#
# @BEGIN LICENSE
#
# Psi4: an open-source quantum chemistry software package
#
# Copyright (c) 2007-2022 The Psi4 Developers.
#
# The copyrights for code used from other parties are included in
# the corresponding files.
#
# This file is part of Psi4.
#
# Psi4 is free software; you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation, version 3.
#
# Psi4 is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License along
# with Psi4; if not, write to the Free Software Foundation, Inc.,
# 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
#
# @END LICENSE
#
__all__ = [
"cg_solver",
"davidson_solver",
"DIIS",
"hamiltonian_solver",
"SolverEngine",
]
import time
from abc import ABC, abstractmethod
from typing import Any, Callable, Dict, List, Optional, Type
import numpy as np
from psi4 import core
from .exceptions import ValidationError
"""
Generalized iterative solvers for Psi4.
"""
[docs]def cg_solver(
rhs_vec: List[core.Matrix],
hx_function: Callable,
preconditioner: Callable,
guess: Optional[List[core.Matrix]] = None,
printer: Optional[Callable] = None,
printlvl: int = 1,
maxiter: int = 20,
rcond: float = 1.e-6) -> List[core.Matrix]:
"""
Solves the :math:`Ax = b` linear equations via Conjugate Gradient. The `A` matrix must be a hermitian, positive definite matrix.
Parameters
----------
rhs_vec
The RHS vector in the Ax=b equation.
hx_function
Takes in a list of :py:class:`~psi4.core.Matrix` objects and a mask of active indices. Returns the Hessian-vector product.
preconditioner
Takes in a list of :py:class:`~psi4.core.Matrix` objects and a mask of active indices. Returns the preconditioned value.
guess
Starting vectors. If None, use a preconditioner (rhs) guess
printer
Takes in a list of current x and residual vectors and provides a print function. This function can also
return a value that represents the current residual.
printlvl
The level of printing provided by this function.
maxiter
The maximum number of iterations this function will take.
rcond
The residual norm for convergence.
Returns
-------
ret : List[Matrix]
Solved `x` vectors and `r` vectors.
Notes
-----
This is a generalized cg solver that can also take advantage of solving multiple RHS's simultaneously when
it is advantageous to do so.
"""
tstart = time.time()
if printlvl:
core.print_out("\n -----------------------------------------------------\n")
core.print_out(" " + "Generalized CG Solver".center(52) + "\n")
core.print_out(" " + "by Daniel. G. A. Smith".center(52) + "\n")
core.print_out(" -----------------------------------------------------\n")
core.print_out(" Maxiter = %11d\n" % maxiter)
core.print_out(" Convergence = %11.3E\n" % rcond)
core.print_out(" Number of equations = %11ld\n\n" % len(rhs_vec))
core.print_out(" %4s %14s %12s %6s %6s\n" % ("Iter", "Residual RMS", "Max RMS", "Remain", "Time [s]"))
core.print_out(" -----------------------------------------------------\n")
nrhs = len(rhs_vec)
active_mask = [True for x in range(nrhs)]
# Start function
if guess is None:
x_vec = preconditioner(rhs_vec, active_mask)
else:
if len(guess) != len(rhs_vec):
raise ValidationError("CG Solver: Guess vector length does not match RHS vector length.")
x_vec = [x.clone() for x in guess]
Ax_vec = hx_function(x_vec, active_mask)
# Set it up
r_vec = [] # Residual vectors
for x in range(nrhs):
tmp_r = rhs_vec[x].clone()
tmp_r.axpy(-1.0, Ax_vec[x])
r_vec.append(tmp_r)
z_vec = preconditioner(r_vec, active_mask)
p_vec = [x.clone() for x in z_vec]
# First RMS
grad_dot = [x.sum_of_squares() for x in rhs_vec]
resid = [(r_vec[x].sum_of_squares() / grad_dot[x])**0.5 for x in range(nrhs)]
if printer:
resid = printer(0, x_vec, r_vec)
elif printlvl:
# core.print_out(' CG Iteration Guess: Rel. RMS = %1.5e\n' % np.mean(resid))
core.print_out(" %5s %14.3e %12.3e %7d %9d\n" %
("Guess", np.mean(resid), np.max(resid), len(z_vec), time.time() - tstart))
rms = np.mean(resid)
rz_old = [0.0 for x in range(nrhs)]
alpha = [0.0 for x in range(nrhs)]
active = np.where(active_mask)[0]
# CG iterations
for rot_iter in range(maxiter):
# Build old RZ so we can discard vectors
for x in active:
rz_old[x] = r_vec[x].vector_dot(z_vec[x])
# Build Hx product
Ap_vec = hx_function(p_vec, active_mask)
# Update x and r
for x in active:
alpha[x] = rz_old[x] / Ap_vec[x].vector_dot(p_vec[x])
if np.isnan(alpha)[0]:
core.print_out("CG: Alpha is NaN for vector %d. Stopping vector." % x)
active_mask[x] = False
continue
x_vec[x].axpy(alpha[x], p_vec[x])
r_vec[x].axpy(-alpha[x], Ap_vec[x])
resid[x] = (r_vec[x].sum_of_squares() / grad_dot[x])**0.5
# Print out or compute the resid function
if printer:
resid = printer(rot_iter + 1, x_vec, r_vec)
# Figure out active updated active mask
for x in active:
if (resid[x] < rcond):
active_mask[x] = False
# Print out if requested
if printlvl:
core.print_out(" %5d %14.3e %12.3e %7d %9d\n" %
(rot_iter + 1, np.mean(resid), np.max(resid), sum(active_mask), time.time() - tstart))
active = np.where(active_mask)[0]
if sum(active_mask) == 0:
break
# Update p
z_vec = preconditioner(r_vec, active_mask)
for x in active:
beta = r_vec[x].vector_dot(z_vec[x]) / rz_old[x]
p_vec[x].scale(beta)
p_vec[x].axpy(1.0, z_vec[x])
if printlvl:
core.print_out(" -----------------------------------------------------\n")
return x_vec, r_vec
[docs]class DIIS:
"""
An object to assist in the DIIS extrpolation procedure.
Parameters
----------
max_vec
The maximum number of error and state vectors to hold. These are pruned based off the removal policy.
removal_policy
{"OLDEST", "LARGEST"}
How the state and error vectors are removed once at the maximum. OLDEST will remove the oldest vector while
largest will remove the residual with the largest RMS value.
"""
def __init__(self, max_vec: int = 6, removal_policy: str = "OLDEST"):
self.error = []
self.state = []
self.max_vec = max_vec
self.removal_policy = removal_policy.upper()
if self.removal_policy not in ["LARGEST", "OLDEST"]:
raise ValidationError("DIIS: removal_policy must either be oldest or largest.")
[docs] def add(self, state: core.Matrix, error: core.Matrix):
"""
Adds a DIIS state and error vector to the DIIS object.
Parameters
----------
state
The current state vector.
error
The current error vector.
"""
self.error.append(error.clone())
self.state.append(state.clone())
def _diag_print_heading(title_lines, solver_name, max_ss_size, nroot, r_convergence, maxiter, verbose=1):
"""Print a message to the output file when the solver has processed all options and is ready to begin"""
if verbose < 1:
# no printing
return
# show title if not silent
core.print_out("\n\n")
core.print_out("\n".join([x.center(77) for x in title_lines]))
core.print_out("\n")
core.print_out("\n ==> Options <==\n\n")
core.print_out(f" Max number of iterations = {maxiter:<5d}\n")
core.print_out(f" Eigenvector tolerance = {r_convergence:.4e}\n")
core.print_out(f" Max number of expansion vectors = {max_ss_size:<5d}\n")
core.print_out("\n")
# show iteration info headings if not silent
core.print_out(" => Iterations <=\n")
if verbose == 1:
# default printing one line per iter max delta value and max residual norm
core.print_out(f" {' ' * len(solver_name)} Max[D[value]] Max[|R|] # vectors\n")
else:
# verbose printing, value, delta, and |R| for each root
core.print_out(f" {' ' * len(solver_name)} value D[value] |R| # vectors\n")
def _diag_print_info(solver_name, info, verbose=1):
"""Print a message to the output file at each iteration"""
if verbose < 1:
# no printing
return
elif verbose == 1:
# print iter maxde max|R| conv/restart
flags = []
if info['collapse']:
flags.append("Restart")
if info['done']:
flags.append("Converged")
m_de = np.max(info['delta_val'])
m_r = np.max(info['res_norm'])
nvec = info["nvec"]
flgs = "/".join(flags)
core.print_out(
f" {solver_name} iter {info['count']:3d}: {m_de:-11.5e} {m_r:12.5e} {nvec:>6d} {flgs}\n")
else:
# print iter / ssdim folowed by de/|R| for each root
core.print_out(f" {solver_name} iter {info['count']:3d}: {info['nvec']:4d} guess vectors\n")
for i, (e, de, rn) in enumerate(zip(info['val'], info['delta_val'], info['res_norm'])):
s = " " * len(solver_name)
core.print_out(f" {i+1:2d}: {s:} {e:-11.5f} {de:-11.5e} {rn:12.5e}\n")
if info['done']:
core.print_out(" Solver Converged! all roots\n\n")
elif info['collapse']:
core.print_out(" Subspace limits exceeded restarting\n\n")
def _diag_print_converged(solver_name, stats, vals, verbose=1, **kwargs):
"""Print a message to the output file when the solver is converged."""
if verbose < 1:
# no printing
return
if verbose > 1:
# print values summary + number of iterations + # of "big" product evals
core.print_out(" Root # eigenvalue\n")
for (i, vi) in enumerate(vals):
core.print_out(f" {i+1:^6} {vi:20.12f}\n")
max_nvec = max(istat['nvec'] for istat in stats)
core.print_out(f"\n {solver_name} converged in {stats[-1]['count']} iterations\n")
core.print_out(f" Computed a total of {stats[-1]['product_count']} large products\n\n")
def _print_array(name, arr, verbose):
"""print a subspace quantity (numpy array) to the output file
Parameters
----------
name : str
The name to print above the array
arr : :py:class:`np.ndarray`
The array to print
verbose : int
The amount of information to print. Only prints for verbose > 2
"""
if verbose > 2:
core.print_out(f"\n\n{name}:\n{str(arr)}\n")
def _gs_orth(engine, U, V, thresh: float = 1.0e-8):
"""Perform Gram-Schmidt orthonormalization of a set V against a previously orthonormalized set U
Parameters
----------
engine : object
The engine passed to the solver, required to define vector algebraic operations needed
U : list of `vector`
A set of orthonormal vectors, len(U) = l; satisfies ||I^{lxl}-U^tU|| < thresh
V : list of `vectors`
The vectors used to augment U
thresh
If the orthogonalized vector has a norm smaller than this value it is considered LD to the set
Returns
-------
U_aug : list of `vector`
The orthonormal set of vectors U' with span(U') = span(U) + span(V), len(U) <= len(U_aug) <= len(U) + len(V)
"""
for vi in V:
for j in range(len(U)):
dij = engine.vector_dot(vi, U[j])
Vi = engine.vector_axpy(-1.0 * dij, U[j], vi)
norm_vi = np.sqrt(engine.vector_dot(vi, vi))
if norm_vi >= thresh:
U.append(engine.vector_scale(1.0 / norm_vi, vi))
return U
def _best_vectors(engine, ss_vectors: np.ndarray, basis_vectors: List) -> List:
r"""Compute the best approximation of the true eigenvectors as a linear combination of basis vectors:
..math:: V_{k} = \Sum_{i} \tilde{V}_{i,k}X_{i}
Where :math:`\tilde{V}` is the matrix with columns that are eigenvectors of the subspace matrix. And
:math:`X_{i}` is a basis vector.
Parameters
----------
engine : object
The engine passed to the solver, required to define vector algebraic operations needed
ss_vectors
Numpy array {l, k}.
The k eigenvectors of the subspace problem, l = dimension of the subspace basis, and k is the number of roots
basis_vectors
list of `vector` {l}.
The current basis vectors
Returns
-------
new_vecs
list of `vector` {k}.
The approximations of the k true eigenvectors.
"""
l, n = ss_vectors.shape
new_vecs = []
for i in range(n):
cv_i = engine.new_vector()
for j in range(l):
cv_i = engine.vector_axpy(ss_vectors[j, i], basis_vectors[j], cv_i)
new_vecs.append(cv_i)
return new_vecs
[docs]class SolverEngine(ABC):
"""Abstract Base Class defining the API for a matrix-vector product object
required by solvers.
Engines implement the correct product functions for iterative solvers that
do not require the target matrix be stored directly.
Classes intended to be used as an `engine` for :func:`davidson_solver` or
:func:`hamiltonian_solver` should inherit from this base class to ensure
that the required methods are defined.
.. note:: The `vector` referred to here is intentionally vague, the solver
does not care what it is and only holds individual or sets of
them. In fact an individual `vector` could be split across two
elements in a list, such as for different spin.
Whatever data type is used and individual vector should be a
single element in a list such that len(list) returns the number
of vector-like objects.
"""
[docs] @abstractmethod
def compute_products(self, X):
r"""Compute a Matrix * trial vector products
Parameters
----------
X : List[`vector`]
Trial vectors.
Returns
-------
Expected by :func:`davidson_solver`
AX : List[`vector`]
The product :math:`A x X_{i}` for each `X_{i}` in `X`, in that
order. Where `A` is the hermitian matrix to be diagonalized.
`len(AX) == len(X)`
n : int
The number of products that were evaluated. If the object implements
product caching this may be less than len(X)
Expected by :func:`hamiltonian_solver`
H1X : List[`vector`]
The product :math:`H1 x X_{i}` for each `X_{i}` in `X`, in that
order. Where H1 is described in :func:`hamiltonian_solver`.
``len(H1X) == len(X)``
H2X : List[`vector`]
The product :math:`H2 x X_{i}` for each `X_{i}` in `X`, in that
order. Where H2 is described in :func:`hamiltonian_solver`.
``len(H2X) == len(X)``
"""
pass
[docs] @abstractmethod
def precondition(self, R_k, w_k):
r"""Apply the preconditioner to a Residual vector
The preconditioner is usually defined as :math:`(w_k - D_{i})^-1` where
`D` is an approximation of the diagonal of the matrix that is being
diagonalized.
Parameters
----------
R_k : single `vector`
The residual vector
w_k : float
The eigenvalue associated with this vector
Returns
-------
new_X_k : single `vector`
The preconditioned residual vector, a correction vector that will be
used to augment the guess space
"""
pass
[docs] @abstractmethod
def new_vector(self):
"""Return a new `vector` object.
The solver is oblivious to the data structure used for a `vector` this
method provides the engine with a means to create `vector` like
quantities.
The engine calls this method with no arguments. So any defined by the
engine for its own use should be optional
Returns
-------
X : singlet `vector`
This should be a new vector object with the correct dimensions,
assumed to be zeroed out
"""
pass
[docs] def vector_dot(X, Y) -> float:
"""Compute a dot product between two `vectors`
Parameters
----------
X : single `vector`
Y : single `vector`
Returns
-------
a : float
The dot product (X x Y)
"""
pass
# cython doesn't like static+ decorators https://github.com/cython/cython/issues/1434#issuecomment-608975116
vector_dot = staticmethod(abstractmethod(vector_dot))
[docs] @abstractmethod
def vector_axpy(a: float, X, Y):
"""Compute scaled `vector` addition operation `a*X + Y`
Parameters
----------
a
The scale factor applied to `X`
X : singlet `vector`
The `vector` which will be scaled and added to `Y`
Y : single `vector`
The `vector` which the result of `a*X` is added to
Returns
-------
Y : single `vector`
The solver assumes that Y is updated, and returned. So it is safe to
avoid a copy of Y if possible
"""
pass
[docs] @abstractmethod
def vector_scale(a: float, X):
"""Scale a vector by some factor
Parameters
----------
a
The scale facor
X : single `vector`
The vector that will be scaled
Returns
-------
X : single `vector`
The solver assumes that the passed vector is modifed. So it is save
to avoid a copy of X if possible.
"""
pass
[docs] @abstractmethod
def vector_copy(X):
"""Make a copy of a `vector`
Parameters
----------
X : single `vector`
The `vector` to copy
Returns
-------
X' : single `vector`
A copy of `X` should be distinct object that can be modified
independently of the passed object, Has the same data when returned.
"""
pass
[docs] @abstractmethod
def residue(self, X, so_prop_ints):
"""Compute residue
Parameters
----------
X
The single `vector` to use to compute the property.
so_prop_ints :
Property integrals in SO basis for the desired transition property.
prefactor
Optional float scaling factor.
Returns
-------
residue : Any
The transition property.
"""
pass
[docs]def davidson_solver(
engine: Type[SolverEngine],
guess: List,
*,
nroot: int,
r_convergence: float = 1.0E-4,
max_ss_size: int = 100,
maxiter: int = 60,
verbose: int = 1,
nonneg_only: bool = False) -> Dict[str, Any]:
"""Solves for the lowest few eigenvalues and eigenvectors of a large problem emulated through an engine.
If the large matrix `A` has dimension `{NxN}` and N is very large, and only
a small number of roots, `k` are desired this algorithm is preferable to
standard methods as uses on the order of `N * k` memory. One only needs to
have the ability to compute the product of a times a vector.
For non-hermitan `A` the basis of the algorithm breaks down. However in
practice, for strongly diagonally-dominant `A` such as the
similarity-transformed Hamiltonian in EOM-CC this algorithm is commonly still
used.
Parameters
----------
engine
The engine drive all operations involving data structures that have at
least one "large" dimension. See :class:`SolverEngine` for requirements
guess
list {engine dependent}
At least `nroot` initial expansion vectors
nroot
Number of roots desired
r_convergence
Convergence tolerance for residual vectors
max_ss_size
The maximum number of trial vectors in the iterative subspace that will
be stored before a collapse is done.
maxiter
The maximum number of iterations
verbose
The amount of logging info to print (0 -> none, 1 -> some, >1 -> everything)
nonneg_only
Should eigenpairs with eigenvalue < 0 be ignored?
Returns
-------
best_values : numpy.ndarray
(nroots, ) The best approximation of the eigenvalues of A, computed on the last iteration of the solver
best_vectors: List[`vector`]
(nroots) The best approximation of the eigenvectors of A, computed on the last iteration of the solver
stats : List[Dict]
Statistics collected on each iteration
- count : int, iteration number
- res_norm : np.ndarray (nroots, ), the norm of residual vector for each roots
- val : np.ndarray (nroots, ), the eigenvalue corresponding to each root
- delta_val : np.ndarray (nroots, ), the change in eigenvalue from the last iteration to this ones
- collapse : bool, if a subspace collapse was performed
- product_count : int, the running total of product evaluations that was performed
- done : bool, if all roots were converged
Notes
-----
The solution vector is normalized to 1/2
The solver will return even when `maxiter` iterations are performed without convergence.
The caller **must check** ``stats[-1]['done']`` for failure and handle each case accordingly.
"""
nk = nroot
iter_info = {
"count": 0,
"res_norm": np.zeros((nk)),
"val": np.zeros((nk)),
"delta_val": np.zeros((nk)),
# conv defaults to true, and will be flipped when a non-conv root is hit
"done": True,
"nvec": 0,
"collapse": False,
"product_count": 0,
}
print_name = "DavidsonSolver"
title_lines = ["Generalized Davidson Solver", "By Ruhee Dcunha"]
_diag_print_heading(title_lines, print_name, max_ss_size, nroot, r_convergence, maxiter, verbose)
vecs = guess
stats = []
best_eigvecs = []
best_eigvals = []
while iter_info['count'] < maxiter:
# increment iteration/ save old vals
iter_info['count'] += 1
old_vals = iter_info['val'].copy()
# reset flags
iter_info['collapse'] = False
iter_info['done'] = True
# get subspace dimension
l = len(vecs)
iter_info['nvec'] = l
# check if ss dimension has exceeded limits
if l >= max_ss_size:
iter_info['collapse'] = True
# compute A times trial vector products
Ax, nprod = engine.compute_products(vecs)
iter_info['product_count'] += nprod
# Build Subspace matrix
G = np.zeros((l, l))
for i in range(l):
for j in range(i):
G[i, j] = G[j, i] = engine.vector_dot(vecs[i], Ax[j])
G[i, i] = engine.vector_dot(vecs[i], Ax[i])
_print_array("SS transformed A", G, verbose)
# diagonalize subspace matrix
lam, alpha = np.linalg.eigh(G)
_print_array("SS eigenvectors", alpha, verbose)
_print_array("SS eigenvalues", lam, verbose)
if nonneg_only:
# remove zeros/negatives
alpha = alpha[:, lam > 1.0e-10]
lam = lam[lam > 1.0e-10]
# sort/truncate to nroot
idx = np.argsort(lam)
lam = lam[idx]
alpha = alpha[:, idx]
# update best_solution
best_eigvecs = _best_vectors(engine, alpha[:, :nk], vecs)
best_eigvals = lam[:nk]
# check convergence of each solution
new_vecs = []
for k in range(nk):
# residual vector
Rk = engine.new_vector()
lam_k = lam[k]
for i in range(l):
Axi = Ax[i]
Rk = engine.vector_axpy(alpha[i, k], Axi, Rk)
Rk = engine.vector_axpy(-1.0 * lam_k, best_eigvecs[k], Rk)
iter_info['val'][k] = lam_k
iter_info['delta_val'][k] = abs(old_vals[k] - lam_k)
iter_info['res_norm'][k] = np.sqrt((engine.vector_dot(Rk, Rk)))
# augment guess vector for non-converged roots
if (iter_info["res_norm"][k] > r_convergence):
iter_info['done'] = False
Qk = engine.precondition(Rk, lam_k)
new_vecs.append(Qk)
# print iteration info to output
_diag_print_info(print_name, iter_info, verbose)
# save stats for this iteration
stats.append(iter_info.copy())
if iter_info['done']:
# finished
_diag_print_converged(print_name, stats, best_eigvals, verbose)
break
elif iter_info['collapse']:
# restart needed
vecs = best_eigvecs
else:
# Regular subspace update, orthonormalize preconditioned residuals and add to the trial set
vecs = _gs_orth(engine, vecs, new_vecs)
# always return, the caller should check ret["stats"][-1]['done'] == True for convergence
return {"eigvals": best_eigvals, "eigvecs": list(zip(best_eigvecs, best_eigvecs)), "stats": stats}
[docs]def hamiltonian_solver(
engine: Type[SolverEngine],
guess: List,
*,
nroot: int,
r_convergence: float = 1.0E-4,
max_ss_size: int = 100,
maxiter: int = 60,
verbose: int = 1):
"""Finds the smallest eigenvalues and associated right and left hand
eigenvectors of a large real Hamiltonian eigenvalue problem emulated
through an engine.
A Hamiltonian eigenvalue problem (EVP) has the following structure::
[A B][X] = [1 0](w)[X]
[B A][Y] [0 -1](w)[Y]
with A, B of some large dimension N, the problem is of dimension 2Nx2N.
The real, Hamiltonian EVP can be rewritten as the NxN, non-hermitian EVP:
:math:`(A+B)(A-B)(X+Y) = w^2(X+Y)`
With left-hand eigenvectors:
:math:`(X-Y)(A-B)(A+B) = w^2(X-Y)`
if :math:`(A-B)` is positive definite, we can transform the problem to arrive at the hermitian NxN EVP:
:math:`(A-B)^{1/2}(A+B)(A-B)^{1/2} = w^2 T`
Where :math:`T = (A-B)^{-1/2}(X+Y)`.
We use a Davidson like iteration where we transform :math:`(A+B)` (H1) and :math:`(A-B)`
(H2) in to the subspace defined by the trial vectors.
The subspace analog of the NxN hermitian EVP is diagonalized and left :math:`(X-Y)`
and right :math:`(X+Y)` eigenvectors of the NxN non-hermitian EVP are approximated.
Residual vectors are formed for both and the guess space is augmented with
two correction vectors per iteration. The advantages and properties of this
algorithm are described in the literature [stratmann:1998]_ .
Parameters
----------
engine
The engine drive all operations involving data structures that have at
least one "large" dimension. See :class:`SolverEngine` for requirements
guess
list {engine dependent}
At least `nroot` initial expansion vectors
nroot
Number of roots desired
r_convergence
Convergence tolerance for residual vectors
max_ss_size
The maximum number of trial vectors in the iterative subspace that will
be stored before a collapse is done.
maxiter
The maximum number of iterations
verbose
The amount of logging info to print (0 -> none, 1 -> some, 2 -> all but matrices, >2 -> everything)
Returns
-------
best_values : numpy.ndarray
(nroots, ) The best approximation of the eigenvalues of `w`, computed on the last iteration of the solver
best_R: List[`vector`]
(nroots) The best approximation of the right hand eigenvectors, :math:`X+Y`, computed on the last iteration of the solver.
best_L: List[`vector`]
(nroots) The best approximation of the left hand eigenvectors, :math:`X-Y`, computed on the last iteration of the solver.
stats : List[Dict]
Statistics collected on each iteration
- count : int, iteration number
- res_norm : np.ndarray (nroots, ), the norm of residual vector for each roots
- val : np.ndarray (nroots, ), the eigenvalue corresponding to each root
- delta_val : np.ndarray (nroots, ), the change in eigenvalue from the last iteration to this ones
- collapse : bool, if a subspace collapse was performed
- product_count : int, the running total of product evaluations that was performed
- done : bool, if all roots were converged
Notes
-----
The solution vector is normalized to 1/2
The solver will return even when `maxiter` iterations are performed without convergence.
The caller **must check** ``stats[-1]['done']`` for failure and handle each case accordingly.
References
----------
R. Eric Stratmann, G. E. Scuseria, and M. J. Frisch, "An efficient
implementation of time-dependent density-functional theory for the
calculation of excitation energies of large molecules." J. Chem. Phys.,
109, 8218 (1998)
"""
nk = nroot
iter_info = {
"count": 0,
"res_norm": np.zeros((nk)),
"val": np.zeros((nk)),
"delta_val": np.zeros((nk)),
# conv defaults to true, and will be flipped when a non-conv root is hit
"conv": True,
"nvec": 0,
"product_count": 0,
}
print_name = "HamiltonianSolver"
title_lines = ["Generalized Hamiltonian Solver", "By Andrew M. James"]
_diag_print_heading(title_lines, print_name, max_ss_size, nroot, r_convergence, maxiter, verbose)
vecs = guess
best_L = []
best_R = []
best_vals = []
stats = []
while iter_info['count'] < maxiter:
# increment iteration/ save old vals
iter_info['count'] += 1
old_w = iter_info['val'].copy()
# reset flags
iter_info['collapse'] = False
iter_info['done'] = True
# get subspace dimension
l = len(vecs)
iter_info['nvec'] = l
# check if subspace dimension has exceeded limits
if l >= max_ss_size:
iter_info['collapse'] = True
# compute [A+B]*v (H1x) and [A-B]*v (H2x)
H1x, H2x, nprod = engine.compute_products(vecs)
iter_info['product_count'] += nprod
# form x*H1x (H1_ss) and x*H2x (H2_ss)
H1_ss = np.zeros((l, l))
H2_ss = np.zeros((l, l))
for i in range(l):
for j in range(l):
H1_ss[i, j] = engine.vector_dot(vecs[i], H1x[j])
H2_ss[i, j] = engine.vector_dot(vecs[i], H2x[j])
_print_array("Subspace Transformed (A+B)", H1_ss, verbose)
_print_array("Subspace Transformed (A-B)", H2_ss, verbose)
# Diagonalize H2 in the subspace (eigen-decomposition to compute H2^(1/2))
H2_ss_val, H2_ss_vec = np.linalg.eigh(H2_ss)
_print_array("eigenvalues H2_ss", H2_ss_val, verbose)
_print_array("eigenvectors H2_ss", H2_ss_vec, verbose)
# Check H2 is PD
# NOTE: If this triggers failure the SCF solution is not stable. A few ways to handle this
# 1. Use davidson solver where product function evaluates (H2 * (H1 * X))
# - Poor convergence
# 2. Switch to CIS/TDA
# - User would probably not expect this
# 3. Perform Stability update and restart with new reference
if np.any(H2_ss_val < 0.0):
msg = ("The H2 matrix is not Positive Definite. " "This means the reference state is not stable.")
raise RuntimeError(msg)
# Build H2^(1/2)
H2_ss_half = np.einsum("ik,k,jk->ij", H2_ss_vec, np.sqrt(H2_ss_val), H2_ss_vec, optimize=True)
_print_array("SS Transformed (A-B)^(1/2)", H2_ss_half, verbose)
# Build Hermitian SS product (H2)^(1/2)(H1)(H2)^(1/2)
Hss = np.einsum('ij,jk,km->im', H2_ss_half, H1_ss, H2_ss_half, optimize=True)
_print_array("(H2)^(1/2)(H1)(H2)^(1/2)", Hss, verbose)
#diagonalize Hss -> w^2, Tss
w2, Tss = np.linalg.eigh(Hss)
_print_array("Eigenvalues (A-B)^(1/2)(A+B)(A-B)^(1/2)", w2, verbose)
_print_array("Eigvectors (A-B)^(1/2)(A+B)(A-B)^(1/2)", Tss, verbose)
# pick positive roots
Tss = Tss[:, w2 > 1.0e-10]
w2 = w2[w2 > 1.0e-10]
# check for invalid eigvals
with np.errstate(invalid='raise'):
w = np.sqrt(w2)
# sort roots
idx = w.argsort()[:nk]
Tss = Tss[:, idx]
w = w[idx]
# Extract Rss = H2^{1/2}Tss
Rss = np.dot(H2_ss_half, Tss)
# Extract Lss = (H1 R)/ w
Lss = np.dot(H1_ss, Rss).dot(np.diag(1.0 / w))
# Biorthonormalize R/L solution vectors
inners = np.einsum("ix,ix->x", Rss, Lss, optimize=True)
Rss = np.einsum("x,ix->ix", 1. / np.sqrt(inners), Rss, optimize=True)
Lss = np.einsum("x,ix->ix", 1. / np.sqrt(inners), Lss, optimize=True)
# Save best R/L vectors and eigenvalues
best_R = _best_vectors(engine, Rss[:, :nk], vecs)
best_L = _best_vectors(engine, Lss[:, :nk], vecs)
best_vals = w[:nk]
# check convergence of each solution
new_vecs = []
for k in range(nk):
# residual vectors for right and left eigenvectors
WR_k = engine.new_vector()
WL_k = engine.new_vector()
wk = w[k]
for i in range(l):
H1x_i = H1x[i]
H2x_i = H2x[i]
WL_k = engine.vector_axpy(Rss[i, k], H1x_i, WL_k)
WR_k = engine.vector_axpy(Lss[i, k], H2x_i, WR_k)
WL_k = engine.vector_axpy(-1.0 * wk, best_L[k], WL_k)
WR_k = engine.vector_axpy(-1.0 * wk, best_R[k], WR_k)
norm_R = np.sqrt(engine.vector_dot(WR_k, WR_k))
norm_L = np.sqrt(engine.vector_dot(WL_k, WL_k))
norm = norm_R + norm_L
iter_info['res_norm'][k] = norm
iter_info['delta_val'][k] = np.abs(old_w[k] - w[k])
iter_info['val'][k] = w[k]
# augment the guess space for non-converged roots
if (iter_info['res_norm'][k] > r_convergence):
iter_info['done'] = False
new_vecs.append(engine.precondition(WR_k, w[k]))
new_vecs.append(engine.precondition(WL_k, w[k]))
# print iteration info to output
_diag_print_info(print_name, iter_info, verbose)
# save stats for this iteration
stats.append(iter_info.copy())
if iter_info['done']:
# Finished
_diag_print_converged(print_name, stats, w[:nk], rvec=best_R, lvec=best_L, verbose=verbose)
break
elif iter_info['collapse']:
# need to orthonormalize union of the Left/Right solutions on restart
vecs = _gs_orth(engine, [], best_R + best_L)
else:
# Regular subspace update, orthonormalize preconditioned residuals and add to the trial set
vecs = _gs_orth(engine, vecs, new_vecs)
# always return, the caller should check ret["stats"][-1]['done'] == True for convergence
return {"eigvals": best_vals, "eigvecs": list(zip(best_R, best_L)), "stats": stats}
