petrbalvin/tensor
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 2 Wiki Activity

fix(linalg): refuse non-finite input in dense Cholesky, Eigen and rank-one updates

Browse Source 
Assisted-by: GLM 5.3 Flash
This commit is contained in:
petrbalvin
2026-09-27 11:12:43 +02:00
parent af4ee19703
commit 7bc030c78b
10 changed files with 144 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/API.md
+4 -4
View File
@@ -629,7 +629,7 @@ neighbours are the way in.
| `Inv(a)` | returns the inverse of a square matrix, through the same LU kernel. |
| `Det(a)` | returns the determinant of a square real matrix; a singular matrix gives 0, not an error. |
| `DetComplex(a)` | the same for a square complex matrix, as a `complex128`. |
| `Cholesky(a)` | returns the lower factor L with a = L·Lᵀ for a symmetric positive definite a; a non-positive pivot is refused. |
| `Cholesky(a)` | returns the lower factor L with a = L·Lᵀ for a symmetric positive definite a; a non-positive pivot and a non-finite entry are refused. |
| `CholeskyUpdate(l, x)` | returns the lower factor of A + x·xᵀ from the factor of A, by orthogonal rotations. |
| `CholeskyDowndate(l, x)` | returns the lower factor of A − x·xᵀ, by hyperbolic rotations; leaving the positive definite cone is an error. |
| `QR(a)` | returns q (m×m orthogonal) and r (m×n upper triangular) with a = q·r, for m ≥ n. |
@@ -641,8 +641,8 @@ neighbours are the way in.
| Call | What it does |
|---|---|
| `Eigen(a)` | returns the eigenvalues of a real symmetric matrix, ascending, and the orthonormal eigenvectors as columns; asymmetry beyond a scale-relative 1e-12 is refused. |
| `EigenComplex(a)` | the same for a complex Hermitian matrix; values are real, vectors complex, and the same 1e-12 Hermitian tolerance applies. |
| `Eigen(a)` | returns the eigenvalues of a real symmetric matrix, ascending, and the orthonormal eigenvectors as columns; asymmetry beyond a scale-relative 1e-12 and a non-finite entry are refused. |
| `EigenComplex(a)` | the same for a complex Hermitian matrix; values are real, vectors complex, the same 1e-12 Hermitian tolerance applies, and a non-finite entry is refused. |
| `EigenGeneral(a)` | the eigenvalues and eigenvectors of a square matrix of any dtype; real and int inputs promote to `complex128`, values descend by magnitude, and a real matrix may carry conjugate pairs. |
| `EigenGeneralised(a, b)` | solves A·v = λ·B·v for symmetric a and symmetric positive definite b, through the Cholesky factor of b; values ascend, eigenvectors are B-orthonormal columns. |
| `SVD(a)` | the thin decomposition a = U·Σ·Vᵀ for m ≥ n: U is m×n with orthonormal columns, Σ a length-n vector descending, Vᵀ n×n orthogonal; a wide matrix is transposed first and the factors are swapped back. |
@@ -898,7 +898,7 @@ The conditions this package reports:
- a modification that needs fill the factor does not hold:
`SparseCholesky.Update` and `Downdate` refuse it and leave the
factor as it was, and `CholeskyUpdate` refuses a rank-1 factor, a
mismatched vector or a complex input.
mismatched vector, a complex input or a non-finite entry.
- a complex or wrong-length right-hand side to a sparse solve, and a
preconditioner built for another dimension.
- non-finite input to `NewSparseLU`, `NewSparseILU` and the sparse