Quasi-Monte Carlo Latent Variable Models

• Quasi-Monte Carlo Latent Variable Models (QLVMs) are deep generative frameworks that use lattice-based RQMC integration to directly optimize marginal likelihood without an encoder network.
• They achieve accelerated convergence and lower reconstruction error compared to VAEs/IWAEs, especially in low-dimensional (d ≤ 3) latent spaces.
• QLVMs, while providing clearer embeddings and faster integration error reduction, face scalability challenges as computational cost rises exponentially with higher latent dimensions.

Quasi-Monte Carlo Latent Variable Models (QLVMs) constitute a class of deep generative models designed to derive extremely low-dimensional, interpretable embeddings from high-dimensional datasets by leveraging randomized quasi-Monte Carlo (RQMC) integration. Unlike canonical latent variable frameworks such as variational autoencoders (VAEs) and importance weighted autoencoders (IWAEs), QLVMs dispense with encoder networks and variational bounds, directly optimizing the empirical estimate of the marginal likelihood via lattice-based RQMC point sets. The QLVM approach provides significant advantages in terms of integration accuracy and embedding transparency in latent spaces of dimension $d \leq 3$ and demonstrably improves both quantitative and qualitative generative performance in these regimes (Martinez et al., 26 Jan 2026).

1. Model Definition and Marginal Likelihood Estimation

Given $n$ observations $x_i \in \mathbb{R}^D$ modeled via a deep generative decoder $p_\theta(x_i|z)$ and a $d$-dimensional latent $z_i \sim p(z)$, QLVMs target the marginal likelihood

$$p_\theta(x_1, ..., x_n) = \prod_{i=1}^n p_\theta(x_i),\qquad p_\theta(x_i) = \int p_\theta(x_i | z_i)\, p(z_i)\, dz_i$$

Standard latent variable models (VAEs, IWAEs) introduce an encoder $q_\phi(z|x)$ and maximize a lower bound (ELBO or IWAE) on $\log p_\theta(x)$. QLVMs eschew this, instead fixing the proposal to $p(z)$, sampling $M$ joint lattice points $\{z_j\}_{j=1}^M$ via a rank-1 lattice rule with a random uniform shift $\Delta \sim U([0,1]^d)$,

$z_j = (u_j + \Delta) \bmod 1$

and directly target the log marginal likelihood with

$$\ell_{\mathrm{RQMC}}(x_i; \theta) = \log\left[ \frac{1}{M} \sum_{j=1}^M p_\theta(x_i | z_j) \right]$$

The RQMC log-empirical average $\ell_{\mathrm{RQMC}}$ converges in expectation from below to $\log p_\theta(x_i)$ as $M \to \infty$.

2. RQMC Construction and Algorithmic Workflow

The construction of lattice points utilizes well-known low-discrepancy rules: Fibonacci lattices in 2D, Korobov lattices in 3D, with each sample batch sharing a random shift $\Delta$ drawn from the uniform $d$-cube. For a minibatch of $B$ data points and $M$ lattice samples, the computational steps are:

1. Sample $\Delta \sim U([0,1]^d)$.
2. Compute lattice samples $z_j = (u_j + \Delta) \bmod 1,\ j=1,...,M$.
3. For each $x_i$ in the batch: a. Evaluate log-likelihoods $\ell_{ij} = \log p_\theta(x_i|z_j)$ via the decoder $f_\theta$. b. Aggregate with $\ell_{\mathrm{RQMC}}(x_i; \theta) = \log$-sum-exp$\{\ell_{i1},...,\ell_{iM}\} - \log M$.
4. Set batch loss $= - (1/B) \sum_{i=1}^B \ell_{\mathrm{RQMC}}(x_i)$.
5. Backpropagate gradients $\nabla_\theta$ batch loss through all $M$ forward evaluations.
6. Update $\theta$ via stochastic optimizer (e.g., Adam).

Backpropagation is straightforward since $z_j$ depends only on $\Delta$ (not $\theta$), and $f_\theta(z_j)$ is reparameterized as in standard decoders.

3. Theoretical Properties: QMC versus MC in Deep Generative Models

Monte Carlo error in marginal likelihood estimation scales as $O(M^{-1/2})$ under i.i.d. sampling. In contrast, QMC lattice rules provide worst-case integration error $O(M^{-1+\varepsilon})$ (for any $\varepsilon > 0$) for smooth, periodic integrands, with randomized QMC delivering unbiasedness and variance $O(M^{-2+\delta})$. For $d \leq 3$, this translates empirically into much faster convergence to the true integral than i.i.d. MC. Both QLVMs and IWAE can recover the exact marginal likelihood in the limit of infinite samples, but QLVMs yield much tighter lower bounds at practical sample sizes in low dimensions (Martinez et al., 26 Jan 2026).

4. Empirical Evaluation: Performance and Latent Space Analysis

QLVMs have been benchmarked on standard and specialized datasets in 2D and 3D latent regimes:

• MNIST $(D=784)$: 2D, 3D latents
• CelebA grayscale $(80 \times 80)$: 2D latent
• Zebra finch syllables, Mongolian gerbil vocalizations: 2D latent
• 3dShapes (6 ground-truth factors): 2D latent
• CMU motion capture: 2D latent

Key findings include:

• QLVMs exceed matched VAE and IWAE baselines in test marginal log-likelihood across all tasks (Figure 2A, $p < 0.05$).
• Pareto curves demonstrate QLVM achieves lower reconstruction error for fixed wall-clock time compared to VAE/IWAE.
• 2D QLVMs provide sharper reconstructions, more diverse and realistic samples, clear cluster boundaries, and interpretable traversals in latent space, outperforming nonlinear embeddings such as UMAP in preserving continuous generative factors (Martinez et al., 26 Jan 2026).

5. Computational Complexity and Scalability

Each QLVM training iteration requires $M$ forward and backward passes through $f_\theta$, with per-batch cost $O(B \cdot M)$. Unlike IWAE, which assigns $M/B$ samples per datapoint under fixed sample budgets, QLVM can reuse the same $M$ lattice points for all $B$ in a batch. The exponential growth in required $M$ for maintaining low discrepancy restricts scalability: QLVMs are practical and effective for $d \leq 3$, but are computationally prohibitive at higher latent dimensions.

6. Limitations and Extensions

QLVMs exhibit several limitations:

• Inferior sample quality on complex, high-resolution datasets (e.g., Celeb-A) relative to high-d VAEs or latent diffusion models.
• Exponential increase in integration error and computational cost as $d$ rises.
• Forced low-dimensional embeddings ($d=2$) may blend heterogeneous generative factors, reducing latent interpretability.
• Identifiability is not guaranteed in absence of strong structural priors.

Potential extensions include adaptive importance sampling initiated by lattices for high-$d$ regimes; conditional QLVMs via covariate-conditioned decoders; advanced QMC constructions (scrambled nets, interlaced sequences) for $d=3$; Lipschitz or information-theoretic regularization to improve latent smoothness; and theoretical development of posterior contraction and identifiability bounds under QMC training objectives.

7. Contextualization and Best Practices in Latent Variable Modelling

QLVMs are related to broader QMC applications in latent variable contexts, including particle filtering (SQMC) and exact simulation of financial stochastic volatility models (Gerber et al., 2014, Baldeaux et al., 2012, Chopin et al., 2017). The best-practice in QMC-driven latent variable models emphasizes reformulation of all simulation steps in terms of inverse CDFs/quantile maps; adoption of scrambled digital nets (e.g., Sobol’ sequences) for low-discrepancy sampling; dimensionality reduction (via bridging and orthogonal transforms); and verification of empirical standard errors against theoretical error bounds. Empirical evidence across both generative modelling and filtering confirms the efficiency gains and low-discrepancy integration accuracy afforded by QMC, provided the effective dimension is small (Martinez et al., 26 Jan 2026, Gerber et al., 2014, Chopin et al., 2017).

Plausible implications for future research include extending QLVMs to higher-dimensional settings through structured QMC designs and adaptive sampling, and deploying them as transparent, interpretable low-dimensional tools in multimodal generative and inferential pipelines.