#### Examples where we know the result "exactly" library(expm) options(digits = 4, width = 99, keep.source = FALSE) mSource <- function(file, ...) source(system.file(file, ..., package = "expm", mustWork=TRUE)) mSource("test-tools.R")## -> assertError(), rMat(), .. doExtras mSource("demo", "exact-fn.R")# -> nilA3(), facMat(), m2ex3(), .... doExtras re.nilA3 <- function(xyz, EXPMlist) { stopifnot(is.list(EXPMlist)) r <- do.call(nilA3, as.list(xyz)) sapply(EXPMlist, function(Efn) relErr(r$expA, Efn(r$A))) } re.facMat <- function(n, EXPMlist, rFUN = rnorm, ...) { stopifnot(is.list(EXPMlist)) r <- facMat(n, rFUN, ...) vapply(EXPMlist, function(EXPM) { ct <- system.time(E <- EXPM(r$A), gcFirst = FALSE)[[1]] c(relErr = relErr(r$expA, E), c.time = ct) }, double(2)) } re.m2ex3 <- function(eps, EXPMlist) { stopifnot(is.list(EXPMlist)) r <- m2ex3(eps) sapply(EXPMlist, function(EXPM) relErr(r$expA, EXPM(r$A))) } ## check 1x1 matrices stopifnot( ## these had failed before 2017-03-28 (= Liselotte's 58-th birthday): all.equal(as.matrix(sqrtm(matrix(4))), matrix(2)) , all.equal(as.matrix(logm (matrix(pi))), matrix(log(pi))) , ## these had "always" worked : all.equal(as.matrix(expm (matrix(0))), matrix(1)) , all.equal(as.matrix(expm (matrix(1))), matrix(exp(1))) ) set.seed(321) re <- replicate(1000, c(re.nilA3(rlnorm(3),list(function(x)expm(x,"Pade"))), re.nilA3(rnorm(3), list(function(x)expm(x,"Pade"))))) summary(t(re)) stopifnot(rowMeans(re) < 1e-15, apply(re, 1, quantile, 0.80) < 1e-16, apply(re, 1, quantile, 0.90) < 2e-15, apply(re, 1, max) < c(4e-14, 6e-15)) showProc.time() ## Check *many* random nilpotent 3 x 3 matrices: set.seed(321) RE <- replicate(1000, c(re.nilA3(rlnorm(3), list(function(x) expm(x, "Ward77"))), re.nilA3(rnorm(3), list(function(x) expm(x, "Ward77"))))) stopifnot(rowMeans(RE) < 1e-15, apply(RE, 1, quantile, 0.80) < 1e-16, apply(RE, 1, quantile, 0.90) < 2e-15, apply(RE, 1, max) < c(4e-14, 6e-15)) print(summary(t(RE))) epsC <- .Machine$double.eps cat("relErr(expm(.,Pade)) - relErr(expm(.,Ward77)) in Machine_eps units:\n") print(summary(c(re - RE)) / epsC) ## Min. 1st Qu. Median Mean 3rd Qu. Max. ## -0.6183442 0.0000000 0.0000000 1.3650410 0.1399719 94.9809161 ## nb-mm3; ditto lynne (both x64), 2014-09-11: ## Min. 1st Qu. Median Mean 3rd Qu. Max. ## -0.8422 0.0000 0.0000 0.0725 0.1067 1.2205 ## 32-bit [i686, florence, Linx 3.14.8-100.fc19..]: ## Min. 1st Qu. Median Mean 3rd Qu. Max. ## -0.62 0.00 0.00 1.36 0.14 95.93 showProc.time() ###--- A second group --- where we know the diagonalization of A --- if(!require("Matrix")) q('no') ## ------ the rest really uses 'Matrix' ##---> now use expm::expm() since Matrix has its own may mask the expm one ## ^^^^^^^^^^^^ osV <- abbreviate(sub("\\(.*", "", osVersion), 10) if(!dev.interactive(TRUE)) pdf(paste0("expm_exact-ex_", osV, ".pdf"), width = 9, height=5) ## myRversion <- paste(R.version.string, "--", osVersion) if((mach <- Sys.info()[["machine"]]) != "x86_64") myRversion <- paste0(myRversion, "_", mach) if(!capabilities("long.double")) myRversion <- paste0(myRversion, "_no_LDbl") myRversion ## in plots use mtext(myRversion, adj=1, cex=3/4) ## rMat() relies on Matrix::rcond(): ## Now with the change default rcondMin, this "works" R40 <- rMat(40) R80 <- rMat(80) showProc.time() expm.safe.Eigen <- function(x, silent = FALSE) { r <- try(expm::expm(x, "R_Eigen"), silent = silent) if(inherits(r, "try-error")) NA else r } ## the S4 generic Matrix::expm ## the dgeMatrix method - adapted to Matrix changes, had *more versatile* ..2dge() : expm.Matr.dge <- function(x) getDataPart(getMethod("expm", "dgeMatrix"))(Matrix::.m2dense(x)) expmList <- list(Matr = Matrix::expm, Matr.d = expm.Matr.dge, Ward = function(x) expm::expm(x, "Ward77"), Ward1b = function(x) expm::expm(x, "Ward77", preconditioning = "1bal"), RWard6 = function(x) expm::expm(x, "R_Ward77", order = 6), RWard7 = function(x) expm::expm(x, "R_Ward77", order = 7), RWard8 = function(x) expm::expm(x, "R_Ward77", order = 8), # default RWard9 = function(x) expm::expm(x, "R_Ward77", order = 9), s.P.s7 = function(x) expm::expm(x, "Pade", order = 7), s.POs7 = function(x) expm::expm(x, "PadeO",order = 7), s.P.s8 = function(x) expm::expm(x, "Pade" ,order = 8), # default s.POs8 = function(x) expm::expm(x, "PadeO",order = 8), # default s.P.s9 = function(x) expm::expm(x, "Pade", order = 9), s.POs9 = function(x) expm::expm(x, "PadeO",order = 9), s.P.sRBS = function(x) expm::expm(x, "PadeRBS"), Rs.P.s7 = function(x) expm::expm(x, "R_Pade", order = 7), Rs.P.s8 = function(x) expm::expm(x, "R_Pade", order = 8), # default Rs.P.s9 = function(x) expm::expm(x, "R_Pade", order = 9), sPs.H08. = function(x) expm:: expm.Higham08(x, balancing=FALSE), sPs.H08b = function(x) expm:: expm.Higham08(x, balancing= TRUE), ## AmHi09.06= function(x) expm:::expm.AlMoHi09(x, p = 6), AmHi09.07= function(x) expm:::expm.AlMoHi09(x, p = 7), AmHi09.08= function(x) expm:::expm.AlMoHi09(x, p = 8), # default -- really sub optimal AmHi09.09= function(x) expm:::expm.AlMoHi09(x, p = 9), AmHi09.10= function(x) expm:::expm.AlMoHi09(x, p = 10), AmHi09.11= function(x) expm:::expm.AlMoHi09(x, p = 11), AmHi09.12= function(x) expm:::expm.AlMoHi09(x, p = 12), AmHi09.13= function(x) expm:::expm.AlMoHi09(x, p = 13), s.T.s7 = function(x) expm::expm(x, "Taylor", order = 7), s.TOs7 = function(x) expm::expm(x, "TaylorO",order = 7), s.T.s8 = function(x) expm::expm(x, "Taylor", order = 8), # default s.TOs8 = function(x) expm::expm(x, "TaylorO",order = 8), # default s.T.s9 = function(x) expm::expm(x, "Taylor", order = 9), s.TOs9 = function(x) expm::expm(x, "TaylorO",order = 9), Eigen = expm.safe.Eigen, # "R_Eigen" hybrid = function(x) expm::expm(x, "hybrid") ) ## set.seed(12) ## facMchk <- replicate(if(doExtras) 100 else 20, facMat(20, rnorm)) set.seed(12) fRE <- replicate(if(doExtras) 100 else 20, re.facMat(20, expmList)) # if(doExtras) gives one "No Matrix found ..." warning nDig <- -log10(t(fRE["relErr",,])) cat("Number of correct decimal digits for facMat(20, rnorm):\n") t(apply(nDig, 2, summary)) ## Now look at that: eaxis <- if(requireNamespace("sfsmisc")) sfsmisc::eaxis else axis op <- par(mar=.1 + c(5,4 + 1.5, 4,2)) boxplot(t(fRE["relErr",,]), log="x", xaxt="n", notch=TRUE, ylim = c(8e-16, 4e-9), horizontal=TRUE, las = 1, main = "relative errors for 'random' eigen-ok 20 x 20 matrix") eaxis(1); grid(lty = 3); mtext(myRversion, adj=1, cex=3/4) par(op) showProc.time() if(doExtras) withAutoprint({ # also "large" n = 100 ------------------------------------------ str(rf100 <- replicate(20, re.facMat(100, expmList))) 1000*t(apply(rf100["c.time",,], 1, summary)) ## lynne {Linux 2.6.34.7-56.fc13.x86_64 --- AMD Phenom II X4 925}: ## Min. 1st Qu. Median Mean 3rd Qu. Max. ## Ward 23 24 24.5 24.4 25.0 25 ## s.P.s 107 109 109.0 109.0 109.0 112 ## s.POs 188 190 191.0 192.0 193.0 198 ## s.P.sRBS 17 18 19.0 18.9 19.2 21 ## sPs.H08. 15 17 18.0 17.6 18.0 19 ## sPs.H08b 18 18 19.0 23.4 20.0 107 ## s.T.s 44 45 45.0 45.6 46.0 48 ## s.TOs 96 98 99.0 100.0 100.0 116 ## Eigen 18 19 20.0 24.4 21.0 109 ## hybrid 40 42 42.0 47.1 44.0 133 nDig <- -log10(t(rf100["relErr",,])) cat("Number of correct decimal digits for facMat(100, rnorm):\n") t(apply(nDig, 2, summary)) ##--> take out the real slow ones for the subsequent tests: (not.slow <- grep("^s\\.[PT]", names(expmList), invert = TRUE, value = TRUE)) set.seed(18) ## 12 replicates is too small .. but then it's too slow otherwise: rf400 <- replicate(12, re.facMat(400, expmList[not.slow])) showProc.time() 1000*t(apply(rf400["c.time",,], 1, summary)) ## lynne: ## Min. 1st Qu. Median Mean 3rd Qu. Max. ## Ward 1740 1790 1830 1820 1860 1900 ## s.P.sRBS 1350 1420 1440 1430 1450 1460 ## sPs.H08. 1020 1030 1130 1140 1210 1290 ## sPs.H08b 1120 1130 1220 1220 1300 1390 ## Eigen 962 977 989 992 1000 1030 ## hybrid 2740 2800 2840 2840 2890 2910 nDig <- -log10(t(rf400["relErr",,])) cat("Number of correct decimal digits for (12 rep. of) facMat(400, rnorm):\n") t(apply(nDig, 2, summary)) }) else { # *not* (doExtras) ----------------------------------------------------------------- ## times (in millisec): print(1000*t(apply(fRE["c.time",,], 1, summary))) } ## Now try an example with badly conditioned "random" M matrix... ## ... ## ... (not yet -- TODO?) ### m2ex3() --- The 2x2 example with bad condition , see A3 in ./ex2.R RE <- re.m2ex3(1e-8, expmList) sort(RE)# Ward + both sps.H08 are best; s.P.s fair, Eigen (and hybrid): ~1e-9 eps <- 10^-(1:18) t.m2 <- t(sapply(eps, re.m2ex3, EXPMlist = expmList)) ## --> 3 error messages from solve(V), 5 error messages from try(. "R_Eigen" ...) showProc.time() cbind(sort(apply(log(t.m2),2, median, na.rm=TRUE))) ## 'na.rm=TRUE' needed for Eigen which blows up for the last 3 eps t.m2.ranks <- sort(rowMeans(apply(t.m2, 1, rank))) cbind(signif(t.m2.ranks, 3)) ## lynne (x86_64, Linux 3.14.4-100; Intel i7-4765T), 2014-09: ## sPs.H08. 2.67 ## sPs.H08b 2.67 ## s.P.sRBS 3.06 ## Ward 4.03 ## AmHi09.13 4.33 <<- still not close to H08 ! ## AmHi09.12 5.86 ## s.T.s 8.33 ## s.TOs 8.33 ## s.P.s 9.11 ## s.POs 9.11 ## hybrid 10.80 ## AmHi09.10 11.70 << astonishingly bad ## Eigen 12.60 ## AmHi09.08 13.10 ## AmHi09.06 14.40 print(t.m2[, names(t.m2.ranks)[1:8]], digits = 3) ## ==> 1st class: H08 (both) and (but slightly better than) Ward ## 2nd class s.T.s and s.P.s ## "bad" : hybrid and Eigen ## ??? AmHi09 - methods, up to order = 10 are worse ! if(require(RColorBrewer)) { ## Bcol <- brewer.pal(ncol(t.m2),"Dark2") Bcol <- brewer.pal(min(9, ncol(t.m2)), "Set1") Bcol <- Bcol[sqrt(colSums(col2rgb(Bcol)^2)) < 340] ## FIXME: more colors ==> ~/R/MM/GRAPHICS/color-palettes.R } else { ## 7 from Dark2 ## Bcol <- c("#1B9E77", "#D95F02", "#7570B3", "#E7298A", ## "#66A61E", "#E6AB02", "#A6761D") ## Rather: those from "Set1" Bcol <- c("#E41A1C", "#377EB8", "#4DAF4A", "#984EA3", "#FF7F00", # too bright: "#FFFF33", "#A65628", "#F781BF", "#999999") } matplot(eps, t.m2, type = "b", log = "xy", col=Bcol, lty = 1:9, pch=1:15, axes=FALSE, frame = TRUE, xlab = expression(epsilon), ylab = "relative error", main = expression(expm(A, method == "*") *" relative errors for " * A == bgroup("[", atop({-1} *" "* 1, {epsilon^2} *" "*{-1}), "]"))) legend("bottomright",colnames(t.m2), col=Bcol, lty = 1:9, pch=1:15, inset = 0.02) if(requireNamespace("sfsmisc")) { sfsmisc::eaxis(1, labels=FALSE) sfsmisc::eaxis(1, at = eps[c(TRUE,FALSE)]) sfsmisc::eaxis(2) ## sfsmisc::eaxis(2, labels=FALSE) ## op <- par(las=2) ## sfsmisc::eaxis(2, at = axTicks(2,log=TRUE)[c(TRUE,FALSE,FALSE)]) ## par(op) } else { axis(1) axis(2) } ## typical case: ep <- 1e-10 (me <- m2ex3(ep)) me$expA * exp(1) ## the correct value ; numerically identical to simple matrix: ## identical() not fulfilled e.g. on Solaris stopifnot(all.equal(me$expA * exp(1), rbind(c( 1, 1), c(ep^2, 1)), tolerance = 1e-14)) ## The relative error (matrices): lapply(expmList, function(EXPM) 1 - EXPM(me$A)/me$expA) ## Average number of correct digits [less "extreme" than plot above] nDig <- sapply(expmList, function(EXPM) -log10(mean(abs(1 - EXPM(me$A)/me$expA)))) round(nDig, 2) ## Ward s.P.s s.POs s.T.s s.TOs Eigen hybrid ## 16.26 14.65 14.65 14.65 14.65 6.20 6.39 [AMD Opteron 64-bit] ## Inf 14.65 14.65 14.65 14.65 6.74 6.33 [Pentium-M (32-bit)] showProc.time() ###--- rnilMat() : random upper triangular (zero-diagonal) nilpotent n x n matrix set.seed(17) m <- rnilMat(10) (m. <- as(m,"sparseMatrix"))# for nicer printing - and more *below* E.m <- expm::expm(m, method="Pade") as(E.m, "sparseMatrix") (dN <- 9*7*320) # 20160 stopifnot(abs(round(E.m * dN) - (E.m * dN)) < 9e-6) EmN <- matrix(c(dN, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3*dN, dN, 0, 0, 0, 0, 0, 0, 0, 0, 352800, 5*dN, dN, 0, 0, 0, 0, 0, 0, 0, 1018080, 332640, 5*dN, dN, 0, 0, 0, 0, 0, 0, 2235240, 786240, 292320, 3*dN, dN, 0, 0, 0, 0, 0, 9368520, 3483480, 1582560, 413280, 181440, dN, 0, 0, 0, 0, 24676176, 9598680, 5073600, 1562400, 826560, 161280, dN, 0,0,0, 43730160, 17451000, 10051440, 3430560, 1955520, 504000, 5*dN, dN, 0, 0, 68438436, 27747480, 16853760, 6036240, 3638880, 1038240, 252000, 3*dN, dN, 0, 119725855, 49165892, 31046760, 11652480, 7198800, 2264640, 614880, 191520, 3*dN, dN), 10, 10) Em.xct <- EmN / dN all.equal(E.m, Em.xct, check.attributes = FALSE, tolerance = 0) stopifnot(all.equal(E.m, Em.xct, check.attributes = FALSE, tolerance= 1e-13)) re.x <- sapply(expmList, function(EXPM) relErr(Em.xct, EXPM(m))) ## with error message from "safe.Eigen" --> Eigen is NA here ## result depends quite a bit on platform which(is.na(re.x)) # gives c(Eigen = 16L) (but not everywhere ?!?) (re.x <- re.x[!is.na(re.x)]) ## Pentium-M 32-bit ubuntu gave ## Ward s.P.s s.POs sPs.H08. sPs.H08b s.T.s s.TOs hybrid ## 1.079e-16 4.505e-14 4.503e-14 9.379e-17 9.379e-17 3.716e-17 7.079e-18 1.079e-16 ## 32-bit Quad-Core AMD Opteron 2380 (Linux 2.6.30.10-105.2.23.fc11.i686.PAE): ## Ward s.P.s s.POs sPs.H08. sPs.H08b s.T.s s.TOs hybrid ## 1.079e-16 4.505e-14 4.503e-14 9.379e-17 9.379e-17 3.716e-17 7.079e-18 1.079e-16 ## "Ward77": again more accurate than s+Pade+s, but s+Taylor+s is even more accurate ## but on 64-bit AMD Opterons ## Ward s.P.s s.POs sPs.H08. sPs.H08b s.T.s s.TOs hybrid ## 4.42e-17 3.99e-17 3.99e-17 1.10e-16 1.10e-16 8.44e-17 8.44e-17 4.42e-17 ## ## even more astonishing the result on Mac OSX (x86_32_mac; R-forge, R 2.9.0 patch.) ## Ward s.P.s s.POs sPs.H08. sPs.H08b s.T.s s.TOs hybrid ## 5.13e-17 3.99e-17 3.99e-17 1.84e-15 1.84e-15 8.44e-17 8.44e-17 5.13e-17 ## 2014-09: AmHi09 are very good (64bit: 8e-17) for p >= 8 (p=6 has 1.5e-11) not.09.06 <- which(names(re.x) != "AmHi09.06") stopifnot(re.x[c("Ward", "s.T.s8", "s.TOs8")] < 3e-16, ## re.x[["AmHi09.06"]] < 9e-11, # x64 & 686(lnx): = 1.509e-11 re.x[not.09.06] < 4e-13)# max: 686(32b): 4.52e-14, x64(lnx): 1.103e-16 ##-- Looking at *sparse* matrices: [C,Fortran "dense" code based methods will fail]: (meths <- eval(formals(expm)$method)) ems <- sapply(meths, function(met) tryCatch(expm::expm(m., method=met), error=identity)) ok <- !sapply(ems, is, class2="error") meths[ok] # now most... are showProc.time() ## Complex Matrices re.facMat.Z <- function(n, EXPMlist, rFUN = function(n) rnorm(n) + 1i*rnorm(n), ...) { stopifnot(is.list(EXPMlist)) r <- facMat(n, rFUN, ...) vapply(EXPMlist, function(EXPM) { ct <- system.time(E <- EXPM(r$A), gcFirst = FALSE)[[1]] c(relErr = relErr(r$expA, E), c.time = ct) }, double(2)) } (nmL <- names(expmList)) ## [1] "Matr" "Matr.d" "Ward" "Ward1b" "s.P.s" "s.POs" "s.P.s7" ## [8] "s.POs7" "s.P.s9" "s.POs9" "s.P.sRBS" "sPs.H08." "sPs.H08b" "AmHi09.06" ## [15] "AmHi09.07" "AmHi09.08" "AmHi09.09" "AmHi09.10" "AmHi09.12" "AmHi09.13" "s.T.s" ## [22] "s.TOs" "s.T.s7" "s.TOs7" "s.T.s9" "s.TOs9" "Eigen" "hybrid" ## dropping "Matr", "Matr.d" (which gives "dgeMatrix" currently --> mean(.) fails ... ## "Ward" "Ward1b" and "hybrid" error "not a numeric Matrix" ## "AmHi09": C code currently only for double precision ((FIXME)) (cplxN <- grep("^(Matr|Ward|hybr|AmHi09|s\\.[PT])", nmL, invert = TRUE, value = TRUE)) rr <- re.facMat.Z(4, expmList[cplxN]) set.seed(47) fREZ <- replicate(if(doExtras) 64 else 12, re.facMat.Z(15, expmList[cplxN])) nDig <- -log10(t(fREZ["relErr",,])) cat("Number of correct decimal digits for facMat(20, rnorm + i*rnorm):\n") t(apply(nDig, 2, summary)) ## times (in millisec): print(1000*t(apply(fREZ["c.time",,], 1, summary))) ## Now look at that: op <- par(mar=.1 + c(5,4 + 1.5, 4,2)) boxplot(t(fREZ["relErr",,]), log="x", xaxt="n", notch=TRUE, # ylim = c(8e-16, 4e-9), horizontal=TRUE, las = 1, main = "relative errors for 'random' eigen-ok 20 x 20 matrix") eaxis(1); grid(lty = 3); mtext(myRversion, adj=1, cex=3/4) par(op) showProc.time()