Rcpp: Syntactic sugar for * produces unexpected results when dealing with NumericMatrix

A recently asked question has lead me to believe the syntactic sugar for * by Rcpp does not work as intended. In the linked question, the user is trying to multiply a matrix by a scalar.

R code

Here's what we're trying to achieve in Rcpp, but for now in plain R:

> m <- matrix(0:3, 2, 2)
> m * 3
     [,1] [,2]
[1,]    0    6
[2,]    3    9

Rcpp Code

I've created some minimal examples demonstrating both the problem above, and also some unexpected behaviour along the way. First note that I'm consistently using List as a return type, because it removes the need for me to declare the appropriate type in advance:

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]

List FooMat() {
  // Create a fill a 2x2 matrix
  NumericMatrix tmp(2,2);
  for (int i = 0; i < 4; i++) {
    tmp[i] = i;
  }

  return List::create(tmp);
}

// [[Rcpp::export]]
List FooMat2() {
  // Create a fill a 2x2 matrix
  NumericMatrix tmp(2,2);
  for (int i = 0; i < 4; i++) {
    tmp[i] = i;
  }

  NumericVector x(1);                                                                                                                                                                                                                      
  x[1] = 3;

  return List::create(tmp * x); 
}

// [[Rcpp::export]]

List FooMat3() {
  // Create a fill a 2x2 matrix
  NumericMatrix tmp(2,2);
  for (int i = 0; i < 4; i++) {
    tmp[i] = i;
  }

  NumericVector x(1);
  x[1] = 3;

  return List::create(tmp * x[1]);
}

// [[Rcpp::export]]

List FooMat4() {
  // Create a fill a 2x2 matrix
  NumericMatrix tmp(2,2);
  for (int i = 0; i < 4; i++) {
    tmp[i] = i;
  }

  return List::create(tmp * 3); 
}

Now if we source the file we get some odd behaviour:

# Proof that we can return a NumericMatrix in a List:
> FooMat()
[[1]]
     [,1] [,2]
[1,]    0    2
[2,]    1    3

# Multiply the whole NumericMatrix by a whole NumericVector
# whose size is 1. Unsafe behaviour?
> FooMat2()
[[1]]
[1]  0.000000e+00  3.000000e+00 1.388988e-309 2.083483e-309

# Multiply the whole NumericMatrix by the first element of
# The NumericVector. Results are correct, but `*` converts
# the answer to a NumericVector instead of a NumericMatrix
> FooMat3()
[[1]]
[1] 0 3 6 9

# Same as FooMat3() except now we just multiply the NumericMatrix
# by an integer
> FooMat4()
[[1]]
[1] 0 3 6 9

One, the syntactic sugar for * provided by Rcpp does not seem to correctly handle multiplication of Matrices with scalars. Two, multiplying by a whole NumericVector, as in FooMat2() leads to unsafe behaviour.

As I stated in previous answers, when I need to do actual math on matrices, I use Armadillo objects:

R> cppFunction('arma::mat scott(arma::mat x, double z) { 
+                 return(x*z); }', 
+              depends="RcppArmadillo")
R> scott(matrix(1:4,2), 2)
     [,1] [,2]
[1,]    2    6
[2,]    4    8
R>

Sugar operations are nice, but not complete. We will certainly take patches, though.

And as we said a few times before: rcpp-devel is the proper support channel.

Edit (Oct 2016 or 2 1/2 years later): Searching for something else just got me back here. In the Rcpp 0.12.* series, some work when into operations between matrix and vector so the basic 'matrix times scalar' now works as you'd expect:

R> cppFunction("NumericMatrix testmat(NumericMatrix m, double multme) { 
+               NumericMatrix n = m * multme; 
+               return n; }") 
R> testmat(matrix(1:4,2), 1)
     [,1] [,2]
[1,]    1    3
[2,]    2    4
R> testmat(matrix(1:4,2), 3)
     [,1] [,2]
[1,]    3    9
[2,]    6   12
R>

I'd probably still use RcppArmadillo for math on matrices though.

This is an unfortunate consequence of a bad design decision, namely making Rcpp matrices derive from Rcpp vectors.

I'm likely to revert this decision in Rcpp implementations I now maintain: Rcpp11 and Rcpp98. I don't think anymore that there any benefit of having Matrix derive from Vector and it gets in the way of CRTP that is used at the end of this file.

来源：https://stackoverflow.com/questions/20950880/rcpp-syntactic-sugar-for-produces-unexpected-results-when-dealing-with-numeri

标签

rcpp