Fast and Elegant Numerical Linear Algebra Using the RcppEigen Package
Abstract
The RcppEigen package provides access from R (R Core Team, 2026a) to the Eigen (Guennebaud, Jacob and others, 2012) C++ template library for numerical linear algebra. Rcpp (Eddelbuettel et al., 2026a; Eddelbuettel, 2013) classes and specializations of the C++ templated functions as and wrap from Rcpp provide the “glue” for passing objects from R to C++ and back. Several introductory examples are presented. This is followed by an in-depth discussion of various available approaches for solving least-squares problems, including rank-revealing methods, concluding with an empirical run-time comparison. Last but not least, sparse matrix methods are discussed.
Keywords: linear algebra, template programming, R, C++, Rcpp
This vignette corresponds to a paper published in the Journal of Statistical Software. Currently still identical to the paper, this vignette version may over time receive minor updates. For citations, please use Bates and Eddelbuettel (2013) as provided by citation("RcppEigen").
This version corresponds to RcppEigen version 0.3.4.0.2 and was typeset on November 21, 2026.
1 Introduction
Linear algebra is an essential building block of statistical computing. Operations such as matrix decompositions, linear program solvers, and eigenvalue/eigenvector computations are used in many estimation and analysis routines. As such, libraries supporting linear algebra have long been provided by statistical programmers for different programming languages and environments. Because it is object-oriented, C++, one of the central modern languages for numerical and statistical computing, is particularly effective at representing matrices, vectors and decompositions, and numerous class libraries providing linear algebra routines have been written over the years.
As both the C++ language and standards have evolved (Meyers, 2005; Meyers, 1995; ISO/IEC, 2011), so have the compilers implementing the language. Relatively modern language constructs such as template meta-programming are particularly useful because they provide overloading of operations (allowing expressive code in the compiled language similar to what can be done in scripting languages) and can shift some of the computational burden from the run-time to the compile-time. (A more detailed discussion of template meta-programming is, however, beyond the scope of this paper). Veldhuizen (1998) provided an early and influential implementation of numerical linear algebra classes that already demonstrated key features of this approach. Its usage was held back at the time by the limited availability of compilers implementing all the necessary features of the C++ language.
This situation has greatly improved over the last decade, and many more libraries have been made available. One such C++ library is Eigen by Guennebaud, Jacob and others (2012) which started as a sub-project to KDE (a popular Linux desktop environment), initially focussing on fixed-size matrices to represent projections in a visualization application. Eigen grew from there and has over the course of about a decade produced three major releases with Eigen3 being the current major version. To check the minor and patch version numbers, load the RcppEigen package and call this (internal) helper function:
R> RcppEigen:::eigen_version(FALSE)
major minor patch 3 3 5Eigen is of interest as the R system for statistical computation and graphics (R Core Team, 2026a) is itself easily extensible. This is particular true via the C language that most of R’s compiled core parts are written in, but also for the C++ language which can interface with C-based systems rather easily. The manual “Writing R Extensions” (R Core Team, 2026b) is the basic reference for extending R with either C or C++.
The Rcpp package by Eddelbuettel and François (2011); Eddelbuettel et al. (2026a) facilitates extending R with C++ code by providing seamless object mapping between both languages. As stated in the Rcpp (Eddelbuettel et al., 2026a) vignette, “Extending Rcpp” as well as in Eddelbuettel (2013)
Rcpp facilitates data interchange between R and C++ through the templated functions Rcpp::as (for conversion of objects from R to C++) and Rcpp::wrap (for conversion from C++ to R).
The RcppEigen package provides the header files composing the Eigen C++ template library, along with implementations of Rcpp::as and Rcpp::wrap for the C++ classes defined in Eigen.
The Eigen classes themselves provide high-performance, versatile and comprehensive representations of dense and sparse matrices and vectors, as well as decompositions and other functions to be applied to these objects. The next section introduces some of these classes and shows how to interface to them from R.
2 Eigen classes
Eigen is a C++ template library providing classes for many forms of matrices, vectors, arrays and decompositions. These classes are flexible and comprehensive allowing for both high performance and well structured code representing high-level operations. C++ code based on Eigen is often more like R code, working on the “whole object”, than like compiled code in other languages where operations often must be coded in loops.
As in many C++ template libraries using template meta-programming (Abrahams and Gurtovoy, 2004), the templates themselves can be very complicated. However, Eigen provides typedefs for common classes that correspond to R matrices and vectors, as shown in Table 1, and this paper will use these typedefs.
| R object type | Eigen class typedef |
|---|---|
| numeric matrix | MatrixXd |
| integer matrix | MatrixXi |
| complex matrix | MatrixXcd |
| numeric vector | VectorXd |
| integer vector | VectorXi |
| complex vector | VectorXcd |
| Matrix::dgCMatrix | SparseMatrix<double> |
Here, Vector and Matrix describe the dimension of the object. The X signals that these are dynamically-sized objects (as opposed to fixed-size matrices such as matrices also available in Eigen). Lastly, the trailing characters i, d and cd denote, respectively, storage types integer, double and complex double.
The C++ classes shown in Table 1 are in the Eigen namespace, which means that they must be written as Eigen::MatrixXd. However, if one prefaces the use of these class names with a declaration like
using Eigen::MatrixXd;
then one can use these names without the namespace qualifier.
2.1 Mapped matrices in Eigen
Storage for the contents of matrices from the classes shown in Table 1 is allocated and controlled by the class constructors and destructors. Creating an instance of such a class from an R object involves copying its contents. An alternative is to have the contents of the R matrix or vector mapped to the contents of the object from the Eigen class. For dense matrices one can use the Eigen templated class Map, and for sparse matrices one can deploy the Eigen templated class MappedSparseMatrix.
One must, of course, be careful not to modify the contents of the R object in the C++ code. A recommended practice is always to declare mapped objects as const.
2.2 Arrays in Eigen
For matrix and vector classes Eigen overloads the `*' operator as matrix multiplication. Occasionally component-wise operations instead of matrix operations are desired, for which the Array templated classes are used in Eigen. Switching back and forth between matrices or vectors and arrays is usually done via the array() method for matrix or vector objects and the matrix() method for arrays.
2.3 Structured matrices in Eigen
Eigen provides classes for matrices with special structure such as symmetric matrices, triangular matrices and banded matrices. For dense matrices, these special structures are described as “views”, meaning that the full dense matrix is stored but only part of the matrix is used in operations. For a symmetric matrix one must specify whether the lower triangle or the upper triangle is to be used as the contents, with the other triangle defined by the implicit symmetry.
3 Some simple examples
C++ functions to perform simple operations on matrices or vectors can follow a pattern of:
- 1.
Map the R objects passed as arguments into Eigen objects.
- 2.
Create the result.
- 3.
Return Rcpp::wrap applied to the result.
An idiom for the first step is
using Eigen::Map;
using Eigen::MatrixXd;
using Rcpp::as;
const MapMatrixXd A(asMapMatrixXd (AA));
where AA is the name of the R object (of type SEXP in C and C++) passed to the C++ function.
An alternative to the using declarations is to declare a typedef as in
typedef Eigen::MapEigen::MatrixXd MapMatd;
const MapMatd A(Rcpp::asMapMatd(AA));
The cxxfunction function from the inline package (Sklyar, Murdoch, Smith, Eddelbuettel, François, Soetaert and Ranke, 2025) for R and its RcppEigen plugin provide a convenient method of developing and debugging the C++ code. For actual production code one generally incorporates the C++ source code files in a package and includes the line LinkingTo: Rcpp, RcppEigen in the package’s DESCRIPTION file. The RcppEigen.package.skeleton function provides a quick way of generating the skeleton of a package that will use RcppEigen.
The cxxfunction with the "Rcpp" or "RcppEigen" plugins has the as and wrap functions already defined as Rcpp::as and Rcpp::wrap. In the examples below these declarations are omitted. It is important to remember that they are needed in actual C++ source code for a package.
The first few examples are simply for illustration as the operations shown could be more effectively performed directly in R. Finally, the results from Eigen are compared to those from the direct R results.
3.1 Transpose of an integer matrix
The next R code snippet creates a simple matrix of integers
R> (A <- matrix(1:6, ncol = 2))
[,1] [,2][1,] 1 4[2,] 2 5[3,] 3 6
R> str(A)
int [1:3, 1:2] 1 2 3 4 5 6and, in Figure 1, the transpose() method for the Eigen::MatrixXi class is used to return the transpose of the supplied matrix. The R matrix in the SEXP AA is first mapped to an Eigen::MatrixXi object, and then the matrix At is constructed from its transpose and returned to R.
using Eigen::Map;
using Eigen::MatrixXi;
// Map the integer matrix AA from R
const MapMatrixXi A(asMapMatrixXi (AA));
// evaluate and return the transpose of A
const MatrixXi At(A.transpose());
return wrap(At);
The R snippet below compiles and links the C++ code segment. The actual work is done by the function cxxfunction from the inline package which compiles, links and loads code written in C++ and supplied as a character variable. This frees the user from having to know about compiler and linker details and options, which makes “exploratory programming” much easier. Here the program piece to be compiled is stored as a character variable named transCpp, and cxxfunction creates an executable function which is assigned to ftrans. This new function is used on the matrix shown above, and one can check that it works as intended by comparing the output to an explicit transpose of the matrix argument.
R> ftrans <- cxxfunction(signature(AA = "matrix"), transCpp,+ plugin = "RcppEigen")R> (At <- ftrans(A))
[,1] [,2] [,3][1,] 1 2 3[2,] 4 5 6
R> stopifnot(all.equal(At, t(A)))For numeric or integer matrices the adjoint() method is equivalent to the transpose() method. For complex matrices, the adjoint is the conjugate of the transpose. In keeping with the conventions in the Eigen documentation the adjoint() method is used in what follows to create the transpose of numeric or integer matrices.
3.2 Products and cross-products
As mentioned in Section 2.2, the `*' operator is overloaded as matrix multiplication for the various matrix and vector classes in Eigen. The C++ code in Figure 2 produces a list containing both the product and cross-product (in the sense of the R function call crossproduct(A, B) evaluating ) of its two arguments
typedef Eigen::MapEigen::MatrixXi MapMati;
const MapMati B(asMapMati(BB));
const MapMati C(asMapMati(CC));
return List::create(Named("B %*% C") = B * C,
Named("crossprod(B, C)") = B.adjoint() * C);
R> fprod <- cxxfunction(signature(BB = "matrix", CC = "matrix"),+ prodCpp, "RcppEigen")R> B <- matrix(1:4, ncol = 2)R> C <- matrix(6:1, nrow = 2)R> str(fp <- fprod(B, C))
List of 2 $ B %*% C : int [1:2, 1:3] 21 32 13 20 5 8 $ crossprod(B, C): int [1:2, 1:3] 16 38 10 24 4 10
R> stopifnot(all.equal(fp[[1]], B %*% C), all.equal(fp[[2]], crossprod(B, C)))Note that the create class method for Rcpp::List implicitly applies Rcpp::wrap to its arguments.
3.3 Crossproduct of a single matrix
As shown in the last example, the R function crossprod calculates the product of the transpose of its first argument with its second argument. The single argument form, crossprod(X), evaluates . One could, of course, calculate this product as
t(X) %*% X
but crossprod(X) is roughly twice as fast because the result is known to be symmetric and only one triangle needs to be calculated. The function tcrossprod evaluates crossprod(t(X)) without actually forming the transpose.
To express these calculations in Eigen, a SelfAdjointView—which is a dense matrix of which only one triangle is used, the other triangle being inferred from the symmetry—is created. (The characterization “self-adjoint” is equivalent to symmetric for non-complex matrices.)
The Eigen class name is SelfAdjointView. The method for general matrices that produces such a view is called selfadjointView. Both require specification of either the Lower or Upper triangle.
For triangular matrices the class is TriangularView and the method is triangularView. The triangle can be specified as Lower, UnitLower, StrictlyLower, Upper, UnitUpper or StrictlyUpper.
For self-adjoint views the rankUpdate method adds a scalar multiple of to the current symmetric matrix. The scalar multiple defaults to 1. The code in Figure 3 produces both and from a matrix .
using Eigen::Map;
using Eigen::MatrixXi;
using Eigen::Lower;
const MapMatrixXi A(asMapMatrixXi (AA));
const int m(A.rows()), n(A.cols());
MatrixXi AtA(MatrixXi(n, n).setZero().
selfadjointViewLower().rankUpdate(A.adjoint()));
MatrixXi AAt(MatrixXi(m, m).setZero().
selfadjointViewLower().rankUpdate(A));
return List::create(Named("crossprod(A)") = AtA,
Named("tcrossprod(A)") = AAt);
R> fcprd <- cxxfunction(signature(AA = "matrix"), crossprodCpp, "RcppEigen")R> str(crp <- fcprd(A))
List of 2 $ crossprod(A) : int [1:2, 1:2] 14 32 32 77 $ tcrossprod(A): int [1:3, 1:3] 17 22 27 22 29 36 27 36 45
R> stopifnot(all.equal(crp[[1]], crossprod(A)),+ all.equal(crp[[2]], tcrossprod(A)))To some, the expressions in Figure 3 to construct AtA and AAt are compact and elegant. To others they are hopelessly confusing. If you find yourself in the latter group, you just need to read the expression left to right. So, for example, we construct AAt by creating a general integer matrix of size (where is ), ensuring that all its elements are zero, regarding it as a self-adjoint (i.e., symmetric) matrix using the elements in the lower triangle, then adding to it and converting back to a general matrix form (i.e., the strict lower triangle is copied into the strict upper triangle).
In more detail:
- 1.
MatrixXi(n, n) creates an integer matrix with arbitrary contents
- 2.
.setZero() zeros out the contents of the matrix
- 3.
.selfAdjointView<Lower>() causes what follows to treat the matrix as a symmetric matrix in which only the lower triangle is used, the strict upper triangle being inferred by symmetry
- 4.
.rankUpdate(A) forms the sum where is the symmetric matrix of zeros created in the previous steps.
The assignment of this symmetric matrix to the (general) MatrixXi object AAt causes the result to be symmetrized during the assignment.
For these products one could define the symmetric matrix from either the lower triangle or the upper triangle as the result will be symmetrized before it is returned.
To cut down on repetition of using statements we gather them in a character variable, incl, that will be given as the includes argument in the calls to cxxfunction. We also define a utility function, AtA, that returns the crossproduct matrix as shown in Figure 4
using Eigen::LLT;
using Eigen::Lower;
using Eigen::Map;
using Eigen::MatrixXd;
using Eigen::MatrixXi;
using Eigen::Upper;
using Eigen::VectorXd;
typedef MapMatrixXd MapMatd;
typedef MapMatrixXi MapMati;
typedef MapVectorXd MapVecd;
inline MatrixXd AtA(const MapMatd& A) {
int n(A.cols());
return MatrixXd(n,n).setZero().selfadjointViewLower()
.rankUpdate(A.adjoint());
}
3.4 Cholesky decomposition of the crossprod
The Cholesky decomposition of the positive-definite, symmetric matrix, , can be written in several forms. Numerical analysts define the “LLt” form as the lower triangular matrix, , such that and the “LDLt” form as a unit lower triangular matrix and a diagonal matrix with positive diagonal elements such that . Statisticians often write the decomposition as where is an upper triangular matrix. Of course, this is simply the transpose of from the “LLt” form.
The templated Eigen classes for the LLt and LDLt forms are called LLT and LDLT and the corresponding methods are .llt() and .ldlt(). Because the Cholesky decomposition involves taking square roots, we pass a numeric matrix, , not an integer matrix.
const LLTMatrixXd llt(AtA(asMapMatd(AA)));
return List::create(Named("L") = MatrixXd(llt.matrixL()),
Named("R") = MatrixXd(llt.matrixU()));
R> storage.mode(A) <- "double"R> fchol <- cxxfunction(signature(AA = "matrix"), cholCpp, "RcppEigen", incl)R> (ll <- fchol(A))
$L [,1] [,2][1,] 3.74166 0.00000[2,] 8.55236 1.96396$R [,1] [,2][1,] 3.74166 8.55236[2,] 0.00000 1.96396
R> stopifnot(all.equal(ll[[2]], chol(crossprod(A))))
3.5 Determinant of the cross-product matrix
The “D-optimal” criterion for experimental design chooses the design that maximizes the determinant, , for the model matrix (or Jacobian matrix), . The determinant, , of the lower Cholesky factor , defined so that , is the product of its diagonal elements, as is the case for any triangular matrix. By the properties of determinants,
Alternatively, if using the “LDLt” decomposition, where is unit lower triangular and is diagonal then is the product of the diagonal elements of . Because it is known that the diagonal elements of must be non-negative, one often evaluates the logarithm of the determinant as the sum of the logarithms of the diagonal elements of . Several options are shown in Figure 6.
const MatrixXd ata(AtA(asMapMatd(AA)));
const double detL(MatrixXd(ata.llt().matrixL()).diagonal().prod());
const VectorXd Dvec(ata.ldlt().vectorD());
return List::create(Named("d1") = detL * detL,
Named("d2") = Dvec.prod(),
Named("ld") = Dvec.array().log().sum());
R> fdet <- cxxfunction(signature(AA = "matrix"), cholDetCpp, "RcppEigen",+ incl)R> unlist(ll <- fdet(A))
d1 d2 ld54.00000 54.00000 3.98898Note the use of the array() method in the calculation of the log-determinant. Because the log() method applies to arrays, not to vectors or matrices, an array must be created from Dvec before applying the log() method.
4 Least squares solutions
A common operation in statistical computing is calculating a least squares solution, , defined as
where the model matrix, , is () and is an -dimensional response vector. There are several ways, based on matrix decompositions, to determine such a solution. Earlier, two forms of the Cholesky decomposition were discussed: “LLt” and “LDLt”, which can both be used to solve for . Other decompositions that can be used are the QR decomposition, with or without column pivoting, the singular value decomposition and the eigendecomposition of a symmetric matrix.
Determining a least squares solution is relatively straightforward. However, statistical computing often requires additional information, such as the standard errors of the coefficient estimates. Calculating these involves evaluating the diagonal elements of and the residual sum of squares, .
4.1 Least squares using the “LLt” Cholesky
const MapMatd X(asMapMatd(XX));
const MapVecd y(asMapVecd(yy));
const int n(X.rows()), p(X.cols());
const LLTMatrixXd llt(AtA(X));
const VectorXd betahat(llt.solve(X.adjoint() * y));
const VectorXd fitted(X * betahat);
const VectorXd resid(y - fitted);
const int df(n - p);
const double s(resid.norm() / std::sqrt(double(df)));
const VectorXd se(s * llt.matrixL().solve(MatrixXd::Identity(p, p))
.colwise().norm());
return List::create(Named("coefficients") = betahat,
Named("fitted.values") = fitted,
Named("residuals") = resid,
Named("s") = s,
Named("df.residual") = df,
Named("rank") = p,
Named("Std. Error") = se);
Figure 7 shows a calculation of the least squares coefficient estimates (betahat) and the standard errors (se) through an “LLt” Cholesky decomposition of the crossproduct of the model matrix, . Next, the results from this calculation are compared to those from the lm.fit function in R (lm.fit is the workhorse function called by lm once the model matrix and response have been evaluated).
R> lltLS <- cxxfunction(signature(XX = "matrix", yy = "numeric"), lltLSCpp,+ "RcppEigen", incl)R> data("trees", package = "datasets")R> str(lltFit <- with(trees, lltLS(cbind(1, log(Girth)), log(Volume))))
List of 7 $ coefficients : num [1:2] -2.35 2.2 $ fitted.values: num [1:31] 2.3 2.38 2.43 2.82 2.86 ... $ residuals : num [1:31] 0.0298 -0.0483 -0.1087 -0.0223 0.0727 ... $ s : num 0.115 $ df.residual : int 29 $ rank : int 2 $ Std. Error : num [1:2] 0.2307 0.0898R> str(lmFit <- with(trees, lm.fit(cbind(1, log(Girth)), log(Volume))))List of 8 $ coefficients : Named num [1:2] -2.35 2.2 ..- attr(*, "names")= chr [1:2] "x1" "x2" $ residuals : num [1:31] 0.0298 -0.0483 -0.1087 -0.0223 0.0727 ... $ effects : Named num [1:31] -18.2218 2.8152 -0.1029 -0.0223 0.0721 ... ..- attr(*, "names")= chr [1:31] "x1" "x2" "" "" ... $ rank : int 2 $ fitted.values: num [1:31] 2.3 2.38 2.43 2.82 2.86 ... $ assign : NULL $ qr :List of 5 ..$ qr : num [1:31, 1:2] -5.57 0.18 0.18 0.18 0.18 ... ..$ qraux: num [1:2] 1.18 1.26 ..$ pivot: int [1:2] 1 2 ..$ tol : num 1e-07 ..$ rank : int 2 ..- attr(*, "class")= chr "qr" $ df.residual : int 29
R> for(nm in c("coefficients", "residuals", "fitted.values", "rank"))+ stopifnot(all.equal(lltFit[[nm]], unname(lmFit[[nm]])))R> stopifnot(all.equal(lltFit[["Std. Error"]],+ unname(coef(summary(lm(log(Volume) ~ log(Girth), trees)))[,2])))There are several aspects of the C++ code in Figure 7 worth mentioning. The solve method for the LLT object evaluates, in this case, but without actually evaluating the inverse. The calculation of the residuals, , can be written, as in R, as y - fitted. (But note that Eigen classes do not have a “recycling rule” as in R. That is, the two vector operands must have the same length.) The norm() method evaluates the square root of the sum of squares of the elements of a vector. Although one does not explicitly evaluate one does evaluate to obtain the standard errors. Note also the use of the colwise() method in the evaluation of the standard errors. It applies a method to the columns of a matrix, returning a vector. The Eigen colwise() and rowwise() methods are similar in effect to the apply function in R.
In the descriptions of other methods for solving least squares problems, much of the code parallels that shown in Figure 7. The redundant parts are omitted, and only the evaluation of the coefficients, the rank and the standard errors is shown. Actually, the standard errors are calculated only up to the scalar multiple of , the residual standard error, in these code fragments. The calculation of the residuals and and the scaling of the coefficient standard errors is the same for all methods. (See the files fastLm.h and fastLm.cpp in the RcppEigen source package for details.)
4.2 Least squares using the unpivoted QR decomposition
A QR decomposition has the form
where is an orthogonal matrix, which means that , and the matrix is zero below the main diagonal. The matrix is the first columns of and the upper triangular matrix is the top rows of . There are three Eigen classes for the QR decomposition: HouseholderQR provides the basic QR decomposition using Householder transformations, ColPivHouseholderQR incorporates column pivots and FullPivHouseholderQR incorporates both row and column pivots.
Figure 8 shows a least squares solution using the unpivoted QR decomposition. The calculations in Figure 8 are quite similar to those in Figure 7. In fact, if one had extracted the upper triangular factor (the matrixU() method) from the LLT object in Figure 7, the rest of the code would be nearly identical.
using Eigen::HouseholderQR;
const HouseholderQRMatrixXd QR(X);
const VectorXd betahat(QR.solve(y));
const VectorXd fitted(X * betahat);
const int df(n - p);
const VectorXd se(QR.matrixQR().topRows(p).triangularViewUpper()
.solve(MatrixXd::Identity(p,p)).rowwise().norm());
4.3 Handling the rank-deficient case
One important consideration when determining least squares solutions is whether is , a situation described by saying that has “full column rank”. When does not have full column rank it is said to be “rank deficient”.
Although the theoretical rank of a matrix is well-defined, its evaluation in practice is not. At best one can compute an effective rank according to some tolerance. Decompositions that allow to estimation of the rank of the matrix in this way are said to be “rank-revealing”.
Because the model.matrix function in R does a considerable amount of symbolic analysis behind the scenes, one usually ends up with full-rank model matrices. The common cases of rank-deficiency, such as incorporating both a constant term and a full set of indicators columns for the levels of a factor, are eliminated. Other, more subtle, situations will not be detected at this stage, however. A simple example occurs when there is a “missing cell” in a two-way layout and the interaction of the two factors is included in the model.
R> dd <- data.frame(f1 = gl(4, 6, labels = LETTERS[1:4]),+ f2 = gl(3, 2, labels = letters[1:3]))[-(7:8), ]R> xtabs(~ f2 + f1, dd)
f1f2 A B C D a 2 0 2 2 b 2 2 2 2 c 2 2 2 2
R> mm <- model.matrix(~ f1 * f2, dd)R> kappa(mm)
[1] 4.30923e+16This yields a large condition number, indicating rank deficiency. Alternatively, the reciprocal condition number can be evaluated.
R> rcond(mm)
[1] 2.3206e-17
R> (c(rank = qr(mm)$rank, p = ncol(mm)))
rank p 11 12
R> set.seed(1)R> dd$y <- mm %*% seq_len(ncol(mm)) + rnorm(nrow(mm), sd = 0.1)R> fm1 <- lm(y ~ f1 * f2, dd)R> writeLines(capture.output(print(summary(fm1),+ signif.stars = FALSE))[9:22])
Coefficients: (1 not defined because of singularities) Estimate Std. Error t value Pr(>|t|)(Intercept) 0.9779 0.0582 16.8 3.4e-09f1B 12.0381 0.0823 146.3 < 2e-16f1C 3.1172 0.0823 37.9 5.2e-13f1D 4.0685 0.0823 49.5 2.8e-14f2b 5.0601 0.0823 61.5 2.6e-15f2c 5.9976 0.0823 72.9 4.0e-16f1B:f2b -3.0148 0.1163 -25.9 3.3e-11f1C:f2b 7.7030 0.1163 66.2 1.2e-15f1D:f2b 8.9643 0.1163 77.1 < 2e-16f1B:f2c NA NA NA NAf1C:f2c 10.9613 0.1163 94.2 < 2e-16f1D:f2c 12.0411 0.1163 103.5 < 2e-16The lm function for fitting linear models in R uses a rank-revealing form of the QR decomposition. When the model matrix is determined to be rank deficient, according to the threshold used in R’s qr function, the model matrix is reduced to columns by pivoting selected columns (those that are apparently linearly dependent on columns to their left) to the right hand side of the matrix. A solution for this reduced model matrix is determined and the coefficients and standard errors for the redundant columns are flagged as missing.
An alternative approach is to evaluate the “pseudo-inverse” of from the singular value decomposition (SVD) of or the eigendecomposition of . The SVD is of the form
where is an orthogonal matrix and is its leftmost columns, is and zero off the main diagonal so that is a diagonal matrix with non-increasing, non-negative diagonal elements, and is a orthogonal matrix. The pseudo-inverse of , written is a diagonal matrix whose first diagonal elements are the inverses of the corresponding diagonal elements of and whose last diagonal elements are zero. The tolerance for determining if an element of the diagonal of is considered to be (effectively) zero is a multiple of the largest singular value (i.e., the element of ). The pseudo-inverse, , of is defined as
inline ArrayXd Dplus(const ArrayXd& d) {
ArrayXd di(d.size());
double comp(d.maxCoeff() * threshold());
for (int j = 0; j d.size(); ++j) di[j] = (d[j] comp) ? 0. : 1./d[j];
return di;
}
In Figure 9 a utility function, Dplus, is defined to return the diagonal of the pseudo-inverse, , as an array, given the singular values (the diagonal of ) as an array. Calculation of the maximum element of (the method is called .maxCoeff()) and the use of a threshold() function provides greater generality for the function. It can be used on the eigenvalues of , as shown in Section 4.5, even though these are returned in increasing order, as opposed to the decreasing order of the singular values.
4.4 Least squares using the SVD
const Eigen::JacobiSVDMatrixXd
UDV(X.jacobiSvd(Eigen::ComputeThinU|Eigen::ComputeThinV));
const ArrayXd Dp(Dplus(UDV.singularValues()));
const int r((Dp 0).count());
const MatrixXd VDp(UDV.matrixV() * Dp.matrix().asDiagonal());
const VectorXd betahat(VDp * UDV.matrixU().adjoint() * y);
const VectorXd se(s * VDp.rowwise().norm());
With these definitions the code for least squares using the singular value decomposition can be written as in Figure 10. Even in the rank-deficient case this code will produce a complete set of coefficients and their standard errors. The user must be careful to check if the computed rank is less than , the number of columns in , in which case the estimated coefficients are just one of an infinite number of coefficient vectors that produce the same fitted values. It happens that the solution from this pseudo-inverse is the minimum norm solution (in the sense that is minimized among those that minimize ).
The standard errors of the coefficient estimates in the rank-deficient case must be interpreted carefully. The solution with one or more missing coefficients, as returned by the lm.fit function in R and by the column-pivoted QR decomposition described in Section 4.6, does not provide standard errors for the missing coefficients. That is, both the coefficient and its standard error are returned as NA because the least squares solution is performed on a reduced model matrix. It is also true that the solution returned by the SVD method is with respect to a reduced model matrix but the coefficient estimates and their standard errors don’t show this. They are, in fact, linear combinations of a set of coefficient estimates and their standard errors.
4.5 Least squares using the eigendecomposition
The eigendecomposition of is defined as
where , the matrix of eigenvectors, is a orthogonal matrix and is a diagonal matrix with non-decreasing, non-negative diagonal elements, called the eigenvalues of . When the eigenvalues are distinct, this is the same (up to reordering of the columns) as that in the SVD and the eigenvalues of are the (reordered) squares of the singular values of .
With these definitions one can adapt much of the code from the SVD method for the eigendecomposition, as shown in Figure 11.
const Eigen::SelfAdjointEigenSolverMatrixXd VLV(AtA(X));
const ArrayXd Dp(Dplus(eig.eigenvalues()).sqrt());
const int r((Dp 0).count());
const MatrixXd VDp(VLV.eigenvectors() * Dp.matrix().asDiagonal());
const VectorXd betahat(VDp * VDp.adjoint() * X.adjoint() * y);
const VectorXd se(s * VDp.rowwise().norm());
4.6 Least squares using the column-pivoted QR decomposition
The column-pivoted QR decomposition provides results similar to those from R in both the full-rank and the rank-deficient cases. The decomposition is of the form
where, as before, is and orthogonal and is and upper triangular. The matrix is a permutation matrix. That is, its columns are a permutation of the columns of . It serves to reorder the columns of so that the diagonal elements of are non-increasing in magnitude.
An instance of the class Eigen::ColPivHouseholderQR has a rank() method returning the computational rank of the matrix. When is of full rank one can use essentially the same code as in the unpivoted decomposition except that one must reorder the standard errors. When is rank-deficient, the coefficients and standard errors are evaluated for the leading columns of only.
In the rank-deficient case the straightforward calculation of the fitted values, as , cannot be used because some of the estimated coefficients, , are NA so the product will be a vector of NA’s. One could do some complicated rearrangement of the columns of X and the coefficient estimates but it is conceptually (and computationally) easier to employ the relationship
The vector is called the “effects” vector in R.
typedef Eigen::ColPivHouseholderQRMatrixXd CPivQR;
typedef CPivQR::PermutationType Permutation;
const CPivQR PQR(X);
const Permutation Pmat(PQR.colsPermutation());
const int r(PQR.rank());
VectorXd betahat, fitted, se;
if (r == X.cols()) { // full rank case
betahat = PQR.solve(y);
fitted = X * betahat;
se = Pmat * PQR.matrixQR().topRows(p).triangularViewUpper()
.solve(MatrixXd::Identity(p, p)).rowwise().norm();
} else {
MatrixXd Rinv(PQR.matrixQR().topLeftCorner(r, r)
.triangularViewUpper().solve(MatrixXd::Identity(r, r)));
VectorXd effects(PQR.householderQ().adjoint() * y);
betahat.fill(::NA_REAL);
betahat.head(r) = Rinv * effects.head(r);
betahat = Pmat * betahat;
se.fill(::NA_REAL);
se.head(r) = Rinv.rowwise().norm();
se = Pmat * se;
// create fitted values from effects
effects.tail(X.rows() - r).setZero();
fitted = PQR.householderQ() * effects;
}
Just to check that the code in Figure 12 does indeed provide the desired answer
R> print(summary(fmPQR <- fastLm(y ~ f1 * f2, dd)), signif.st = FALSE)
Call:fastLm.formula(formula = y ~ f1 * f2, data = dd) Estimate Std. Error t value Pr(>|t|)(Intercept) 0.977859 0.058165 16.812 3.413e-09f1B 12.038068 0.082258 146.346 < 2.2e-16f1C 3.117222 0.082258 37.896 5.221e-13f1D 4.068523 0.082258 49.461 2.833e-14f2b 5.060123 0.082258 61.516 2.593e-15f2c 5.997592 0.082258 72.912 4.015e-16f1B:f2b -3.014763 0.116330 -25.916 3.266e-11f1C:f2b 7.702999 0.116330 66.217 1.156e-15f1D:f2b 8.964251 0.116330 77.059 < 2.2e-16f1B:f2c NA NA NA NAf1C:f2c 10.961326 0.116330 94.226 < 2.2e-16f1D:f2c 12.041081 0.116330 103.508 < 2.2e-16Residual standard error: 0.2868 on 11 degrees of freedomMultiple R-squared: 0.9999,Adjusted R-squared: 0.9999
R> all.equal(coef(fm1), coef(fmPQR))
[1] TRUE
R> all.equal(unname(fitted(fm1)), fitted(fmPQR))
[1] TRUE
R> all.equal(unname(residuals(fm1)), residuals(fmPQR))
[1] TRUEThe rank-revealing SVD method produces the same fitted values but not the same coefficients.
R> print(summary(fmSVD <- fastLm(y ~ f1 * f2, dd, method = 4)),+ signif.st = FALSE)
Call:fastLm.formula(formula = y ~ f1 * f2, data = dd, method = 4) Estimate Std. Error t value Pr(>|t|)(Intercept) 0.977859 0.058165 16.812 3.413e-09f1B 7.020458 0.038777 181.049 < 2.2e-16f1C 3.117222 0.082258 37.896 5.221e-13f1D 4.068523 0.082258 49.461 2.833e-14f2b 5.060123 0.082258 61.516 2.593e-15f2c 5.997592 0.082258 72.912 4.015e-16f1B:f2b 2.002847 0.061311 32.667 2.638e-12f1C:f2b 7.702999 0.116330 66.217 1.156e-15f1D:f2b 8.964251 0.116330 77.059 < 2.2e-16f1B:f2c 5.017610 0.061311 81.838 < 2.2e-16f1C:f2c 10.961326 0.116330 94.226 < 2.2e-16f1D:f2c 12.041081 0.116330 103.508 < 2.2e-16Residual standard error: 0.2868 on 11 degrees of freedomMultiple R-squared: 0.9999,Adjusted R-squared: 0.9999
R> all.equal(coef(fm1), coef(fmSVD))
[1] "'is.NA' value mismatch: 0 in current 1 in target"
R> all.equal(unname(fitted(fm1)), fitted(fmSVD))
[1] TRUE
R> all.equal(unname(residuals(fm1)), residuals(fmSVD))
[1] TRUEThe coefficients from the symmetric eigendecomposition method are the same as those from the SVD, hence the fitted values and residuals must be the same for these two methods.
R> summary(fmVLV <- fastLm(y ~ f1 * f2, dd, method = 5))R> all.equal(coef(fmSVD), coef(fmVLV))
[1] TRUE
4.7 Comparative speed
In the RcppEigen package the R function to fit linear models using the methods described above is called fastLm. It follows an earlier example in the Rcpp package which was carried over to both RcppArmadillo and RcppGSL. The natural question to ask is, “Is it indeed fast to use these methods based on Eigen?”. To this end, the package provides benchmarking code for these methods, R’s lm.fit function and the fastLm implementations in the RcppArmadillo (Eddelbuettel et al., 2026b) and RcppGSL (Eddelbuettel and François, 2026) packages, if they are installed. The benchmark code, which uses the rbenchmark (Kusnierczyk et al., 2026) package, is in a file named lmBenchmark.R in the examples subdirectory of the installed RcppEigen package.
It can be run as
R> source(system.file("examples", "lmBenchmark.R", package = "RcppEigen"))Results will vary according to the speed of the processor and the implementation of the BLAS (Basic Linear Algebra Subroutines) used. (Eigen methods do not use the BLAS but the other methods do.) The Eigen3 template library does not use multi-threaded code for these operations but does use the graphics pipeline instructions (SSE and SSE2, in this case) in some calculations.
| Method | Relative | Elapsed | User | Sys |
|---|---|---|---|---|
| LDLt | 1.00 | 1.18 | 1.17 | 0.00 |
| LLt | 1.01 | 1.19 | 1.17 | 0.00 |
| SymmEig | 2.76 | 3.25 | 2.70 | 0.52 |
| QR | 6.35 | 7.47 | 6.93 | 0.53 |
| arma | 6.60 | 7.76 | 25.69 | 4.47 |
| PivQR | 7.15 | 8.41 | 7.78 | 0.62 |
| lm.fit | 11.68 | 13.74 | 21.56 | 16.79 |
| GESDD | 12.58 | 14.79 | 44.01 | 10.96 |
| SVD | 44.48 | 52.30 | 51.38 | 0.80 |
| GSL | 150.46 | 176.95 | 210.52 | 149.86 |
Results obtained on a desktop computer, circa 2010, are shown in Table 2. The processor used for these timings is a 4-core processor but almost all the methods are single-threaded and not affected by the number of cores. Only the arma, lm.fit, GESDD and GSL methods benefit from the multi-threaded BLAS implementation provided by OpenBLAS, and the relative speed increase is modest for this problem size and number of cores (at 7.76 seconds relative to 10.29 seconds for arma, 13.74 seconds relative to 16.91 seconds for lm.fit, and 176.95 seconds relative to 193.29 seconds for the GSL). Parallel computing approaches will always have to trade-off increased communication and overhead costs against the potential gains from running multiple execution threads.
These results indicate that methods based on forming and decomposing (LDLt, LLt and SymmEig) are considerably faster than the others. The SymmEig method, using a rank-revealing decomposition, would be preferred, although the LDLt method could probably be modified to be rank-revealing. However, the dimensions of the problem will influence the comparative results. Because there are 100,000 rows in , methods that decompose the whole matrix (all the methods except those named above) will be at a disadvantage.
The pivoted QR method is 1.6 times faster than R’s lm.fit on this test and provides nearly the same information as lm.fit. Methods based on the singular value decomposition (SVD and GSL) are much slower but, as mentioned above, this is caused in part by having many more rows than columns. The GSL method from the GNU Scientific Library uses an older algorithm for the SVD and is clearly out of contention.
The GESDD method provides an interesting hybrid: It uses the Eigen classes, but then deploys the LAPACK routine dgesdd for the actual SVD calculation. The resulting time is much faster than using the SVD implementation of Eigen which is not a particularly fast SVD method.
5 Delayed evaluation
A form of delayed evaluation is used in Eigen. That is, many operators and methods do not evaluate the result but instead return an “expression object” that is evaluated when needed. As an example, even though one writes the evaluation as .rankUpdate(X.adjoint()) the X.adjoint() part is not evaluated immediately. The rankUpdate method detects that it has been passed a matrix that is to be used in its transposed form and evaluates the update by taking inner products of columns of instead of rows of .
As Eigen has developed some of these unevaluated expressions have been modified. In Eigen 3.1, which is incorporated in version 0.3.1 of RcppEigen, the .adjoint() method applied to a real dense matrix copies the contents of the original matrix, flags it as row-major and interchanges the number of rows and columns. This is indeed the adjoint of the original matrix but, at one time, the wrap method for the Eigen dense matrix classes would fail on row-major matrices.
const MapMati A(asMapMati(AA));
return wrap(A.transpose());
In the code for the transpose of an integer matrix shown in Figure 1 the transpose is assigned to a MatrixXi object before applying wrap to it. The assignment forces the evaluation as a column-major matrix. In early versions of the RcppEigen package if this step is skipped, as in Figure 13, the result would have been incorrect.
R> Ai <- matrix(1:6, ncol = 2L)R> ftrans2 <- cxxfunction(signature(AA = "matrix"), badtransCpp,+ "RcppEigen", incl)R> (At <- ftrans2(Ai))
[,1] [,2] [,3][1,] 1 2 3[2,] 4 5 6Although the problem no longer persists for this particular example, the recommended practice is to first assign objects before wrapping them for return to R.
6 Sparse matrices
using Eigen::MappedSparseMatrix;
using Eigen::SparseMatrix;
const MappedSparseMatrixdouble A(asMappedSparseMatrixdouble (AA));
const MapVecd y(asMapVecd(yy));
const SparseMatrixdouble At(A.adjoint());
return List::create(Named("At") = At,
Named("Aty") = At * y);
Eigen provides sparse matrix classes. An R object of class dgCMatrix (from the Matrix package by Bates and Maechler (2026)) can be mapped as in Figure 14.
R> sparse1 <- cxxfunction(signature(AA = "dgCMatrix", yy = "numeric"),+ sparseProdCpp, "RcppEigen", incl)R> data("KNex", package = "Matrix")R> rr <- sparse1(KNex$mm, KNex$y)R> stopifnot(all.equal(rr$At, t(KNex$mm)),+ all.equal(rr$Aty, as.vector(crossprod(KNex$mm, KNex$y))))Sparse Cholesky decompositions are provided by the SimplicialLLT and SimplicialLDLT classes in the RcppEigen package for R. These are subclasses of the SimplicialCholesky templated. A sample usage is shown in Figure 15.
typedef Eigen::MappedSparseMatrixdouble MSpMat;
typedef Eigen::SparseMatrixdouble SpMat;
typedef Eigen::SimplicialLDLTSpMat SpChol;
const SpMat At(asMSpMat(AA).adjoint());
const VectorXd Aty(At * asMapVecd(yy));
const SpChol Ch(At * At.adjoint());
if (Ch.info() != Eigen::Success) return R_NilValue;
return List::create(Named("betahat") = Ch.solve(Aty),
Named("perm") = Ch.permutationP().indices());
R> sparse2 <- cxxfunction(signature(AA = "dgCMatrix", yy = "numeric"),+ sparseLSCpp, "RcppEigen", incl)R> str(rr <- sparse2(KNex$mm, KNex$y))
List of 2 $ betahat: num [1:712] 823 340 473 349 188 ... $ perm : int [1:712] 572 410 414 580 420 425 417 426 431 445 ...
R> res <- as.vector(solve(Ch <- Cholesky(crossprod(KNex$mm)),+ crossprod(KNex$mm, KNex$y)))R> stopifnot(all.equal(rr$betahat, res))R> all(rr$perm == Ch@perm)
[1] FALSEThe fill-reducing permutations are different.
7 Summary
This paper introduced the RcppEigen package which provides high-level linear algebra computations as an extension to the R system. RcppEigen is based on the modern C++ library Eigen which combines extended functionality with excellent performance, and utilizes Rcpp to interface R with C++. Several illustrations covered common matrix operations and several approaches to solving a least squares problem—including an extended discussion of rank-revealing approaches. Sparse matrix computations were also discussed, and a short example provided an empirical demonstration of the excellent run-time performance of the RcppEigen package.
References
- C++ Template Metaprogramming: concepts, tools and techniques from Boost and beyond. Addison-Wesley, Boston. Cited by: §2.
- Fast and elegant numerical linear algebra using the RcppEigen package. Journal of Statistical Software 52 (5), pp. 1–24. External Links: Link, Document Cited by: Fast and Elegant Numerical Linear Algebra Using the RcppEigen Package.
- Matrix: sparse and dense matrix classes and methods. Note: R package version 1.7-5 External Links: Document Cited by: §6.
- Rcpp: seamless R and C++ integration. Note: R package version 1.1.2 External Links: Document Cited by: §1, Abstract.
- RcppArmadillo: rcpp integration for the armadillo templated linear algebra library. Note: R package version 15.4.0-1 External Links: Document Cited by: §4.7.
- Rcpp: seamless R and C++ integration. Journal of Statistical Software 40 (8), pp. 1–18. External Links: Link, Document Cited by: §1.
- RcppGSL: rcpp integration for gnu gsl vectors and matrices. Note: R package version 0.3.14 External Links: Document Cited by: §4.7.
- Seamless r and c++ integration with rcpp. Use R!, Springer, New York. External Links: ISBN 978-1-4614-6867-7, Document Cited by: §1, Abstract.
- Eigen v3. External Links: Link Cited by: §1, Abstract.
- C++ 2011 standard document 14882:2011. International Organization for Standardization. Note: ISO/IEC Standard Group for Information Technology / Programming Languages / C++ External Links: Link Cited by: §1.
- rbenchmark: benchmarking routine for r. Note: R package version 1.0.1 External Links: Document Cited by: §4.7.
- More effective c++: 35 new ways to improve your programs and designs. Addison-Wesley, Boston. Note: ISBN 020163371X Cited by: §1.
- Effective c++: 55 specific ways to improve your programs and designs. 3rd edition, Addison-Wesley, Boston. Note: ISBN 978-0321334879 Cited by: §1.
- R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. External Links: Document, Link Cited by: §1, Abstract.
- Writing r extensions. R Foundation for Statistical Computing, Vienna, Austria. Note: ISBN 3-900051-11-9 External Links: Document, Link Cited by: §1.
- Inline: functions to inline c, c++, fortran function calls from r. Note: R package version 0.3.21 External Links: Document Cited by: §3.
- Arrays in Blitz++. In ISCOPE ’98: Proceedings of the Second International Symposium on Computing in Object-Oriented Parallel Environments, London, pp. 223–230. Note: ISBN 3-540-65387-2 Cited by: §1.
Affiliation
Douglas Bates
Department of Statistics
University of Wisconsin-Madison
Madison, WI, United States of America
E-mail: bates@stat.wisc.edu
URL: http://www.stat.wisc.edu/~bates/
Dirk Eddelbuettel
Debian Project
River Forest, IL, United States of America
E-mail: edd@debian.org
URL: http://dirk.eddelbuettel.com