Ratio-Free Algorithm Cuts Hidden Bias in Likelihood-Free Parameter Estimation

One of the most stubborn problems in modern statistics and machine learning is deceptively simple to state: how do you fit a model to data when you cannot even write down the probability of the data? Many of the most realistic simulation models in science, from epidemic dynamics to financial volatility to physical systems, can generate synthetic data given parameters, yet their likelihood functions, the mathematical quantities on which classical inference depends, are analytically intractable. A new study published in the journal Machine Learning by Zehao Li, Zhouchen Lin, and Yijie Peng tackles this likelihood-free inference problem head-on, and its central finding is striking: a fundamental source of bias that has plagued existing methods can be eliminated entirely, and doing so improves estimation accuracy by one to two orders of magnitude at the same computational cost.

The heart of the difficulty lies in how many likelihood-free methods estimate the score, the gradient of the log-likelihood that drives parameter updates. When the likelihood itself cannot be computed, practitioners often estimate both the numerator and denominator of the score using noisy Monte Carlo simulations and take their ratio. This seemingly innocuous step is a mathematical trap. The ratio of two random variables is not equal to the ratio of their expectations, so the estimator carries an intrinsic asymptotic bias on the order of the square root of one over the batch size N. No matter how long you run the optimization, this bias does not vanish; it is baked into the geometry of the estimator itself. The result is systematic error and instability that caps the achievable accuracy regardless of how much compute is thrown at the problem.

The new paper, titled Beyond Likelihood Ratio Bias: Nested Multi-time Scale Stochastic Approximation for Likelihood-Free Parameter Estimation, introduces a ratio-free alternative built on a nested multi-time-scale stochastic approximation scheme, which the authors abbreviate as NMTS. Instead of forming a noisy ratio of noisy estimates, the algorithm simultaneously tracks two coupled stochastic recursions. One recursion, running on a faster time scale, maintains a running estimate of the score function; the other, operating on a slower time scale, uses that estimate to drive the parameter update itself. The key insight is that the inner score-tracking process can converge to the true score as a fixed point, so the outer parameter update effectively consumes a far more accurate signal than a raw Monte Carlo ratio could ever provide.

The machinery behind this idea draws on a rich tradition of two-time-scale stochastic approximation, a framework long used in reinforcement learning and adaptive control, where a fast inner loop learns a critic or value function while a slow outer loop updates a policy or decision variable. What distinguishes the present work is the nesting: multiple inner score estimates are maintained in parallel blocks, each associated with a simulated sample, and the slow process aggregates them into a variational-gradient signal. The step sizes governing the two loops are chosen so that the fast step size decays more slowly than the slow one, ensuring that the inner estimate remains locked onto the true score while the outer parameters settle down.

The theoretical analysis is unusually comprehensive for an algorithm paper of this kind. The authors establish strong convergence of the iterates, asymptotic normality, and explicit convergence rates. They prove that the scheme eliminates the original asymptotic bias of order the square root of one over N, and that the convergence rate accelerates from a combination of the slow step size and that Monte Carlo bias term to a new expression involving the ratio of the two step sizes and the square root of the fast step size divided by N. With a proper choice of step-size schedules, this rate matches the optimal rate known in the multi-time-scale stochastic approximation literature. The proofs proceed by embedding the discrete iterations into continuous-time dynamical systems and analyzing their asymptotic stability, a classical technique that connects noisy algorithms to deterministic ordinary differential equations.

Convergence guarantees alone can be cold comfort if the gains are negligible in practice, so the numerical experiments are where the paper becomes genuinely compelling. Across benchmark stochastic systems, the NMTS algorithm improved estimation accuracy by one to two orders of magnitude compared with a standard single-time-scale approach at identical computational cost. The comparison is apples to apples in every sense that matters: the same observed data, the same initialization, the same batch sizes, the same number of iterations, and the same number of independent repetitions. The improvement is not bought with extra compute; it is extracted from the structure of the algorithm itself.

Particularly persuasive is the robustness of the result across noise distributions. Many methods shine in tidy Gaussian settings and collapse elsewhere, so the authors deliberately stress-tested their approach with exponential, uniform, and logistic latent noise, situations where the maximum likelihood estimate has no closed form and must be computed numerically as a reference. In the uniform-noise experiment with a batch size of one thousand, the absolute bias of NMTS was roughly 0.008, compared with 0.139 for the conventional method, a gap of more than an order of magnitude that persisted across batch sizes spanning four powers of ten. Similar patterns held for posterior mean and variance estimation in a partial differential equation example, confirming that the advantage is structural rather than an artifact of convenient test cases.

The paper also pushes the framework into the deep generative modeling territory that dominates contemporary simulation-based inference. In one experiment, the conditional density of observations given parameters is represented by a Masked Autoregressive Flow network, while the approximate posterior is modeled by an Inverse Autoregressive Flow network, both built from autoregressive MADE layers that allow efficient parallel evaluation of densities and samples. The NMTS scheme trains these two networks in coupled iterations, using thousands of outer-layer simulations per round to refine the posterior of the parameter given the observed sequence. This demonstrates that ratio-free multi-time-scale optimization is compatible with the flexible neural density models that practitioners actually want to use, not just low-dimensional textbook examples.

For a broader audience, the significance of this work is best understood as a reminder that in simulation-based science, the accuracy bottleneck is often not raw computational power but the mathematics of the estimator. A ratio of noisy quantities looks innocuous and has been a workhorse technique, yet it silently imposes a bias floor that no amount of simulation can breach. By redesigning the estimation loop so that the ratio is never formed, and by proving theoretically that the resulting algorithm converges at an optimal rate, the authors have converted a hidden source of systematic error into recoverable accuracy. In fields where parameter estimates feed directly into predictions, such as the spread of an epidemic, the pricing of a financial derivative, or the calibration of a physical model, an order-of-magnitude reduction in error at no extra cost is not a technical nicety; it is the difference between answers that mislead and answers that can be trusted.

The study, published in volume 115 of Machine Learning, arrives at a moment when likelihood-free or simulation-based inference is exploding in popularity across disciplines that increasingly rely on complex simulators rather than tractable formulas. As those simulators grow richer, the demand grows for inference engines that are simultaneously fast, stable, and provably correct. The NMTS framework offers a template for how to get all three: exploit the time-scale separation between what needs to be tracked and what needs to be decided, and refuse to divide one noisy estimate by another. If the pattern of results here carries over to applied domains as the theory predicts, the ratio bias that once seemed an unavoidable tax on likelihood-free estimation may simply cease to be paid.

Subject of Research: Likelihood-free parameter estimation using nested multi-time-scale stochastic approximation to eliminate likelihood-ratio bias

Article Title: Beyond Likelihood Ratio Bias: Nested Multi-time Scale Stochastic Approximation for Likelihood-Free Parameter Estimation

Article References: Li, Z., Lin, Z., & Peng, Y. (2026). Beyond Likelihood Ratio Bias: Nested Multi-time Scale Stochastic Approximation for Likelihood-Free Parameter Estimation. Machine Learning, 115(11), Article 250. https://doi.org/10.1007/s10994-026-07175-y

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07175-y

Keywords: likelihood-free inference, stochastic approximation, ratio bias, Monte Carlo estimation, score estimation, convergence analysis, simulation-based inference, variational inference, autoregressive flows, parameter estimation, machine learning, Beyond