Many models of fracture and anomalous diffusion replace derivatives with integrals over a small neighborhood of radius $\delta$, called the horizon. Computing these integrals on a point cloud with spacing $h$ is hard. A scheme counts as asymptotically compatible (AC) if it converges to the right answer however $h$ and $\delta$ go to zero together, including the regime where $\delta$ shrinks and the classical PDE should reappear. The working AC schemes all choose quadrature weights by solving a small optimization at each node that forces exact reproduction of low-degree polynomials. They return one deterministic answer and no estimate of how wrong that answer is. Probabilistic numerics offers a principled source of error bars. Bayes-Sard cubature is a Gaussian-process quadrature rule that is exact on polynomials and still reports a posterior variance, so it looks like a natural replacement for these weight solvers.
We argue that the substitution is easy and the uncertainty is not. A Gaussian-process prior with a fixed length-scale becomes miscalibrated as $\delta \to 0$, because the horizon shrinks relative to the prior's correlation length. The error bars then collapse or explode depending on the path $(h,\delta)\to 0$ takes, while the weights themselves stay AC. Our proposal defines asymptotically compatible uncertainty: per-node and solution-level credible intervals that stay calibrated uniformly over coupling ratios $h/\delta$. We construct it with a horizon-rescaled Bayes-Sard prior, and test it against the cheap ways of attaching error bars to existing AC solvers.
Core Claim
At a node $x$, the nonlocal diffusion operator is
$$ L_\delta u(x) = \int_{B(x,\delta)} K_\delta(|y-x|)\,\big(u(y)-u(x)\big)\,dy \approx \sum_{j} w_j\,\big(u(x_j)-u(x)\big). $$
Here $K_\delta$ is a kernel normalized so that $L_\delta \to \Delta$ as $\delta\to 0$, the $x_j$ are the cloud points inside the horizon ball $B(x,\delta)$, and the $w_j$ are the weights to be chosen.
OBMeshfree and the GMLS variational quadrature pick weights by least squares subject to exact reproduction of quadratics. That exactness is what makes them AC.
Bayes-Sard cubature models the integrand as a polynomial trend plus a Gaussian-process residual and puts a flat prior on the trend coefficients. Its weights are then exact on that polynomial space, and they minimize the worst-case error of the residual in the kernel's function space. So a Bayes-Sard stencil with a quadratic trend inherits AC by the same argument OBMeshfree uses. This step is close to free, and we treat it as a lemma rather than the contribution.
The contribution concerns the variance. Rescale the ball to the unit ball by $z=(y-x)/\delta$. The posterior standard deviation of the per-node truncation error then factorizes as
$$ \mathrm{sd}\big[L_\delta u(x) - \textstyle\sum_j w_j(\cdot)\big] = \delta^{-2}\,\sigma_\delta\,\varepsilon(h/\delta). $$
In this expression, $\varepsilon(h/\delta)$ is the worst-case error of the rescaled stencil. By the covering-radius argument of Optimal Monte Carlo integration on closed manifolds, it decays with the rescaled covering radius of the points. The factor $\sigma_\delta$ is the prior amplitude of the residual of $u$ beyond the trend at scale $\delta$. The $\delta^{-2}$ comes from the kernel normalization.
A prior with fixed length-scale and fixed amplitude gets $\sigma_\delta$ wrong by a power of $\delta$. Its intervals therefore drift as $\delta\to 0$, which can be checked directly on the coupled-path benchmarks below. A prior whose length-scale is tied to $\delta$, with its amplitude fitted by marginal likelihood across nodes, should give a $z$-score distribution that is stable across $\delta\propto h$, $\delta\propto h^2$ and $\delta\propto\sqrt h$. This is our falsifiable claim.
Research Program
We will build Bayes-Sard stencils inside an OBMeshfree-style strong-form solver for 2D nonlocal diffusion. Each node gets a local Gaussian-process prior on $u$ restricted to its horizon, a quadratic trend (cubic as an ablation), and a Matérn residual kernel with length-scale $c\,\delta$. Each stencil costs only a small local solve, which sidesteps the cubic global cost flagged by Ehler et al..
Stencils of neighboring nodes share points, so their truncation errors are correlated. We will compute the cross-covariances between neighboring nodes from the shared prior. This gives a sparse covariance for the vector of truncation errors.
We will push that covariance through the linear solve to get a Gaussian over solution error. The update is linearized, $e \approx A^{-1}\tau$, where $A$ is the assembled discrete operator and $\tau$ is the truncation-error vector. We will implement it as a belief update in ProbNum.
The theory work proves the scaling relation above for quasi-uniform point clouds. It states the amplitude-scaling condition under which the coverage of the intervals is independent of the path $(h,\delta)\to 0$. It also measures how far marginal-likelihood fitting of the amplitude, which Bayes-Sard leaves without guarantees, departs from that condition in practice.
A secondary thread addresses a tension the file documents. Positivity of the weights is essential for monotone schemes (nonlocal relaxation), and it costs nothing in convergence rate for manifold quadrature (Ehler et al.). Yet Bayes-Sard weights can turn negative on perturbed clouds. We will study weights constrained to be non-negative and reinterpret their worst-case error as a conservative interval width. This thread is a constraint on the method, not its core.
Baselines
The naive reference point is uniform Monte Carlo weighting of the horizon integral. It converges like $n^{-1/2}$, which is the $-0.5$ log-log slope reported on the sphere by Ehler et al.. It also lacks polynomial exactness, so it should fail to be AC. We include it only to anchor the error axis.
The strongest accuracy baseline is the deterministic AC solver with exactness-constrained least-squares weights: OBMeshfree for strong form and GMLS for the variational setting. It runs on identical point clouds, horizons and coupling paths. The reported numbers we must match are a $10^{-14}$ patch-test error with second-order $L^2$ convergence on uniform grids, and first-order convergence with second-order pre-asymptotic behavior on perturbed grids (GMLS quadrature). Further reported figures are second order for $\delta\propto h$ and $\delta\propto h^2$ and first order for $\delta\propto\sqrt h$ in RK collocation, and $\mathcal{O}(\delta^2)$ local convergence in OBMeshfree. Bayes-Sard stencils need not beat these. Matching them within a constant factor is the bar.
The strongest uncertainty baselines are two cheap ways of bolting error bars onto this solver. The first is Richardson-style extrapolation from a second solve at $h/2$. The second keeps the GMLS weights and reports their worst-case error in the same rescaled kernel space as a standard deviation. Both get the same kernel and the same amplitude-fitting budget as our method.
No reported result on Lacuna exists for calibration of discretization error bars in nonlocal problems. Bayes-Sard reports only qualitatively that its intervals contained the true integral on Gauss–Patterson tests. "Beating" therefore means the following. Empirical 95% coverage should stay within a few points of nominal on every coupling path and grid type. The $z$-score variance should stay near one where the baselines drift. All of this should hold at no more than one solve's cost.
Two ablations isolate the key ingredient. One is standard zero-mean Bayesian quadrature, which has no polynomial trend and so should break AC. The other is Bayes-Sard with a fixed length-scale prior, which should keep AC while losing calibration.
Why not just…?
Why not just run the deterministic solver twice and extrapolate? This is the most serious competitor. On uniform grids with a known rate it may be as well calibrated as our method. Its weakness is that it needs the convergence rate, and the rate changes with the coupling path: RK collocation reports second versus first order depending on whether $\delta\propto h$ or $\delta\propto\sqrt h$. The rate also changes with grid perturbation, where GMLS drifts from pre-asymptotic second order toward first (GMLS quadrature). Extrapolation also gives a global error number, not a spatial map. If our perturbed-grid and $\sqrt h$-path experiments show extrapolation calibrated within a few points of ours, we will concede that the value of our method is limited to single-solve cost and node-resolved error maps.
Why not keep the GMLS weights and just report their worst-case error? Any weights have a computable worst-case error in a chosen kernel space, so this attaches uncertainty to OBMeshfree unchanged. The Bayes-Sard weights minimize that same quantity under the same exactness constraint. Hence the post-hoc interval can only be wider, and the gap measures the cost of not reoptimizing the weights. The calibration question, however, is shared. Getting the amplitude scaling right is what both approaches need. If the post-hoc version is well calibrated, that confirms our central claim about horizon-rescaled priors. It would also show those priors can be retrofitted to existing solvers, which we would report as the practical outcome.
Why not just use standard Bayesian quadrature? Zero-mean priors revert to zero away from the data. Bayes-Sard reports a two-orders-of-magnitude error gap against standard Bayesian cubature on the Vasicek bond-pricing integrals for exactly this reason. More decisively for our setting, standard Bayesian quadrature does not reproduce quadratics, so it should not recover the Laplacian as $\delta\to 0$. Our ablation shows this directly on the $\delta\propto h^2$ path.
Proposed Experiments
The first experiment uses the OBMeshfree 2D nonlocal diffusion manufactured problem. We sweep $h$ along $\delta$ fixed, $\delta\propto h$, $\delta\propto h^2$ and $\delta\propto\sqrt h$, on uniform clouds and on randomly perturbed clouds. We report $L^2$ error and observed order against the deterministic baseline. We then report per-node truncation-error $z$-scores, solution-error coverage at 50% and 95%, and interval sharpness for every uncertainty method.
The decisive check comes from plotting $z$-score variance against $\delta$ along each path. The fixed-length-scale Bayes-Sard ablation should drift by a predictable power of $\delta$. The rescaled prior should stay flat. If the rescaled prior also drifts, the scaling analysis is wrong and the core claim fails.
A second check is the patch test. Bayes-Sard stencils must reproduce quadratics to near machine precision, as GMLS does at $10^{-14}$. They must also report a near-zero posterior variance, because the residual is zero on quadratics.
We then repeat the variational version on the GMLS 1D interval and 2D square benchmarks, including the perturbed grids, to test the theory beyond uniform Cartesian clouds. That theory remains open for the deterministic schemes. For the positivity thread, we compare unconstrained and non-negative Bayes-Sard stencils on the unit-disk problem of the monotone meshfree paper. We measure maximum-principle violations, observed order (that paper reports an empirical second order), and coverage. Cost is reported throughout as wall-clock per stencil and per solve relative to OBMeshfree.
References
- OBMeshfree: An Optimization-Based Meshfree Solver for Nonlocal Diffusion and Peridynamics Models
- Efficient optimization-based quadrature for variational discretization of nonlocal problems
- A Bayes-Sard Cubature Method
- Optimal Monte Carlo integration on closed manifolds
- Numerically Optimizing Weights for Monte Carlo Integration on Smooth Compact Manifolds
- ProbNum: Probabilistic Numerics in Python
- Monotone Meshfree Methods for Linear Elliptic Equations in Non-divergence Form via Nonlocal Relaxation
- Asymptotically Compatible Reproducing Kernel Collocation and Meshfree Integration for Nonlocal Diffusion