matematija / continuous-vmc

Research code for neural-network variational quantum states operating on continuous degrees of freedom..

Variational Monte Carlo with continuous variables

Research code for neural-network variational quantum states operating on continuous degrees of freedom. This code implements the standard Variational Monte Carlo (VMC) algorithm for lattice systems with a focus on real-time evolution.

Code author: Matija Medvidović

Installation

Optional but recommended: If you want GPU support through JAX, install the CUDA-enabled version of JAX. At the time of writing this, this can be done by running

pip install --upgrade pip
# Installs the wheel compatible with CUDA 11 and cuDNN 8.6 or newer.
# Note: wheels only available on linux.
pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

To install the package itself, clone simply the repo and run

pip install ./continuous-vmc

A quick example

import jax
from jax import numpy as jnp

from continuousvmc import RotorCNN, VariationalHMC, StochasticReconfiguration, QuantumRotors

Define the variational wavefunction and initialize its the parameters:

logpsi = RotorCNN(dims=(4,4), kernels=[3,3], K=4, param_dtype=jnp.complex64)

key = jax.random.PRNGKey(42)
params = logpsi.initialize(key)

and the local energy associated with the Hamiltonian:

eloc = QuantumRotors(logpsi, g=6.0, pbc=False, chunk_size=4000)

Local energies can be computed by simply calling eloc or alongside gradients w.r.t the variational parameters params:

dummy_inputs = np.random.randn(100, *logpsi.dims)
value, grad = eloc.value_and_grad(params, dummy_inputs)

Constructing the Hamiltonian Monte Carlo (HMC) sampler is equally simple:

sampler = VariationalHMC(logpsi, n_samples=200, n_chains=50, warmup=600, n_leaps=40, target_acc_rate=0.8)

Samples can be obtained with:

key, = jax.random.split(key, 1)
samples, observables, info = sampler(params, key)

Finally, we can find the ground state of the corresponding Hamiltonian by using one of the optimizers on offer:

init, kernel = StochasticReconfiguration(
    logpsi, eloc, sampler, lr=1e-2, solver='shift', eps=1e-3
)

kernel = jax.jit(kernel)
state = init(params)

One optimization step can then be performed with: state = kernel(state, key).

Alternatively, real-time evolution from the initial state described by params can be performed by defining:

init, integrator = Propagator(
    logpsi,
    eloc,
    sampler,
    solver='svd',
    eps=0.0,
    integrator='rk23',
    integrator_dt=1e-3,
    integrator_atol=1e-3,
    integrator_rtol=1e-2,
    solver_acond=1e-3,
    solver_rcond=1e-2,
)

key = random.PRNGKey(1337)
state = init(params, key=key)

and then performing one step of the integrator with state = integrator(state, key).

More examples can be found in the examples folder.

Main features overview

This code implements the time-independent and time-dependent versions of the VMC. Automatic differentiation and GPU support is provided through JAX.

Roughly, the code is subdivided into three parts:

1. Sampling of variational wavefunctions

We offer a fast and jax.jit-able implementation of:

• Hamiltonian Monte Carlo (HMC)
• Random Walk Metropolis (RWM)

sampling algorithms, allowing for efficient evaluations of quantum averages required to run Quantum Monte Carlo (QMC) algorithms. Our HMC implementation includes more advanced features described in the original No-U-Turn Sampler paper and this overview:

• Mass-matrix/metric tensor adaptation during warmup to eliminate global covariance in the momentum distribution. Support for diagonal and dense mass-matrices.
• Step-size adaptation during warmup to achieve a target acceptance rate using the Nesterov Dual-Averaging algorithm.
• Support for periodic variables by automatic wrapping of the proposed HMC move and using directional statistics to compute mass matrices.
• Leapfrog step-size clipping for numerical stability with near-uniform target distributions.

Some or all of these features can be combined to mimic the robust multi-stage HMC chain warm-up scheme of Stan.

2. Dense and matrix-free linear solvers for Quantum natural gradient optimization and regularization

Quantum averages computed using HMC are used to compute the Quantum Geometric Tensor (QGT) to construct the Quantum Natural Gradient. We offer a fast and jax.jit-able implementation of the following solvers for inverting the resulting linear system. Some options include:

• shift: Apply a diagonal shift to the QGT and invert using the Cholesky decomposition.
• svd: Solve the corresponding least-squares problem by singular value decomposition (SVD).
• cg: Solve the linear system using conjugate gradient, without ever materializing the full QGT matrix.
• More options available in the documentation.

3. Runge-Kutta integrators for t-VMC time-stepping

We implement many fixed-step and adaptive-step Runge-Kutta integrators in Jax. Options include:

• Fixed-step methods: euler, midpoint, heun, rk4
• Adaptive methods: rk12, rk12_fehlberg, rk23, rk45_fehlberg, rk45_dopri

4. AD tools

We implement a memory-efficient Hessian-diagonal operator using Automatic differentiation (AD), materializing only one row of the full Hessian matrix at a time. The API is similar to JAX:

def f(x):
    return jax.numpy.log(x[0]) * jax.numpy.sin(x[1:]).prod()

x = np.random.rand(10)

grad_and_diag_hess_fn = grad_and_diag_hess(f)
# Any Jax-differentiable scalar function

grad, diag_hess = grad_and_diag_hess_fn(x)

Related code libraries

• Overlapping VMC implementation details in this code are inspired by NetKet.
• Neural-network and parameter management is done using Flax.
• Overlapping Hamiltonian Monte Carlo implementation details have been inspired by BlackJax and TensorFlow Probability.