PaperlayerView exact revision on arXiv

Source-linked AI summary

A priori error analysis for one-parameter Lagrange and POD reduced order methods

Antti Hannukainen, Vigdis Toresen

arXiv:2609.34865v1math.NAcs.NA

TL;DR

Repeated solutions of parametric elliptic PDEs motivate reduced-order models, but their error depends on the snapshot points. The paper constructs a stable approximation operator from snapshot subspaces and proves error bounds, with sub-exponential convergence under mild point-set assumptions and exponential convergence for structured sets.

  • Problem

    Parametric PDE applications require repeated solutions, while a priori error estimates for Lagrange and POD ROMs as functions of snapshot selection are the target of this work.

    Source: §1 · p. 1

  • Method

    The authors construct an approximation operator from the snapshot subspace and use it to derive a priori error estimates for Lagrange and POD ROMs.

    Source: §1 · p. 2, §6 · p. 25

  • Results

    The error is bounded by the best degree-n polynomial approximation error, with sub-exponential convergence under mild snapshot-point assumptions and exponential convergence for structured point sets.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The operator’s stability grows at rate O(n^1/2) under mild assumptions, unlike the exponentially growing Lebesgue constant for equidistant points.

    Source: §6 · p. 26

    The analysis concerns a one-parameter model problem and assumes a positive-weight snapshot-point quadrature rule with relative accuracy below one for squared degree-n polynomials.

    Source: §3.2.4 · p. 13, §4 · p. 17

Abstract

from arXiv · show

This work deals with the numerical solution of parametric linear elliptic partial differential equations using the finite element method for spatial discretization and a proper orthogonal decomposition or Lagrange subspace method based ROM for treating the parameter dependency. We study a simple one-parameter model problem and prove an error estimate for these ROMs as a function of the snapshot point set. Sub-exponential convergence with respect to the number of snapshot points is established under very mild assumptions on the point set. Exponential convergence is proven for structured point sets. In contrast to existing error analysis of Lagrange subspace methods, our estimates are not based on interpolation or perturbation arguments. Rather, we define a new approximation operator from the snapshot subspace and prove error and stability estimates for it.

1 Introduction

Parametric elliptic PDEs require repeated solves, motivating reduced models that accelerate evaluation while retaining accuracy. The paper derives snapshot-set-dependent error bounds for Lagrange and POD ROMs using a new snapshot-subspace approximation operator.

1 Introduction

  • POD uses singular value decomposition of snapshot solutions to form its reduced subspace, while the Lagrange method uses their span.Both methods are designed for faster evaluation and are straightforward to implement in existing finite-element codes.Source: §1 · p. 2
  • The paper studies a one-parameter model and estimates ROM error as a function of the snapshot point set, aiming to help select snapshots while limiting solution evaluations.The error analysis concerns repeated solutions needed in applications such as uncertainty quantification, design optimisation, and inverse problems.Source: §1 · pp. 1–2
  • A positive-weight quadrature rule on the snapshot points that integrates squared degree-n polynomials with relative accuracy below one supports the approximation bound.Under mild point-set assumptions, the operator’s stability grows as √n rather than exponentially as the Lebesgue constant does for equidistant points.Source: §1 · p. 2
  • The approximation operator bounds Lagrange and POD ROM errors by the best degree-n polynomial approximation error in a weighted L∞-norm.It yields sub-exponential convergence under mild point-set assumptions and exponential convergence for structured sets.Source: §1 · p. 3
  • The new operator constructs approximations from the snapshot subspace and establishes stability and error estimates without interpolation or perturbation arguments.Earlier point-evaluation approaches do not express their L2 approximations as linear combinations of snapshot solutions, limiting direct use for these ROM analyses.Source: §1 · p. 3

2 Model problem

The paper studies a one-parameter elliptic PDE with an affine parameter-dependent bilinear form, discretizes it by finite elements, and derives an explicit parameter-to-solution formula using a generalized eigenbasis.

2 Model problem

  • The model is a one-parameter elliptic PDE on a Lipschitz domain, with assumptions ensuring a coercive parameter-dependent form and a unique solution.The assumptions include a symmetric continuous bilinear form, with a0 coercive and a1 positive semidefinite; such problems can model conductivity with an inclusion.Source: §2 · pp. 3–4

2.1 Finite element discretization

  • The finite-element discretization solves the weak problem in a mesh-defined finite-dimensional subspace, producing a linear system whose matrices are symmetric positive definite.The FE solution satisfies the same bound as the exact solution; affine parameter dependence gives A(y)=A0+yA1.Source: §2.1 · pp. 4–5

2.2 Solution formula

  • An A0-orthonormal eigenbasis is also A1-orthogonal, with eigenvalues bounded by 0 ≤ λk ≤ 1−r.The parameter-to-solution map is expressed in that basis using coefficients γk(y)=1/(1+yλk).Source: §2.2 · pp. 5–6

3 Review of subspace methods

The section surveys established subspace results, their application to the model problem, and prior approaches to a priori error bounds for Lagrange and POD ROMs.

3 Review of subspace methods

  • The review covers known subspace methods, existing Lagrange and POD ROM error-bound approaches, and snapshot-based ROM analysis.It first recalls prior results and their use for the model problem, then reviews error-analysis approaches.Source: §3 · p. 6

3.1 Subspace methods

A reduced subspace yields a Galerkin approximation with a best-approximation property, and the same low-dimensional space is reused across parameter values for faster solves.

3.1 Subspace methods

  • A subspace method represents the approximate solution as Qz(y), where z(y) solves a reduced surrogate system.The reduced space is shared across parameter values, allowing faster computation than the full solve when it approximates the solution adequately.Source: §3.1 · p. 6
  • The subspace solution is the A(y)-orthogonal projection of the full solution and therefore satisfies a best-approximation property.Its accuracy depends on the chosen subspace, which is intended to approximate the solution sufficiently well while keeping the surrogate system small.Source: §3.1 · p. 6

3.2 Lagrange subspace method

The section reviews Lagrange subspace ROM error estimates and compares how perturbation and interpolation techniques depend on the snapshot point set.

3.2.1 Perturbation analysis

  • The perturbation bound depends on the distance from each evaluation point to a snapshot and on the solution map’s Lipschitz constant K = Cr^-2||L||.Consequently, reaching a target accuracy depends on snapshot proximity and K.Source: §3.2.1 · p. 8

3.2.2 Lagrange interpolation error analysis

  • For Chebyshev interpolation points, the interpolation remainder is bounded by 1/(2^(N−1)N!) times the L∞ norm of the Nth derivative.The resulting ROM bound is derived under the Chebyshev-point assumption, which considerably limits its applicability.Source: §3.2.2 · pp. 9–10

3.2.3 Near-best Lagrange interpolation error analysis

  • The near-best interpolation estimate gives exponential convergence when the Lebesgue constant grows sub-exponentially, but does not explain equidistant-point performance.The estimate relates error to the Lebesgue constant and best polynomial approximation error.Source: §3.2.3 · pp. 10, 12

3.2.4 Local interpolation error analysis

  • Local interpolation removes the need for structured interpolation points, but its convergence condition can require rapidly growing point counts for small r.The section notes that the earlier positive-perturbation analysis does not address this issue.Source: §3.2.4 · p. 13

3.3 Proper orthogonal decomposition

The section reviews POD subspaces formed from snapshots and relates their ROM error to interpolation accuracy and the truncated snapshot approximation.

3.3 Proper orthogonal decomposition

  • The POD subspace is motivated as an r-dimensional subspace minimizing the discrete weighted L2 snapshot error.The cited earlier analysis also bounds that error using singular values of the weighted snapshot matrix.Source: §3.3 · p. 13
  • POD selects an r-dimensional subspace through a truncated singular value decomposition of the weighted snapshot matrix.The weighted best-rank approximation is computed by transforming with a Cholesky factor, truncating the SVD, and transforming back.Source: §3.3 · pp. 13–14
  • For W = I, the POD ROM error estimate links Lagrange interpolation accuracy, the Lebesgue constant, and the discarded singular value σr+1.This is the special case used to analyze POD through Lagrange interpolation; other weights are addressed using the paper’s approximation operator.Source: §3.3 · p. 14

3.4 Greedy method

The section describes greedy reduced-subspace construction and notes the sampling burden in its convergence analysis.

3.4 Greedy method

  • The greedy method builds subspaces from sampled parameters by adding solutions that address the current approximation error.The process is analyzed using Lipschitz continuity and an ε-net requirement for the sample set.Source: §3.4 · p. 15
  • Reaching ε-accuracy can require a very large number of sample points unless the Kolmogorov n-width of the PDE coefficient set is small.The weak greedy variant reduces the cost of evaluating the exact error by using an a posteriori error estimator.Source: §3.4 · p. 15

4 Approximation from the snapshot subspace

The section constructs an approximation operator from snapshot values that is exact for degree-n polynomials and stable under a condition on the snapshot points. It then uses that operator to obtain near-best approximation results and examine Gaussian and arbitrary point sets.

4 Approximation from the snapshot subspace

  • The point-set condition requires positive quadrature weights that integrate squared degree-n polynomials with sufficient relative accuracy.The standing assumption requires the discrete inner product to be positive on the polynomial space; the quadrature rule is exact for constants.Source: §4 · pp. 16–17
  • The operator uses coefficient functions that depend only on the polynomial degree and snapshot points, not on the function being approximated.Its coefficients are degree-n polynomials, forming a weighted sum of snapshot values.Source: §4 · pp. 15–16
  • If the point set satisfies the stated quadrature condition, the operator is degree-n polynomially exact and has a stability constant growing as O(√n).The stability constant grows sublinearly with polynomial degree under the section’s point-set assumptions.Source: §4 · pp. 17, 19, 20, 21
  • The operator’s stability and polynomial exactness yield a near-best approximation bound in terms of the best degree-n polynomial approximation.The section compares its O(√n) stability factor with the exponentially growing Lebesgue constant for equidistant-point Lagrange interpolation.Source: §4 · pp. 20–21

4.1 Gaussian and Arbitrary Snapshot point sets

  • For Gaussian quadrature points, exactness through degree 2N−1 gives zero quadrature error when 2n ≤ 2N−1.The section also treats arbitrary points using weights chosen for a Riemann-sum rule and derives a point-spacing condition for the quadrature criterion.Source: §4 · p. 16, §4.1 · pp. 21–22

5 Error analysis

The section applies the snapshot-subspace approximation operator to derive a priori error bounds for Lagrange and POD reduced models. The bounds relate ROM error to polynomial approximation of the parameter-to-solution map and conditions on the snapshot set.

5 Error analysis

  • When the snapshot condition e < 1 holds, the Lagrange ROM error is bounded using a stability factor and the best degree-n polynomial approximation of the solution map.The section gives estimates for Gaussian quadrature points and arbitrary point sets; for arbitrary points, the degree is tied to the spacing parameter h.Source: §5 · pp. 23–24
  • The Lagrange error depends on both snapshot-quadrature accuracy and the best polynomial approximation of the parameter-to-solution map.The quadrature rule must be sufficiently accurate for the error to be controlled by the degree-n best approximation.Source: §5 · p. 24

5.1 Error Estimate for the POD Method

  • The POD error estimate uses the same point-set condition and includes the POD truncation quantity σr+1 alongside the stability factor Cs(n).The POD subspace is defined using the same weights as the approximation operator.Source: §5.1 · pp. 24–25

6 Conclusions

The conclusions present a new snapshot-subspace approximation operator as the basis for a priori error estimates for Lagrange and POD ROMs. They state its point-set condition and identify high-dimensional parameter dependence as outside the study’s scope.

6 Conclusions

  • The paper’s Lagrange and POD ROM error estimates apply to arbitrary snapshot point sets under the stated quadrature condition.The condition requires positive weights and relative accuracy below one when integrating squared degree-n polynomials.Source: §6 · pp. 25–26
  • For arbitrary points, the condition is satisfied when point spacing h = O(n^-2), while the operator’s stability constant grows as O(n^1/2).The conclusion contrasts this with the exponentially growing Lebesgue constant for equidistant-point Lagrange interpolation.Source: §6 · p. 26
  • The analysis addresses a one-parameter setting and leaves the greater complexity of high-dimensional parametric PDEs for future work.Source: §6 · p. 26

A Best uniform approximation by polynomials

The appendix recalls a Chebyshev-polynomial result concerning best uniform approximation of a rational function. It defines the Chebyshev polynomial used in that discussion.

A Best uniform approximation by polynomials

Loading 2609.34865v1…