DIIS solver

DIIS module for iterative solvers

This module provides the DIIS (Direct Inversion in the Iterative Subspace) method for iterative solvers.

The DIIS method is used to accelerate the convergence of iterative solvers by combining previous solutions to the problem to minimize the residual. The vectors and residuals are stored in files as Vector{Vector{Float64}}.

Usage

diis = Diis(EC)
for it = 1:maxit
  # compute Vec = [Vec1,Vec2,...] and Res = [Res1,Res2,...]
  # ...
  perform!(diis, Vec, Res)
  # ...
end

One can also provide a tuple of custom dot-product functions for the residuals components as customdots argument in perform! function.

DIIS object

perform!(diis::Diis, Amps, Res, customdots=())

Perform DIIS.

Amps is an array of vectors and Res is an array of residuals. The vectors Amps will be replaced by the DIIS optimized vectors. customdots is a tuple of functions for each residual component to calculate the dot-product. The functions should have the signature f(ten1::Array{T,N}, ten2::Array{T,N}).

combine(diis::Diis, vecfiles, coeffs)

Combine vectors from files with coefficients.

custom_dot(diis::Diis, customdots, tens, vecs)

Compute weighted (with diis.weights) dot product of vectors using custom dot-product functions customdots::Tuple. vecs are reshaped to the shape of tensors tens.

loadamps(diis::Diis, ipos)

Load vectors from file at position ipos as Vector{Vector{Float64}}.

loadres(diis::Diis, ipos)

Load residuals from file at position ipos as Vector{Vector{Float64}}.

loadvecs(file)

Load vectors from file as Vector{Vector{Float64}}.

saveamps(diis::Diis, vecs, ipos)

Save vectors to file (replacing previous vectors at position ipos).

saveres(diis::Diis, vecs, ipos)

Save residuals to file (replacing previous residuals at position ipos).

tuple_reshape(vecs, tens)

Reshape vectors vecs to the shape of tensors tens.

Returns Tuple of reshaped vectors.

update_Bmat(diis::Diis, nDim, Res, ithis, customdots=())

Update B matrix with new residual (at the position ithis).

B matrix is defined as:

\[{\bf B} = \begin{pmatrix} \langle {\bf R}_1, {\bf R}_1 \rangle & \langle {\bf R}_1, {\bf R}_2 \rangle & \cdots & \langle {\bf R}_1, {\bf R}_{\rm nDim} \rangle & -1 \\ \langle {\bf R}_2, {\bf R}_1 \rangle & \langle {\bf R}_2, {\bf R}_2 \rangle & \cdots & \langle {\bf R}_2, {\bf R}_{\rm nDim} \rangle & -1 \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ \langle {\bf R}_{\rm nDim}, {\bf R}_1 \rangle & \langle {\bf R}_{\rm nDim}, {\bf R}_2 \rangle & \cdots & \langle {\bf R}_{\rm nDim}, {\bf R}_{\rm nDim} \rangle & -1 \\ -1 & -1 & \cdots & -1 & 0 \end{pmatrix}\]

Returns the dot product of the new residual with itself, $\langle {\bf R}_{\rm ithis}, {\bf R}_{\rm ithis} \rangle$.

weighted_dot(diis::Diis, vecs1, vecs2)

Compute weighted (with diis.weights) dot product of vectors.