Spectra v1.2.0

Following v1.1.0, this version of Spectra further improves the support for complex eigen solvers. The main update in v1.2.0 is that GenEigsSolver now supports general complex-valued matrices. See the Complex-valued Matrices section in the README for an example.

Added

• Added the support for complex-valued matrices in the GenEigsSolver eigen solver
• Added a separate Givens class for computing Givens rotations in a numerically stable way
• Added the implementation of UpperHessenbergQR and UpperHessenbergEigen for complex-valued matrices
• Added testing code for complex eigen solvers based on the GenEigsSolver class
• Added testing code for the Givens class
• Added UpperHessenbergQR and UpperHessenbergEigen tests for complex-valued matrices

Changed

• UpperHessenbergQR now uses the Givens class to compute Givens rotations
• Removed the redundant Scalar template paramter in internal classes such as ArnoldiOp, Arnoldi, and Lanczos
• The CMake script now uses CPM.cmake to manage the dependence on Eigen. By default, it will first search for Eigen in the system, and if it is not found, the script will automatically download the Eigen library. This functionality was contributed by @eduardz1 (#178)
• The minimum version for Clang-Format is now 16, and the configuration file is also updated
• Updated Doxygen configuration to v1.14.0
• Updated the included Catch2 to v2.13.10
• Updated author information
• Various documentation updates and improvements

Spectra v1.1.0

Happy New Year 2025! 🎉

This version of Spectra has a major update: the support for complex Hermitian matrices is finally added, and now you can use the HermEigsSolver eigen solver to compute real eigenvalues of Hermitian matrices. This functionality has been requested for a long time (e.g. #6 dating back to ten years ago!), and I am excited to make this first step. 🥳

Added

• Added the eigen solver HermEigsSolver for complex-valued Hermitian matrices
• Added the support for complex Hermitian matrices to various linear algebra classes, such as Arnoldi, Lanczos, and BKLDLT
• SimpleRandom can now generate random complex values
• Added testing code for Arnoldi factorization on both real matrices and complex Hermitian matrices
• Added BKLDLT tests for complex Hermitian matrices
• Added testing code for the HermEigsSolver solver
• Added test examples from previous bug reports

Changed

• Better heuristics for basis restarting in the Lanczos factorization, which fixes #159
• Fixed the support for non-literal data types (#150)
• Various CMake configuration improvements (#127, #140, #152), thanks to @LTLA, @JensWehner, @shivupa, @jschueller, @alecjacobson, and @jdumas
• Miscellaneous GitHub Actions updates
• Various internal implementation changes to accommodate operations on complex-valued matrices
• Improved the numerical stability of BKLDLT linear system solver
• Change SymEigsBase to HermEigsBase as the base class for both real symmetric and complex Hermitian eigen solvers

Spectra v1.0.1

Added

• Added SIMD support for UpperHessenbergSchur. This should accelerate general eigen solvers such as GenEigsSolver
• Added test code for UpperHessenbergSchur

Changed

• Fixed several bugs in the examples caused by the const keyword, reported by @alexpghayes (#135, #137)
• Updated the included Catch2 to v2.13.8

Spectra v1.0.0

Finally, Spectra v1.0.0 is released!

This new version is the accumulation of dozens of new features and bug fixes, with the help of many users and contributors. The most visible changes compared to v0.9.0 include fully refactored code base using C++11, a redesigned API, and the newly added Davidson eigen solver by Jens Wehner and his group. More details can be found below and in the migration guide.

Added

• Added version macros SPECTRA_MAJOR_VERSION, SPECTRA_MINOR_VERSION, SPECTRA_PATCH_VERSION, and SPECTRA_VERSION that are included by all eigen solvers
• Added the wrapper class SparseGenComplexShiftSolve for eigen solver with complex shifts
• Added the SymGEigsShiftSolver class for symmetric generalized eigen solver with real shifts
• Added the wrapper class SymShiftInvert that can be used with SymGEigsShiftSolver
• Added test code for symmetric generalized eigen solver with real shifts
• Added an internal class UpperHessenbergSchur to compute the Schur decomposition of upper Hessenberg matrices more efficiently
• Added a Flags template parameter to every matrix operation class (e.g. DenseCholesky and DenseSymMatProd), whose possible values are Eigen::ColMajor and Eigen::RowMajor. This parameter allows these wrapper classes to handle row-major matrices. If the input matrix is inconsistent with the Flags parameter (e.g., if Flags is Eigen::ColMajor but the input matrix is row-major), a compiler error will occur
• Added the member function info() and convergence tests to SparseRegularInverse, suggested by @Spammed (#111)
• Added symmetric Davidson eigen solver DavidsonSymEigsSolver, written by Felipe Zapata, Nicolas Renaud, Victor Azizi, Pablo Lopez-Tarifa, and Jens Wehner from the Netherlands eScience Center
• Extended matrix operations in DenseGenMatProd, DenseSymMatProd, SparseGenMatProd, and SparseSymMatProd to handle matrix-matrix products and coefficient-wise accessors

Changed

• API change: Spectra now requires C++11
• API change: All enumerations have been converted to enum classes (e.g. LARGEST_MAGN is now SortRule::LargestMagn)
• API change: Selection rules are no longer template parameters. They are now specified in the compute() member function as arguments
• API change: The Scalar template parameter has been removed from eigen solvers. Instead, matrix operation classes now need to define a public type named Scalar
• API change: Constructors of solvers now request references of matrix operators instead of pointers
• Clang-Format now uses the C++11 standard to format code
• Updated documentation to reflect the new API
• Many internal changes to make use of C++11 features
• Added a SPECTRA_ prefix to each header guard to prevent potential name clash
• Changed the default value of the Flags template parameter that exists in various class templates from 0 to the more readable constant Eigen::ColMajor
• Renamed the function mat_prod to perform_op in the SparseRegularInverse wrapper class. This makes the API more consistent when implementing new generalized eigen solvers
• Improved the precision of UpperHessenbergQR and TridiagQR by computing the Givens rotations in a more stable way
• Added a deflation test to TridiagQR to accelerate the convergence of eigen solvers
• Improved the precision of TridiagQR::matrix_QtHQ() by directly applying rotations to the original input matrix
• Improved the precision of DoubleShiftQR by computing the Householder reflectors in a more stable way
• Improved the deflation test in DoubleShiftQR
• More careful computation of residual vectors in the Lanczos process
• Initial vectors in the Lanczos and Arnoldi processes are now forced to be in the range of the A matrix
• More sensible test for orthogonality in generating new random vectors in the Lanczos and Arnoldi processes
• In symmetric eigen solvers large shifts are applied first to increase precision
• Updated the included Catch2 to v2.13.6

Spectra v0.9.0

Added

• Added support for CMake build, contributed by Guillaume Acke and Jens Wehner (#70, #88)
• Spectra can now be installed via conda-forge, thanks to Guillaume Acke and Julien Schueller (#81, #85)
• The source code of Spectra is now formatted using ClangFormat, suggested by Jens Wehner

Changed

• Fixed a compiler warning caused by unused parameter, contributed by Julien Schueller (#80)
• Changed the implementation of BKLDLT solver to improve precision in some tests

Spectra v0.8.1

Changed

• Fixed a bug in BKLDLT in which a wrong type was used, thanks to @jdbancal for the issue #64
• Fixed a bug in BKLDLT that caused segmentation fault in some edge cases, also reported by @jdbancal in issue #66
• The type Eigen::Index is now globally used for indices and sizes, in order to handle potentially large matrices. This was suggested by Yuan Yao in issue #19

Spectra v0.8.0

Added

• Added a BKLDLT class that implements the Bunch-Kaufman LDLT decomposition for symmetric indefinite matrices. According to the Eigen documentation, currently Eigen::LDLT cannot handle some special indefinite matrices such as [0, 1; 1, 0], but BKLDLT is applicable to any symmetric matrices as long as it is not singular. LDLT decomposition is used in shift-and-invert solvers (see below)
• Added a unit test for BKLDLT

Changed

• DenseSymShiftSolve now uses the newly added BKLDLT class to do the decomposition. This change broadens the class of matrices that DenseSymShiftSolve can handle, reduces memory use, and should also improve the numerical stability of the solver
• Replaced Eigen::SimplicialLDLT with Eigen::SparseLU in the SparseSymShiftSolve class, as some edge-case indefinite matrices may break Eigen::SimplicialLDLT
• SparseSymShiftSolve and SparseGenRealShiftSolve will throw an error if the factorization failed, for example, on singular matrices
• Fixed a missing #include in DenseCholesky.h, thanks to Lennart Trunk for the issue #59
• Fixed errors in examples (#60), thanks to @linuxfreebird
• Updated the included Catch2 to v2.7.0

Spectra v0.7.0

Added

• Added a directory contrib to include code contributed by users. It is not formally a part of the Spectra library, but it may contain useful solvers and applications based on Spectra. Code in contrib may not be fully tested, so please use with caution. Feedback and report of issues are always welcome
• Added an eigen solver LOBPCGSolver in the contrib directory using the LOBPCG algorithm, contributed by Anna Araslanova
• Added a partial SVD solver PartialSVDSolver in the contrib directory
• Added two internal classes Arnoldi and Lanczos to compute the Arnoldi/Lanczos factorization in eigen solvers
• Added a few other internal classes to refactor the eigen solver classes (see below)

Changed

• API change: Spectra now requires Eigen >= 3.3
• API change: The library header files are moved into a directory named Spectra. Hence the recommended include directive would look like #include <Spectra/SymEigsSolver.h>
• All eigen solvers have been refactored using a cleaner class hierarchy. It may potentially make the implementation of new eigen solvers easier, especially for generalized eigen problems
• The matrix operation classes (e.g. DenseSymMatProd and SparseSymMatProd) are now internally using an Eigen::Ref object to wrap the user matrices, thanks to Dario Mangoni who raised this issue in #16
• Fixed inappropriate range of random numbers in the tests

Spectra v0.6.2

Changed

• Fixed regressions in v0.6.0 on some edge cases
• Improved the accuracy of restarting processes in SymEigsSolver and GenEigsSolver
• Updated the included Catch2 to v2.2.2
• Code and documentation cleanup

Spectra v0.6.1

Changed

• Fixed a bug of uninitialized memory
• Updated the included Catch2 to v2.1.2