Source-linked AI summary
Bayesian Numerical Homogenization
Houman Owhadi
TL;DR
The paper addresses how to identify accurate finite-dimensional bases for numerical homogenization with arbitrary rough coefficients. It reformulates the task as Bayesian inference with noisy sources and finite observations, then applies conditioning to integro-differential equations. The framework recovers Rough Polyharmonic Splines and establishes optimal recovery properties for the resulting bases.
Problem
Numerical homogenization needs accurate basis elements for finite-dimensional approximation of solution spaces with arbitrary rough coefficients, but their identification is often laborious and guess-based.
Method
The paper places a prior on system degrees of freedom or source terms, selects finite observations or coarse variables, and computes the posterior through Bayesian conditioning.
Results
The framework identifies bases for arbitrary linear integro-differential equations and re-discovers Rough Polyharmonic Splines as the optimal solution of a Gaussian filtering problem.
Takeaways & Limitations
Bayesian conditioning provides a general guiding principle for designing basis elements and coarse-graining multiscale systems.
Takeaways & Limitations
The general setup assumes operator mappings between specified Hilbert spaces, a source in a strict subspace of HL(Ω), and a unique solution with a Green’s function.
Abstract
from arXiv · showhide
Numerical homogenization, i.e. the finite-dimensional approximation of solution spaces of PDEs with arbitrary rough coefficients, requires the identification of accurate basis elements. These basis elements are oftentimes found after a laborious process of scientific investigation and plain guesswork. Can this identification problem be facilitated? Is there a general recipe/decision framework for guiding the design of basis elements? We suggest that the answer to the above questions could be positive based on the reformulation of numerical homogenization as a Bayesian Inference problem in which a given PDE with rough coefficients (or multi-scale operator) is excited with noise (random right hand side/source term) and one tries to estimate the value of the solution at a given point based on a finite number of observations. We apply this reformulation to the identification of bases for the numerical homogenization of arbitrary integro-differential equations and show that these bases have optimal recovery properties. In particular we show how Rough Polyharmonic Splines can be re-discovered as the optimal solution of a Gaussian filtering problem.
1 Bayesian Numerical Analysis
The paper reframes numerical homogenization as Bayesian inference: randomize the source, condition on finite observations, and use the posterior to identify basis elements. This recovers established spline-based approximations and generalizes the recipe to other priors, observations, and multiscale systems.
- Illustrative quadrature example: For Brownian-motion priors, conditioning on function values produces piecewise linear interpolation and re-discovers the trapezoidal quadrature rule.Integrated Brownian-motion priors similarly produce cubic and higher-order spline interpolants.
- Motivation: Numerical homogenization approximates solution spaces for arbitrary rough coefficients with finite-dimensional spaces, where standard piecewise linear finite elements can perform arbitrarily badly.Accurate basis elements must be adapted to the coefficient microstructure, but their identification has traditionally required substantial investigation and trial and error.
- Bayesian formulation: Bayesian Numerical Analysis places a prior on the source term, excites the PDE with white noise, and estimates solution values from finite observations.For Gaussian fields, conditional expectations are computed by linear projection, yielding a Gaussian-filtering interpretation of basis construction.
- Basis identification: With white noise and point observations, the Bayesian conditioning theorem yields Rough Polyharmonic Splines as basis elements for PDEs with rough coefficients.These splines were previously identified as accurate basis elements with variational, optimal recovery, and localization properties.
- Generality: The proposed framework also considers different priors and linear observations to identify bases for arbitrary linear integro-differential equations and guide coarse-graining of multiscale systems.Its recipe is to assign priors to system degrees of freedom, select coarse variables, and compute the posterior conditioned on them.
2 General setup
The general setup extends the Bayesian homogenization perspective from a prototypical PDE to linear integro-differential equations. It specifies operator, solution, source, and Green’s-function assumptions while targeting a basis for finite-dimensional approximation or coarse-graining.
- General setup: The framework models linear integro-differential operators L and B between Hilbert spaces of generalized functions, with H(Ω) contained in L2(Ω) and HL(Ω) containing L2(Ω).These mapping assumptions define the function-space setting for the operator equation and boundary operator.
- General setup: Numerical homogenization of the integro-differential equation requires the source g to belong to a strict subspace of HL(Ω).The source-space restriction is stated as an assumption for the approximation problem.
- Well-posedness and Green’s function: The operators are assumed to give a unique solution in H(Ω) and possess a Green’s function satisfying LG(x, y) = δ(x −y) and BG(x, y) = 0 on the boundary.The Delta mass at y supplies the Green’s-function source condition.
- Objective: The stated purpose is to identify a good basis for numerical homogenization or coarse-graining of the integro-differential equation.This generalizes the basis-identification objective beyond the prototypical PDE.
3 Bayesian Numerical Homogenization
The paper reformulates numerical homogenization through Gaussian conditioning: random source terms induce Gaussian solution fields whose conditional expectations identify projection bases. For point observations and white noise, this recovers Rough Polyharmonic Splines and provides posterior error information.
- Bayesian formulation: The approach replaces the deterministic source term with a centered Gaussian field, including white noise, and studies the resulting stochastic integro-differential equation.For non-white noise, the paper models the field through an auxiliary equation driven by white noise.
- Bayesian formulation: The stochastic solution is a Gaussian field, with covariance determined by the operator, boundary conditions, and covariance of the source noise.The adjoint Green’s function is used to relate the covariance construction to the adjoint problem.
- Noise and regularity: Noise choice determines the regularity assumption in accuracy estimates: white noise corresponds to g ∈L2(Ω), while regularized noise supports g ∈Hs(Ω) for s ≥0 or s < 0.For non-white noise, estimates depend on the transformed source LΛg.
- Basis identification: Conditioning on N linear observations of the solution yields a projection basis whose linear combination is the conditional mean of u(x).The observations may be generated by linearly independent generalized functions, including Dirac masses and indicator functions.
- Basis identification: For point observations and white noise, the identified basis elements are Rough Polyharmonic Splines, generalizing Polyharmonic Splines to PDEs with rough coefficients.This recovers the previously accurate basis elements directly through one Bayesian conditioning step.
- Uncertainty and accuracy: The posterior distribution supplies both a conditional mean and variance, enabling deviation probabilities and guidance for adding interpolation points.The conditional variance vanishes at interpolation points and also controls pointwise approximation error in the original equation.
4 Variational properties of basis elements
The Bayesian basis elements possess variational and reproducing-kernel structure, with each element characterized as a unique constrained minimizer. They also satisfy orthogonality and optimal recovery properties, while practical computation favors quadratic optimization over covariance inversion.
- Section 4 Variational properties of basis elements: The Bayesian basis elements have variational and optimal recovery properties that support their computation and the derivation of accuracy estimates.These properties are established in analogy with the recovery behavior of conditionally positive definite kernels.
- Reproducing-kernel structure: When Γ(x,x) is finite, the space V with reproducing kernel Γ(x,y) forms a Reproducing Kernel Hilbert Space.The kernel structure provides the functional setting for the basis and its associated norm.
- Variational characterization: Each constrained space Vi is a non-empty closed affine subspace, and its optimization problem is strictly convex with unique minimizer φi.The construction verifies feasibility and strict convexity before identifying the minimizer.
- Practical computation: In numerical applications, solving the quadratic optimization problem is preferred because identifying the covariance Γ is more expensive than solving the associated linear systems.The same basis element can also be represented through nested equations.
- Orthogonality and recovery: The basis φi is orthogonal to V0 under the defined product, and its recovery property is stated for every v ∈V.The constraints defining V0 encode functions whose observation functionals vanish.
- Extension beyond white noise: For non-white noise, the reproducing-kernel, variational, and optimization results remain valid after redefining the space V and scalar product.The extension is conditional on the corresponding modified definitions.
5 Accuracy of the basis elements φi
The section establishes pointwise accuracy bounds for basis reconstructions, with the error controlled by the conditional variance σ²(x). It connects this quantity to radial-basis interpolation and gives concrete geometric conditions for the bounds.
- Pointwise estimates: Theorem 5.1 bounds pointwise reconstruction errors for v and for solutions u using σ(x) multiplied by an appropriate function norm.The displayed estimates include bounds involving ∥v∥V, ∥g∥L2(Ω), and ∥LΛg∥L2(Ω).
- Pointwise estimates: σ²(x) is the conditional variance of u(x) given the observations, linking estimation uncertainty to approximation error.
- Pointwise estimates: σ²(x) is also called the Power function in radial basis interpolation and the Kriging function in geostatistics.
- Examples: For white noise and Dirac observations at points x_i, the pointwise estimate depends on coefficient ellipticity bounds and the mesh norm H.The constant depends only on λmin(a) and λmax(a).
- Examples: For localized observable functions supported within radius H′ of centers x_i with mesh norm H′′, the resulting geometric parameter satisfies H ≤ H′ + H′′.
- Examples: The estimates extend to linearly independent generalized probability densities, including overlapping supports, and can be justified through zeros induced by the observations.For v in V0, the observation constraints yield points where v vanishes, allowing a mesh-norm argument.
6 Pseudo-algorithm
The pseudo-algorithm constructs homogenization bases from selected measurements and a Gaussian noise model. Each basis function is characterized by observation constraints and an equivalent constrained minimization problem, with possible localization.
- Selection: Select N linearly independent measurement functions ψ1, …, ψN in L2(Ω).
- Noise model: Choose a mean-zero Gaussian field ξ with covariance function Λ and a non-degenerate inverse covariance function.
- Basis identification: Identify φ1, …, φN from the stochastic solution u so that their measurements against ψj equal δi,j.
- Optimization formulation: Each φi is equivalently the unique minimizer of a constrained optimization problem enforcing the same measurement conditions.
- Localization: With suitable measurement functions and covariance Λ, the optimization problems can be localized to subdomains of Ω.
7 Statistical Decision Theory and Practical Applications
The section recasts numerical homogenization as a repeated minimax game between data-based estimation and adversarial source selection. It also describes a generalized Bayesian framework in which prior design can yield optimal bases and solvers.
- Game formulation: Player B chooses an estimator from linear measurements, while Player A chooses a source term g in the unit ball of L2(Ω).
- Game formulation: Player B minimizes the error and Player A maximizes it, forming a deterministic min-max formulation of numerical homogenization.
- Game formulation: Under weak regularity conditions, optimal strategies may use probability distributions over candidate sources and estimators.
- Bayesian interpretation: The estimator may be Bayesian even though the game itself is not, and identifying an optimal prior requires min-max optimization over the players’ distributions.
- Practical applications: A generalized framework selects the prior distribution of ξ to obtain optimal numerical homogenization basis functions.
- Practical applications: For Gaussian priors, identifying an optimal distribution for ξ leads to automated discovery of multigrid and multiresolution solvers for rough-coefficient PDEs.