Randomized singular value decomposition (SVD) written in C++14 / Eigen.

Let A be a real or complex matrix of size m ⨉ n, rank(A) = r. Its economic singular value decomposition (SVD) is given by A = UΣV* with the following factor matrices:

• left singular vectors matrix U of size m ⨉ r
• singular values matrix Σ of size r ⨉ r
• right singular vectors matrix V of size n ⨉ r

The left and right singular vector matrices have orthonormal columns: U*U = I, V*V = I. The singular values matrix Σ is diagonal and has sorted singular values on its principal diagonal.

When m and n are very large, we want to approximate this SVD using the randomized range approximation Q of size m ⨉ r such that || A - QQ*A ||₂ < ε. Then, the problem can be projected onto a smaller subspace as follows: B = Q*A, B is of size r ⨉ n. After the SVD of the smaller problem B = ŨΣV*, the solution to the original problem can be recovered as U = QŨ.

The range Q is approximated using random sampling. In order to capture the largest singular values, randomized subspace iterations can be used. However, as any power iteration methods, they suffer from numerical problems. To mitigate this problem, the user can select an appropriate conditioner based on the modified Gram–Schmidt process, the LU decomposition, and the QR decomposition. The conditioner choice is a trade-off between runtime and numerical properties.

Well, first of all make sure you need to use C++. RSVD implementations in Python (scikit-learn/scikit-learn, facebook/fbpca) are competitive performance-wise and much easier to use.

Still want to use C++?

Then you might want to use mp4096/rsvd because:

• 🚀 it is compile-time generic over real and complex, single and double precision matrices
• 🎓 it supports one-shot range approximation as well as randomized subspace iterations with a choice of conditioners (none, MGS, LU, QR)
• 📖 it has nice Doxygen documentation
• 🔬 it is well-tested
• 🎲 it offers you fine control over random number generation for the sampling matrix Ω, thus alleviating any potential problems with non-deterministic approximation results due to different PRNG seeds
• 👍 it is written in idiomatic C++14 style

• gcc ≥ 5.4 or clang ≥ 3.8 (clang 6.0 is recommended)
• CMake ≥ 3.5
• Eigen ≥ 3.3

This is a header-only library, just make sure you have Eigen installed and add the following includes to your program:

#include <rsvd/Constants.hpp>
#include <rsvd/ErrorEstimators.hpp>
#include <rsvd/RandomizedSvd.hpp>

Unfortunately, a header-only library requires longer compile times, since it is recompiled every time.

See examples/SimpleUsage/main.cpp. You can compile it as follows:

$ mkdir -p build
$ cd build
$ cmake ..
$ make -j2 example_simple_usage

• Nathan Halko, Per-Gunnar Martinsson, Joel A. Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. 2010.

• Markus Herb @herb for the kalman library which served as an example for good C++ style.
• Kazuya Otani @kazuotani14 for the RandomizedSvd C++ / Eigen implementation.
• scikit-learn and fbpca contributors for the reference randomized SVD implementations in Python.