Flow Matching ISQMC: Neural QMC Integration

• FM-ISQMC is a framework that integrates deep generative transport maps, importance sampling, and quasi-Monte Carlo sequences to achieve robust high-dimensional integration.
• It employs a continuous-time flow matching model with a logistic base transformation and Euler integration, ensuring unbiased estimates and refined error bounds.
• The method bridges generative models with rigorous numerical integration, attaining error rates near O(N⁻¹+ε) while mitigating discretization bias.

Flow Matching Importance Sampling Quasi-Monte Carlo (FM-ISQMC) is a framework for integrating complex high-dimensional target measures by combining deep generative transport maps, importance sampling, and quasi-Monte Carlo (QMC) sequences. It leverages continuous-time flow matching, low-discrepancy QMC points, and an unbiased importance-sampling correction to achieve high-precision integration and robust convergence properties. FM-ISQMC extends classical QMC theory to generative models, bridging advances in @@@@1@@@@ and neural ODEs with rigorous numerical integration methodology (Andral, 2024, Zeng et al., 3 Jan 2026).

1. Algorithmic Construction

FM-ISQMC begins with the generation of QMC points, followed by transport through a neural flow, and estimation via importance sampling correction.

1. Base distribution and QMC sampling: QMC points $\{u_i\}_{i=1}^N \subset [0,1]^d$ are generated using scrambled Sobol’ or Halton sequences to obtain low discrepancy coverage of the unit hypercube.
2. Logistic base and normalizing flow transport: The QMC point $u \in (0,1)^d$ is mapped component-wise via the logit function $G(u) = (\phi(u_1), ..., \phi(u_d))$, where $\phi(u) = \log\frac{u}{1-u}$ yields a standard logistic prior. Transport to the target is achieved by a learned vector field $v_\theta(x, t)$ and discretized with forward Euler integration:

$$x_0 = G(u), ~ x_{k+1} = x_k + h v_\theta(x_k, t_k), ~ T(u) = x_N$$

where $h = 1/N$ and $N$ Euler steps yield $T: (0,1)^d \to \mathbb{R}^d$.

1. Importance sampling correction: For each sample, the pushforward density $q_T(x)$ is calculated via the Jacobian determinant:

$$q_T(x) = \left| \det \nabla_u T(u) \right|^{-1} \Big|_{u = T^{-1}(x)}$$

The importance weight is:

$w(u) = \pi(T(u)) \left| \det \nabla_u T(u) \right|$

2. Estimator: The unbiased (self-normalized) estimator using QMC is

$$I_N = \frac{\sum_{i=1}^N w(u_i) f(T(u_i))}{\sum_{i=1}^N w(u_i)}$$

2. Mathematical Foundation

FM-ISQMC exploits change-of-variables and QMC error bounds to provide theoretical guarantees.

• Push-forward density:

If $U \sim \mathrm{Uniform}([0,1]^d)$ and $x = f_\theta(u)$, the pushforward is $$q_\theta(x) = p_U(f_\theta^{-1}(x)) |\det J_{f_\theta^{-1}}(x)|$$.

• QMC error bound (Koksma–Hlawka):

For composite integrand $\psi(u) = h(f_\theta(u)) w(f_\theta(u))$ and sequence $\{u_i\}$,

$$\left| \frac{1}{N}\sum_{i=1}^N \psi(u_i) - \int_{[0,1]^d} \psi(u) du \right| \le D_N^*(\{u_i\}) V_{\mathrm{HK}}(\psi)$$

where $D_N^*$ is the star discrepancy and $V_{\mathrm{HK}}$ is the Hardy–Krause variation. For “good” QMC sequences, $D_N^* = O(N^{-1}(\log N)^d)$.

• Variance and error analysis:
  • Plain importance sampling with MC has $\mathrm{Var}[I_N] = O(N^{-1})$.
  • QMC–IS under sufficient smoothness achieves $O(N^{-1}(\log N)^d)$ error.
  • With mixed partial-derivative bounds (Owen’s boundary-growth condition), FM-ISQMC achieves $O(N^{-1+\varepsilon})$ RMSE for any $\varepsilon > 0$.

3. Theoretical Convergence

FM-ISQMC is backed by rigorous convergence analysis.

• Assumptions:
  • Integrand $f$ and derivatives grow at most like $\min(F(x_k), 1 - F(x_k))^{-B}$.
  • Target density $\pi(x)$ and derivatives decay exponentially: $$|\partial^\lambda \pi(x)| \le C e^{-\alpha \|x\|_1}$$.
  • Transport map $T$ is a $C^d$ diffeomorphism with explicit value-growth, derivative-growth, and log-lower-bound constraints.
• Owen’s boundary-growth condition:

Composite integrand $h(u) = f(T(u)) \pi(T(u)) \det\nabla_u T(u)$ satisfies the condition such that, by Owen’s Theorem, FM-ISQMC attains RMSE

$$\mathrm{RMSE}(I_N) = O(N^{-1+\epsilon}), \forall~\epsilon > 0$$

regardless of domain unboundedness.

• Architectural verification:

The logistic base transformation and bounded Euler flow mathematically ensure the necessary boundary-growth and regularity properties, verified in Theorems 5.2 and 5.3 of (Zeng et al., 3 Jan 2026).

4. Pseudocode and Workflow

FM-ISQMC involves two key stages: flow training and QMC-driven estimation.

Input: QMC points {u_i}_{i=1}^N ⊂ (0,1)^d (scrambled Sobol′) Flow field v_θ(·,t), N steps, step size h=1/N Target unnormalized density π(x) Output: Estimate of E_π[f(X)] for i=1…N: u = u_i x = logistic_base_map(u) # G(u) log_det = 0 for k in 0…N-1: Jk = jacobian_x v_θ(x, t_k) log_det += log_abs_det(I + h * Jk) x = x + h * v_θ(x, t_k) w = π(x) * exp(log_det) store f(x), w return I_N = sum(w * f) / sum(w)

Editor’s term: "logistic_base_map" denotes the component-wise logit applied to $u$.

5. Practical Implementation and Empirical Performance

• QMC sequences and transformation choices:
  • Scrambled Sobol’ outperforms Halton, with choice of $N=2^m$ recommended for uniformity.
  • Component-wise Box–Muller versus inverse-CDF for uniform-to-Gaussian transformation: Box–Muller slightly preferred (empirical error).
• Flow architectures:
  • RealNVP: Stacked $L$ affine coupling layers; dimension split $d/2$.
  • RQ-NSF: Rational-quadratic neural spline flows with $K$ bins per dimension; conditioners are MLPs (e.g., two hidden layers of size 32).
  • FM-ISQMC also supports neural ODE field parametrizations (Euler-discretized).
• Empirical variance reduction:
  • 2D mixtures of 40 Gaussians: Sobol'-QMC yields MC/QMC std-ratio $\approx 2.2$–$2.8$.
  • Dual-moon targets $d=2…10$: Gain up to $40\times$ in $d=2$, decays to $\approx 1.4\times$ in $d=10$.
  • Log–log RMSE plots: FM-ISQMC slope approaches $-1$, outperforming MC’s $-1/2$ and surpassing naïve QMC’s error floor.
  • For MRTH–MCMC, naïve QMC can introduce bias; the independent importance Markov chain (iIMC) retains QMC variance reduction.
• Error-floor phenomenon:

Direct QMC through a biased generative flow plateaus in error due to discretization bias, which FM-ISQMC corrects via importance sampling, preserving unbiasedness and QMC rates.

6. Context, Significance, and Outlook

FM-ISQMC bridges generative modeling and numerical integration. By correcting transport-bias through importance sampling and exploiting boundary-growth conditions, it leverages neural flows and QMC theory for robust, high-precision integration of complex target measures. The demonstrated $O(N^{-1+\epsilon})$ RMSE and absence of error floors show superior precision and scalability for applications involving multimodal or high-dimensional targets, subject to smoothness and regularity assumptions on the learned transport. This methodology enables deep networks to serve as unbiased samplers for rigorous statistical estimation, formalizing the connection between flow-based machine learning and scientific computation (Zeng et al., 3 Jan 2026, Andral, 2024).

A plausible implication is wider applicability to statistical inference, probabilistic numerics, and domains requiring efficient, unbiased integration under challenging distributions, provided the flow mapping satisfies the necessary technical conditions.