Source-linked AI summary
Weighted Laplace Spaces for Spectral Measures and Rational Approximation
Stefan Jakobsson, Alice Kozakevicius, Stig Larsson
TL;DR
Spectral comparison in SPDE applications calls for combined estimates of discrete spectra relative to the continuous model. The paper uses weighted Laplace spaces and heat-trace representation to establish a FEM spectral-measure norm bound, alongside conditional rational-approximation results for shifted symbols.
Problem
SPDE applications need combined estimates of discrete spectra relative to the continuous model.
Source: §1 · p. 2Method
The paper introduces a weighted Laplace RKHS and represents its dual norm through heat traces to compare continuous and conforming-FEM spectral measures.
Source: §1 · p. 1, §8 · p. 22Results
For conforming FEM eigenvalues, Z_h(t) ≤ Z(t) implies the dual-norm inequality when the nonnegative weight and spectral measures satisfy the stated conditions.
Source: §4.7 · pp. 10–11Takeaways & Limitations
The dual-norm inequality applies to conforming finite element discretizations and nonnegative weights whenever both spectral measures belong to H'_w.
Source: §8 · p. 22The analysis is for the Dirichlet Laplacian and a structured rational ansatz; no universal rational convergence rate is proved, and rate statements are conditional on weighted exponential-sum estimates.
Source: §8 · p. 23
Abstract
from arXiv · showhide
We introduce the weighted Laplace space $H_w$, an RKHS of Laplace transforms on $(0,\infty)$, and study spectral measures in its dual space $H'_w$. For conforming FEM discretizations of the Dirichlet Laplacian on bounded Lipschitz domains, we prove the dual-norm inequality $\|\mu_h\|_{H'_w} \leq \|\mu\|_{H'_w}$, where $\mu = \sum_k\delta_{\lambda_k}$ and $\mu_h = \sum_k \delta_{\lambda_{k,h}}$. The proof combines min-max monotonicity of FEM eigenvalues with a heat-trace representation of the dual norm. We then analyze $H_w$-adapted rational approximation of shifted symbols $\phi(x)=(x+\kappa^2)^{-\beta}$ and give a conditional transfer principle for estimates proved in the corresponding weighted Laplace pre-image norm. Via dual pairing, the norm inequality yields uniform bounds for finite spectral sums and related transformed observables.
1. Introduction
The paper develops weighted Laplace spaces to compare continuous and finite-element spectral measures through heat traces. It connects this framework to observable bounds and weighted rational approximation of shifted symbols.
1. Introduction
- The weighted Laplace space H_w is an RKHS of Laplace transforms whose dual contains continuous and discrete spectral measures and admits a heat-trace representation.This representation makes the space suitable for comparing the spectra.Source: §1 · p. 1
- Dual pairing transfers spectral-measure control into uniform bounds for finite spectral sums and related transformed quantities.The pairing also controls observables represented by ψ in H_w when the spectral pairing is defined.Source: §1 · p. 2
- For conforming FEM eigenvalues, heat-trace monotonicity yields a dual-norm bound for the discrete spectral measure under the stated weight and membership conditions.The inequality is sharp under mesh refinement and needs no further domain or eigenfunction assumptions beyond FEM eigenvalue monotonicity.Source: §1 · p. 2
- For shifted symbols ϕκ(x) = (x + κ^2)^−β, external approximation estimates can transfer to optimal rational approximants in the H_w norm.The paper frames the transfer as conditional on estimates in the corresponding weighted Laplace pre-image norm.Source: §1 · p. 2
- For weights w(t) = t^αe^−γt with α > d − 1, the paper gives a large-γ asymptotic and shows the exact heat-trace integral remains finite as γ ↓ 0.At γ ↓ 0 it converges to the value for w(t) = t^α, with large-time heat-trace decay controlling that regime.Source: §1 · pp. 2–3
2. Conforming FEM Spectral Preliminaries
For conforming finite element discretizations, min–max monotonicity places discrete eigenvalues above the corresponding continuous eigenvalues. Their heat traces therefore provide a comparison between the discrete and continuous spectra.
Conforming FEM eigenvalue monotonicity
- λk ≤ λk,h for each computed eigenvalue, by the min–max monotonicity principle for conforming finite element spaces.The discrete problem has Nh eigenvalues.Source: §2 · p. 3
Heat-trace comparison
- 0 < Zh(t) < Z(t) for every t > 0, so the discrete heat trace is strictly below the continuous one.The proof uses min–max monotonicity and the positive tail of the continuous heat trace.Source: §2 · p. 4
Heat-trace asymptotics
- The continuous heat trace has a small-time Weyl-law asymptotic and satisfies Z(t) ∼ e^-λ1t as t → ∞.The small-time asymptotic is identified as the classical Weyl/heat-kernel asymptotic.Source: §2 · p. 4
Spectral measures and heat traces
- The continuous and discrete spectral measures have heat traces equal to their Laplace transforms, providing the spectral comparison’s transform-side representation.The section frames the objective as bounding the discrete spectral measure in a test-function space’s dual norm.Source: §2 · p. 4
3. Weighted Transform Spaces and Their Duals
The section builds an abstract weighted-transform construction that turns measure pairings into an RKHS and identifies its continuous dual. Its mechanism is an injective synthesis transform paired with a weighted transform-side Hilbert space.
Framework and construction
- An injective synthesis transform is used to construct an RKHS from a weighted transform-side pairing and identify the associated dual norm.The framework is presented as the abstract basis for the Laplace-space construction and the later heat-trace norms.Source: §3 · p. 5
Kernel and duality
- The kernel is strictly positive definite and determines the RKHS, whose continuous dual is isometrically identified through the natural pairing.The proof identifies the pairing’s representing element and extends from finite Dirac combinations by completion.Source: §3 · p. 6
Transform-side representation
- Injectivity of the Laplace synthesis transform makes each function’s pre-image unique on the admissible weighted class.For the Laplace kernel, the synthesis operator is the Laplace transform, so the pre-image is its inverse Laplace transform.Source: §3 · pp. 5–6
4. Weighted Laplace Spaces for Spectral Measures
The weighted Laplace construction supplies dual norms for comparing continuous and conforming FEM spectral measures through their heat traces. The section also relates the resulting spaces to classical RKHS and Sobolev structures and describes weight selection and eigenvalue-free evaluation.
4.1. Laplace Measure Spaces.
- The reproducing kernel is the Laplace transform of the weight evaluated at λ + η, and the table relates example weights to their kernels and classical RKHS families.Examples include Cauchy-type, fractional-resolvent, and Gaussian kernels.Source: §4.1 · p. 7, Table 1 · p. 7
4.2. The Laplace Spectral Measure Spaces.
- The dual norm of the conforming FEM spectral measure is bounded by that of the continuous measure for non-negative weights when both measures belong to the dual space.The comparison follows from heat-trace ordering; the summary states no further domain or eigenfunction assumptions beyond FEM eigenvalue monotonicity.Source: §4.2 · p. 7, §4.7 · p. 11
4.2.1. Sharpness and Gap Dependence.
- For a fixed mesh the norm inequality is strict when the weight is positive on a set of positive measure, but its constant 1 is sharp under refinement.When eigenvalue errors are O(h^2p), the squared-norm gap is also O(h^2p) if the weight is compatible with the heat trace.Source: §4.2.1 · pp. 7–8
4.3. Choosing weights.
- The power-exponential weight w(t) = t^αe^-δt is proposed, with dual-space membership tied to heat-trace asymptotics and the weight’s behavior near zero and infinity.The summary notes the finiteness condition for this weight holds when α > d − 1.Source: §4.3 · p. 8, §4.7 · p. 10
4.4. Eigenvalue-Free Evaluation of ∥µh∥H′
- An eigenvalue-free evaluation strategy approximates the scalar heat kernel rationally, then uses shifted linear solves and quadrature for the weighted time integral.The passages also describe direct or randomized trace estimation and note that approximation error depends on rational approximation, quadrature, and trace estimation.Source: §4.4 · p. 9
4.5. Connection to Weighted Sobolev Spaces.
- For w(t) = t^(2s−1)e^-αt, Hw has a weighted Sobolev interpretation, with point evaluations continuous when s > 1/2.The Sobolev-type norm is expressed through spectral multipliers, with the Laplace transform playing the role of the Fourier transform.Source: §4.5 · pp. 9–10
4.6. Interpretation and Remarks.
- The dual norm measures weighted heat-trace energy; the discrete heat trace is lower because FEM eigenvalues shift upward, while the inequality remains one-directional.Under the stated eigenvalue convergence rate, the squared-norm gap has the same rate when the weight is compatible; the inequality may fail for sign-changing weights or a different space.Source: §4.6.1 · p. 10, §4.6.2 · p. 10, §4.6.3 · p. 10
4.7. Summary.
- The section’s summary states that the inequality holds for admissible non-negative weights, is strict on a fixed finite mesh, and is sharp under mesh refinement.It requires spectral-measure membership in the dual space and no assumptions beyond FEM eigenvalue monotonicity.Source: §4.7 · pp. 10–11
5. Rational Approximation of Shifted Symbols in Hw: Optimal and Classical Constructions
The section develops rational approximation of shifted symbols in the weighted Laplace space Hw, combining direct Hw-error optimization with classical comparison constructions. It establishes admissibility conditions and gives a conditional, rather than proved, transfer of external approximation rates.
5.1. Membership of shifted symbols in Hw.
- Joint membership requires d −1 < α < 2β −1 and 0 ≤ δ < 2κ²; this range is nonempty exactly when d < 2β, or ν > 0 under the Matérn parameterization.These conditions ensure simultaneous membership of the spectral measure in H′w and the shifted symbol in Hw.Source: §5.1 · p. 12
- The unshifted symbol x−β is not in this setting because its pre-image lacks exponential decay and its unweighted endpoint conditions are incompatible.Thus a positive shift is essential for membership in Hw.Source: §5.1 · p. 12
5.2.1. The Correspondence.
- Restricting rational functions to simple real poles shifted by −κ² yields real sums of decaying exponentials, and Hw error becomes weighted L2(w−1) pre-image error.This admissible class excludes complex-conjugate poles, which instead produce damped oscillatory terms.Source: §5.2.1 · p. 13
5.3. Optimal Hw Approximation by Variable Projection.
- Variable projection directly minimizes the Hw objective over poles and eliminates linear coefficients, with closed-form formulas and gradients supporting evaluation without numerical quadrature.The resulting pipeline is used for pole optimization.Source: §5.3 · pp. 13–14
5.4. Convergence as the Number of Poles Increases.
- The rate-transfer principle does not prove a best-approximation rate: it requires an external estimate for the admissible shifted exponential-sum class.The constants and parameter range for such a rate remain undetermined.Source: §5.4 · p. 14, §5.4.2 · p. 15
5.4.1. Best Uniform Rational Approximation (BURA).
- BURA controls relative error in L∞ on a finite spectral interval, which does not imply an Hw estimate without a separate stability theorem linking the norms.“Best” is relative to each method’s native target metric; the Hw construction and classical methods optimize different objectives.Source: §5 · p. 11, §5.4.2 · p. 14, §5.4.1 · p. 14
5.4.2. Plausible asymptotics and conditional rate transfer.
- A global Hw rate must account for the target’s algebraic behavior near t = 0, exponential tail governed by κ², and admissible pole class; finite-interval estimates alone do not suffice.The exact rate also depends on how both endpoints are treated.Source: §5.4.2 · p. 15
5.4.3. Stieltjes Integral as a Possible Proof Route.
- For 0 < β < 1, Gauss quadrature of a Stieltjes representation offers a possible route, but a global Hw bound still needs quadrature-error and tail control.For larger β, this route requires a generalized Stieltjes representation or another external approximation theorem.Source: §5.4.3 · p. 15
6. Application to Gaussian Matérn Fields
The section connects rational approximation of Matérn covariance symbols to resolvent-field mixtures and spectral-statistic error estimates. The duality bound controls the FEM contribution, while balancing it against rational error depends on an approximation rate not established here.
6.1. Matérn Fields via the SPDE Representation.
- A Matérn field with smoothness ν and correlation length κ−1 has covariance symbol ϕκ(x) = (x + κ²)−β, where β = ν + d/2.Its covariance operator is (κ² − Δ)−β.Source: §6.1 · p. 16
6.2. Rational Approximation →Mixture of Simple SPDEs.
- Rational approximation replaces the covariance operator by a finite mixture of resolvent covariance operators, each corresponding to a standard elliptic PDE.With nonnegative coefficients, the mixture is the covariance of independent resolvent fields and has sparse precision matrices.Source: §6.2 · p. 16
6.3. Error in Spectral Statistics via the Duality Bound.
- The spectral-statistics error separates into FEM and rational-approximation contributions, and the dual-norm inequality controls the FEM term.The trace expressions represent expected solution energy.Source: §6.3 · pp. 16–17
- Choosing n so that En ≲ h2p balances rational and FEM errors, but a specific relation between n and h requires a convergence theorem for En.The rational-approximation analysis provides only a conditional rate-transfer statement, not that rate.Source: §6.3 · p. 17
Key Parameter Choices.
- The parameter choices require d −1 < α < 2ν + d −1 and 0 ≤ δ < 2κ²; n is selected to balance rational and FEM errors.Here α is the weight exponent and δ the weight decay parameter.Source: Key Parameter Choices · p. 17
7. Numerical Verification
The unit-square experiments numerically support the FEM dual-norm inequality and show that optimizing rational approximants for the H_w metric improves both norm-level and spectral-observable accuracy. The tests cover mesh refinement and rational degrees from small to moderate.
7.1. Computational Setup.
- Experiments use the unit square with Dirichlet conditions, conforming piecewise-linear FEM meshes, and matched-degree rational approximants evaluated through the same H_w error pipeline.The dual-norm test uses w(t) = t^2e^-t and N = 200; rational tests report two representative weights.Source: §7 · p. 18, §7.1 · p. 18
7.2. Verification of the Dual Norm Inequality.
- The discrete dual norm stays below the continuous reference on tested meshes, and the gap decreases under mesh refinement.The computations support the conforming FEM inequality in the unit-square setting.Source: §7.2 · pp. 19–20, Figure 1 · p. 19, Figure 2 · p. 19
7.3. Rational Approximation Experiments.
- Metric-adapted optimization substantially lowers H_w error versus fixed-grid construction and remains competitive with joint nonlinear least squares across tested pole counts.The advantage is already visible at small and moderate rational degrees.Source: §7.3 · p. 20, Figure 3 · p. 20
7.4. Application-facing Spectral-sum Test.
- Increasing pole count reduces spectral-observable error across tested meshes, and the approximation families remain effective under mesh refinement.This connects the H_w error comparison to a discrete spectral quantity.Source: §7.4 · p. 21, Figure 4 · p. 21
7.5. Discussion.
- The experiments jointly show FEM dual-norm control and improved rational and observable accuracy from target-metric optimization at comparable degrees.The numerical narrative emphasizes usefulness in the low- to moderate-order regime.Source: §7.5 · pp. 21–22
8. Conclusions and Outlook
The paper develops weighted Laplace spaces to compare continuous and conforming-FEM spectral measures and to analyze rational approximation of shifted symbols. It proves dual-norm control with observable bounds, while rational rate transfer remains conditional and the analysis has stated scope boundaries.
8. Conclusions and Outlook
- For conforming FEM, the dual-norm inequality follows from eigenvalue monotonicity and the heat-trace representation, and dual pairing transfers it to uniform bounds for finite spectral sums and related observables.The result holds for non-negative weights for which the spectral measures belong to H′_w.Source: §8 · p. 22
- Explicit variable-projection optimization approximates shifted symbols in H_w, and verified weighted Laplace pre-image estimates transfer conditionally to H_w error estimates.Potential covariance-approximation applications are identified, but developing them is beyond the paper’s scope.Source: §8 · pp. 22–23
- The analysis is for the Dirichlet Laplacian, relies on conforming FEM eigenvalue monotonicity, and asserts no universal rational-approximation rate.Rate statements depend on corresponding weighted exponential-sum estimates.Source: §8 · p. 23
8.1. Extensions to Other Operators and Spectral Measures.
- Extensions to other operators require suitable monotone discretizations, a heat-trace or related kernel representation, and spectral measures in H′_w.The text presents these as conditions for expected validity, not proved extensions.Source: §8.1 · p. 23
8.2. Final Remarks.
- The framework unifies spectral-measure comparison and rational approximation, with numerical behavior assessed at both norm and observable levels.Broader high-dimensional and non-standard applications are framed as future work.Source: §8.2 · p. 24
Appendix A. RKHS and Green’s-Function Interpretation
Appendix A explains the RKHS and Green’s-function structure relevant to weighted Laplace spaces. It connects point evaluation, inverse-operator kernels, and a weighted Sobolev interpretation of H_w.
Appendix A. RKHS and Green’s-Function Interpretation
- The appendix places the weighted Laplace construction in an RKHS and Green’s-function framework.It describes this perspective as underlying the weighted Sobolev interpretation of H_w.Source: Appendix A · p. 24
A.1. Reproducing Kernel Spaces.
- An RKHS has bounded point evaluation, represented at each domain point by a unique kernel function.The reproducing kernel is symmetric and positive definite.Source: §A.1 · p. 24
A.2. When the Kernel Is a Green’s Function.
- When m > d/2, the operator-domain space embeds continuously into C0, making it an RKHS whose reproducing kernel is the inverse operator’s Green’s function.The Green’s function represents point evaluation.Source: §A.2 · p. 24
A.3. Weighted Sobolev Spaces and Laplace-Transform RKHSs.
- For weights w(t) = t^(2s−1)e^(−αt), the H_w kernel has a Green’s-function interpretation and H_w is naturally a weighted Sobolev space of order s in the Laplace variable.The weight e^(−αλ) is built into the inner product.Source: §A.3 · p. 25
A.4. Why This Matters.
The weighted Laplace dual norm has a squared heat-trace integral representation, and the corresponding RKHS kernel has a differential-operator interpretation.
A.4. Why This Matters.
- The heat-trace integral makes the dual norm concrete, while the RKHS kernel links the Laplace-transform space to differential operators.This connects spectral-measure norms to the Green’s-function structure of elliptic operators.Source: §A.4 · p. 25
Appendix B. Randomized Trace Estimation for the Matrix Dual Norm
The appendix approximates the discrete dual norm through randomized estimates of shifted FEM traces, avoiding explicit generalized eigenpairs.
Appendix B. Randomized Trace Estimation for the Matrix Dual Norm
- For each rational heat-kernel shift, m random probes estimate tr((K + σM)^−1M), and the trace estimates are combined to form the heat-trace quantity.The same probes can be reused across shifts when possible.Source: Appendix B · p. 26
- The approach avoids full diagonalization, relying instead on shifted sparse solves with K + σM and mass-matrix multiplications.Multi-shift solvers and a small number of random probes can make trace estimation efficient for large problems.Source: Appendix B · p. 26
- The trace estimator is unbiased, with variance decreasing as m^-1 and RMS error as m^-1/2; Rademacher probes minimize single-sample variance in the stated i.i.d. class.Source: Appendix B · pp. 25–26
- Reusing probes across shifts typically reduces noise in the final quadrature sum.Source: Appendix B · p. 26
B.1. Convergence and Error Estimate.
The section bounds stochastic trace-estimation error as one component of the norm approximation and gives closed-form pole-optimization formulas and a tensor-product FEM spectral construction.
B.1. Convergence and Error Estimate.
- Trace estimation has O_P(m^-1/2) error, up to a prefactor depending on the shifted inverse; total norm error also includes outer-quadrature and heat-kernel rational-approximation errors.Source: §B.1 · pp. 26–27
Appendix C. Closed Forms and Gradients for Pole Optimization
- Under the stated parameter restrictions, model-weight quadratic-form entries have closed forms and the Gram matrix is invertible, yielding a reduced gradient for pole optimization.The explicit gradient supports gradient-based pole search without repeated numerical quadrature of the objective or its derivatives.Source: Appendix C · p. 27
Appendix D. Tensor-product FEM Spectral Construction
- For aligned Cartesian meshes, tensor-product FEM reduces the global eigenproblem to 1D eigenproblems whose eigenvalues combine additively and yield separable heat-trace formulas.Source: Appendix D · p. 28