MC squared: Monte Carlo Correction for Fast Elliptic PDE Solving
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "MC squared: Monte Carlo Correction for Fast Elliptic PDE Solving".
Jane: Here's the gist of the paper. Partial differential equation (PDE) solvers underpin scientific computing, but realworld deployment is bounded by compute.
Tom: MC2, a hybrid WoS-NN PDE solver that treats a low-budget Monte Carlo solution as a structured estimator of the true field and learns a single-pass neural correction to recover a high-fidelity solution, matches the accuracy of solutions using over 1000× more Monte Carlo compute.
Jane: The Core Problem and Proposed Solution Classical Monte Carlo solvers like Walk-on-Spheres (WoS) are unbiased but suffer from high variance, requiring orders of magnitude more samples to reach acceptable accuracy.
Tom: Learned solvers are fast but biased and brittle under distribution shift.
Title and authors: Tom: So we're looking at the paper "MC squared: Monte Carlo Correction for Fast Elliptic PDE Solving". The title tells you exactly what they did—they used a Monte Carlo correction for fast elliptic PDE solving—and it targets the exact area where compute is a big limit.
Jane: It’s about taking problems where traditional solvers are too slow, like those Monte Carlo methods, and applying a specific correction to speed things up without losing the accuracy you get from using a lot of samples.
Lu: The authors are using this hybrid approach they call MC squared, which treats a low budget Monte Carlo solution as an estimator and then learns a neural network correction. It’s essentially treating that rough estimate as something meaningful, not just garbage data.
The paper's summary: Tom: The summary basically says the key innovation is learning a neural network correction that reduces the error in just one forward pass. They are using this to recover solution quality that would otherwise need around one thousand times more Monte Carlo iterations.
Jane: So, they are showing that you can get a better answer much faster by feeding the initial, rough estimate into this learned operator instead of running more samples.
Lu: They’re not modifying the original Monte Carlo solver or just swapping it for a learned surrogate; they are applying this transformer-based field to field model to early Walk-on-Spheres outputs instead.
Meng: That sounds like a lot of training data is needed to teach a model how to correct the noise structure in that output, which I worry about from a deployment standpoint.
Lalam: The paper argues that this approach keeps the generality and robustness of Monte Carlo solvers while dramatically accelerating convergence and lowering the bias of standalone neural solvers.
The paper's improvements: Tom: One big improvement is that MC squared matches accuracy using over one thousand times more Monte Carlo compute, which is a massive speedup for real-world applications, and it outperforms all the evaluated classical, denoising, and neural operator baselines.
Jane: They’re also showing it outperforms all those other methods across various metrics when you look at things like PSNR or mean squared error.
Lu: They found that the correction works because if this corrector is contractive on the Monte Carlo noise directions, it strictly reduces conditional variance in a way that can compensate for some introduced squared bias.
Meng: So they claim this method improves over raw Monte Carlo in mean squared error whenever the variance reduction more than compensates for the introduced squared bias.
Lalam: They even show that MC squared generalizes well to out of distribution PDE families, which means it learns features of how the Walk-on-Spheres estimate behaves rather than just memorizing one specific PDE type.
Conclusion: Tom: So to wrap up, the main point is that finite sample Monte Carlo error is structured and correctable in a single forward pass, which allows them to get accuracy using over one thousand times less compute.
Jane: It’s about showing that the noise from low budget Monte Carlo isn't just random noise; it carries a structure that can be learned and corrected by a neural operator.
Lu: The framework suggests a broader idea where any unbiased sampler can be converted into a fast, single pass estimator while still keeping bounded unbiasness from the original solver.
Meng: But I have to say, the paper only claims this works for the class of PDEs that Walk-on-Spheres can actually solve, and that's a limitation you have to keep in mind when trying to apply it elsewhere.
Lalam: Exactly, and they acknowledge that the model is trained using data from Monte Carlo budgets up to two to n for n equals two, and they use an error correction objective based on their analytical ground truth as the target output.
Tom: So while this is a huge win for speed and accuracy within those specific PDE families, it opens up a new way to use Monte Carlo methods in practice by using them as structured estimators.
Ethan Hsu, Ivan Ge, Hong Meng Yam
Stanford University
cs.LG, cs.AI, cs.CE, cs.CV, cs.NA, math.NA
Submitted: 2026-05-10
Updated: 2026-11-10
Comments: Accepted to NeurIPS 2026
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 96/100
The gist: Partial differential equation (PDE) solvers underpin scientific computing, but realworld deployment is bounded by compute.
Key concepts
- Walk-on-Spheres (WoS)
- A classical Monte Carlo solver that estimates a PDE solution by averaging function values sampled at random points. It is unbiased but suffers from high variance, requiring many samples for good accuracy.
- MC² Methodology
- This approach treats the finite output of a WoS solver as noisy data and trains a neural network to learn a correction. This correction aims to minimize the error in a single forward pass, trading perfect unbiasedness for massive variance reduction.
- PDEZoo
- A large-scale benchmark infrastructure providing diverse elliptic PDEs with exact solutions and multi-budget Monte Carlo trajectories. It allows researchers to test solvers under real compute constraints and compare methods across different PDE types.
Terminology
Summary
Partial differential equation (PDE) solvers underpin scientific computing, but realworld deployment is bounded by compute. p. 0 edu
Hong Meng Yam*
Stanford University
hongmeng@stanford. edu
Abstract
Partial differential equation (PDE) solvers underpin scientific computing, but realworld deployment is bounded by compute. We present MC2
, a
hybrid WoS-Neural Network (WoS-NN) PDE solver that treats a low-budget Monte
Carlo solution as a structured estimator of the true field and learns a single-pass
neural correction to recover a high-fidelity solution. MC2 matches the accuracy of
solutions using over 1000× more Monte Carlo compute, outperforming all evaluated classical, denoising, and neural-operator baselines.Where the paper says it
Where the paper says it
The Core Problem and Proposed Solution
Classical Monte Carlo solvers like Walk-on-Spheres (WoS) are unbiased but suffer from high variance, requiring orders of magnitude more samples to reach acceptable accuracy. Learned solvers are fast but biased and brittle under distribution shift. p. 0 Classical Monte Carlo solvers such as
Walk-on-Spheres (WoS) are unbiased and geometry-agnostic but are slow. Learned
solvers are fast but biased and brittle under distribution shift. , neural caches
[38]) or condition generative models on sparse numerical samples [16], but require per-problem
retraining, slow iterative inference, or both. In this work, we propose MC² (Monte Carlo error correction with Monte Carlo solvers): a new
approach to hybrid PDE solving that treats the low-iteration output of a Monte Carlo method as
Preprint. Rather than modifying a Monte Carlo solver or replacing it with a
learned surrogate, MC² applies a transformer-based field-to-field model to early Walk-on-Spheres
outputs, recovering solution quality that would otherwise require around 1000× more Monte Carlo
iterations [10]. This preserves the generality and robustness of Monte Carlo solvers while dramatically
accelerating convergence and lowering the bias of standalone neural solvers.Where the paper says it
Where the paper says it
Where the paper says it
The Benchmark Infrastructure
To enable reproducible study, PDEZoo is introduced, which offers nearly limitless diversity within the elliptic PDE regime. The benchmark spans five elliptic PDE families and unlimited geometric compositions with analytic ground truth and multi-budget Monte Carlo trajectories. p. 0 MC2 matches the accuracy of
solutions using over 1000× more Monte Carlo compute, outperforming all evaluated classical, denoising, and neural-operator baselines. To enable reproducible
study of finite-compute PDE solving, we additionally release PDEZoo, the largest
standardized elliptic PDE benchmark to date: 2M PDEs spanning five elliptic
families and unlimited geometric compositions, with analytic ground truth and
multi-budget Monte Carlo trajectories. PDEZoo spans Laplace, Poisson,
Yukawa, Biharmonic, and Helmholtz equations with explicit forcing functions, boundary conditions,
and dense reference solutions. Unlike existing PDE benchmarks, PDEZoo is designed for finitecompute evaluation: it includes Monte Carlo solver trajectories at multiple compute budgets, exposing
how solvers behave before asymptotic convergence and allowing methods to be compared under
practical resource constraints.Where the paper says it
Where the paper says it
The MC2 Methodology
MC2 is formalized by treating the finite-budget output as a noisy observation uˆK = u + εK, where εK is zero-mean, heteroskedastic, and spatially correlated noise induced by finite sampling. The goal is to find a corrected estimator u˜K that minimizes mean-squared error under fixed compute. p. 3 PDEs Fixed grids ✗ ✗ ✗ ∼5–10k
FNO datasets
[24]
Canonical PDEs Unit square ✗ ✗ ✗ ∼ 7k
Poseidon [14] Multi-PDE Fixed grids ✗ ✗ ✗ 1. 6M
PDEZoo Elliptic (5 families) Unlimited
(8 primitives) ✓ ✓ ✓ 2M
u˜ = Φ(ˆuK) that minimizes mean-squared error under fixed compute. Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction. Since Φ
⋆
is
intractable, we learn an approximation Φθ via:
min
θ
E
∥Φθ(ˆuK) − u∥
2
2
, (2)
using ground truth solutions as supervision. Since Φ
⋆
is
intractable, we learn an approximation Φθ via:
min
θ
E
∥Φθ(ˆuK) − u∥
2
2
, (2)
using ground truth solutions as supervision. Crucially, Φθ operates purely as a post-processing
correction: it does not need PDE coefficients, boundary conditions, or solver internals at inference
time.Where the paper says it
Where the paper says it
Where the paper says it
Empirical Results and Generalization
MC2 matches the accuracy of solutions using over 1000× more Monte Carlo compute, outperforming all evaluated classical, denoising, and neural-operator baselines. MC2 achieves 44. p. 6 3 Evaluation Results
We evaluate on the 5,000-instance PDEZoo test split and present our quantitative results across all
methods in Table 2. Our MC2
corrector achieves 44. Our MC2
corrector achieves 44. 10dB PSNR from budget B = 8 input, which
is an improvement of 31. 10dB PSNR from budget B = 8 input, which
is an improvement of 31. 6 dB over the raw B = 8 WoS estimate (25. 6 dB over the raw B = 8 WoS estimate (25. 54 dB) and surpassing WoS at
B = 16,384 (2048× more compute). MC2 outperforms all other methods such as FNO [24], PINO
[25], and CoCoGen [17] across all metrics. We also show strong scaling at every metric
Generalization to Out-of-Distribution (OOD) PDE Families: Additionally, we show that MC2
generalizes well to OOD PDE families, as seen in Table 3, which is a huge advantage over previous
SOTA diffusion-based methods who are unable to generalize outside single PDE families.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Future Directions
The central empirical claim is that the residual εK = ˆuK − u produced by a low-budget WoS solver is not well-modeled as spatially independent noise, but carries structure that is consistent across PDE instances and recoverable by a learned operator. The framework suggests a broader paradigm where any unbiased sampler can be converted into a fast, single-pass estimator while retaining bounded unbisness from the underlying unbiased solver. p. 8 We view MC2
as the first instance of a broader paradigm:
train once across the problem distribution, then convert any unbiased sampler into a fast, single-pass
estimator while retaining bounded unbisness from the underlying unbiased solver. Extending the
framework to these other domains is an immediate direction for future work. 7 Conclusion
We make two contributions to elliptic PDE solving in the finite-compute regime. MC2
establishes that
finite-sample Monte Carlo error is structured and correctable in a single forward pass, achieving the
accuracy of over 1000x more sampling compute and outperforming all evaluated baselines, including
on out-of-distribution PDE families. 7 Conclusion
We make two contributions to elliptic PDE solving in the finite-compute regime. MC2
establishes that
finite-sample Monte Carlo error is structured and correctable in a single forward pass, achieving the
accuracy of over 1000x more sampling compute and outperforming all evaluated baselines, including
on out-of-distribution PDE families. MC2
establishes that
finite-sample Monte Carlo error is structured and correctable in a single forward pass, achieving the
accuracy of over 1000x more sampling compute and outperforming all evaluated baselines, including
on out-of-distribution PDE families. PDEZoo provides the infrastructure the field has so far lacked:
2M elliptic PDEs with analytic ground truth and multi-budget Monte Carlo trajectories. MC2
establishes that
finite-sample Monte Carlo error is structured and correctable in a single forward pass, achieving the
accuracy of over 1000x more sampling compute and outperforming all evaluated baselines, including
on out-of-distribution PDE families. PDEZoo provides the infrastructure the field has so far lacked:
2M elliptic PDEs with analytic ground truth and multi-budget Monte Carlo trajectories. PDEZoo provides the infrastructure the field has so far lacked:
2M elliptic PDEs with analytic ground truth and multi-budget Monte Carlo trajectories. A main
limitation is that we can only solve the class of PDEs that Walk-on-Spheres can solve, which are
broadly elliptic problems on domains where the Green’s-function machinery applies, and extending
the framework to hyperbolic, parabolic, or strongly nonlinear regimes will require pairing it with a
different unbiased estimator. PDEZoo provides the infrastructure the field has so far lacked:
2M elliptic PDEs with analytic ground truth and multi-budget Monte Carlo trajectories. A main
limitation is that we can only solve the class of PDEs that Walk-on-Spheres can solve, which are
broadly elliptic problems on domains where the Green’s-function machinery applies, and extending
the framework to hyperbolic, parabolic, or strongly nonlinear regimes will require pairing it with a
different unbiased estimator.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
For MC2, the model is trained using data from MC budgets of B ∈ 2 n for n=2 and their respective source terms, with an error-correction objective using our analytical ground truth as the target output. During inference, the B = 8 WoS solutions (noisy estimates) and the source term are inputted into the model. p. 5 For MC2
, we train using data from MC budgets of B ∈ {2
n}
5
n=2 and their respective source terms,
with an error-correction objective using our analytical ground truth as the target output. During
inference, the B = 8 WoS solutions (noisy estimates) and the source term are inputted into the model. During
inference, the B = 8 WoS solutions (noisy estimates) and the source term are inputted into the model. For baselines trained to learn operators, the inputs are the forcing functions f evaluated on the same
grid. For baselines trained to learn operators, the inputs are the forcing functions f evaluated on the same
grid. For diffusion baselines, we train a separate EDM teacher per PDE family on the joint distribution
6Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The variance reduction is achieved because if the corrector is contractive on the Monte Carlo noise directions (Lθ < 1 locally on the relevant input distribution), MC2 strictly reduces conditional variance. MC2 improves over raw WoS in MSE whenever the variance reduction more than compensates for the introduced squared bias. p. 3 6M
PDEZoo Elliptic (5 families) Unlimited
(8 primitives) ✓ ✓ ✓ 2M
u˜ = Φ(ˆuK) that minimizes mean-squared error under fixed compute. Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction. 6M
PDEZoo Elliptic (5 families) Unlimited
(8 primitives) ✓ ✓ ✓ 2M
u˜ = Φ(ˆuK) that minimizes mean-squared error under fixed compute. Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction. Rather than modifying a Monte Carlo solver or replacing it with a
learned surrogate, MC² applies a transformer-based field-to-field model to early Walk-on-Spheres
outputs, recovering solution quality that would otherwise require around 1000× more Monte Carlo
iterations [10]. This preserves the generality and robustness of Monte Carlo solvers while dramatically
accelerating convergence and lowering the bias of standalone neural solvers.Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper introduces MC2 as a hybrid WoS–NN solver that corrects structured Monte Carlo error across diverse elliptic PDEs. This approach combines the generality and unbiasedness of Monte Carlo PDE solvers with learned error correction to yield a more than 1000x wall-clock speedup on PDE solving without sacrificing robustness. p. 1 MC², a hybrid WoS–NN solver that corrects structured Monte Carlo error across diverse
elliptic PDEs, matching the quality of using 512–2,048× more Monte Carlo compute and
outperforming all evaluated baselines. In other words, it combines the generality and
unbiasedness of Monte Carlo PDE solvers with learned error correction to yield a more than
1000x wall-clock speedup on PDE solving without sacrificing robustness. a structured estimator of the true solution and learns a neural network correction that reduces
error in a single forward pass. Rather than modifying a Monte Carlo solver or replacing it with a
learned surrogate, MC² applies a transformer-based field-to-field model to early Walk-on-Spheres
outputs, recovering solution quality that would otherwise require around 1000× more Monte Carlo
iterations [10].Where the paper says it
Where the paper says it
A PDEZoo: A Large-Scale Benchmark for Elliptic PDEs
PDEZoo is the only benchmark combining exact analytic ground truth, diverse per-instance geometries, multi-budget stochastic solver data, and a scalable procedural generator. It includes five elliptic PDE families: Laplace, Poisson, Yukawa, Biharmonic, and Helmholtz. p. 4 1 Problem Specification
Each PDE instance is defined by a tuple P = (Ω, E, f, g, u), where Ω ⊂ R
d
is a bounded domain
with boundary ∂Ω, E denotes an elliptic PDE family, f : Ω → R is a forcing term, g : ∂Ω → R
specifies Dirichlet boundary conditions, and u : Ω → R is the analytical solution. PDEZoo includes
five elliptic PDE families:
Laplace: ∆u = 0, Poisson: ∆u = f, Yukawa: ∆u − λu = f,
Biharmonic: ∆2u = f, Helmholtz: ∆u + k
2u = f. Using these atoms, PDEZoo contains
2 million elliptic PDE instances across Laplace, Poisson, Yukawa, Helmholtz, and Biharmonic
families. Domains are represented implicitly through signed distance functions (SDFs), which
support geometry evaluation, boundary projection, and interior/boundary sampling.Where the paper says it
Where the paper says it
A.1 Solution Atom Types
Each solution u(x, y) is constructed as a sum of n randomly chosen atoms, where n is sampled uniformly from [2, 6]. For Poisson and Yukawa instances, atoms are drawn from the general pool (20 types). p. 4 The atom pool is designed to span the spectrum of features that arise in real
elliptic PDE solutions: smooth low-frequency modes (polynomials, plane waves), localized structures
(Gaussian bumps, radial basis functions), singularity-like behavior (log sources, rational functions),
and sharp gradients (tanh transitions, high-frequency products). Using these atoms, PDEZoo contains
2 million elliptic PDE instances across Laplace, Poisson, Yukawa, Helmholtz, and Biharmonic
families. The atom pool is designed to span the spectrum of features that arise in real
elliptic PDE solutions: smooth low-frequency modes (polynomials, plane waves), localized structures
(Gaussian bumps, radial basis functions), singularity-like behavior (log sources, rational functions),
and sharp gradients (tanh transitions, high-frequency products). Using these atoms, PDEZoo contains
2 million elliptic PDE instances across Laplace, Poisson, Yukawa, Helmholtz, and Biharmonic
families.Where the paper says it
Where the paper says it
A.2 Domain Geometry
Domains are represented implicitly through signed distance functions (SDFs), which support geometry evaluation, boundary projection, and interior/boundary sampling. Composed SDFs are provably conservative, ensuring every WoS ball lies entirely within omega. p. 4 PDEZoo includes
eight primitive domain families (disk, square, rectangle, ellipse, annulus, triangle, hexagon, and
stadium) as well as composed domains formed by Boolean operations (union, intersection, difference)
on two or more primatives. Composed SDFs are conservative underestimates of the true boundary
distance, preserving WoS correctness (Appendix A.Where the paper says it
A.4 Data Format
Each instance is stored as a JSONL record containing case ID, PDE kind, screening parameter λ (nonzero only for Yukawa), a PyTorch expression string for the analytic solution, domain name and parameters, number of atom terms, and hardness metadata. Each packaged.npz file contains three (256 × 256) float32 arrays: noisy, clean, and mask. p. 5 Helmholtz and Biharmonic instances are reserved for testing
to evaluate generalization across PDE types. Each instance stores three 256 × 256 float32 arrays per
budget: the WoS estimate, analytic ground truth, and domain mask.Where the paper says it
A.7 PDEZoo
Each PDE instance is defined by a tuple P = (omega, E, f, g, u) where omega ⊂ R d is a bounded domain with boundary ∂omega. PDEZoo includes five elliptic PDE families: Laplace, Poisson, and Yukawa. p. 4 1 Problem Specification
Each PDE instance is defined by a tuple P = (Ω, E, f, g, u), where Ω ⊂ R
d
is a bounded domain
with boundary ∂Ω, E denotes an elliptic PDE family, f : Ω → R is a forcing term, g : ∂Ω → R
specifies Dirichlet boundary conditions, and u : Ω → R is the analytical solution. PDEZoo includes
five elliptic PDE families:
Laplace: ∆u = 0, Poisson: ∆u = f, Yukawa: ∆u − λu = f,
Biharmonic: ∆2u = f, Helmholtz: ∆u + k
2u = f.Where the paper says it
B Monte Carlo Walk-On-Spheres Solver
The Walk-on-Spheres solver expresses the solution u(x) as the expectation of a random process derived from Green’s functions. In practice, this expectation is approximated with K samples: uˆK(x) = 1/K Σ k=1 u(Yk), which satisfies E[ˆuK(x)] = u(x). p. 2 Monte Carlo PDE Solvers as Unbiased Stochastic Estimators
Monte Carlo solvers such as Walk-on-Spheres (WoS) express the solution u(x) as the expectation
of a random process derived from Green’s functions: u(x) = S(x) + E[u(Y )], where Y denotes
the next random walk location and S(x) aggregates source terms. In practice, this expectation is
approximated with K samples: uˆK(x) = 1
K
PK
k=1 u(Yk), which satisfies E[ˆuK(x)] = u(x) and
Var[ˆuK(x)] = O(1/K).Where the paper says it
C MC2 Architecture
Our MC2 corrector is instantiated as a Swin-UNet (SUNet) architecture. The model processes 256 × 256 inputs with patch size 4, embedding dimension 96, and four encoder stages with depths [8, 8, 8, 8] and [8, 8, 8, 8] attention heads.
D Baseline Models
Traditional denoisers include Gaussian filtering (σ = 1.0), heavy Gaussian filtering (σ = 2. p. 6 Traditional Denoisers. We apply six classical denoising algorithms to budget B = 8 WoS outputs:
gaussian filtering, median filtering, bilateral filtering, total variation minimization, and non-local
means. Traditional Denoisers. We apply six classical denoising algorithms to budget B = 8 WoS outputs:
gaussian filtering, median filtering, bilateral filtering, total variation minimization, and non-local
means. Neural Operator Models. We evaluate FNO [24] and PINO [25] in the operator learning setting,
where the input is the forcing function f and the target is the solution u. Diffusion Models. We evaluate DiffusionPDE and CoCoGen, two SOTA PDE solvers built on a
per-PDE-family EDM.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
E Extended Results
MC2 from a single B = 8 WoS input matches WoS quality at B ∼ 16,384–32,768 on in-distribution families and at B ∼8,192–16,384 on out-of-distribution families. MC2 outperforms all other methods such as FNO [24], PINO [25], and CoCoGen [17] across all metrics. p. 6 54 dB) and surpassing WoS at
B = 16,384 (2048× more compute). MC2 outperforms all other methods such as FNO [24], PINO
[25], and CoCoGen [17] across all metrics. (b) We evaluate performance of applying MC2
error corrector to different WoS budgets. At every WoS budget, MC2
achieves substantially lower
MSE than input WoS budget, matching the accuracy of ∼1,000× more MC compute. Table 3: Per-family PSNR (dB, ↑) on the PDEZoo test set. MC2
achieves the highest PSNR on every
PDE family compared to other Neural Network methods, including the held-out Biharmonic and
Helmholtz families absent from the training distribution.Where the paper says it
Where the paper says it
Where the paper says it
F Bias–Variance Analysis of PDE Solvers
The conditional mean-squared error decomposes as E[(T − u(x))2 P] = Bias2(TP) + Var(T P). Pure WoS has a variance that decays only as O(1/K). p. 1 These properties make
WoS attractive for graphics and simulation [30, 31]. However, the estimator variance decays only as
O(1/K), meaning that the typical estimation error decays only as O(1/
√
K). Our evidence establishes that finite-sample WoS error on
2D elliptic problems contains learnable structure sufficient to recover ∼30 dB from a single B = 8
estimate; we do not claim the structure admits a closed-form description, that the learned corrector is
Bayes-optimal, or that comparable gains transfer to arbitrary stochastic solvers. The bias–variance
analysis in Appendix F formalizes when a Lipschitz corrector strictly reduces MSE relative to the
underlying unbiased estimator, but the empirical magnitude is a property of the WoS–elliptic-PDE
pair we study. 6M
PDEZoo Elliptic (5 families) Unlimited
(8 primitives) ✓ ✓ ✓ 2M
u˜ = Φ(ˆuK) that minimizes mean-squared error under fixed compute. Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction.Where the paper says it
Where the paper says it
Where the paper says it
F.3 MC2
MC2 takes a finite-budget Monte Carlo estimate and optional auxiliary PDE information as input, and outputs u˜K = Φθ(XK). The population-optimal corrector under squared loss is the conditional expectation, also known as the regression function or Bayes estimator under squared loss. p. 3 6M
PDEZoo Elliptic (5 families) Unlimited
(8 primitives) ✓ ✓ ✓ 2M
u˜ = Φ(ˆuK) that minimizes mean-squared error under fixed compute. Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction.Where the paper says it
F.4 Variance Reduction from a Lipschitz Corrector
If the corrector is Lθ-Lipschitz in its first argument (holding a fixed) on the support of low-budget Monte Carlo inputs, MC2 strictly reduces conditional variance. The cost is that the corrected estimator is no longer guaranteed to be unbiased. p. 8 A main
limitation is that we can only solve the class of PDEs that Walk-on-Spheres can solve, which are
broadly elliptic problems on domains where the Green’s-function machinery applies, and extending
the framework to hyperbolic, parabolic, or strongly nonlinear regimes will require pairing it with a
different unbiased estimator. Within that scope, however, MC2
closes most of the gap to high-budget
Monte Carlo at a fraction of the cost, and we see the same recipe as a template for the broader family
of unbiased Monte Carlo solvers.Where the paper says it
F.5 Comparison and Asymptotic Behavior
As K → ∞, the raw WoS estimator converges to the true solution up to the ε-shell bias. By continuity of Φθ, u˜K = Φθ(ˆuK, a) → Φθ(u, a), so asymptotic consistency requires the corrector to approach the identity on clean inputs: Φθ(u, a) ≈ u. p. 3 Under squared loss, the optimal
mapping is Φ
⋆
(ˆuK) = E[u | uˆK], which trades unbiasedness for variance reduction. Since Φ
⋆
is
intractable, we learn an approximation Φθ via:
min
θ
E
∥Φθ(ˆuK) − u∥
2
2
, (2)
using ground truth solutions as supervision. Our evidence establishes that finite-sample WoS error on
2D elliptic problems contains learnable structure sufficient to recover ∼30 dB from a single B = 8
estimate; we do not claim the structure admits a closed-form description, that the learned corrector is
Bayes-optimal, or that comparable gains transfer to arbitrary stochastic solvers. The bias–variance
analysis in Appendix F formalizes when a Lipschitz corrector strictly reduces MSE relative to the
underlying unbiased estimator, but the empirical magnitude is a property of the WoS–elliptic-PDE
pair we study.Where the paper says it
Where the paper says it
Conclusion
MC2 establishes that finite-sample Monte Carlo error is structured and correctable in a single forward pass, achieving the accuracy of over 1000x more sampling compute and outperforming all evaluated baselines, including on out-of-distribution PDE families. PDEZoo provides the infrastructure the field has so far lacked. The main limitation is that we can only solve the class of PDEs that Walk-on-Spheres can solve. Within that scope, however, MC2 closes most of the gap to high-budget Monte Carlo at a fraction of the cost. p. 8 A main
limitation is that we can only solve the class of PDEs that Walk-on-Spheres can solve, which are
broadly elliptic problems on domains where the Green’s-function machinery applies, and extending
the framework to hyperbolic, parabolic, or strongly nonlinear regimes will require pairing it with a
different unbiased estimator. Within that scope, however, MC2
closes most of the gap to high-budget
Monte Carlo at a fraction of the cost, and we see the same recipe as a template for the broader family
of unbiased Monte Carlo solvers. In each case, finite-sample error exhibits the same kind of structured, distribution-specific
patterns that a learned corrector can exploit. We view MC2
as the first instance of a broader paradigm:
train once across the problem distribution, then convert any unbiased sampler into a fast, single-pass
estimator while retaining bounded unbisness from the underlying unbiased solver.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can solve elliptic PDEs in finite compute regimes by treating low-iteration Monte Carlo solutions as a structured estimator and applying a single forward pass neural correction to recover high-fidelity solutions. This enables solving PDEs approximately 1000 times faster than with just Walk-on-Spheres while matching the accuracy of using over 1000 times more Monte Carlo compute.
-
Improved AI system can generalize effectively to out-of-distribution PDE families, such as Biharmonic and Helmholtz equations, by learning features of the WoS estimator’s finite sample behavior rather than relying solely on family-specific solution priors. This demonstrates that
the corrector has learned features of the WoS estimator’s finite-sample behavior
which allows for transfer between different PDE types. -
Improved AI system can provide a standardized evaluation infrastructure through PDEZoo, which offers nearly limitless diversity within the elliptic PDE regime, including multi-budget Monte Carlo trajectories and analytic ground truth. This benchmark allows researchers to
compare solvers across papers in this space
and evaluate methods under practical resource constraints rather than just converged solutions.
Abstract
Partial differential equation (PDE) solvers underpin modern computer graphics, geometry processing, and scientific computing. Classical Monte Carlo PDE solvers such as Walk-on-Spheres (WoS) provide mesh-free estimations that are unbiased, but with accuracy limited by sampling and numerical approximation. Learned solvers are fast but biased and brittle under distribution shift. We therefore present MC squared, a hybrid Neural Network-Monte Carlo PDE solver that treats a low-budget Monte Carlo solution as a structured estimator of the true field and learns a single-pass neural correction to recover a high-fidelity solution. MC squared matches the accuracy of solutions using over 1000 times more Monte Carlo samples, outperforming all evaluated mesh-free classical, denoising, and neural-operator baselines. To enable reproducible study of finite-compute PDE solving, we additionally release PDEZoo, the largest standardized elliptic PDE benchmark to date: 2 million PDEs spanning 5 elliptic families and unlimited geometric compositions, with analytic ground truth and multi-budget Monte Carlo trajectories. Together, MC squared and PDEZoo (1) empirically establish that finite-sample Monte Carlo error is structured, learnable, and correctable in a single forward pass, (2) show that we can solve PDEs about 218 times faster than WoS at matched accuracy and achieve a about 171 times reduction in inference time relative to state-of-the-art diffusion methods while attaining higher accuracy, and (3) provide the reproducible evaluation infrastructure the field has so far lacked. Code & Data: https://mc-error-correction.github.io
Sources
- Towards Multi-spatiotemporal-scale Generalized PDE Modeling
- Physics-Informed Neural Operator for Learning Partial Differential Equations
Related papers
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis
- (Mis)generalization of Helpful-Only Fine-tuning