28)

Bug fixes

Make sure that the object is a uncompressed rxode2 ui for solving with rxSolve (See #661)
Fix #670 by using the last simulated observation residual when there are trailing doses.

New features

Create a function to see if a rxode2 solve is loaded in memory (rxode2::rxSolveSetup())
Create a new function that fixes the rxode2 population values in the model (and drops them in the initial estimates); rxFixPop()

Other changes

Pendantic no-remap (as requested by CRAN)
gcc USBAN fix (as requested by CRAN)

rxode2 2.1.2 (2024-01-30)

Other changes

rxUi compression now defaults to fast compression
Fixes String literal formatting issues as identified by CRAN (#643)
Removes linear compartment solutions with gradients for intel c++ compiler (since they crash the compiler).

rxode2 2.1.0 (2023-12-11)

Breaking changes

Steady state with lag times are no longer shifted by the lag time and then solved to steady state by default. In addition the steady state at the original time of dosing is also back-calculated. If you want the old behavior you can bring back the option with ssAtDoseTime=FALSE.
"dop853" now uses the hmax/h0 values from the rxControl() or rxSolve(). This may change some ODE solving using "dop853"
When not specified (and xgxr is available), the x axis is no longer assumed to be in hours

New features

User defined functions can now be R functions. For many of these R functions they can be converted to C with rxFun() (you can see the C code afterwards with rxC("funName"))
Parallel solving of models that require sorting (like modeled lag times, modeled duration etc) now solve in parallel instead of downgrading to single threaded solving
Steady state infusions with a duration of infusions greater than the inter-dose interval are now supported.
Added $symengineModelNoPrune and $symengineModelPrune for loading models into rxode2 with rxS()
When plotting and creating confidence intervals for multiple endpoint models simulated from a rxode2 ui model, you can plot/summarize each endpoint with sim. (ie. confint(model, "sim") or plot(model, sim)).

If you only want to summarize a subset of endpoints, you can focus on the endpoint by pre-pending the endpoint with sim. For example if you wanted to plot/summarize only the endpoint eff you would use sim.eff. (ie confint(model, "sim.eff") or plot(model, sim.eff))
Added model$simulationIniModel which prepend the initial conditions in the ini({}) block to the classic rxode2({}) model.
Now model$simulationModel and model$simulationIniModel will save and use the initialization values from the compiled model, and will solve as if it was the original ui model.
Allow ini(model) <- NULL to drop ini block and as.ini(NULL) gives ini({}) (Issue #523)
Add a function modelExtract() to extract model lines to allow modifying them and then changing the model by piping or simply assigning the modified lines with model(ui) <- newModifiedLines
Add Algebraic mu-referencing detection (mu2) that allows you to express mu-referenced covariates as:

cl <- exp(tcl + eta.cl + wt_cl * log(WT/70.5))

Instead of the

cl <- exp(tcl + eta.cl + wt_cl * log.WT.div.70.5)

That was previously required (where log.WT.div.70.5 was calculated in the data) for mu expressions. The ui now has more information to allow transformation of data internally and transformation to the old mu-referencing style to run the optimization.

Allow steady state infusions with a duration of infusion greater than the inter-dose interval to be solved.
Solves will now possibly print more information when issuing a "could not solve the system" error
The function rxSetPipingAuto() is now exported to change the way you affect piping in your individual setup
Allow covariates to be specified in the model piping, that is mod %>% model(a=var+3, cov="var") will add "var" as a covariate.
When calculating confidence intervals for rxode2 simulated objects you can now use by to stratify the simulation summary. For example you can now stratify by gender and race by: confint(sim, "sim", by=c("race", "gender"))
When calculating the intervals for rxode2 simulated objects you can now use ci=FALSE so that it only calculates the default intervals without bands on each of the percentiles; You can also choose not to match the secondary bands limits with levels but use your own ci=0.99 for instance
A new function was introduced meanProbs() which calculates the mean and expected confidence bands under either the normal or t distribution
A related new function was introduced that calculates the mean and confidence bands under the Bernoulli/Binomial distribution (binomProbs())
When calculating the intervals for rxode2 simulated objects you can also use mean=TRUE to use the mean for the first level of confidence using meanProbs(). For this confidence interval you can override the n used in the confidence interval by using n=#. You can also change this to a prediction interval instead using pred=TRUE.
Also when calculating the intervals for rxode2 simulated object you can also use mean="binom" to use the binomial distributional information (and ci) for the first level of confidence using binomProbs(). For this confidence interval you can override the n used in the confidence interval by using n=#. You can also change this to a prediction interval instead using pred=TRUE. With pred=TRUE you can override the number of predicted samples with m=#
When plotting the confint derived intervals from an rxode2 simulation, you can now subset based on a simulated value like plot(ci, Cc) which will only plot the variable Cc that you summarized even if you also summarized eff (for instance).
When the rxode2 ui is a compressed ui object, you can modify the ini block with $ini <- or modify the model block with $model <-. These are equivalent to ini(model) <- and model(model) <-, respectively. Otherwise, the object is added to the user defined components in the function (ie $meta). When the object is uncompressed, it simply assigns it to the environment instead (just like before).
When printing meta information that happens to be a lotri compatible matrix, use lotri to express it instead of the default R expression.
Allow character vectors to be converted to expressions for piping (#552)
rxAppendModel() will now take an arbitrary number of models and append them together; It also has better handling of models with duplicate parameters and models without ini() blocks (#617 / #573 / #575).
keep will now also keep attributes of the input data (with special handling for levels); This means a broader variety of classes will be kept carrying more information with it (for example ordered factors, data frame columns with unit information, etc)
Piping arguments append for ini() and model() have been aligned to perform similarly. Therefore ini(append=) now can take expressions instead of simply strings and model(append=) can also take strings. Also model piping now can specify the integer line number to be modified just like the ini() could. Also model(append=FALSE) has been changed to model(append=NULL). While the behavior is the same when you don't specify the argument, the behavior has changed to align with ini() when piping. Hence model(append=TRUE) will append and model(append=FALSE) will now pre-pend to the model. model(append=NULL) will modify lines like the behavior of ini(append=NULL). The default of model(line) modifying a line in-place still applies. While this is a breaking change, most code will perform the same.
Labels can now be dropped by ini(param=label(NULL)). Also parameters can be dropped with the idiom model(param=NULL) or ini(param=NULL) changes the parameter to a covariate to align with this idiom of dropping parameters
rxRename has been refactored to run faster

Internal new features

Add as.model() for list expressions, which implies model(ui) <- ui$lstExpr will assign model components. It will also more robustly work with character vectors
Simulated objects from rxSolve now can access the model variables with $rxModelVars
Simulation models from the UI now use rxerr.endpoint instead of err.endpoint for the sigma residual error. This is to align with the convention that internally generated variables start with rx or nlmixr
Sorting only uses timsort now, and was upgraded to the latest version from Morwenn

Bug fixes

Simulating/solving from functions/ui now prefers params over omega and sigma in the model (#632)
Piping does not add constants to the initial estimates
When constants are specified in the model({}) block (like k <- 1), they will not be to the ini block
Bug fix for geom_amt() when the aes transformation has x
Bug fix for some covariate updates that may affect multiple compartment models (like issue #581)

Maintenance fixes

Modify plot code to work with development xgxr

rxode2 2.0.14 (2023-10-07)

CRAN requested that FORTRAN kind be changed as it was not portable; This was commented code, and simply removed the comment.
Bug-fix for geom_amt(); also now uses linewidth and at least ggplot2 3.4.0
Some documentation was cleaned up from rxode2 2.0.13

rxode2 2.0.13 (2023-04-22)

Bug fixes

A bug was fixed so that the zeroRe() function works with correlated omega values.
A bug was fixed so that the rename() function works with initial conditions for compartments (cmt(0))

New features

A new function zeroRe() allows simple setting of omega and/or sigma values to zero for a model (#456)
Diagonal zeros in the omega and sigma matrices are treated as zeros in the model. The corresponding omega and sigma matrices drop columns/rows where the diagonals are zero to create a new omega and sigma matrix for simulation. This is the same idiom that NONMEM uses for simulation from these matrices.
Add the ability to pipe model estimates from another model by parentModel %>% ini(modelWithNewEsts)
Add the ability to append model statements with piping using %>% model(x=3, append=d/dt(depot)), still supports appending with append=TRUE and pre-pending with append=NA (the default is to replace lines with append=FALSE)
rxSolve's keep argument will now maintain character and factor classes from input data with the same class (#190)
Parameter labels may now be modified via ini(param = label("text")) (#351).
Parameter order may be modified via the append argument to ini() when piping a model. For example, ini(param = 1, append = 0) or ini(param = label("text"), append = "param2") (#352).

Internal changes

If lower/upper bounds are outside the required bounds, the adjustment is displayed.
When initial values are piped that break the model's boundary condition reset the boundary to unbounded and message which boundary was reset.
Added as.rxUi() function to convert the following objects to rxUi objects: rxode2, rxModelVars, function. Converting nlmixr2 fits to rxUi will be placed in the s3 method in the corresponding package.
assertRxUi(x) now uses as.rxUi() so that it can be extended outside of rxode2/nlmixr2.
rxode2 now supports addl with ss doses
Moved rxDerived to rxode2parse (and re-exported it here).
Added test for transit compartment solving in absence of dosing to the transit compartment (fixed in rxode2parse but solving tested here)
Using ini() without any arguments on a rxode2 type function will return the ini() block. Also added a method ini(mod) <- iniBlock to modify the ini block is you wish. iniBlock should be an expression.
Using model() without any arguments on a rxode2 type function will return the model() block. Also added a new method model(mod) <- modelBlock
Added a new method rxode2(mod) <- modFunction which allows replacing the function with a new function while maintaining the meta information about the ui (like information that comes from nonmem2rx models). The modFunction should be the body of the new function, the new function, or a new rxode2 ui.
rxode2 ui objects now have a $sticky item inside the internal (compressed) environment. This $sticky tells what variables to keep if there is a "significant" change in the ui during piping or other sort of model change. This is respected during model piping, or modifying the model with ini(mod)<-, model(mod)<-, rxode2(mod)<-. A significant change is a change in the model block, a change in the number of estimates, or a change to the value of the estimates. Estimate bounds, weather an estimate is fixed or estimate label changes are not considered significant.
Added as.ini() method to convert various formats to an ini expression. It is used internally with ini(mod)<-. If you want to assign something new that you can convert to an ini expression, add a method for as.ini().
Added as.model() method to convert various formats to a model expression. It is used internally with model(mod)<-. If you want to assign something new that you can convert to a model expression, add a method for as.model().

rxode2 2.0.11 (2022-11-01)

Give a more meaningful error for 'rxode2' ui models with only error expressions
Break the ABI requirement between roxde2() and rxode2parse()
The new rxode2parse will fix the sprintf exclusion shown on CRAN.

rxode2 2.0.10 (2022-10-20)

Time invariant covariates can now contain 'NA' values.
When a column has 'NA' for the entire id, now 'rxode2' warns about both the id and column instead of just the id.
To fix some CRAN issues in 'nlmixr2est', make the version dependency explicit.

rxode2 2.0.9 (2022-10-19)

Remove log likelihoods from 'rxode2' to reduce compilation time and increase maintainability of 'rxode2'. They were transferred to 'rxode2ll' (requested by CRAN).
Remove the parsing from 'rxode2' and solved linear compartment code and move to 'rxode2parse' to reduce the compilation time (as requested by CRAN).
Remove the random number generation from 'rxode2' and move to 'rxode2random' to reduce the compilation time (as requested by CRAN).
Remove the event table translation and generation from 'rxode2' and move to 'rxode2et' to reduce the compilation time (as requested by CRAN).
Change the rxode2 ui object so it is a compressed, serialized object by default. This could reduce the C stack size problem that occurs with too many environments in R.
Warn when ignoring items during simulations
Export a method to change 'rxode2' solve methods into internal integers
Bug fix for time invariant covariates identified as time variant covariate when the individual's time starts after 0.

rxode2 2.0.8 (2022-09-23)

Breaking changes

rxgamma now only allows a rate input. This aligns with the internal rxode2 version of rxgamma and clarifies how this will be used. It is also aligned with the llikGamma function used for generalized likelihood estimation.
ui cauchy simulations now follow the ui for normal and t distributions, which means you can combine with transformations. This is because the cauchy is a t distribution with one degree of freedom.
ui dnorm() and norm() are no longer equivalent to add(). Now it allows you to use the loglik llikNorm() instead of the standard nlmixr2 style focei likelihood. This is done by adding dnorm() at the end of the line. It also means dnorm() now doesn't take any arguments.
Vandercorput normal removed (non-random number generator)

New features

Allow models in the nlmixr2 form without an ini({}) block
Allow model piping of an omega matrix by f %>% ini(omegaMatrix)
Standard models created with rxode2() can no be piped into a model function
Families of log-likelihood were added to rxode2 so that mixed likelihood nonlinear mixed effects models may be specified and run.
The memory footprint of a rxode2 solving has been reduced
Piping now allow named strings (issue #249)

Bug fixes

rxode2's symengine would convert sqrt(2) to M_SQRT_2 when it should be M_SQRT2. This has been fixed; it was most noticeable in nlmixr2 log-likelihood estimation methods
rxode2 treats DV as a non-covariate with etTran (last time it would duplicate if it is in the model). This is most noticeable in the nlmixr2 log-likelihood estimation methods.

New features

A new flag (rxFlag) has been created to tell you where in the rxode2 solving process you are. This is useful for debugging. If outputting this variable it will always be 11 or calculating the left handed equations. If you are using in conjunction with the printf() methods, it is a double variable and should be formatted with "%f".
An additional option of fullPrint has been added to rxode2() which allows rprintf() to be used in almost all of rxode2() steps (inductive linearization and matrix exponential are the exception here) instead of just the integration ddt step. It defaults to FALSE.

rxode2 2.0.7 (2022-05-17)

Removed accidental ^S from news as requested by CRAN.
Bug fix for more complicated mu-referencing.
Change rxode2 md5 to only depend on the C/C++/Fortran code and headers not the R files. That way if there is binary compatibility between nlmixr2est and rxode2, a new version of nlmixr2est will not need to be submitted to CRAN.

rxode2 2.0.6 (2022-05-09)

Breaking changes

Solving controls

The options for rxControl and rxSolve are more strict. camelCase is now always used. Old options like add.cov and transit_abs are no longer supported, only addCov is supported.
A new option, sigdig has been added to rxControl(), which controls some of the more common significant figure options like atol, rtol, ssAtol, ssRtol, with a single option.

Simulations

For simulations, $simulationSigma now assumes a diagonal matrix. The sigma values are assumed to be standard normal, and uncorrelated between endpoints. Simulation with uncertainty will still draw from this identity diagonal matrix
Parallel solving now seeds each simulation per each individual based on the initial seed plus the simulation id. This makes the simulation reproducible regardless of the number of cores running the simulation.

Other breaking changes

Solved objects now access the underlying rxode model with $rxode2 instead of $rxode
Since this change names, rxode2, rxode and RxODE all perform the same function.
Options were changed from RxODE.syntax to rxode2.syntax.
Assigning states with rxode2.syntax.assign.state (was RxODE.syntax.assign.state) is no longer supported.
Enforcing "pure" assignment syntax with = syntax is no longer supported so rxode2.syntax.assign is no longer supported (was RxODE.syntax.assign).
Since R supports ** as an exponentiation operator, the pure syntax without ** can no longer be enabled. Hence rxode2.syntax.star.pow (was RxODE.syntax.star.pow) no longer has any effect.
The "pure" syntax that requires a semicolon can no longer be enabled. Therefore rxode2.syntax.require.semicolon (was RxODE.syntax.require.semicolon) no longer has any effect.
The syntax state(0) can no longer be turned off. rxode2.syntax.allow.ini0 (was RxODE.syntax.allow.ini0) has been removed.
Variable with dots in variable and state names like state.name works in R. Therefore, "pure" syntax of excluding . values from variables cannot be enforced with rxode2.syntax.allow.dots (was RxODE.syntax.allow.dots).
The mnemonic et(rate=model) and et(dur=model) mnemonics have been removed. rate needs to be set to -1 and -2 manually instead.
The function rxode2Test() has been removed in favor of using testthat directly.
Transit compartments need to use a new evid, evid=7. That being said, the transitAbs option is no longer supported.
ID columns in input parameter data frames are not sorted or merged with original dataset any more; The underlying assumption of ID order should now be checked outside of rxode2(). Note that the event data frame is still sorted.

Additional features

The UI functions of nlmixr have been ported to work in rxode2 directly.
rxModelVars({}) is now supported.
You may now combine 2 models in rxode2 with rxAppendModel(). In fact, as long as the first value is a rxode2 evaluated ui model, you can use c/rbind to bind 2 or more models together.
You may now append model lines with piping using %>% model(lines, append=TRUE) you can also pre-pend lines by %>% model(lines, append=NA)
You may now rename model variables, states and defined parameters with %>% rxRename(new=old) or if dplyr is loaded: %>% rename(new=old)
You can fix parameters with %>% ini(tcl=fix) or %>% ini(fix(tcl)) as well as unfix parameters with %>% ini(tcl=unfix) or %>% ini(unfix(tcl))

Internal changes

Strict R headers are enforced more places
Since there are many changes that could be incompatible, this version has been renamed to rxode2
rxode2() printout no longer uses rules and centered headings to make it display better on a larger variety of systems.

Bug fixes

tad() and related time features only reset at the start of an infusion (as opposed to starting at the beginning and end of an infusion)

rxode2 1.1.3

Change handling of missing covariates while interpolating "nocb" so that the time-varying covariates use "nocb" interpolation (#469)

rxode2 1.1.2

Fix subject initialization of focei problem (#464)
Fix LHS offset to allow internal threading and more parallel processing in the future.
Remove warnings for duration and rate
Don't export pillar methods any more (simply register at load if present)
As requested by CRAN, change fortran and C binding for BLAS an LINPACK

rxode2 1.1.1

Fix the LTO issue that CRAN identified.
Move the omp files so they come first to support clang13, as identified by CRAN.
For now, be a little more conservative in dur() and rate() warnings because linCmt() models in nlmixr currently produce irrelevant warnings.

rxode2 1.1.0

Always calculate "nolhs" for using numeric differences when the inner problem. This allows the inner problem to fallback to a finite difference approximation to the focei objective function.
Updated the parser C code grammar using latest dparser CRAN package
Added a new cbind function that is used to mix data frame input with simulated individual parameters and residual parameters, rxCbindStudyIndividual().
Now data frame input can be mixed with simulating from omega and sigma matrices (though not yet in nested simulations)
Race conditions when simulating random numbers is solved by chunking each simulation into groups that will always be performed per each thread. This way the simulation is now reproducible regardless of load. Because of the chunking, simulations with random numbers generated inside of it are now threaded by default (though a warning is produced about the simulation only be reproducible when run with the same number of threads)
Simulations were double checked and made sure to use the engine reserved for each core run in parallel; Some of the random generators were not taking random numbers from the correct engine, which was corrected. Therefore, simulations from this version are expected to be different (in parallel) than previous versions.
Added function rxSetSeed() to set the internal RxODE seed instead of grabbing it from a uniform random number tied to the original R seed. This will avoid the possibility of duplicate seeds and is the best practice.
Updating parameter pointers is done once per ID and locked based on ID to remove the recursion in #399, but still have the correct behavior see #430
Parsing updated to retain "param()" in normalized model, #432.
Handle edge case of interpolation at first index correctly, fixes #433
Instead of storing each dose information sequentially, store dose information at the same index of the evid defining the dose. This memory rewrite is to fix the issue #435.
Start using strict headers as it is required for the forthcoming release of Rcpp. Thanks to Dirk Eddelbuettel for some of the fixes and alerting us to this change.
Check arguments for add.dosing() more strictly. See Issue #441
Issue a warning when either dur() or rate() is in the model but the modeled rate and duration is not included in the event table.
When the data requires a modeled rate and modeled duration but it is not in the model, warn about the mismatch in data
Added a back-door for debugging. If you specify options(RxODE.debug=TRUE) then each solve saves the solving information to the file "last-rxode.qs" before actually solving the system.
Only will try to solve RxODE problems on compatible models; If the model is not supported it will throw an error instead of crashing (See #449)
Turn off parallel ODE solving whenever the system needs to sort times based on model dosing. Currently this type of solving is not thread safe.
Update timsort headers to latest version.

rxode2 1.0.9

At the request of CRAN, stripping the debugging symbols for the CRAN release is no longer performed. This means a larger binary size for RxODE in this release.
At the request of CRAN the liblsoda code has been changed so that the memory in C defined by _C() is now defined by _rxC(). This will be seen in some of the error messages, which will no longer match the error messages of unmodified liblsoda.
iCov behavior has shifted to merge on the input event dataset. See Issue #409; This is more in line with expectations of iCov behavior, and reduces the amount of code needed to maintain iCov.

The iCov in the pipeline is no longer supported because it simply is a merge with the event dataset.

This can be a breaking change depending on the code you use. Note that clinical trial simulations, resampling is likely better than trying to fill out iCov for every individual which was the prior use.
Bug fix for crashes with string covariates or factor covariates, issue #410. Also factor column names are compared with case insensitivity just like the rest of the column names for event tables or data sets in RxODE.

rxode2 1.0.8

Fix issue #399

rxode2 1.0.7

Change syntax vignette to use markdown option screenshot.force=FALSE. This should get rid of the webshot error
Change to depend on dparser 1.3.0, which has some memory fixes

rxode2 1.0.6

RxODE imports but does not link to checkmate any longer. This change should make recompilation of RxODE to work with different releases of checkmate unnecessary.
Default Solaris solver changed back to "lsoda"
Fix Bug #393, where in certain circumstances rxSolve(...,theta=) did not solve for all subjects.
Will not ignore NEWS and README when building the package so that they will show up on CRAN. You can also access the news by news(package="RxODE")
Changed ODR model names from time id to _rx followed by the md5 hash id and a per-session counter id; For packages the id is _rxp followed by the md5 hash and a per-session counter id.
Changed qs to be more conservative in hash creation. Add a check hash as well as NOT using altrep stringfish representation.

rxode2 1.0.5

Maintenance release -- use std::floor and cast variables to double for internal C functions. This should allow a successful compile on Solaris CRAN.
Changed units from an Imports to a Suggests to allow testing on Solaris rhub
Changed ODR model names from time id to _rx followed by the md5 hash id; For packages the id is _rxp followed by the md5 hash.
Removed AD linear compartment solutions for Windows R 3.6, though they still work for Windows R 4.0 (You can get them back for Windows R 3.6 if you install BH 1.66.0-1 and then recompile from source).
- This will cause nlmixr to fail with solved systems on Windows 3.6. Currently the Stan Headers do not compile on this system so they are disabled at this time.
RxODE imports but does not link to qs any longer; This change should make recompilation of RxODE to work with different releases of qs unnecessary.
RxODE now checks for binary compatibility for Rcpp, dparser, checkmate, and PreciseSums

rxode2 1.0.4

Breaking changes

RxODE can only use supported functions (could be breaking); You may add your own functions with rxFun and their derivatives with rxD
RxODE now uses its own internal truncated multivariate normal simulations based on the threefry sitmo library. Therefore random numbers generated within RxODE like providing rxSolve(...,omega=) will have different results with this new random number generator. This was done to allow internal re-sampling of sigmas/etas with thread-safe random number generators (calling R through mvnfast or R's simulation engines are not thread safe).
RxODE now moved the precise sum/product type options for sum() and prod() to rxSolve or rxControl
cvPost now will returned a named list of matrices if the input matrix was named
rxSolve will now return an integer id instead of a factor id when id is integer or integerish (as defined by checkmate). Otherwise a factor will be returned.
When mixing ODEs and linCmt() models, the linCmt() compartments are 1 and possibly 2 instead of right after the last internal ODE. This is more aligned with how PK/PD models are typically defined.
EVID=3 and EVID=4 now (possibly) reset time as well. This occurs when the input dataset is sorted before solving.
When EVID=2 is present, an evid column is output to distinguish evid=0 and evid=2

New features

Add the ability to order input parameters with the param() pseudo-function
Add the ability to resample covariates with resample=TRUE or resample=c("SEX", "CRCL"). You can resample all the covariates by ID with resampleID=TRUE or resample the covariates without respect to ID with resampleID=FALSE
Comparison of factors/strings is now supported in RxODE; Therefore ID=="Study-1" is now allowed.
Completion for elements of rxSolve() objects, and et() objects have been added (accessed through $)
Completion of rxSolve() arguments are now included since they are part of the main method
Allow simulation with zero matrices, that provide the simulation without variability. This affects rxSolve as well as rxMvnrnd and cvPost (which will give a zero matrix whenever one is specified)
et() can dose with length(amt) > 1 as long as the other arguments can create a event table.
Rstudio notebook output makes more sense
Printing upgraded to cli 2.0
Caching of internal C data setup is now supported increasing speed of optim code when:
- Event Table doesn't change
- The size of the parameters doesn't change
- inits do not change (though you can specify them as cmt(0)=... in the model and change them by parameters)
- See Issue #109
Allow while(logical) statements with ability to break out if them by break. The while has an escape valve controlled by maxwhere which by default is 10000 iterations. It can be change with rxSolve(..., maxwhere = NNN)
Allow accessing different time-varying components of an input dataset for each individual with:
- lag(var, #)
- lead(var, #)
- first(var)
- last(var)
- diff(var)

Each of these are similar to the R lag, lead, first, last and diff. However when undefined, it returns NA

Allow sticky left-handed side of the equation; This means for an observation the left handed values are saved for the next observations and then reassigned to the last calculated value.

This allows NONMEM-style of calculating parameters like tad:

mod1 <-RxODE({
    KA=2.94E-01;
    CL=1.86E+01;
    V2=4.02E+01;
    Q=1.05E+01;
    V3=2.97E+02;
    Kin=1;
    Kout=1;
    EC50=200;
    C2 = centr/V2;
    C3 = peri/V3;
    d/dt(depot) =-KA*depot;
    d/dt(centr) = KA*depot - CL*C2 - Q*C2 + Q*C3;
    d/dt(peri)  =                    Q*C2 - Q*C3;
    d/dt(eff)  = Kin - Kout*(1-C2/(EC50+C2))*eff;
    if (!is.na(amt)){
        tdose <- time
    } else {
        tad <- time - tdose
    }
})

It is still simpler to use:

mod1 <-RxODE({
    KA=2.94E-01;
    CL=1.86E+01;
    V2=4.02E+01;
    Q=1.05E+01;
    V3=2.97E+02;
    Kin=1;
    Kout=1;
    EC50=200;
    C2 = centr/V2;
    C3 = peri/V3;
    d/dt(depot) =-KA*depot;
    d/dt(centr) = KA*depot - CL*C2 - Q*C2 + Q*C3;
    d/dt(peri)  =                    Q*C2 - Q*C3;
    d/dt(eff)  = Kin - Kout*(1-C2/(EC50+C2))*eff;
    tad <- time - tlast
})

If the lhs parameters haven't been defined yet, they are NA

Now the NONMEM-style newind flag can be used to initialize lhs parameters.
Added tad(), tad(cmt) functions for time since last dose and time since last dose for a compartment; Also added time after first dose and time after first dose for a compartment tafd(), tafd(cmt); time of last dose tlast(), tlast(cmt) and dose number dosenum() (currently not for each compartment)
Changed linear solved systems to use "advan" style linCmt() solutions, to allow correct solutions of time-varying covariates values with solved systems; As such, the solutions may be slightly different. Infusions to the depot compartment are now supported.
Added sensitivity auto-differentiation of linCmt() solutions. This allows sensitivities of linCmt() solutions and enables nlmixr focei to support solved systems.
- One solution is to use Stan's auto-differentiation which requires C++14
When calculating the empirical Bayesian estimates for with rxInner (used for nlmixr's 'focei') ignore any variable beginning with rx_ and nlmixr_ to hide internal variables from table output. This also added tad=tad() and dosenum=dosenum() to the ebe output allowing grouping by id, dose number and use TAD for individual plot stratification.
Added ability to prune branching with rxPrune. This converts if/else or ifelse to single line statements without any if/then branching within them.
Added ability to take more complex conditional expressions, including:
- ifelse(expr, yes, no)
- x = (x==1)*1 + (!(x==1))*2
- if (logic){ expr} else if (logic) {expr} else {}. The preferred syntax is still only if/else and the corresponding parsed code reflects this preference.
  - Note ifelse is not allowed as an ODE compartment or a variable.
Switched to symengine instead of using sympy
- Remove dependence on python.
- Since symengine is C-based and doesn't require the python interface it is much faster than sympy, though some functions in sympy are no longer accessible.
- Also symengine requires R 3.6, so now RxODE requires R 3.6
Added new ODE solving method "indLin", or inductive linearization. When the full model is a linear ODE system this becomes simply the matrix exponential solution. Currently this requires a different setup.
Added arbitrary function definition to RxODE using rxFun
- Requires function, arguments and corresponding C-code
- Derivatives (if required) can be added to the derivative table rxD. When taking deviates without a derivative function, RxODE will use numerical differences.
Will error if RxODE does not know of a function that you are trying to use; This could be a breaking change. Currently:
- C's functions from math.h are supported
- R's function returning and taking doubles are supported
- Other functions can be added using rxFun and rxD
Added NA, NaN, Inf and +Inf handling to a RxODE model. Can be useful to diagnose problems in models and provide alternate solutions. In addition, added R-like functions is.nan, is.na, is.finite and is.infinite which can be called within the RxODE block.
Allowed the following data variables can be accessed (but not assigned or used as a state):
- cmt
- dvid
- addl
- ss
- amt
- rate
- id which requires calling the id as factor ID=="1" for instance.
Kept evid and ii as restricted items since they are not part of the covariate table and are restricted in use.
Added the following random number generators; They are thread safe (based on threefry sitmo and c++11) and your simulations with them will depend on the number of cores used in your simulation (Be careful about reproducibility with large number of threads; Also use parallel-solve type of RxODE simulations to avoid the birthday problem).

During ODE solving, the values of these are 0, but while calculating the final output the variable is randomized at least for every output. These are:
- rxnorm() and rxnormV() (low discrepancy normal)
- rxcauchy()
- rxchisq()
- rxexp()
- rxf()
- rxgamma()
- rxbeta()
- rxgeom()
- rxpois()
- rxt()
- rxunif()
- rxweibull()
In addition, while initializing the system, the following values are simulated and retained for each individual:
- rinorm() and rinormV() (low discrepancy normal)
- ricauchy()
- richisq()
- riexp()
- rif()
- rigamma()
- ribeta()
- rigeom()
- ripois()
- rit()
- riunif()
- riweibull()
Added simeta() which simulates a new eta when called based on the possibly truncated normal omega specified by the original simulation. This simulation occurs at the same time as the ODE is initialized or when an ODE is missing, before calculating the final output values. The omega will reflect whatever study is being simulated.
Added simeps() which simulates a new eps from the possibly truncated normal sigma at the same time as calculating the final output values. Before this time, the sigma variables are zero.

All these change the solving to single thread by default to make sure the simulation is reproducible. With high loads/difficult problems the random number generator may be on a different thread and give a different number than another computer/try.

Also please note that the clang and gcc compiler use different methods to create the more complex random numbers. Therefore MacOS random numbers will be different than Linux/Windows at this time (with the exception of uniform numbers).

These numbers are still non-correlated random numbers (based on the sitmo test) with the exception of the vandercorput distributions, so if you increase the number of threads (cores=...) the results should still be valid, though maybe harder to reproduce. The faster the random number generation, the more likely these results will be reproduced across platforms.

Added the ability to integrate standard deviations/errors of omega diagonals and sigma diagonals. This is done by specifying the omega diagonals in the theta matrix and having them represent the variabilities or standard deviations. Then these standard deviations are simulated along with the correlations using the IJK correlation matrix (omega dimension < 10) or a correlation matrix or Inverse Wishart-based correlation matrix (omega dimension > 10). The information about how to simulate this is in the variability simulation vignette.
Now have a method to use lotri to simulate between occasion variability and other levels of nesting.
Added lower gamma functions See Issue #185
Upgraded comparison sort to timsort 2.0.1
Changed in-place sort to a modified radix sort from data.table. The radix search was modified to:

Work directly with RxODE internal solved structures
Assume no infinite values or NA/NaN values of time
Always sort time in ascending order
Changed sorting to run in a single thread instead of taking over all the threads like data.table

Changed method for setting/getting number of threads based on data.table's method
Added function rxDerived which will calculate derived parameters for 1, 2, and 3 compartment models
More descriptive errors when types of input are different than expected

Engine changes

Moved many C functions to C++. CRAN OpenMP support requires C++ only when C and C++ are mixed. See:

https://stackoverflow.com/questions/54056594/cran-acceptable-way-of-linking-to-openmp-some-c-code-called-from-rcpp
No longer produces C code that create the model variables. Instead, use qs to serialize, compress and encode in base91 and then write the string into the C file. The qs package then decodes all of that into the model variables. This also increases the compilation speed for models in RxODE.
Pre-compile RxODE headers once (if cache is enabled), which increases compilation speed for models in RxODE
RxODE's translation from the mini-language to C has been refactored

Bug fixes:

Occasionally RxODE misidentified dual lhs/param values. An additional check is performed so that this does not happen.
For solved matrices with similar names (like "tadd" and "tad") RxODE will now prefer exact matches instead of the first match found when accessing the items with $tad.
A fix where all ID information is kept with keep=c(""..."")
Transit compartment models using the transit ODE or variable are now allowed. Also check for more internally parsed items (see Issue #145).
Bug fix for etSeq and etRep where greater than 2 items were mis-calculated

rxode2 0.9.2-0

New plotting engine
Various bug fixes for upcoming R 4.0 release:
- Dropped some imports for 21 imports restriction
- Fixed incompatibility with new ggplot2 3.3.0
- Fixed allowing NAs in RxODE dataset
- Fixed setting all compartment default values for bioavailability, rate, etc.
- Added additional protection against floating point -> NaN for power functions

rxode2 0.9.1-9

Minor namespace/documentation changes for R 4.0 compatibility

rxode2 0.9.1-8

Added the ability to have an input parameter to be assigned to a new value (Issue #135)
Added LINPACK authors as contributors
Added a NEWS.md file to track changes to the package

NEWS

rxode2 5.1.8

New features

Bug fixes

rxode2 5.1.7.1 (2026-09-29)

rxode2 5.1.7 (2026-09-14)

New features

Bug fixes

Event translation

New features

Bug fixes

Compilation

Mixture models

Model piping

Estimation / symengine translation

Event translation

Simulation

Initial conditions data frame

Parsing

Model interface

Solving

Sensitivities

Delay differential equations

Matrix exponential / inductive linearization

Serialization

Installation / linking

rxode2 5.1.6 (2026-08-04)

New features

Bug fixes

Compilation cache

Dependencies

Installation / linking

Solving

rxode2 5.1.5 (2026-07-28)

New features

Bug fixes

Serialization

Error models / transformations

Estimation / symengine translation

Solving

Event tables

Model compilation

Installation / linking

rxode2 5.1.4 (2026-07-20)

Bug fixes

Model piping

Compilation

rxode2 5.1.3 (2026-07-19)

New features

Bug fixes

Estimation / symengine translation (rxFromSE())

Model parsing / mu-referencing

Delay models

linCmt() models

Solving

rxode2 5.1.2 (2026-06-02)

rxode2 5.1.1 (2026-05-28)

rxode2 5.0.2 (2026-03-20)

rxode2 5.0.1 (2025-12-09)

rxode2 5.0.0 (2025-11-28)

rxode2 4.1.1 (2025-10-08)

rxode2 4.1.0 (2025-08-29)

rxode2 4.0.3 (2025-07-24)

rxode2 4.0.2 (2025-07-21)

rxode2 4.0.1 (2025-07-17)

rxode2 4.0.0 (2025-07-16)

rxode2 3.0.4 (2025-02-14)

rxode2 3.0.3 (2024-12-15)

rxode2 3.0.2 (2024-10-30)

rxode2 3.0.1 (2024-09-22)

rxode2 3.0.0 (2024-09-18)

Breaking Changes

Possible breaking changes (though unlikely)

New features

Bug fixes

Big change

rxode2et (no changes before merge)

rxode2et 2.0.13

rxode2et 2.0.12

rxode2et 2.0.11

Estimation / symengine translation (`rxFromSE()`)

`linCmt()` models