Krylov methods#

Implemented in yastn#

We provide a high-level, backend-agnostic implementation of some Krylov-based algorithms used in yastn.tn.mps. They assume a linear operation acting on a generalized vector, with the vector being an instance of a class that includes methods norm, vdot, linear_combination, expand_krylov_space, among others. Examples of such a vector include yastn.Tensor (see methods).

yastn.expmv(f, v, t=1.0, tol=1e-12, ncv=10, hermitian=False, normalize=False, return_info=False, **kwargs) → vector[source]#

Calculate \(e^{(tF)}v\), where \(v\) is a vector, and \(F(v)\) is linear operator acting on \(v\).

The algorithm of: J. Niesen, W. M. Wright, ACM Trans. Math. Softw. 38, 22 (2012), Algorithm 919: A Krylov subspace algorithm for evaluating the phi-functions appearing in exponential integrators.

Parameters:

  • f (Callable[[vector], vector]) – defines an action of a ‘square matrix’ on vector.

  • v (vector) – input vector to apply exponential map onto.

  • t (number) – exponent amplitude.

  • tol (number) – targeted tolerance; it is used to update the time-step and size of Krylov space. The returned result should have better tolerance, as correction is included.

  • ncv (int) – Initial guess for the size of the Krylov space.

  • hermitian (bool) – Assume that f is a hermitian operator, in which case Lanczos iterations are used. Otherwise Arnoldi iterations are used to span the Krylov space.

  • normalize (bool) – The result is normalized to unity using 2-norm.

  • return_info (bool) – if True, returns (vector, info), where

    • info.ncv : guess of the Krylov-space size,

    • info.error : estimate of error (likely over-estimate)

    • info.krylov_steps : number of execution of f(x),

    • info.steps : number of steps to reach t,

  • kwargs (any) – Further parameters that are passed to expand_krylov_space() and linear_combination().

yastn.eigs(f, v0, k=1, which='SR', ncv=10, maxiter=None, tol=1e-13, hermitian=False, **kwargs) → tuple[array, Sequence[vectors]][source]#

Search for dominant eigenvalues of linear operator f using Arnoldi algorithm. Economic implementation (without restart) for internal use within yastn.tn.dmrg_().

Parameters:

  • f (function) – define an action of a ‘square matrix’ on the ‘vector’ v0. f(v0) should preserve the signature of v0.

  • v0 (Tensor) – Initial guess, ‘vector’ to span the Krylov space.

  • k (int) – Number of desired eigenvalues and eigenvectors. Default is 1.

  • which (str) – One of [‘LM’, ‘LR’, ‘SR’] specifying which k eigenvectors and eigenvalues to find: ‘LM’ : largest magnitude, ‘SM’ : smallest magnitude, ‘LR’ : largest real part, ‘SR’ : smallest real part.

  • ncv (int) – Dimension of the employed Krylov space. Default is 10. Must be greated than k.

  • maxiter (int) – Maximal number of restarts; (not implemented for now)

  • tol (float) – Stopping criterion for an expansion of the Krylov subspace. Default is 1e-13. (not implemented for now)

  • hermitian (bool) – Assume that f is a hermitian operator, in which case Lanczos iterations are used. Otherwise Arnoldi iterations are used to span the Krylov space.

Other libraries#

With NumPy backend, it is possible to link to algorithms in sparse.sparse.linalg, employing yastn.Tensor.compress_to_1d() and yastn.decompress_from_1d(). See, an example.