Entropy-Cut Metropolis-Hastings

• Entropy-Cut Metropolis-Hastings is defined by two formulations: an exact randomized cumulant correction and a learned, entropy-aware acceptance modulation for noisy ratio estimates.
• The cumulant correction uses the moment generating function of log-ratio noise to counteract bias, ensuring detailed balance and stationarity under stochastic conditions.
• The learned discriminator approach employs density-ratio estimation with entropy attenuation to manage uncertainty, improving robustness in implicit Metropolis-Hastings sampling.

Searching arXiv for the cited papers and closely related terminology to ground the article in published work. Entropy-Cut Metropolis-Hastings denotes, in the cited literature, two related classes of Metropolis-Hastings acceptance corrections. In one formulation, the Metropolis-Hastings log-ratio is replaced by a noisy estimator and then corrected by subtracting a cumulant term so that the randomized acceptance remains exact for the target distribution. In the other, arising from Implicit Metropolis-Hastings, acceptance is modulated by the entropy of a learned discriminator when the requisite density ratio is unavailable analytically and must be estimated from samples. Both constructions address the same structural problem: acceptance decisions based on noisy or unreliable ratios can introduce bias, degrade stationarity, or reduce mixing unless the acceptance rule is explicitly corrected or conservatively attenuated [(Nicholls et al., 2012); (Neklyudov et al., 2019)].

1. Randomized Metropolis-Hastings with a cumulant cut

The exact randomized formulation starts from the ordinary Metropolis-Hastings setting with current state $x \in E$, proposal $y \in E$, target density $\pi(x)$, and proposal density $q(y \mid x)$ such that $q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$. The exact Metropolis-Hastings log-ratio is

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$

In expensive or doubly intractable settings, $\Delta(x,y)$ is replaced by an estimator

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$

where $\xi_m$ is independent of $(x,y)$ and has moment generating function $y \in E$0 when it exists. Entropy-Cut Metropolis-Hastings uses the corrected acceptance rule

$y \in E$1

with

$y \in E$2

The correction is chosen so that

$y \in E$3

which exactly offsets the exponential bias induced by the noisy exponentiation. In the terminology of the source, $y \in E$4 is an “entropy/cumulant cut” (Nicholls et al., 2012).

This formulation is embedded in the randomized acceptance framework through an auxiliary randomization variable $y \in E$5. With forward density $y \in E$6, backward density

$y \in E$7

and identity involution $y \in E$8, the randomized Metropolis-Hastings kernel satisfies “very detailed balance” pointwise in $y \in E$9 and therefore detailed balance after integrating over $\pi(x)$0. The resulting acceptance kernel recovers the exact Metropolis-Hastings kernel marginally over the noise. This establishes exactness under the stated independence and moment assumptions, rather than merely asymptotic small-bias behavior.

The Gaussian case yields the Penalty Method. If $\pi(x)$1, then

$\pi(x)$2

and the acceptance becomes

$\pi(x)$3

With $\pi(x)$4 and $\pi(x)$5, this is exactly the Ceperley-Dewing penalty correction. The same framework also encompasses the Universal Algorithm of Ball et al. and the Single Variable Exchange algorithm of Murray et al., each as a particular randomized acceptance construction rather than as unrelated procedures.

2. Implicit Metropolis-Hastings and learned acceptance ratios

Implicit Metropolis-Hastings considers a target distribution represented only by samples, typically an empirical distribution on a finite set

$\pi(x)$6

The proposal may be independent, with an implicit generator $\pi(x)$7 defining $\pi(x)$8, or Markov, with an implicit kernel $\pi(x)$9 generated for example through a latent-space Markov transition composed with a generator. In both cases, the proposal is sampleable but not analytically evaluable. The central objective is therefore to learn the Metropolis-Hastings ratio from samples and then run a Metropolis-Hastings chain whose stationary distribution $q(y \mid x)$0 remains close to $q(y \mid x)$1 (Neklyudov et al., 2019).

For independent proposals, a single-argument discriminator $q(y \mid x)$2 is trained by the class-balanced logistic cross-entropy

$q(y \mid x)$3

Under the standard classification-calibration assumption for the logistic loss, the optimal discriminator satisfies

$q(y \mid x)$4

so the density ratio

$q(y \mid x)$5

is recovered as

$q(y \mid x)$6

In practice, the plug-in estimator

$q(y \mid x)$7

is used, optionally with temperature scaling or Platt scaling to improve calibration.

For Markov proposals, the discriminator is pairwise. A function $q(y \mid x)$8 is trained so that the ratio

$q(y \mid x)$9

approximates

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$0

The core objective is

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$1

whose minimizer satisfies

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$2

A convenient surrogate is the pairwise cross-entropy

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$3

When $q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$4 is constrained to $q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$5 for some $q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$6, minimizing $q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$7 also decreases the total variation upper bound derived for the stationary distribution.

The learned acceptance rules follow the ordinary Metropolis-Hastings form. The exact acceptance is

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$8

For independent proposals this reduces to

$q(y \mid x) > 0 \Leftrightarrow q(x \mid y) > 0$9

and with the discriminator estimate,

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$0

For Markov proposals,

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$1

Because the ratio is estimated rather than exact, the resulting chain is generally biased; the main theoretical task is to control the discrepancy between $$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$2 and $$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$3.

3. Entropy as uncertainty and the “cut” interpretation

The two cited constructions use “entropy cut” in different but connected senses. In randomized Metropolis-Hastings, the cut is the cumulant correction $$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$4 applied directly to a noisy log-ratio. In Implicit Metropolis-Hastings, the paper does not explicitly define an “entropy cut,” but it states that the framework naturally admits entropy-aware acceptance modulation based on classifier uncertainty [(Nicholls et al., 2012); (Neklyudov et al., 2019)].

For the independent-proposal discriminator, the entropy is

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$5

which is maximized at $$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$6. High entropy correlates with large cross-entropy loss and poor ratio accuracy. In the total variation bound for Implicit Metropolis-Hastings, this means that uncertain discriminator outputs loosen the guarantee connecting the stationary distribution to the target.

Two entropy-aware formulations are given. The hard-cut version rejects proposals whenever the maximum entropy exceeds a threshold $$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$7:

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$8

The soft version uses an attenuation weight

$$\Delta(x,y)=\log \pi(y)-\log \pi(x)+\log q(x \mid y)-\log q(y \mid x).$$9

and sets

$\Delta(x,y)$0

for independent proposals, or

$\Delta(x,y)$1

for Markov proposals. Low-entropy decisions keep acceptance near standard Metropolis-Hastings, while high-entropy decisions are conservatively downweighted.

The rationale is explicit: high-entropy classifier outputs indicate unreliable ratio estimates, and entropy-aware attenuation reduces the effect of those unreliable estimates on acceptance. This suggests a conceptual continuity between the two usages of “entropy cut.” In one case, the cut removes exponential inflation caused by stochastic noise in the log-ratio; in the other, it suppresses acceptance when the learned ratio is uncertain. The first is exact under its assumptions, whereas the second is a conservative robustness device within an approximate Metropolis-Hastings construction.

4. Stationarity, total variation control, and exactness conditions

The exact randomized construction preserves the target distribution by detailed balance. The paper’s randomized Metropolis-Hastings formalism proves “very detailed balance” pointwise in the randomization variable and then integrates it to obtain ordinary detailed balance. Under mild conditions, the randomized kernel inherits $\Delta(x,y)$2-irreducibility and minorization from the underlying Metropolis-Hastings kernel. The same work also proves a Peskun ordering bound,

$\Delta(x,y)$3

so randomized acceptance cannot improve statistical efficiency relative to exact Metropolis-Hastings; increased noise variance lowers acceptance and worsens mixing (Nicholls et al., 2012).

Implicit Metropolis-Hastings addresses a different problem: the chain is not exact because the acceptance ratio is learned. The analysis therefore targets the stationary distribution of the approximate chain. Under continuity and strict positivity of $\Delta(x,y)$4 and $\Delta(x,y)$5, a lower bound $\Delta(x,y)$6 for some $\Delta(x,y)$7, and a minorization condition on $\Delta(x,y)$8, the transition kernel

$\Delta(x,y)$9

defines an irreducible and aperiodic Markov chain with stationary distribution $\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$0. Moreover, if $\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$1, then

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$2

and

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$3

The main theorem for Markov proposals states

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$4

The expectation term is exactly the discriminator loss $\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$5. In the independent case, factorizing $\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$6 yields the bound

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$7

Thus minimizing standard cross-entropy decreases the upper bound on the stationary-distribution error (Neklyudov et al., 2019).

The optimal discriminator recovers exactness in the approximate setting. When

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$8

plugging this into the acceptance step yields the exact Metropolis-Hastings transition, and therefore

$\widehat{\Delta}_m(x,y)=\Delta(x,y)+\xi_m,$9

For entropy-aware attenuation, if the weight is strictly positive and bounded below,

$\xi_m$0

then minorization is retained with $\xi_m$1 and the same type of total variation bound holds with $\xi_m$2 replaced by $\xi_m$3. Hard cuts with $\xi_m$4 can break minorization and degrade mixing.

5. Algorithmic realization and empirical behavior

The randomized cumulant-corrected algorithm is procedurally simple. One proposes $\xi_m$5, computes a noisy estimate $\xi_m$6 from $\xi_m$7 inner samples, evaluates the correction $\xi_m$8, and accepts with probability

$\xi_m$9

In the Gaussian case, $(x,y)$0. In the general case, the source suggests estimating

$(x,y)$1

though this introduces extra randomness and slight bias. Diagnostics include monitoring

$(x,y)$2

to check whether the mean-one correction is being realized empirically, and comparing acceptance behavior as $(x,y)$3 varies (Nicholls et al., 2012).

The Implicit Metropolis-Hastings implementation separates training from sampling. In the independent case, minibatches from the dataset $(x,y)$4 and from the generator-induced proposal $(x,y)$5 are used to minimize $(x,y)$6, after which sampling proceeds by proposing $(x,y)$7, computing $(x,y)$8, and accepting or rejecting according to the learned Metropolis-Hastings rule. In the Markov case, pairwise samples are constructed by drawing $(x,y)$9 from the dataset, obtaining a latent representation $y \in E$00, proposing $y \in E$01 via a latent Markov kernel such as HMC or interpolation,

$y \in E$02

and setting $y \in E$03. The pairwise discriminator is then trained with $y \in E$04 or $y \in E$05, with constraints such as sigmoid parameterization and clamping to enforce $y \in E$06.

The empirical validation reported for Implicit Metropolis-Hastings uses CIFAR-10 and CelebA, with independent proposals from DCGAN, WPGAN, and VAE generators, and a Markov proposal based on WPGAN latent traversal via HMC/interpolation with $y \in E$07. Evaluation uses Inception Score and Frechet Inception Distance on 10k samples per snapshot, averaged over 5 evaluations. The reported trend is monotonic: as discriminator training progresses, Inception Score increases and Frechet Inception Distance decreases consistently for all independent proposals. For comparable rejection rates, Markov Implicit Metropolis-Hastings achieves better Inception Score and lower Frechet Inception Distance than independent MH-GAN, indicating an advantage from exploiting proposal conditioning through the learned pairwise discriminator (Neklyudov et al., 2019).

6. Limitations, failure modes, and relation to adjacent methods

The two formulations fail for different reasons. In the randomized exact scheme, the essential requirement is the existence of the moment generating function at $y \in E$08. If $y \in E$09 does not exist, the entropy cut $y \in E$10 is undefined. The source identifies heavy-tailed noise as the canonical obstruction and suggests redesigning the inner estimator, truncating or winsorizing $y \in E$11, or moving to auxiliary-variable randomized Metropolis-Hastings schemes such as Single Variable Exchange. Exactness also depends on the independence structure and on constructing valid forward and backward randomization densities. If the correction is misspecified, the chain becomes inexact; the paper’s coupling analysis then shows Monte Carlo bias terms of order $y \in E$12 or smaller, compared with Monte Carlo error of order $y \in E$13 (Nicholls et al., 2012).

In Implicit Metropolis-Hastings, the main limitations are ratio unreliability, support mismatch, covariate shift, generator degeneracy, and proposal design. If $y \in E$14 is poorly trained or miscalibrated, $y \in E$15 is inaccurate and acceptance becomes biased. If $y \in E$16 places negligible mass where $y \in E$17 has mass, acceptance is bounded but the chain rarely visits those regions. Minorization may fail if $y \in E$18 somewhere, which is why the method enforces $y \in E$19. If the discriminator is trained under a distributional regime different from inference-time proposals, ratio estimates degrade. Generators with collapsed modes provide poor proposals, so even an acceptance correction may not restore adequate exploration. The source recommends calibration, monitoring cross-entropy and entropy, use of Markov proposals in latent space when feasible, and tuning $y \in E$20 and entropy weights to maintain acceptance in a practical range such as 20–60%.

The relation to prior work is explicit in both sources. The randomized acceptance construction places the Penalty Method, Universal Algorithm, and Single Variable Exchange algorithm inside a single r-MCMC framework. Implicit Metropolis-Hastings generalizes MH-GAN from independent proposals to pairwise Markov proposals and removes the optimal-discriminator assumption by proving total variation bounds tied to discriminator loss. Relative to Discriminator Rejection Sampling, it retains Metropolis-Hastings corrections and accommodates Markov proposals. Relative to pseudo-marginal Metropolis-Hastings, it uses biased plug-in ratios and compensates with an explicit stationary-distribution bound rather than unbiasedness. Relative to classifier-based likelihood-free MCMC, it frames classifier training as density-ratio estimation with minorization guarantees and explicit total variation control (Neklyudov et al., 2019).

Taken together, these developments identify two complementary meanings of Entropy-Cut Metropolis-Hastings. One is an exact randomized acceptance rule based on subtracting the cumulant generating function of log-ratio noise. The other is a conservative acceptance modulation based on discriminator entropy within Implicit Metropolis-Hastings. The first preserves the target exactly under the stated assumptions; the second controls the effect of classifier uncertainty and preserves ergodicity when the entropy weight is bounded away from zero. Both are responses to the same methodological constraint: Metropolis-Hastings remains reliable only when uncertainty in the acceptance ratio is explicitly modeled or explicitly controlled.