PaperlayerView exact revision on arXiv

Source-linked AI summary

Analysis of the Ill-Conditioning of the Discrete Inverse Laplace transform in Monte Carlo Simulations of Quantum Many-Body Systems

Phil-Alexander Hofmann, Thomas Chuna, Tobias Dornheim, Michael Hecht

arXiv:2609.07208v1math.NAcs.NA

TL;DR

Recovering real-frequency spectra from quantum Monte Carlo imaginary-time data requires inverting an ill-posed two-sided Laplace transform, while the conditioning of its finite-dimensional discretization remains insufficiently analyzed. The paper derives conditioning bounds by representing equidistant-grid discretizations with Vandermonde-type systems, including a Chebyshev variant under detailed-balance. The bounds show that detailed-balance improves discrete conditioning without removing super-exponential growth, and identify a manageable low-dimensional regime.

  • Problem

    Analytic continuation of quantum Monte Carlo imaginary-time data requires an ill-posed inverse Laplace transform, and the conditioning of its finite-dimensional discretizations has lacked thorough analysis.

    Source: §1 · p. 2

  • Method

    The authors derive conditioning bounds for equidistant-grid discretizations by expressing the general system as a monomial Vandermonde matrix and the detailed-balanced system as a Chebyshev-Vandermonde-type matrix.

    Source: §1 · p. 2, §5 · p. 18

  • Results

    Detailed-balance reduces the upper-bound exponential dependence on βµ from exp(3βµ) to exp(βµ), but does not remove ill-conditioning or super-exponential growth.

    Source: §5 · p. 18

  • Takeaways & Limitations

    Low-dimensional reconstruction in the t-variable can remain numerically manageable, while oversampling in the s-variable does not substantially improve conditioning.

    Source: §5 · p. 18

    The analysis is formulated for a Dirac-delta representation, and systematic analysis of other problem-specific physical constraints remains future work.

    Source: §5 · pp. 18–19

Abstract

from arXiv · show

Analytic continuation of imaginary-time quantum Monte Carlo data to real-frequency spectra requires the inversion of a severely ill-posed two-sided Laplace transform and arises naturally in quantum many-body calculations of dynamic properties. In this work, we distinguish the intrinsic ill-posedness of the continuous inverse two-sided Laplace problem from the conditioning of its finite-dimensional discretization. For equidistant sampling and reconstruction grids, we express the discrete problem via a diagonally scaled monomial Vandermonde matrix with exponentially distributed nodes. Imposing the physical detailed-balance symmetry transforms the discretization into a diagonally scaled Chebyshev-Vandermonde system. Exploiting these structures, we derive explicit lower and upper bounds on the condition numbers in terms of the physical and discretization parameters. For the unconstrained discretization, our bounds reveal super-exponential growth of the condition number with the reconstruction dimension, which cannot be removed by increasing the number of imaginary-time samples alone. Detailed-balance substantially improves the conditioning, especially in the practically relevant pre-asymptotic regime, although the asymptotic super-exponential dependence remains. In both cases, our bounds identify a low-dimensional regime in which the super-exponential contribution is suppressed and stable reconstruction remains feasible. These results provide a mathematical explanation for the effectiveness of low-dimensional spectral representations and detailed-balance in analytic continuation. While they do not remove the fundamental ill-posedness, they substantially mitigate the ill-conditioning of its finite-dimensional discretization, delay the onset of its catastrophic super-exponential growth and thereby enlarge the range of numerically accessible reconstructions.

1 Introduction

Analytic continuation in quantum Monte Carlo requires recovering real-frequency spectra from imaginary-time data through an ill-posed inverse Laplace transform. The paper addresses the limited analysis of discrete conditioning by relating equidistant-grid discretizations to Vandermonde systems and examining detailed-balance and grid dimension.

1 Introduction

  • Dense discretizations of analytic continuation are ill-conditioned, making direct inversion futile despite its role in recovering spectra from imaginary-time data.The reconstruction connects quantum Monte Carlo correlation data to dynamic material properties.Source: §1 · p. 2

1.3 Related work

  • Existing approaches stabilize inversion through grid adaptation, physical prior information, regularized optimization, or learned mappings from Laplace data to spectra.Regularized approaches use assumptions such as positivity, smoothness, sparsity, or moment constraints.Source: §1 · p. 2, §1.3 · p. 5

1 Introduction

  • The paper fills a conditioning-analysis gap by expressing the general discretization as a monomial Vandermonde system and the detailed-balanced version as a Chebyshev-Vandermonde-type system.The authors report that detailed-balance improves conditioning and that low-dimensional discretizations are more stable; at higher dimensions, analytical bounds reveal super-exponential growth beyond the numerical precision limit.Source: §1 · p. 2

1.1 Mathematical presentation of the analytic continuation problem

  • Uniqueness of analytic continuation from exact data does not ensure stability: small perturbations on [0, β] may be amplified dramatically away from the real axis.The formal contour inversion requires values beyond the available real interval.Source: §1.1 · pp. 3–4

1.2 Physical origin of the analytic continuation problem

  • Quantum Monte Carlo accesses imaginary-time rather than real-time correlations, so recovering the spectrum requires analytic continuation followed by Fourier transformation.The continuation maps z to −is for 0 < s < β; the resulting problem is equivalent to reconstructing f from finite-interval imaginary-time data via the Laplace transform.Source: §1.2 · p. 4

2 The two-sided Laplace transform on a weighted L1 space

The paper separates the continuous inverse problem's ill-posedness from the conditioning of its finite-dimensional discretizations. It establishes the two-sided Laplace operator in a weighted space, then analyzes inverse instability and forward-backward amplification.

2 The two-sided Laplace transform on a weighted L1 space

  • The continuous inverse two-sided Laplace problem is ill-posed, while condition numbers characterize only its finite-dimensional discretizations.The forward operator is bounded, but its inverse is unbounded; singular values decay exponentially.Source: §2 · p. 5, §2.2 · p. 7

2.1 Defining the two-sided Laplace transform

  • Weighting the function space compensates for the kernel's exponential growth as t approaches −∞, making the transform a bounded forward operator.The operator norm is bounded by β.Source: §2.1 · pp. 6–7

2.2 Ill-posedness of the inverse two-sided Laplace transform

  • The inverse is not well-posed in Hadamard's sense, so arbitrarily small data perturbations may produce arbitrarily large reconstruction perturbations.The inverse operator has infinite norm.Source: §2.2 · p. 7

2.3 Forward and backward amplification of the two-sided Laplace transform

  • For localized Dirac perturbations, the unweighted forward-backward amplification grows exponentially in βµ, whereas weighting reduces it to at most linear growth.Imposing detailed-balance similarly yields amplification growing at most linearly in βµ by incorporating the exponential compensation into the model.Source: §2.3 · pp. 9–10

3 Conditioning of the discrete general AC problem

The general discrete AC problem becomes a diagonally scaled monomial Vandermonde system, whose condition number admits explicit bounds in the sampling and reconstruction parameters. The bounds show super-exponential growth with reconstruction dimension, but identify a low-dimensional regime where conditioning may remain manageable.

3.1 Discretizing the continuous general AC problem

  • Equidistant sampling and reconstruction reduce the general AC problem to a diagonally scaled monomial Vandermonde system.The regression matrix factors into a diagonal matrix and a Vandermonde matrix with exponentially distributed nodes.Source: §3.1 · p. 11

3.2 Condition number bounds for the discrete general AC problem

  • For fixed βµ, the condition number grows super-exponentially with reconstruction points, and oversampling in imaginary time cannot remove that growth.The lower bound is independent of the number of samples ns.Source: §3.2 · pp. 11–12
  • In the low-dimensional regime, the super-exponential contribution is suppressed and reconstruction may remain manageable, although the upper bound still grows exponentially with βµ.This supports seeking low-dimensional representations, as in Chuna et al. [16].Source: §3.2 · p. 12

3.3 Numerical validation of the condition number bounds for the discrete general AC problem

  • Numerical condition numbers are consistent with the bounds; the upper bound tracks their behavior more closely across β, though neither bound is tight.The estimates make dependence on ns, nt, and βµ explicit.Source: §3.3 · p. 12

4 Conditioning of the discrete AC problem with detailed-balance enforced on the solution

Imposing detailed-balance converts the discrete AC problem to a Chebyshev-basis regression system and substantially improves conditioning in the low-dimensional regime. The condition number nevertheless remains super-exponentially ill-conditioned as the reconstruction dimension grows.

4.1 Discretizing the continuous detailed-balanced AC problem

  • Detailed-balance symmetry transforms the discretization into polynomial regression in the Chebyshev basis.The symmetry follows from the Kubo-Martin-Schwinger condition for equilibrium correlation functions at inverse temperature β.Source: §4.1 · p. 13

4.2 Condition number bounds for the discrete detailed-balanced AC problem

  • For fixed βµ, the detailed-balanced problem remains super-exponentially ill-conditioned, and oversampling cannot remove that growth.The analysis provides condition-number bounds in ns, nt, and βµ.Source: §4.2 · pp. 15–16
  • In the low-dimensional regime, detailed-balance substantially improves conditioning: the prefactor nearly halves and the exponential rate in βµ falls from 3βµ to βµ.The improvement enlarges the range of numerically manageable discretizations.Source: §4.2 · p. 16

4.3 Numerical validation of the condition number bounds for the discrete detailed-balanced AC problem

  • Numerical condition numbers agree with the bounds, and comparison with the general discretization shows that much larger grids remain below numerical infinity.The upper bound follows numerical behavior more closely than the lower bound.Source: Figure 2 · p. 13, §4.3 · p. 16, Figure 3 · p. 17

5 Conclusion

The conclusion separates the continuous inverse problem’s ill-posedness from the conditioning of its finite-dimensional discretization. Detailed balance improves the discrete conditioning, while low-dimensional reconstruction remains important and broader physical constraints warrant further study.

Continuous inverse problem

  • The continuous inverse two-sided Laplace transform is unbounded and has exponentially decaying singular values; weighting reduces localized forward-backward amplification from exponential to at most linear in βµ.The weighted formulation uses a physically motivated weight, and imposing detailed balance directly accounts for its compensating effect.Source: §5 · p. 17

Discrete problem and detailed balance

  • For the unconstrained discretization, the condition number grows super-exponentially with nt at fixed βµ, and increasing ns cannot compensate for that growth.The upper bound identifies a low-dimensional regime in which the discrete problem can remain numerically manageable.Source: §5 · p. 18
  • Imposing detailed balance changes the discrete regression system to a Chebyshev Vandermonde-type matrix, while its condition number still grows super-exponentially with nt.At a given resolution, its upper bound is substantially improved relative to the unconstrained system.Source: §5 · p. 18
  • At the same resolution, detailed balance improves the condition-number upper bound, reducing its exponential dependence on βµ from exp(3βµ) to exp(βµ).The corresponding exponential base drops from e3 ≈ 20.09 to e ≈ 2.71; numerical experiments confirmed the improvement, but super-exponential growth remains.Source: §5 · p. 18

Kernel-based models

  • For translated-kernel basis functions, the transform adds only a diagonal row scaling, so the conditioning analysis applies up to that prefactor’s condition number.Kernel-based representations may require fewer basis functions to resolve smooth spectra, keeping reconstruction in the better-conditioned low-dimensional regime; the additional scaling may worsen conditioning.Source: §5 · p. 18

Practical implications and future work

  • The authors recommend low-dimensional reconstruction in t, note that oversampling in s does not substantially improve conditioning, and identify systematic analysis of other physical constraints as future work.Restricting reconstruction to a suitably low-dimensional discretization can be viewed as self-regularization.Source: §5 · pp. 18–19

A Proof: Condition number bounds for the general AC problem

The proof establishes condition-number bounds for the general discretized AC regression matrix. It uses a polynomial test for the lower bound and partitions rows into blocks to develop the upper bound.

Theorem 3.1

Lower bound

  • The lower-bound proof uses a polynomial and compares its Euclidean norm with a Rayleigh-quotient evaluation at the first canonical basis vector.The proof frames conditioning through the Gram matrix of R.Source: A Proof: Condition number bounds for… · p. 20

Upper bound

  • For the upper bound, the proof partitions row indices into q disjoint sets, factors the resulting square blocks, and combines eigenvalue estimates.The proof then substitutes ht = 2µ/nt to obtain the stated lower bound and combines estimates to obtain the upper bound.Source: A Proof: Condition number bounds for… · pp. 21–22

B Proof: Condition number bounds for the detailed-balanced AC problem

The proof establishes condition-number bounds for the detailed-balanced regression matrix. Its lower- and upper-bound arguments use block structure and Chebyshev-basis properties.

Theorem 4.1

Lower bound

  • The lower-bound proof uses a polynomial, the leading coefficient of Tnt in the monomial basis, and a Rayleigh-quotient evaluation at the first canonical basis vector.The proof analyzes conditioning through the Gram matrix of Rdb.Source: B Proof: Condition number bounds for… · pp. 23–24

Upper bound

  • For the upper bound, the proof partitions row indices into q disjoint sets, factors the resulting blocks, and applies estimates involving sinh and cosh.The block estimates are combined to obtain the stated upper bound.Source: B Proof: Condition number bounds for… · pp. 24–26

C Auxiliary results

This appendix establishes inverse-norm bounds for monomial and Chebyshev Vandermonde-type matrices at distinct positive nodes, using Lagrange-polynomial coefficient representations.

C Auxiliary results

  • For pairwise distinct positive nodes, Proposition C.1 bounds the 2-norms of the inverse monomial and Chebyshev Vandermonde-type matrices.The proof treats their inverse columns through the coefficients of the corresponding Lagrange polynomials.Source: C Auxiliary results · p. 27
  • For positive nodes, the numerator coefficients of each Lagrange polynomial alternate in sign, so evaluating it at x = -1 gives a bound used in the monomial case.The proof then sums the resulting bounds over the interpolation nodes.Source: C Auxiliary results · pp. 27, 29
  • In the Chebyshev representation, converting from monomial coefficients does not increase the absolute coefficient sum, supporting the stated upper bound.The derivation uses the Chebyshev expansion and the triangle inequality for its coefficients.Source: C Auxiliary results · p. 28
Loading 2609.07208v1…