Implementation of Toeplitz matricies using several algorithms and SciPy's LinearOperators.
- Two of the available operators use an implementation that forms a vector of the elements of the first row and column and indexes out the subsets corresponding to the rows as needed.
- One of these operators uses Cython and SciPy's Cython wrappers for BLAS to speed up the floating-point cases.
- Another operator uses scipy.signal.convolve.
- The fourth uses NumPy's FFTs.
- The fifth uses NumPy stride tricks, which appears to make the array contiguous before doing any products.
All implementations should be lower-memory than scipy.linalg.toeplitz, and the third and fourth implementations also have algorithmic speedups. The first two implementations may be slightly faster due to cache interaction with the smaller memory footprint, but this effect will be small both for matrices small enough to fit entirely in cache and for matrices large enough that even the smaller representation doesn't fit in cache.
| -- | implementation |
|---|---|
|
|
| ========= | ========== ======================== ========== ============ ================== ============= |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
CPython 3.6 with Cython 0.29.5, NumPy 1.16.6, and SciPy 1.2.3. ConvolveToeplitz and FFTToeplitz may be faster with the recent FFT improvements in NumPy.