Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Complex Helmholtz wavefields can be solved with standard real-valued Gaussian-process conditioning by realifying the operator into a coupled 2×2 block; a multiscale prior reaches 0.767 correlation on in vivo brain data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A real-imaginary 'real2' construction lets operator-informed Gaussian processes solve complex Helmholtz problems, reaching 0.77 correlation with measured brain curl wavefields but with uncalibrated posterior uncertainty.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid methodological paper with reproducible synthetic benchmarks; the brain-elastography claim is not yet isolated from boundary interpolation, so read that section with a skeptical eye. the 4 major comments →

arxiv 2607.14193 v2 pith:KUO3EYZV submitted 2026-07-15 stat.ML cs.LGcs.NAmath.NAphysics.med-ph

Operator-Informed Gaussian Processes for Complex Helmholtz Wavefields: From Synthetic Benchmarks to In Vivo Brain Elastography

Boyuan Deng , Kshitiz Upadhyay , Michael Shields This is my paper
classification stat.ML cs.LGcs.NAmath.NAphysics.med-ph
keywords complex Helmholtz equationGaussian process regressionphysics-informed machine learninguncertainty quantificationmagnetic resonance elastographyrealificationcoregionalizationmultiscale kernels
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Physics-informed Gaussian-process regression has been limited to real-valued fields; this paper removes that limit for the Helmholtz equation, which governs time-harmonic waves in dissipative media. The central move is a realification: write the complex operator Δ+κ² with κ²=a+ib as the coupled 2×2 real block [[Δ+a, −b],[b, Δ+a]], so real-valued GP conditioning applies unchanged. On closed-form problems in one to three dimensions the solver is comparable to deterministic baselines while using far fewer interior constraints and returning a posterior rather than a point estimate. On in vivo brain elastography, a proper two-scale prior reconstructs the shear-curl magnitude at Pearson correlation 0.767 versus a 0.75 acceptance target, with the gain coming from the multiscale kernel rather than real–imaginary coupling. The paper also reports a low-frequency accuracy ceiling set by model mismatch and posterior uncertainty that is not yet calibrated.

Core claim

The discovery is that complex-valued Helmholtz inference reduces to real-valued GP inference. Under the real–imaginary isomorphism, the C-linear Helmholtz operator becomes a coupled 2×2 block operator with equal diagonal blocks and opposite off-diagonal blocks; conditioning on PDE residuals and boundary traces through this block yields a posterior over both real and imaginary components. The paper claims this is not merely a mathematical identity: on benchmarks in one, two, and three dimensions, the posterior mean is comparable to finite-difference and physics-informed neural-network baselines at a much smaller constraint budget, and the operator-induced imaginary component is recovered even

What carries the argument

The realification map Φ: u ↦ (Re u, Im u) is the central object; it is an R-linear isomorphism that turns any complex-coefficient linear operator into a 2×2 real block operator. The central identity is the block form of the Helmholtz operator, with a=Re κ² and b=Im κ². The prior family moves from a diagonal proper (circular) prior through an intrinsic coregionalization model with a single scale to a multiscale linear model of coregionalization that sums kernels at fine and coarse length scales. Conditioning uses bounded linear information operators—Dirichlet boundary traces and PDE residuals at collocation points—so the posterior has closed-form mean and covariance. In the brain experiment,

Load-bearing premise

The brain reconstruction assumes each Cartesian component of the measured curl satisfies the scalar homogeneous Helmholtz equation with locally constant complex wavenumber, exact modulus maps, and a true Dirichlet boundary at the mask; the paper's own residual diagnostic (median |r|/|κ²q| ≈ 3–4, with fewer than ~6% of voxels below 1 at 70 Hz) indicates this model is strongly violated.

What would settle it

Solve a one-dimensional complex Helmholtz problem with a closed-form solution using both the real2 block formulation and a direct complex-valued Hermitian GP conditioning, and compare the posterior means and covariances; any difference beyond floating-point round-off would falsify the claimed equivalence.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For any complex-coefficient linear PDE, operator-informed GP regression now supplies a posterior over the complex field, not just a point estimate.
  • The real2 construction avoids implementing complex Hermitian adjoints; standard real-valued transpose-based conditioning suffices.
  • On measured brain data, a proper multiscale prior meets the acceptance target where a single-scale prior does not, and the paper attributes the gain to the multiscale base kernel.
  • The posterior uncertainty is not calibrated and cannot flag regions where the physics model is violated; the paper identifies calibrated uncertainty as the central next step.
  • In the 2D and 3D benchmarks, adding more collocation points barely changes accuracy, indicating that the prior and model specification, not the constraint budget, set the practical ceiling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the realification equivalence is exact, the same block construction should carry over to other complex wave equations, such as frequency-domain Maxwell or Schrödinger-type problems, without new inference machinery; the paper does not claim this.
  • The frequency trend implies a testable extension: augmenting the residual with modulus-gradient coupling terms should raise low-frequency agreement, since the paper's scaling argument identifies this as the dominant neglected term at 30 Hz.
  • Because the multiscale kernel rather than Re/Im coupling drives accuracy, spatially variable length scales matched to tissue type could improve reconstructions further; this is not explored in the paper.
  • The uncalibrated posterior suggests downstream Bayesian inversions should not treat these posterior widths as calibrated error bars until the noise and likelihood model is repaired; the paper does not say this explicitly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper extends operator-informed GP regression (LinPDE-GP) to complex-valued Helmholtz problems by mapping the complex Helmholtz operator to a coupled 2×2 real block ('real2' representation) and conditioning real-valued GPs on PDE residuals and boundary traces. A family of priors is introduced: a diagonal proper prior, an ICM prior with Re/Im coupling, and a multiscale LMC prior. Synthetic benchmarks in 1D, 2D, and 3D validate the construction against closed-form/eigenfunction solutions and compare with FDM and PINN baselines; in higher dimensions LinPDE-GP is less accurate than these baselines but uses far fewer interior constraints and returns a posterior. On in vivo brain MRE data, a multiscale LMC prior with length scales 4/16 mm reaches Pearson correlation 0.767 between predicted and measured curl magnitude, above an acceptance target of 0.75; an ablation attributes the gain to the multiscale kernel rather than Re/Im coupling. The paper also reports substantial model mismatch with the scalar homogeneous Helmholtz model and posterior uncertainties that are not calibrated.

Significance. If the central claim is established, the paper makes a useful contribution: it connects operator-conditioned GP regression, developed for real fields, to complex dissipative wave problems, and it demonstrates the construction on a real biomedical dataset with openly available data and code. The realification of the complex Helmholtz operator is algebraically sound and the synthetic validation against closed-form solutions is convincing. The authors are also commendably explicit about limitations: they report the large empirical Helmholtz residual, the weak calibration, and the constraint-efficiency nature of the baseline comparisons. The main weakness is evidential: the brain experiment does not isolate the contribution of the interior Helmholtz constraints from boundary interpolation, and the reported 'above target' correlation is based on selected hyperparameters and a spatially correlated sample without uncertainty quantification. These issues are fixable and central to the paper's headline real-data claim.

major comments (4)
  1. [§V, Tables V–VI and Fig. 7] The brain experiment lacks a boundary-only control. The claim that Helmholtz operator constraints drive the reconstruction requires comparing the reported 0.767 Pearson against conditioning on the same 2000 boundary-shell observations while omitting the 2000 interior residual constraints. Table VI ablates the prior but always retains the PDE constraints, and the FDM column imposes the full measured boundary in addition to the same Helmholtz model. Because the measured curl is smooth and spatially correlated, a GP interpolating the boundary trace could achieve high interior Pearson correlation without solving the Helmholtz equation. The 1D experiment in Fig. 2 shows that boundary-only conditioning leaves the interior largely prior-driven, so the authors recognize this phenomenon; the same diagnostic must be run on the in vivo data before the 'above target' result can be attributed to the
  2. [§V, Table VI and acceptance target] The reported success appears to depend on prior hyperparameters selected against the evaluation metric on the same data. Table VI shows that the 3/16 mm LMC gives Pearson 0.747 (below 0.75) while the 4/16 mm LMC gives 0.767 (above), and the 4/16 mm LMC is then presented as the champion reconstruction. No validation split, selection rule, or account of multiple comparisons is provided. The arbitrary 0.75 threshold is also pre-specified nowhere. A fair evaluation would report the selection procedure, correct for the selection, or demonstrate that the conclusion is robust across a range of length scales and thresholds. Without this, the conclusion 'clears the acceptance target' is not yet statistically grounded.
  3. [§V, Pearson metric and spatial autocorrelation] The agreement metric corr(|q_hat|, |q|) is computed on n=2000 interior sites that are spatially autocorrelated, so the effective sample size is far below 2000. No confidence interval, block-bootstrap interval, or sensitivity analysis is given. The difference between 0.767 and the 0.75 target, or between 0.767 and 0.747, is not interpretable without such uncertainty quantification. The paper should report a bootstrap or subsampling-based interval, and ideally also a metric less sensitive to smooth large-scale structure (e.g., residual spectra or scale-decomposed correlation).
  4. [§V–§VI, model mismatch diagnostics] The paper's own diagnostic shows the scalar homogeneous Helmholtz model is strongly violated: the median |r|/|κ2q| is roughly 3–4 at 70 Hz and fewer than about 6% of interior voxels satisfy |r|/|κ2q| < 1, with the 30 Hz residual even larger. The Limitations section acknowledges this, but the abstract and introduction state the reconstruction 'reconstructs the shear curl field' as a physics-based result. Given this mismatch, the reconstruction agreement may reflect boundary information and smoothness rather than the validity of the Helmholtz operator. The boundary-only control requested above is therefore not a formality but a necessary check on what the reconstruction is actually testing.
minor comments (5)
  1. [§IV.C–D, Tables III–IV] In the 2D and 3D synthetic benchmarks LinPDE-GP trails the FDM and PINN by roughly two orders of magnitude in relative ℓ2 error. This is acknowledged, but the abstract's phrase 'competitive with finite-difference and neural-network baselines' overstates the observed comparison; consider phrasing that reflects the constraint-efficiency trade-off more precisely.
  2. [§V, Table V] The paper states the full 3×3×3 grid was computed but only OFAT variations about one reference are reported. Since the full grid exists, a compact heatmap or table of all 27 Pearson values would be more informative and would allow readers to see interaction effects, which are currently excluded without analysis.
  3. [§III.B.2, Eq. (23)] The ICM prior is described as 'improper (noncircular)' because kRI = kIR = ρRI k. It may be helpful to give the explicit complex covariance/pseudo-covariance relations, so readers can immediately connect the real2 construction to the standard proper/improper complex-GP literature cited in §I.
  4. [Fig. 7 caption] The caption reports Pearson 0.753 for subject 0003 while the reference configuration in Table V is subject 0001 with Pearson 0.767. The figure and table are consistent, but the relationship should be stated explicitly to avoid confusion.
  5. [§V, Table VI and posterior calibration] The calibration diagnostic corr(σ, r) is said to be identical across the three curl components because the posterior variance is observation-independent. This is an important and unusual property; it deserves a brief explanation in the main text, not only in the paragraph where it is introduced, since it sharply limits the usefulness of the reported credible intervals.

Circularity Check

0 steps flagged

No significant circularity: real2 construction is an exact algebraic equivalence, synthetic benchmarks are externally validated, and brain-experiment concerns are design limitations rather than reductions to inputs.

full rationale

The central construction (realified 2x2 Helmholtz block, Eq. 33) is an exact R-linear isomorphism with the complex Helmholtz operator, so conditioning on the realified system is algebraically identical to conditioning on the complex PDE; no predicted quantity is defined in terms of the target metric. The LinPDE-GP conditioning formulas are imported from an external, code-released framework (Pfoertner et al.), not from the authors' own prior work, and the synthetic benchmarks in Sec. IV are checked against closed-form eigenfunction expansions and independent FDM/PINN baselines, so the solver's accuracy is not an artifact of its own assumptions. In the brain experiment (Sec. V), the reported Pearson correlation is computed between the posterior mean and measured curl magnitudes at interior sites where no field values are directly conditioned; the PDE-residual constraints at those sites are a model input, not the measured quantity, so the agreement is a genuine (if model-dependent) out-of-sample check. The paper explicitly discloses that the 0.75 acceptance target is internal, that the scalar local-homogeneity model is strongly violated (median |r|/|kappa^2 q| ~ 3-4), and that posterior uncertainty is uncalibrated. The absence of a boundary-only control and the possible selection of LMC length scales (4/16 vs 3/16) are experimental-design and statistical concerns, not circularity: no equation reduces to its input, and no load-bearing claim rests solely on a self-citation. Self-citations to Bayly et al. and Upadhyay et al. are for publicly available data and a standard density value, not for the validity of the method.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

No new physical entities are introduced; the 'real2' representation is a mathematical construction. The central claim rests on standard GP conditioning, the realification of the Helmholtz operator, and domain assumptions specific to the brain experiment (local homogeneity, exact modulus maps, Dirichlet boundary at the mask). The free parameters are prior hyperparameters and the acceptance threshold, all chosen by hand.

free parameters (5)
  • Kernel length scales for brain priors (IID 15mm, LMC 3/16mm, 4/16mm) = 15, 3/16, 4/16 mm
    Chosen by hand; the 4/16 LMC clears the acceptance target while 3/16 does not, suggesting selection toward the metric.
  • Re/Im correlation ρRI in ICM/LMC ablation = 0, 0.6
    Hand-set to test coupling effect.
  • Kernel amplitude σ_f and length scales in synthetic benchmarks = σ_f = 2 or 4, ℓ = 1.0 or 0.7
    Set by hand to match expected field magnitude; not fitted to test error.
  • Observation jitter variances (boundary and PDE) = 1e-7 to 1e-4
    Chosen for numerical conditioning of the Gram matrix.
  • Acceptance target Pearson > 0.75 = 0.75
    Arbitrary threshold set by authors; no external standard or pre-registration.
axioms (7)
  • standard math GP conditioning formulas for linear information operators (Eqs. 8-9) are valid
    Taken from Pförtner et al. 23; used throughout the method.
  • standard math The real2 representation Φ : C → R² is an R-linear isomorphism and the Helmholtz operator realifies to [[Δ+a, -b],[b, Δ+a]]
    Standard algebraic fact; explicitly verified in Sec. III.C.
  • standard math The complex GP prior is proper (zero pseudo-covariance) and the noise is Gaussian, so conditioning yields a GP posterior
    Standard complex GP theory (Neeser & Massey, Schreier & Scharf).
  • domain assumption Each Cartesian component of the curl field q satisfies a scalar Helmholtz equation under local homogeneity
    Sec. II and V; limitations state that modulus gradients are neglected.
  • domain assumption The brain mask surface acts as a true Dirichlet boundary with measured trace, and the modulus maps κ²(x) are exact
    Sec. V and Limitations; boundary idealizations are untested.
  • standard math The sine eigenfunction expansion (Eq. 34) is a valid closed-form reference for constant-coefficient benchmarks
    Separable solution of the Helmholtz equation with constant forcing and homogeneous Dirichlet data.
  • domain assumption The viscoelastic constitutive model κ² = ρω²/(G′+iG′′) governs the benchmarks and brain experiment
    Standard linear viscoelastic model; used to set benchmark coefficients and κ²(x).

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Operator-Informed Gaussian Processes for Complex Helmholtz Wavefields: From Synthetic Benchmarks to In Vivo Brain Elastography." pith.science (2026). https://pith.science/paper/KUO3EYZV

@misc{pith2026260714193,
  author       = {Pith},
  title        = {Pith review of: Operator-Informed Gaussian Processes for Complex Helmholtz Wavefields: From Synthetic Benchmarks to In Vivo Brain Elastography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUO3EYZV}},
  note         = {Machine review of arXiv:2607.14193}
}
Share X Bluesky LinkedIn Reddit HN
abstract

The Helmholtz equation governs time-harmonic wave propagation, and in dissipative media a complex modulus renders its squared wavenumber $\kappa^2$ complex. Inferring such fields from sparse, noisy data calls for solvers that also quantify their own uncertainty. Physics-informed Gaussian-process (GP) regression supplies this by returning a posterior over the solution, yet operator-conditioned formulations have been developed almost exclusively for real-valued fields. We extend operator-informed GP regression to complex-valued Helmholtz problems by realifying the complex operator into an equivalent coupled real block, which enables inference with standard real-valued GP conditioning. The construction admits a family of priors, from a proper diagonal prior to coregionalized and multiscale variants, and conditions on PDE residuals and boundary traces. On benchmark problems in one to three dimensions, the solver is competitive with finite-difference and neural-network baselines at a far smaller interior-constraint budget. Unlike those deterministic baselines, it returns a posterior over the complex wavefield rather than a point estimate. Applied to \textit{in vivo} brain magnetic resonance elastography, a proper multiscale prior reconstructs the shear curl field to a correlation of $0.77$ with measurement, above a $0.75$ target. The gain arises from the multiscale kernel rather than from real--imaginary coupling. We further identify a low-frequency accuracy ceiling set by model mismatch and a posterior uncertainty that is not yet calibrated. Calibrated uncertainty therefore emerges as the central next step for probabilistic wavefield inference in dissipative media.

Figures

Figures reproduced from arXiv: 2607.14193 by Boyuan Deng, Kshitiz Upadhyay, Michael Shields.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 2 canonical work pages

  1. [1]

    The Journal of the Acoustical Society of America , volume=

    Helmholtz and parabolic equation solutions to a benchmark problem in ocean acoustics , author=. The Journal of the Acoustical Society of America , volume=. 2003 , publisher=

    2003

  2. [2]

    2005 , publisher=

    Fast multipole methods for the Helmholtz equation in three dimensions , author=. 2005 , publisher=

    2005

  3. [3]

    Rasmussen, Carl Edward and Williams, Christopher K. I. , biburl =

  4. [4]

    2024 , eprint=

    Physics-Informed Gaussian Process Regression Generalizes Linear PDE Solvers , author=. 2024 , eprint=

    2024

  5. [5]

    2008 , publisher=

    Mechanical behavior of materials , author=. 2008 , publisher=

    2008

  6. [6]

    2006 , publisher=

    Pattern recognition and machine learning , author=. 2006 , publisher=

    2006

  7. [7]

    IEEE transactions on information theory , volume=

    Proper complex random processes with applications to information theory , author=. IEEE transactions on information theory , volume=. 2002 , publisher=

    2002

  8. [8]

    2010 , publisher=

    Statistical signal processing of complex-valued data: the theory of improper and noncircular signals , author=. 2010 , publisher=

    2010

  9. [9]

    Complex Gaussian Processes for Regression , year=

    Boloix-Tortosa, Rafael and Murillo-Fuentes, Juan José and Payán-Somet, Francisco Javier and Pérez-Cruz, Fernando , journal=. Complex Gaussian Processes for Regression , year=

  10. [10]

    Gryazin and Michael V

    Yuriy A. Gryazin and Michael V. Klibanov and Thomas R. Lucas , keywords =. Two numerical methods for an inverse problem for the 2-D Helmholtz equation , journal =. 2003 , issn =. doi:https://doi.org/10.1016/S0021-9991(02)00023-2 , url =

    work page doi:10.1016/s0021-9991(02)00023-2 2003

  11. [11]

    2017 , publisher=

    Modern solvers for Helmholtz problems , author=. 2017 , publisher=

    2017

  12. [12]

    Computers & Mathematics with Applications , volume=

    Finite element solution of the Helmholtz equation with high wave number Part I: The h-version of the FEM , author=. Computers & Mathematics with Applications , volume=. 1995 , publisher=

    1995

  13. [13]

    Journal of Computational Physics , volume=

    An optimized CIP-FEM to reduce the pollution errors for the Helmholtz equation on a general unstructured mesh , author=. Journal of Computational Physics , volume=. 2024 , publisher=

    2024

  14. [14]

    Archives of Computational Methods in Engineering , volume=

    Advances in iterative methods and preconditioners for the Helmholtz equation , author=. Archives of Computational Methods in Engineering , volume=. 2008 , publisher=

    2008

  15. [15]

    Journal of Computational Physics , volume=

    A wideband fast multipole method for the Helmholtz equation in three dimensions , author=. Journal of Computational Physics , volume=. 2006 , publisher=

    2006

  16. [16]

    Journal of Computational Physics , volume=

    A fast direct solver for boundary integral equations in two dimensions , author=. Journal of Computational Physics , volume=. 2005 , publisher=

    2005

  17. [17]

    The Journal of the Acoustical Society of America , volume=

    Time-domain analysis of power law attenuation in space-fractional wave equations , author=. The Journal of the Acoustical Society of America , volume=. 2018 , publisher=

    2018

  18. [18]

    Ultrasound in medicine & biology , volume=

    Comparison of fractional wave equations for power law attenuation in ultrasound and elastography , author=. Ultrasound in medicine & biology , volume=. 2014 , publisher=

    2014

  19. [19]

    Geophysical Journal International , volume=

    Wave propagation simulation in a linear viscoelastic medium , author=. Geophysical Journal International , volume=. 1988 , publisher=

    1988

  20. [20]

    Journal of Geophysical Research: Solid Earth , volume=

    Constant Q-wave propagation and attenuation , author=. Journal of Geophysical Research: Solid Earth , volume=. 1979 , publisher=

    1979

  21. [21]

    Nature Reviews Physics , volume=

    Physics-informed machine learning , author=. Nature Reviews Physics , volume=. 2021 , publisher=

    2021

  22. [22]

    Acta Mechanica Sinica , volume=

    Physics-informed neural networks (PINNs) for fluid mechanics: A review , author=. Acta Mechanica Sinica , volume=. 2021 , publisher=

    2021

  23. [23]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  24. [24]

    Journal of Computational Physics , volume=

    B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data , author=. Journal of Computational Physics , volume=. 2021 , publisher=

    2021

  25. [25]

    Journal of Scientific Computing , volume=

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next , author=. Journal of Scientific Computing , volume=. 2022 , publisher=

    2022

  26. [26]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

    Probabilistic numerics and uncertainty in computations , author=. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=. 2015 , publisher=

    2015

  27. [27]

    AIP Conference Proceedings , volume=

    Probabilistic numerical methods for PDE-constrained Bayesian inverse problems , author=. AIP Conference Proceedings , volume=. 2017 , organization=

    2017

  28. [28]

    Proceedings , volume=

    Gaussian processes for data fulfilling linear differential equations , author=. Proceedings , volume=. 2019 , organization=

    2019

  29. [29]

    Signal Processing , volume=

    Complex-valued gaussian process regression for time series analysis , author=. Signal Processing , volume=. 2019 , publisher=

    2019

  30. [30]

    IEEE transactions on neural networks and learning systems , volume=

    Complex Gaussian processes for regression , author=. IEEE transactions on neural networks and learning systems , volume=. 2018 , publisher=

    2018

  31. [31]

    , booktitle=

    Tobar, Felipe and Turner, Richard E. , booktitle=. Modelling of complex signals using gaussian processes , year=

  32. [32]

    Thomas-Seale and L

    L.E.J. Thomas-Seale and L. Hollis and D. Klatt and I. Sack and N. Roberts and P. Pankaj and P.R. Hoskins , keywords =. The simulation of magnetic resonance elastography through atherosclerosis , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.jbiomech.2016.04.013 , url =

    work page doi:10.1016/j.jbiomech.2016.04.013 2016

  33. [33]

    Physics in Medicine & Biology , volume=

    Effects of frequency-and direction-dependent elastic materials on linearly elastic MRE image reconstructions , author=. Physics in Medicine & Biology , volume=

  34. [34]

    Annals of biomedical engineering , volume=

    MR imaging of human brain mechanics in vivo: new measurements to facilitate the development of computational models of brain injury , author=. Annals of biomedical engineering , volume=. 2021 , publisher=

    2021

  35. [35]

    Journal of the Royal Society Interface , volume=

    Development and validation of subject-specific 3D human head models based on a nonlinear visco-hyperelastic constitutive framework , author=. Journal of the Royal Society Interface , volume=

  36. [36]

    1997 , publisher=

    Geostatistics for natural resources evaluation , author=. 1997 , publisher=

    1997

  37. [37]

    Foundations and Trends

    Kernels for vector-valued functions: A review , author=. Foundations and Trends. 2012 , publisher=

    2012

  38. [38]

    Mining geostatistics , author=

  39. [39]

    PloS one , volume=

    Measuring the characteristic topography of brain stiffness with magnetic resonance elastography , author=. PloS one , volume=. 2013 , publisher=

    2013

  40. [40]

    The American Statistician , volume=

    A note on the delta method , author=. The American Statistician , volume=. 1992 , publisher=

    1992

  41. [41]

    DeepXDE: A Deep Learning Library for Solving Differential Equations , volume=

    Lu, Lu and Meng, Xuhui and Mao, Zhiping and Karniadakis, George Em , year=. DeepXDE: A Deep Learning Library for Solving Differential Equations , volume=. SIAM Review , publisher=. doi:10.1137/19m1274067 , number=

    doi:10.1137/19m1274067

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.