Importance-weighted variational inference approximates intractable densities by optimizing multi-sample Monte Carlo objectives whose tightness increases with the number of importance samples N. When reparameterization is unavailable, optimization relies on score-function (REINFORCE) gradient estimators. In a new arXiv preprint, Kamélia Daudel, Minh-Ngoc Tran, and Cheng Zhang analyze these estimators and identify a serious failure mode.

The authors characterize the asymptotic expectation and variance of score-function estimators for importance-weighted objectives. They show that standard variational inference for Monte Carlo objectives (VIMCO) gradient estimators suffer from a vanishing signal-to-noise ratio as N increases, caused by a suboptimal control variate choice. This SNR collapse can impair optimization. To address it, they derive a general mechanism for constructing VIMCO gradient estimators that avoid the problem.

Among the resulting estimators, they identify the High-Effective Sample Size optimal variance importance sampling estimator, which has sqrt(N) SNR scaling. However, it does not have zero variance at optimality, and degenerate importance weights may require prohibitively large N to reach that rate. To overcome these limitations, they introduce VIMCO gradient estimators that attain zero variance at optimality and support importance-weight tempering.

Experiments on discrete latent-variable and state-space models validate the theoretical predictions and illustrate the benefits of the proposed estimators. The paper is available on arXiv as 2602.01412.