Overcoming critical slowing down in frustrated spin systems by learned multiscale sampling

Technical Summary: Overcoming Critical Slowing Down in Frustrated Spin Systems by Learned Multiscale Sampling

Problem Statement
Efficiently sampling high-dimensional probability distributions is a central challenge in computational statistical physics, particularly near continuous phase transitions where correlation lengths diverge. Standard local Markov chain Monte Carlo (MCMC) methods suffer from "critical slowing down," where equilibration and decorrelation times scale as $\tau \sim L^z$ (with $z \approx 2$ for local updates in 2D systems). While cluster algorithms (e.g., Swendsen–Wang, Wolff) successfully mitigate this in unfrustrated systems by updating correlated degrees of freedom collectively, they fundamentally fail in the presence of frustration, even when weak. The constructive identification of efficient clusters in frustrated systems is impossible via standard procedures, necessitating a new approach to sampling.

Methodology: Wavelet Conditional Renormalization Group (WCRG)
The authors propose a "learned" multiscale sampling strategy based on the Wavelet Conditional Renormalization Group (WCRG), which combines machine learning with renormalization group (RG) ideas.

1. Multiscale Decomposition: The method utilizes an orthogonal wavelet transform to decompose a microscopic field configuration $\phi_0$ into a hierarchy of coarse-grained fields ($\phi_j$) and wavelet coefficients ($\bar{\phi}_j$) representing fluctuations at progressively finer scales. This transformation is invertible and has a unit Jacobian.
2. Factorization of Probability: The exact microscopic probability distribution $p_0(\phi_0)$ is factorized into a product of conditional probabilities: the distribution of the coarsest field and the conditional distributions of wavelet variables given the coarser fields at each scale.
   $$p_0(\phi_0) = p_J(\phi_J) \prod_{j=1}^J p_j(\bar{\phi}_j | \phi_j)$$
3. Learning Conditional Distributions: Instead of constructing clusters, the conditional distributions $p_j(\bar{\phi}_j | \phi_j)$ are learned from equilibrium Monte Carlo data. The authors employ score matching to fit unnormalized probability models by matching gradients of log-probabilities, thereby avoiding the direct calculation of normalization constants.
   • The conditional energies $\bar{E}_j$ are approximated using finite-dimensional parametric ansätze (linear in parameters, containing translationally invariant quadratic couplings and local potentials).
   • Different wavelet families are selected based on the phase: Haar wavelets for homogeneous phases (FM, PM) and the Ising critical line; Daubechies (DB4) wavelets for modulated phases (MPM, IC, AF) to better capture finite-wavevector structures.
4. Recursive Sampling: Once trained, new configurations are generated recursively from coarse to fine scales. The coarsest field is sampled first, followed by conditional sampling of wavelet variables at each subsequent scale using a local Metropolis algorithm. The final microscopic configuration is reconstructed via the inverse wavelet transform.

Key Contributions and Results
The study focuses on the soft-spin biaxial next-nearest-neighbor Ising (sBNNNI) model on a 2D square lattice, which exhibits a complex phase diagram including ferromagnetic (FM), paramagnetic (PM), modulated paramagnetic (MPM), incommensurate (IC), and antiphase (AF) regimes, as well as an Ising-like critical line.

1. Statistical Fidelity:

   • The WCRG-generated configurations accurately reproduce local field distributions $P(\phi_0(i))$ and the diagonal structure factor $S_{\text{diag}}(q)$ across the FM, PM, MPM, and IC phases, as well as the critical line.
   • The method successfully captures both homogeneous and finite-wavevector modulated structures.
   • Limitation: The antiphase (AF) regime, characterized by a rigid commensurate modulation tied to the microscopic lattice, remains a challenging case. The coarse-graining process averages out the checkerboard pattern, concentrating relevant information in the finest wavelet degrees of freedom, which hinders the multiscale reconstruction.

2. Elimination of Critical Slowing Down:

   • Local Dynamics: Standard local Monte Carlo updates at the critical point exhibit critical slowing down with a dynamical exponent $z_{\text{loc}} \approx 2.11$, scaling as $\tau \sim L^{2.11}$.
   • WCRG Dynamics: The conditional wavelet updates remain decorrelated within $O(1)$ sweeps at each scale $j$. Since the number of reconstruction levels scales as $J = \log_2 L$, the total number of conditional sweeps required to generate an independent configuration scales as $O(\log_2 L)$.
   • Complexity: While the total number of elementary local update attempts scales as $O(L^2)$ (linear in the number of degrees of freedom), the method avoids the additional critical factor $L^z$ characteristic of local sampling. This represents a significant efficiency gain over local MCMC and circumvents the failure of cluster algorithms in frustrated systems.

3. Configuration-Wide Observables and Trade-offs:

   • Magnetization: The distribution of magnetization $P(m)$ is accurately reproduced. This is largely a consequence of the reconstruction scheme: the coarsest field determines the total magnetization, and wavelet modes are constructed to have zero spatial average.
   • Energy: The histogram of the microscopic energy $E_0$ shows a residual mismatch between original and synthetic samples.
   • Source of Discrepancy: The authors attribute this to the finite-dimensional ansatz used for the conditional energies. The learned conditional energies are approximations, leading to a "synthetic hierarchical measure" that differs from the exact Gibbs measure. The expressiveness of the model limits the accuracy of the energy distribution, even with infinite training data.

Significance and Claims
The paper demonstrates that learned multiscale sampling can overcome critical slowing down in frustrated systems where conventional cluster algorithms fail. The WCRG method offers a fast generative procedure that bypasses the slow propagation of information inherent in local dynamics.

The authors characterize the method's significance through a specific trade-off:

• Standard MCMC: Asymptotically exact but suffers from critical slowing down.
• WCRG: Provides fast sampling ($O(\log L)$ decorrelation depth) but samples from an approximate recursive measure. The accuracy of this approximation is controlled by the expressiveness of the parametrized energy models.

The work positions WCRG as a complementary strategy to Monte Carlo, particularly valuable for frustrated systems. It suggests that approximation errors can be systematically associated with specific reconstruction levels and improved by refining the ansatz (e.g., adding higher-order operators or using convolutional neural networks), though such improvements are left for future work. The paper does not claim to solve the sampling problem for all frustrated systems universally but establishes a proof of principle for soft-spin models with competing interactions.