-
Landscape modification meets spin systems: from torpid to rapid mixing, tunneling and annealing in the low-temperature regime
Authors:
Michael C. H. Choi
Abstract:
Given a target Gibbs distribution $π^0_β \propto e^{-β\mathcal{H}}$ to sample from in the low-temperature regime on $Σ_N := \{-1,+1\}^N$, in this paper we propose and analyze Metropolis dynamics that instead target an alternative distribution $π^{f}_{α,c,1/β} \propto e^{-\mathcal{H}^{f}_{α,c,1/β}}$, where $\mathcal{H}^{f}_{α,c,1/β}$ is a transformed Hamiltonian whose landscape is suitably modified…
▽ More
Given a target Gibbs distribution $π^0_β \propto e^{-β\mathcal{H}}$ to sample from in the low-temperature regime on $Σ_N := \{-1,+1\}^N$, in this paper we propose and analyze Metropolis dynamics that instead target an alternative distribution $π^{f}_{α,c,1/β} \propto e^{-\mathcal{H}^{f}_{α,c,1/β}}$, where $\mathcal{H}^{f}_{α,c,1/β}$ is a transformed Hamiltonian whose landscape is suitably modified and controlled by the parameters $f,α,c$ and $β$ and shares the same set of stationary points as $\mathcal{H}$. With appropriate tuning of these parameters, the major advantage of the proposed Metropolis dynamics on the modified landscape is that it enjoys an $\mathcal{O}(1)$ critical height while its stationary distribution $π^{f}_{α,c,1/β}$ maintains close proximity with the original target $π^0_β$ in the low-temperature. We prove rapid mixing and tunneling on the modified landscape with polynomial dependence on the system size $N$ and the inverse temperature $β$, while the original Metropolis dynamics mixes torpidly with exponential dependence on the critical height and $β$. In the setting of simulated annealing, we prove its long-time convergence under a power-law cooling schedule that is faster than the typical logarithmic cooling in the classical setup.
We illustrate our results on a host of models including the Ising model on various deterministic and random graphs as well as Derrida's Random Energy Model. In these applications, the original dynamics mixes torpidly while the proposed dynamics on the modified landscape mixes rapidly with polynomial dependence on both $β$ and $N$ and find the approximate ground state provably in $\mathcal{O}(N^4)$ time. This paper highlights a novel use of the geometry and structure of the landscape to the design of accelerated samplers or optimizers.
△ Less
Submitted 22 August, 2022;
originally announced August 2022.
-
Improved annealing for sampling from multimodal distributions via landscape modification
Authors:
Michael C. H. Choi,
Jing Zhang
Abstract:
Given a target distribution $μ\propto e^{-\mathcal{H}}$ to sample from with Hamiltonian $\mathcal{H}$, in this paper we propose and analyze new Metropolis-Hastings sampling algorithms that target an alternative distribution $μ^f_{1,α,c} \propto e^{-\mathcal{H}^{f}_{1,α,c}}$, where $\mathcal{H}^{f}_{1,α,c}$ is a landscape-modified Hamiltonian which we introduce explicitly. The advantage of the Metr…
▽ More
Given a target distribution $μ\propto e^{-\mathcal{H}}$ to sample from with Hamiltonian $\mathcal{H}$, in this paper we propose and analyze new Metropolis-Hastings sampling algorithms that target an alternative distribution $μ^f_{1,α,c} \propto e^{-\mathcal{H}^{f}_{1,α,c}}$, where $\mathcal{H}^{f}_{1,α,c}$ is a landscape-modified Hamiltonian which we introduce explicitly. The advantage of the Metropolis dynamics which targets $π^f_{1,α,c}$ is that it enjoys reduced critical height described by the threshold parameter $c$, function $f$, and a penalty parameter $α\geq 0$ that controls the state-dependent effect.
First, we investigate the case of fixed $α$ and propose a self-normalized estimator that corrects for the bias of sampling and prove asymptotic convergence results and Chernoff-type bound of the proposed estimator.
Next, we consider the case of annealing the penalty parameter $α$. We prove strong ergodicity and bounds on the total variation mixing time of the resulting non-homogeneous chain subject to appropriate assumptions on the decay of $α$.
We illustrate the proposed algorithms by comparing their mixing times with the original Metropolis dynamics on statistical physics models including the ferromagnetic Ising model on the hypercube or the complete graph and the $q$-state Potts model on the two-dimensional torus. In these cases, the mixing times of the classical Glauber dynamics are at least exponential in the system size as the critical height grows at least linearly with the size, while the proposed annealing algorithm, with appropriate choice of $f$, $c$, and annealing schedule on $α$, mixes rapidly with at most polynomial dependence on the size. The crux of the proof harnesses on the important observation that the reduced critical height can be bounded independently of the size that gives rise to rapid mixing.
△ Less
Submitted 29 November, 2021; v1 submitted 4 November, 2021;
originally announced November 2021.
-
Improved Metropolis-Hastings algorithms via landscape modifcation with applications to simulated annealing and the Curie-Weiss model
Authors:
Michael C. H. Choi
Abstract:
In this paper, we propose new Metropolis-Hastings and simulated annealing algorithms on finite state space via modifying the energy landscape. The core idea of landscape modification rests on introducing a parameter $c$, in which the landscape is modified once the algorithm is above this threshold parameter to encourage exploration, while the original landscape is utilized when the algorithm is be…
▽ More
In this paper, we propose new Metropolis-Hastings and simulated annealing algorithms on finite state space via modifying the energy landscape. The core idea of landscape modification rests on introducing a parameter $c$, in which the landscape is modified once the algorithm is above this threshold parameter to encourage exploration, while the original landscape is utilized when the algorithm is below the threshold for exploitation purpose. We illustrate the power and benefits of landscape modification by investigating its effect on the classical Curie-Weiss model with Glauber dynamics and external magnetic field in the subcritical regime. This leads to a landscape-modified mean-field equation, and with appropriate choice of $c$ the free energy landscape can be transformed from a double-well into a single-well, while the location of the global minimum is preserved on the modified landscape. Consequently, running algorithms on the modified landscape can improve the convergence to the ground-state in the Curie-Weiss model. In the setting of simulated annealing, we demonstrate that landscape modification can yield improved or even subexponential mean tunneling time between global minima in the low-temperature regime by appropriate choice of $c$, and give convergence guarantee using an improved logarithmic cooling schedule with reduced critical height. We also discuss connections between landscape modification and other acceleration techniques such as Catoni's energy transformation algorithm, preconditioning, importance sampling and quantum annealing. The technique developed in this paper is not only limited to simulated annealing and is broadly applicable to any difference-based discrete optimization algorithm by a change of landscape.
△ Less
Submitted 19 July, 2023; v1 submitted 19 November, 2020;
originally announced November 2020.
-
Nonlinear stability of phase-locked states for the Kuramoto model with finite inertia
Authors:
Young-Pil Choi,
Chulho Choi,
Meesoon Ha,
Seung-Yeal Ha
Abstract:
We discuss the {\it nonlinear stability} of phase-locked states for globally coupled nonlinear oscillators with finite inertia, namely the modified Kuramoto model, in the context of the robust $\ell^{\infty}$-norm. We show that some classes of phase-locked states are orbitally $\ell{\infty}$-stable in the sense that its small perturbation asymptotically leads to only the phase shift of the phase-l…
▽ More
We discuss the {\it nonlinear stability} of phase-locked states for globally coupled nonlinear oscillators with finite inertia, namely the modified Kuramoto model, in the context of the robust $\ell^{\infty}$-norm. We show that some classes of phase-locked states are orbitally $\ell{\infty}$-stable in the sense that its small perturbation asymptotically leads to only the phase shift of the phase-locked state from the original one without changing its fine structures as keeping the same suitable coupling strength among oscillators and the same natural frequencies. The phase shift is uniquely determined by the average of initial phases, the average of initial frequencies, and the strength of inertia. We numerically confirm the stability of the phase-locked state as well as its uniqueness and the phase shift, where various initial conditions are considered. Finally, we argue that some restricted conditions employed in the mathematical proof are not necessary, based on numerical simulation results.
△ Less
Submitted 12 December, 2011;
originally announced December 2011.
