incorrect use of NaN

[conradsnicta]

@simonsays1980 This bit of code in optimize.h is not likely to work correctly:

finmix/src/optimize.h

Lines 67 to 70 in d853be4

	else if (rvalues.has_nan())
	{
	rvalues.elem(arma::find(rvalues == arma::datum::nan)).zeros();
	}

Extract from cppreference.com: "NaN values never compare equal to themselves or to other NaN values".
https://en.cppreference.com/w/cpp/numeric/math/isnan

Suggest to use arma::replace() instead, in order to replace NaN values with zeros.

CC: @eddelbuettel

[eddelbuettel]

Hi guys, coming in here without too much context but the main take away is that R use the standard IEEE approach for NaN on double but then .... goes beyond by a) also defining NA and b) doing so for types besides double. Which is extremely valuable for the statistical context R is in -- but also not portable. So while it may work for NaN on double with Armadillo, it is no passthrough.

We do have documentation (a little scattered, I admit) on how to do withing the R context via Rcpp and hence RcppArmadillo.

[eddelbuettel]

To the point @conradsnicta was making, a more C++-standards compliant way is std::isfinite -- see https://en.cppreference.com/w/cpp/numeric/math/isfinite. But yes the == comparison will not work, just how it does not in R either:

> x <- c(3, NA, NaN, 5)
> x == NA
[1] NA NA NA NA
> sapply(x, is.finite)
[1]  TRUE FALSE FALSE  TRUE
> 

[simonsays1980]

@conradsnicta, thank you very much for inspecting my code. I was surprised that someone even looks at it :) I will try to fix that hopefully at the weekend. Also thanks for the helpful links. And of course thumbs up for this awesome library that has helped me with many problems :) Armadillo with Rcpp runs like a charm. Best, Simon Am Mo., 10. Jan. 2022 um 04:10 Uhr schrieb Conrad Sanderson < ***@***.***>:

…

@simonsays1980 <https://github.com/simonsays1980> This bit of code in optimize.h is not likely to work correctly: https://github.com/simonsays1980/finmix/blob/d853be4b4b208a391b32a160d6b90afecb5d62b0/src/optimize.h#L67-L70 Extract from cppreference.com: "NaN values never compare equal to themselves or to other NaN values". https://en.cppreference.com/w/cpp/numeric/math/isnan Suggest to use arma::replace() <http://arma.sourceforge.net/docs.html#replace> instead, in order to replace NaN values with zeros. CC: @eddelbuettel <https://github.com/eddelbuettel> — Reply to this email directly, view it on GitHub <#4>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANWKR5IQLSTLRPP4QU3NU3UVJEZ7ANCNFSM5LSUT4MA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[simonsays1980]

Dirk, thank you for clarifying Conrad's point. Given your answer and the one of Conrad, would you yourself rather go with std::infinite or with arma::replace(datum::nan, 0)? Best, Simon Am Mo., 10. Jan. 2022 um 04:41 Uhr schrieb Dirk Eddelbuettel < ***@***.***>:

…

To the point @conradsnicta <https://github.com/conradsnicta> was making, a more C++-standards compliant way is std::isfinite -- see https://en.cppreference.com/w/cpp/numeric/math/isfinite. But yes the == comparison will not work, just how it does not in R either: > x <- c(3, NA, NaN, 5)> x == NA [1] NA NA NA NA> sapply(x, is.finite) [1] TRUE FALSE FALSE TRUE> — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANWKRZ5OVMNQQQRAMRZB6DUVJIQDANCNFSM5LSUT4MA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[eddelbuettel]

Hi Simon,

This hit me out of context per Conrad's tag but I think he is right on that you cannot do == comparison for non-finite value but need to use a function. And there is merit to both sides: I have code where I try keep files 'portable' without R headers and if the variable is floating point then maybe one of the STL functions can sweep a std::isfinite(rvalues) across an Armadillo vector (if you need positions). I always have to look up std::find_if and alike. Else arma::replace as Conrad suggested may do what you need.

What is nice about the Rcpp(Armadillo) integration is that you can test that oneliners immediately. And Rcpp helps with a few of the vectorised functions.

> Rcpp::cppFunction("LogicalVector foo(NumericVector x) { return is_finite(x); }")
> foo(c(2, 3, NA, 5, NaN, 7, Inf, 9))
[1]  TRUE  TRUE FALSE  TRUE FALSE  TRUE FALSE  TRUE
>