REVIEW
Dynamic critical exponents of Swendsen-Wang and Wolff algorithms by nonequilibrium relaxation
Statistical Mechanics · Submitted 2 Mar 2006 · 19 references
Pith's verdict: Careful, honest energy-relaxation numerics, but the z=0 claim is about the even sector and the Wolff initial-state dependence rests on a questionable time rescaling.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
In the two-dimensional Ising model, the Swendsen-Wang cluster algorithm has zero dynamic critical exponent, so its exponential relaxation time grows only logarithmically with lattice size, while the Wolff algorithm's apparent exponent…
The reading
What carries the argument
The central object is the exponential relaxation time $\tau$, defined through the largest non-unit eigenvalue $\lambda_1$ of the reversible Markov transition matrix by $\lambda_1=\exp(-1/\tau)$, and measured in simulations from the late-time energy decay $E(t)\approx A e^{-t/\tau}+E(\infty)$ after a quench to $T_c$. For a fair comparison of algorithms, the paper rescales Wolff Monte Carlo steps by $t=\sum_{t''} C(t'')/L^d$, counting one unit whenever the total number of flipped spins reaches $L^d$. In one dimension, the Swendsen-Wang dynamics is expressed in link variables whose transition matrix is a direct product of $2\times 2$ matrices, yielding the exact spectrum $\lambda_m=(p/2)^m$ with decay times $\tau/m$.
Extended reading notes
Core claim
The paper establishes that, for the two-dimensional Ising model at $T_c$, the exponential relaxation time of the Swendsen-Wang dynamics grows as $\ln L$ up to $L=8192$ for both a quench from infinite temperature and a quench from the ground state, giving $z_{\rm exp}=0$. The same logarithmic behavior is found for Wolff dynamics started from an ordered state, with $z_{\rm int}\approx 0$ as well. From a disordered start, Wolff dynamics instead shows a power-law-like growth, and after corrections to scaling the fitted exponents are $z_{\rm exp}=1.19(2)$ and $z_{\rm int}=0.29(3)$ up to $L=2048$. The paper also derives exactly for a one-dimensional Ising chain with open boundaries that the Swendsen-Wang transition matrix has eigenvalues $\lambda_m=(p/2)^m$, so the slowest non-equilibrium mode relaxes with $\tau=1/\ln(2/p)$.
Load-bearing premise
The load-bearing premise is that Eq. (17)'s rescaled time and the intermediate fit window capture the Wolff chain's slowest mode; the paper itself notes the time definition is controversial and that the energy signal is too noisy at extremely long times, so if either fails, the reported $z_{\rm exp}=1.19(2)$ is not the intrinsic exponent.
Editorial extensions
If this is right
- For large 2D Ising lattices, Swendsen-Wang sampling costs grow essentially as $L^2$ times a logarithmic factor, not as $L^{2+z}$ with a positive $z$.
- Wolff from an ordered start is as efficient as Swendsen-Wang, while Wolff from a disordered start should be avoided in practice because its effective $z_{\rm exp}=1.19(2)$ makes relaxation much slower.
- Integrated correlation times tell the same story: $z_{\rm int}\approx 0$ for Swendsen-Wang and ordered Wolff, but $z_{\rm int}=0.29(3)$ for disordered Wolff.
- Equilibrium measurements that report a near-zero Wolff exponent are consistent with the ordered-start nonequilibrium result, not with the disordered-start result.
- In one dimension, the exact spectrum shows well-separated decay times $\tau/m$, so Swendsen-Wang dynamics has no critical slowing down at any finite temperature.
Reading between the lines
- A testable extension is to recompute the disordered-start Wolff exponent under alternative time conventions, such as raw CPU time or the $t=t'\langle C\rangle/L^d$ definition used in earlier work; if $z_{\rm exp}$ changes with the clock, the reported 1.19 is not an intrinsic property of the algorithm.
- The paper's own conjecture of a dense eigenvalue spectrum for Wolff dynamics implies multi-exponential relaxation at long times, which could be probed by exact diagonalization of small systems or by measuring higher-order correlation functions.
- The same nonequilibrium relaxation method could be applied to other cluster algorithms, such as embedded-cluster dynamics, to test whether logarithmic scaling survives beyond Swendsen-Wang.
- Because the paper concedes that the energy signal is too noisy at extremely long times for disordered-start Wolff, a variance-reduced estimator that resolves the asymptotic regime could reveal whether the effective exponent continues to decrease.
Original abstract
With a nonequilibrium relaxation method, we calculate the dynamic critical exponent z of the two-dimensional Ising model for the Swendsen-Wang and Wolff algorithms. We examine dynamic relaxation processes following a quench from a disordered or an ordered initial state to the critical temperature T_c, and measure the exponential relaxation time of the system energy. For the Swendsen-Wang algorithm with an ordered or a disordered initial state, and for the Wolff algorithm with an ordered initial state, the exponential relaxation time fits well to a logarithmic size dependence up to a lattice size L=8192. For the Wolff algorithm with a disordered initial state, we obtain an effective dynamic exponent z_exp=1.19(2) up to L=2048. For comparison, we also compute the effective dynamic exponents through the integrated correlation times. In addition, an exact result of the Swendsen-Wang dynamic spectrum of a one-dimension Ising chain is derived.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
With a nonequilibrium relaxation method, we calculate the dynamic critical exponent z of the two-dimensional Ising model for the Swendsen-Wang and Wolff algorithms. We examine dynamic relaxation processes following a quench from a disordered or an ordered initial state to the critical temperature T_c, and measure the exponential relaxation time of the system energy. For the Swendsen-Wang algorithm with an ordered or a disordered initial state, and for the Wolff algorithm with an ordered initial state, the exponential relaxation time fits well to a logarithmic size dependence up to a lattice size L=8192. For the Wolff algorithm with a disordered initial state, we obtain an effective dynamic exponent z_exp=1.19(2) up to L=2048. For comparison, we also compute the effective dynamic exponents through the integrated correlation times. In addition, an exact result of the Swendsen-Wang dynamic spectrum of a one-dimension Ising chain is derived.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | With a nonequilibrium relaxation method, we calculate the dynamic critical exponent z of the two-dimensional Ising model for the Swendsen-Wang and Wolff algorithms. | contribution 0.99 | 0.97 | 0.18 | asserted 1.00 |
| s1 | We examine dynamic relaxation processes following a quench from a disordered or an ordered initial state to the critical temperature T_c, and measure the exponential relaxation time of the system energy. | method 0.99 | 0.96 | 0.04 | asserted 0.99 |
| s2 | For the Swendsen-Wang algorithm with an ordered or a disordered initial state, and for the Wolff algorithm with an ordered initial state, the exponential relaxation time fits well to a logarithmic size dependence up to a lattice size L=8192. | result 1.00 | 0.96 | 0.29 | asserted 0.99 |
| s3 | For the Wolff algorithm with a disordered initial state, we obtain an effective dynamic exponent z_exp=1.19(2) up to L=2048. | result 1.00 | 0.97 | 0.22 | asserted 1.00 |
| s4 | For comparison, we also compute the effective dynamic exponents through the integrated correlation times. | method 0.86 | 0.95 | 0.03 | asserted 0.92 |
| s5 | In addition, an exact result of the Swendsen-Wang dynamic spectrum of a one-dimension Ising chain is derived. | result 0.85 | 0.96 | 0.05 | asserted 1.00 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:cond-mat/0603038 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.