R-devel change in complex variables and NA assignment

[eddelbuettel]

While trying to re-jig coverage (which I got working again for RcppArmadillo and RcppEigen) I noticed two failures on sugar tests for complex variables. It reproduces locally:

  ----- FAILED[data]: test_sugar.R<1272--1272>
   call| expect_equal(cx_row_means(x, TRUE), rowMeans(x, TRUE), info = "complex / rowMeans / rm NA / mixed input")
   diff| Mean relative Mod difference: 0.7597472                                                                                                                                                                   
   info| complex / rowMeans / rm NA / mixed input
  ----- FAILED[data]: test_sugar.R<1275--1275>
   call| expect_equal(cx_col_means(x, TRUE), colMeans(x, TRUE), info = "complex / colMeans / rm NA / mixed input")
   diff| Mean relative Mod difference: 0.1915024
   info| complex / colMeans / rm NA / mixed input
  Error: 2 out of 1622 tests failed
  Execution halted

This is from the part where we take a 3 x 3 complex matrix and inject NA values. This now differs:

> x <- matrix(rnorm(9) + 2i, 3)
> x
             [,1]         [,2]         [,3]
[1,] -1.862645+2i -2.059943+2i  0.694593+2i
[2,]  1.469971+2i -0.722717+2i -0.367111+2i
[3,] -0.946859+2i  0.468356+2i -0.486049+2i
> x[sample(1:9, 4)] <- NA
> x
             [,1]        [,2]         [,3]
[1,] -1.862645+2i -2.05994+2i           NA
[2,]           NA          NA -0.367111+2i
[3,] -0.946859+2i          NA -0.486049+2i
> rowMeans(x, TRUE)
[1] -1.961294+1.333333i -0.367111+0.666667i -0.716454+1.333333i

We get

> cppFunction("Rcpp::ComplexVector cx_row_means(Rcpp::ComplexMatrix x, bool na_rm = false) { return rowMeans(x, na_rm); }")
> cx_row_means(x, TRUE)
[1] -1.961294+2i -0.367111+2i -0.716454+2i
> 

So a disagreement over the complex part. R scales by 'n_rows' (NA or not), we scale by number of non-NA. One of the two is wrong.

I see

r88444 | maechler | 2025-07-22 08:43:24 -0500 (Tue, 22 Jul 2025) | 1 line

subassignment <complex>[i] <- NA should only touch the real part

which could be related. We may have to mirror that.

[eddelbuettel]

This could well be a bug in R-devel. I just wrote to the mailing list.

[eddelbuettel]

For now, we are switching to using apply(x, DIM, mean, na.rm=TRUE) instead of rowMeans() and colMeans() for these complex-value matrices with NAs as R-devel currently gets this wrong. For row means:

Rcpp/inst/tinytest/test_sugar.R

Lines 1272 to 1277 in 04847eb

	if (getRversion() < "4.6.0") {
	expect_equal(cx_row_means(x, TRUE), rowMeans(x, TRUE), info = "complex / rowMeans / rm NA / mixed input")
	} else {
	## TODO FIXME R-devel has borked rowMeans / colMeans for complex matrices with NA values
	expect_equal(cx_row_means(x, TRUE), apply(x, 1, mean, na.rm=TRUE), info = "complex / rowMeans / rm NA / mixed input")
	}

with similar change for column means:

Rcpp/inst/tinytest/test_sugar.R

Lines 1280 to 1287 in 79ba459

	if (getRversion() < "4.6.0") {
	expect_equal(cx_col_means(x, TRUE), colMeans(x, TRUE), info = "complex / colMeans / rm NA / mixed input")
	} else {
	## TODO FIXME R-devel has borked rowMeans / colMeans for complex matrices with NA values
	expect_equal(cx_col_means(x, TRUE), apply(x, 2, mean, na.rm=TRUE), info = "complex / rowMeans / rm NA / mixed input")
	}
	x[] <- NA_complex_