Pith. sign in

REVIEW 4 major objections 8 minor 68 references

Stability of Electrical Impedance Tomography with Anisotropies and its Application to the Deep Calde\'on Method

T0 review · 4 major / 8 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that electrical impedance tomography with anisotropies is Lipschitz stable for a restricted but infinite-dimensional class of conductivity perturbations dominated by low frequencies, and uses this to explain when…

desk verdict A worthwhile 2D stability theorem for anisotropic EIT, but the numerical application to the deep Calderón method does not test the theorem's admissible set and should be reconsidered. read the letter →

arxiv 2608.03053 v1 pith:JJM57VLD submitted 2026-08-04 math.AP

Tianhao Hu , Bangti Jin , Yiran Wang This is my paper · ORCID
classification math.AP MSC 35R3065N21
keywords electricalimpedancetomographyanisotropicconductivityLipschitzstabilitydeepCalderónmethodkernelawarenessDirichlet-to-Neumannmapconditionalinverseproblems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proves a conditional Lipschitz stability estimate for anisotropic electrical impedance tomography: within an admissible class of smooth, compactly supported conformal perturbations whose amplitude is controlled by the $L^2$ norm of their low-frequency Laplacian, the size of the conductivity perturbation is bounded by a constant times the discrepancy between the boundary Dirichlet-to-Neumann maps, with the constant independent of the perturbation. The same conditional stability is established in higher dimensions through a conformal metric and a Schrödinger-equation reformulation. The authors then use this stability to analyze the deep Calderón method, a two-stage deep-learning reconstruction pipeline: if training data come from such a low-frequency-dominated class with a small stability constant, a well-trained network is Lipschitz stable and kernel-aware, and severe ill-posedness appears only outside the class. Numerical experiments with Gaussian bumps and smoothed disks illustrate the predicted trade-off: broad or small features reconstruct stably, while sharp, large, or out-of-distribution features degrade.

What carries the argument

The load-bearing object is the admissible set $E_{M,N}=\{\delta\in C^2(\Omega): \operatorname{supp}(\delta)\Subset\Omega,\ \|\delta\|_{L^\infty(\Omega)}<M,\ \|\delta\|_{L^\infty(\Omega)}\leq N\|\chi(D)\Delta\delta\|_{L^2(\mathbb{R}^2)}\}$, with $\chi(D)$ the Fourier multiplier by the characteristic function of the unit disc. This condition ensures that what the boundary data can see—the low-frequency Laplacian of the perturbation—controls the perturbation's full amplitude. The proof's workhorse is the boundary identity $\int_{\partial\Omega} u_1(\Lambda_\gamma-\Lambda_{\gamma_0})u_2\,dS=-\frac{|k|^2}{2}\widehat{\delta}(-k)$ plus error terms, where $u_1,u_2$ are complex exponential harmonic functions tuned to wave vector $k$; Plancherel's theorem converts the resulting estimate into a bound on $\|\chi(D)\Delta\delta\|_{L^2}$, and the condition $C_1N^2<1/2$ closes the argument. In dimensions $n\geq3$, the machinery shifts to the conformal metric $g=(\det\gamma)^{1/(n-2)}\gamma^{-1}$, the associated Schrödinger equation with potential $q$, complex geometric optics solutions, and an analogous low-frequency admissibility condition on $q$.

What would settle it

Search numerically over smooth, compactly supported perturbations in $E_{M,N}$ with a fixed small $N$, compute the finite-element Dirichlet-to-Neumann discrepancy for each, and minimize the ratio $\|\delta\|_{L^\infty(\Omega)}/\|\Lambda_\gamma-\Lambda_{\gamma_0}\|_*$; if a sequence inside $E_{M,N}$ makes this ratio tend to infinity, or if two perturbations in the class have arbitrarily small boundary discrepancy but nonnegligible difference, the Lipschitz estimate is false. Verifying that the ratio stays bounded for such a family would support the theorem quantitatively.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central result is Theorem 2.1: for conductivities of the form $\gamma=(1+\delta)(I+A)$, where $A$ is a known elliptic anisotropy and $\delta$ lies in the admissible set $E_{M,N}$, for $N>0$ sufficiently small there holds $\|\delta\|_{L^\infty(\Omega)}\leq C\|\Lambda_\gamma-\Lambda_{\gamma_0}\|_*$ with $C$ independent of $\delta$. The admissible set consists of smooth, compactly supported perturbations satisfying $\|\delta\|_{L^\infty(\Omega)}\leq N\|\chi(D)\Delta\delta\|_{L^2(\mathbb{R}^2)}$, where $\chi(D)$ is the Fourier multiplier cutting to the unit disc; the condition makes the amplitude of $\delta$ visible through its low-frequency Laplacian. The proof reconstructs the low-frequency Fourier transform of $\delta$ from boundary integrals of complex exponential harmonic functions in the manner of Calderón's original method, bounds the error terms arising from the known anisotropy and from the nonlinearity, and applies Plancherel's theorem. In dimensions $n\geq 3$, the analogous statement is proved by transforming the conductivity equation to a Schrödinger equation on a conformally related metric, using complex geometric optics solutions, and placing the admissibility condition on the potential $q$. For the deep Calderón method, the stability estimate is interpreted as a sufficient condition on the training data: a well-trained postprocessing network is Lipschitz stable, and therefore kernel-aware, on the admissible class, with ill-posedness relegated to data outside it.

Load-bearing premise

Everything rests on assuming the unknown perturbation $\delta$ satisfies the a priori bound $\|\delta\|_{L^\infty(\Omega)}\leq N\|\chi(D)\Delta\delta\|_{L^2(\mathbb{R}^2)}$ with $N$ sufficiently small; this low-frequency-amplitude condition is strong, cannot be verified from boundary measurements alone, and the stability claim simply does not apply to perturbations outside it.

Editorial extensions

If this is right

  • On the admissible class $E_{M,N}$ with $N$ sufficiently small, the conductivity perturbation is recovered from boundary measurements with error at most a constant times the measurement discrepancy, so inversion there is Lipschitz stable.
  • A well-trained deep Calderón postprocessor trained on such a class is kernel-aware and stable on that class; the severe ill-posedness of EIT only reappears when the network is evaluated on data outside the class.
  • For Gaussian perturbations, stability requires the width parameter $b$ to be small, meaning smooth broad features are stable while sharper or finer features are unstable, as the numerical experiments show.
  • For smoothed piecewise-constant features, smaller objects keep the stability parameter small more easily, and reconstructions degrade for larger or out-of-distribution objects.
  • The truncation radius $R$ in the initial Calderón reconstruction step should be kept small, because the stability constant grows exponentially with $R$, forcing $N$ to be exponentially small for large $R$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: because the admissibility condition is an a priori smoothness and low-frequency constraint that boundary measurements alone do not reveal, the practical working rule is to train on low-frequency-dominated images and to distrust network outputs on sharper or more oscillatory inputs, even when those inputs resemble the training set.
  • Beyond the paper: the numerical contrast between sign-changing and sign-definite perturbations suggests a testable refinement of the stability theory—single-signed perturbations may admit a smaller stability constant, and computing the ratio $\|\delta\|_{L^\infty(\Omega)}/\|\Lambda_\gamma-\Lambda_{\gamma_0}\|_*$ for the two families would indicate whether the sign restriction relaxes the admissibl
  • Beyond the paper: the paper's explanation for failed two-disk reconstructions (Fourier cancellation between nearby disks) can be tested directly by evaluating the required $N$ in the admissible-set condition for pairs of smoothed disks; if $N$ grows as the separation shrinks, the stability class itself predicts the resolution limit independently of the trained network.
  • Beyond the paper: one could compose the Lipschitz stability on $E_{M,N}$ with the Lipschitz constant of a trained postprocessor to obtain an explicit noise-robustness bound for the deep Calderón method, converting the qualitative kernel-awareness argument into a measurable error estimate that depends on the training distribution.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper studies local inversion for anisotropic electrical impedance tomography in a conformal class, writing the conductivity as γ=(1+δ)(I+A) with a known anisotropy matrix A. The main analytic result, Theorem 2.1 in two dimensions, asserts a conditional Lipschitz estimate ‖δ‖_{L∞(Ω)} ≤ C‖Λγ−Λγ0‖_* for perturbations δ in an admissible set E_{M,N} defined by ‖δ‖_{L∞} ≤ N‖χ(D)Δδ‖_{L2(R^2)} with N sufficiently small. Section 4 extends the result to dimensions n≥3 by transforming the conductivity equation to a Schrödinger equation and using complex geometric optics, with the admissible condition imposed on the potential q rather than on δ. The authors then relate the stability estimates to the robustness of the deep Calderón method, arguing that training data satisfying the admissible condition promote kernel awareness, and they report numerical experiments with Gaussian and piecewise-constant conductivities.

Significance. The two-dimensional proof of Theorem 2.1 is self-contained and appears correct; the paper establishes a genuinely infinite-dimensional, non-finite-dimensional conditional Lipschitz stability class for anisotropic perturbations, which is a useful contribution to the EIT stability literature. The explicit discussion in Remark 3.4 of how the admissible constant degenerates with the Fourier cutoff is also valuable. However, the multi-dimensional extension contains a potentially fatal vacuousness in the admissible set, the numerical experiments do not sample the admissible set of the theorem, and the advertised application to the deep Calderón method is heuristic rather than proven. If the multi-dimensional issues are resolved and the deep-learning link is properly qualified, the two-dimensional theorem alone would be a publishable contribution.

major comments (4)
  1. [§4, Eq. (4.7), Proposition 4.3] The admissible set D_N is effectively trivial for small N. For q∈D_N, the condition ‖q‖_{L∞(Ω)} ≤ N‖χ(D)q‖_{L2(R^n)} combined with the contraction property ‖χ(D)q‖_{L2(R^n)} ≤ ‖q‖_{L2(Ω)} ≤ |Ω|^{1/2}‖q‖_{L∞(Ω)} implies (1 − N|Ω|^{1/2})‖q‖_{L∞(Ω)} ≤ 0, so D_N contains only q=0 whenever N < |Ω|^{-1/2}. The proof of Proposition 4.3 chooses N sufficiently small so that N^2 C e^{Cτ} < 1/2 with τ > M/ε; for a fixed domain such as the unit ball in R^3, where |Ω|^{-1/2} ≈ 0.49, this requirement can force N below the triviality threshold, rendering the proposition vacuous. The paper must quantify the constants and either exhibit a nontrivial admissible set compatible with the smallness condition or modify the admissible condition, for example by using a high-order derivative of q in the low-frequency norm.
  2. [§4, Lemma 4.2 and proof of Proposition 4.3] The proof repeatedly inverts the operator −Δ_{g0}+q with zero Dirichlet boundary condition, both to define v in Lemma 4.2 and in the elliptic estimates of Proposition 4.3. The text asserts that zero is not an eigenvalue for g,q derived from the conductivity problem, but Proposition 4.3 is stated for arbitrary q∈D_N with ‖q‖_{H^s(Ω)}≤M, and no invertibility argument is given for this whole class. Since the smallness of ‖q‖_{L∞} is not guaranteed by D_N, a zero eigenfunction cannot be excluded. In addition, the equation for v in Lemma 4.2 has the wrong sign: since −Δ_{g0}w = −Δw + P(x,∂)w, the right-hand side should be −P(x,∂)u, not P(x,∂)u; the sign error does not affect the subsequent bound, but the missing invertibility argument is load-bearing for the multi-dimensional theorem.
  3. [§5, Examples 5.1 and 5.2] The numerical experiments do not test the admissible set of Theorem 2.1. In Example 5.1 the Gaussian parameters satisfy b∈(30,50) with Fourier cutoff R=1.8; substituting into (2.6) gives ̲N of order 10–100, whereas Theorem 2.1 and Remark 3.4 require N to be small, in fact exponentially small in R. In Example 5.2 the test and training perturbations are discontinuous characteristic functions, so δ∉C^2(Ω) and Δδ∉L^2, violating the hypotheses of (2.4); the Gaussian smoothing described in §2.3 is not applied in the experiments. The observed trends, such as better stability for smaller b or smaller support, may be suggestive, but they are not a validation of the theorem. The paper should either generate training data lying in E_{M,N} or explicitly state that the numerical section is illustrative rather than a test of the theory.
  4. [§2.1–§2.3, Section 6] The link between Theorem 2.1 and the robustness of the deep Calderón method is not a theorem. Theorem 2.1 concerns the map Λγ↦δ, whereas the deep Calderón method composes a truncated Fourier inversion with a trained U-net; the paper does not show that a U-net trained on data from E_{M,N} has a controlled Lipschitz constant or is kernel aware in the precise sense of [20]. The argument in §2.2 relies on an unquantified universal-approximation assumption and on the phrase 'well trained', and no condition on the training distribution or on the network's Lipschitz constant is derived from the stability estimate. Accordingly, the abstract and Section 6 statements that the stability theory 'relates to' or provides 'theoretical underpinnings' for the deep Calderón method should be substantially qualified.
minor comments (8)
  1. [Title and abstract] The name 'Caldeón' appears to be a typo; it should read 'Calderón' throughout the title and abstract.
  2. [§2.2, Eq. (2.4)] The admissible set E_{M,N} is defined for δ∈C^2(Ω) with supp(δ)⋐Ω, but the norm ‖χ(D)Δδ‖_{L2(R^2)} requires an extension of δ to R^2; the paper should state explicitly that δ is extended by zero outside Ω.
  3. [§2.2, Theorem 2.1] The theorem says C is independent of δ, but the proof gives a constant that depends on the chosen N and on M, A, and Ω; the statement should clarify that C may depend on the admissible-set parameters and on the a priori data but not on the particular perturbation.
  4. [§4, Eq. (4.4)] The potential q is written using Δ_g; since the fixed metric in the Schrödinger problem is g0, the formula should be rewritten in terms of Δ_{g0} (or a precise convention stated) so that q is a potential on a fixed metric.
  5. [§4, Proposition 4.3] The admissible set D_N is stated for q∈L∞(Ω), but the proof uses q∈H^s(R^n); the paper should specify the compact extension by zero and the relation between the Sobolev norms on Ω and R^n.
  6. [§2.3, Fig. 1 and Eqs. (2.6), (2.7)] The symbol ̲N is used both as the lower bound and as a variant of the parameter N, which is confusing; a distinct notation such as N_0 or N_min would be clearer.
  7. [§2.2, Remark 3.4 and §3] The constants C0 and C1 in the proof of Theorem 2.1 are not explicitly defined; for the claimed explicit exponential dependence in Remark 3.4, the paper should state how they depend on the cutoff R, on M, and on the a priori bounds for A and Ω.
  8. [§5, Example 5.2] The symbol χ is used both for the Fourier cutoff χ(D) and as the characteristic function of the ball {x:|x−x0|<β}; the notation should be disambiguated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the stability estimates are derived from Calderón identities and elliptic estimates, not from the deep-learning application or fitted parameters.

full rationale

The derivation chain of Theorem 2.1 is self-contained: the proof starts from Calderón's original identity (Section 3, eq. (3.8)) and uses elliptic estimates (Lemmas 3.1 and 3.3) to bound the error terms E1 and E2, obtains the Fourier-domain inequality (3.9), and only then invokes the admissible-set condition (2.4) via Plancherel's theorem to close the estimate. The condition ||δ||_L∞ ≤ N ||χ(D)Δδ||_L2 is an a priori restriction on the perturbation, not an assumption of the conclusion; the final bound ||δ||_L∞ ≤ C ||Λγ - Λγ0||_* is a genuine conditional Lipschitz estimate whose right-hand side is boundary data. The multi-dimensional result (Proposition 4.3 and Theorem 4.4) is similarly derived from CGO estimates of Sylvester and Uhlmann and the conformal Schrödinger reduction, with the same a priori low-frequency control on q rather than on the target boundary data. The deep Calderón method is used as the application context, but no result about the trained U-net is imported into the proof of the stability theorems, and the numerical experiments serve as qualitative illustrations; the paper itself flags in Section 5.2 that sampling from E_{M,N} for neural network training is not clear, which is a scope limitation rather than a circular step. Self-citations such as [16] or [63] identify prior methods and related stability mechanisms, but they are not load-bearing for the paper's own estimates. No fitted parameter is renamed as a prediction, and no known result is merely relabelled. Hence no circular step can be exhibited.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard PDE tools and CGO theory (standard math), plus two domain assumptions: ellipticity of the known background conductivity and absence of zero eigenvalues for the Schrödinger reformulation. No free parameters are fitted to data; M and N are user-chosen thresholds in the admissible set. The training-data assumption is a practical leap not formally established.

assumptions (6)
  • domain assumption The known background conductivity γ0 = I + A is strictly elliptic, i.e., ∑(Iij + Aij)ξiξj ≥ c|ξ|^2 (Eq. 3.1).
    Needed in Lemma 3.1 to ensure well-posedness and estimates for the background problem.
  • standard math Standard elliptic regularity, existence and uniqueness for second-order uniformly elliptic Dirichlet problems (Gilbarg-Trudinger, Theorems 6.14 and 8.12).
    Invoked in Lemma 3.1 and Lemma 3.3 to obtain solutions and H^1/H^2 estimates.
  • standard math Existence of complex geometric optics solutions for the Schrödinger equation -Δu + qu = 0 for small q (Sylvester-Uhlmann, Corollary 2.5).
    Used as Theorem 4.1 in the multidimensional proof to construct CGO solutions.
  • standard math Conformal transformation identities for the Dirichlet-to-Neumann map under a change of metric (Lee-Uhlmann; Dos Santos Ferreira et al., Proposition 8.2).
    Used in Section 4 to rewrite the anisotropic conductivity problem as a Schrödinger equation with a known metric.
  • domain assumption The Schrödinger operator -Δ_{g0} + q obtained from a conductivity satisfies the zero-eigenvalue exclusion condition (0 is not a Dirichlet eigenvalue).
    Asserted in Section 4 just before (4.3) to guarantee well-posedness; no proof is provided in the preprint.
  • domain assumption The training data in the numerical experiments belong (approximately) to the admissible set E_{M,N} with small N.
    Required for the stability theory to apply to the deep Calderón method; the paper states in Section 5.2 that it is 'not clear how to sample functions from E_{M,N}' and does not verify the condition for the generated data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stability of Electrical Impedance Tomography with Anisotropies and its Application to the Deep Calde\'on Method." pith.science (2026). https://pith.science/paper/JJM57VLD

@misc{pith2026260803053,
  author       = {Pith},
  title        = {Pith review of: Stability of Electrical Impedance Tomography with Anisotropies and its Application to the Deep Calde\'on Method},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JJM57VLD}},
  note         = {Machine review of arXiv:2608.03053}
}
read the original abstract

In this work, we establish new conditional Lipschitz stability results for electrical impedance tomography (EIT) with anisotropies, of recovering the conductivity in a conformal class of a known anisotropic conductivity in both two- and multi-dimensional cases. Then we employ the stability theory to understand the property of the deep Calder\'on method, one deep learning-based technique for image reconstruction in EIT that has shown promising empirical results, but still lacks theoretical underpinnings. Specifically, we relate the stability theory to the robustness of the method with the proper choice of the training data, and present numerical results in two-dimension to complement the theoretical analysis.

Figures

Figures reproduced from arXiv: 2608.03053 by the authors.

Figure 1
Figure 1. The relation between the lower bound N of N, and b for the perturbation by a Gaussian bump (given by (2.6)) and perturbation by a characteristic function (given by (2.7)) with various values of a. = (2π) −2 Z 1 0 4π 2a 2 b 2 2πe− r 2 b r 3 dr = 2πa2 Z √1 b 0 e −s 2 s 3 ds. Thus, for δ ∈ EM,N , there holds N ≥ kδkL∞(Ω) kχ(D)∆δkL2(R2) = 1 (2π R √1 b 0 e−s 2 s 3 ds) 1 2 := N. (2.6) We plot the graph of the lower bound … view at source ↗
Figure 2
Figure 2. The structure of U-net fθ. 5.2 Numerical results and discussions In this part, we present numerical results for two examples corresponding to the settings in Section 2.3. The experiments are designed to verify the stability/instability analysis and test the generalization ability of the deep Calderón method. We remark that it is generally not clear how to sample functions from the admissible set EM,N in Theorem 2.1 … view at source ↗
Figure 3
Figure 3. The true image, Calderón reconstruction, deep reconstructions (after 20, 40 and 400 epochs) and slice [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The true image, Calderón reconstruction, deep reconstructions (after 20 and 200 epochs) and slices at [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: The true image, Calderón reconstruction, deep reconstructions and slice at [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: The dynamics of the training error and validation loss for Example [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: The true image, Calderón reconstruction, deep reconstructions (after 100, 200 and 800 epochs) and [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: The true image, Calderón reconstruction, deep reconstructions (after 100 and 400 epochs) and slice at [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: The true image, Calderón reconstruction, deep reconstructions and slice at [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: The dynamics of the training error and validation loss for Example [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 66 canonical work pages

  1. [20]

    M. J. Colbrook, V. Antun, and A. C. Hansen , The difficulty of computing stable and accurate neural networks: On the barriers of deep learning and Smale’s 18th problem , Proc. Natl. Acad. Sci. USA, 119 (2022), p. e2107151119

    work page 2022

  2. [1]

    Alberti, A

    G. Alberti, A. Arroyo, and M. Santacesaria , Inverse problems on low-dimensional manifolds , Nonlin- earity, 36 (2023), pp. 734–808

    work page 2023

  3. [2]

    G. S. Alberti, R. Petit, C. Poon, and I. W aldspurger , On the non-convexity issue in the radial Calderón problem, SIAM J. Imaging Sci., (2026), p. in press. arXiv:2507.03379

    work page arXiv 2026

  4. [3]

    G. S. Alberti and M. Santacesaria , Calderón ’s inverse problem with a finite number of measurements , Forum Math. Sigma, 7 (2019), pp. e35, 20 pp

    work page 2019

  5. [4]

    , Infinite-dimensional inverse problems with finite measurements , Arch. Ration. Mech. Anal., 243 (2022), pp. 1–31

    work page 2022

  6. [5]

    Alessandrini, M

    G. Alessandrini, M. de Hoop, R. Gaburro, and E. Sincich , Lipschitz stability for the electrostatic inverse boundary value problem with piecewise linear conductivities , J. Math. Pures Appl., 107 (2017), pp. 638– 664

    work page 2017

  7. [6]

    Alessandrini and S

    G. Alessandrini and S. Vessella , Lipschitz stability for the inverse conductivity problem , Adv. Appl. Math., 35 (2005), pp. 207–241

    work page 2005

  8. [7]

    Ammari, J

    H. Ammari, J. Garnier, L. Giovangigli, W. Jing, and J.-K. Seo , Spectroscopic imaging of a dilute cell suspension, J. Math. Pures Appl. (9), 105 (2016), pp. 603–661

    work page 2016

Show all 68 references
  1. [8]

    Antun, F

    V. Antun, F. Renna, C. Poon, B. Adcock, and A. C. Hansen , On instabilities of deep learning in image reconstruction and the potential costs of AI , Proc. Natl. Acad. Sci. USA, 117 (2020), pp. 30088–30095

    2020

  2. [9]

    Aspri, E

    A. Aspri, E. Beretta, E. Francini, and S. Vessella , Lipschitz stable determination of polyhedral con- ductivity inclusions from local boundary measurements , SIAM J. Math. Anal., 54 (2022), pp. 5182–5222. 19

    2022

  3. [10]

    Beretta, M

    E. Beretta, M. V. de Hoop, and L. Qiu , Lipschitz stability of an inverse boundary value problem for a Schrödinger-type equation, SIAM J. Math. Anal., 45 (2013), pp. 679–699

    2013

  4. [11]

    Beretta, E

    E. Beretta, E. Francini, and S. Vessella , Lipschitz stable determination of polygonal conductivity inclusions in a two-dimensional layered medium from the Dirichlet-to-Neumann map , SIAM J. Math. Anal., 53 (2021), pp. 4303–4327

    2021

  5. [12]

    Bhata, K

    A. Bhata, K. Chen, and C. W ang , Neural correction operator: a reliable and fast approach for electrical impedance tomography. Preprint, arXiv:2507.18875v1, 2025

    2025

  6. [13]

    Bikowski and J

    J. Bikowski and J. L. Mueller , 2D EIT reconstructions using Calderón ’s method , Inverse Probl. Imaging, 2 (2008), pp. 43–61

    2008

  7. [14]

    Calderón , On an inverse boundary value problem , in Seminar on Numerical Analysis and its Applica- tions to Continuum Physics (Rio de Janeiro, 1980), Soc

    A.-P. Calderón , On an inverse boundary value problem , in Seminar on Numerical Analysis and its Applica- tions to Continuum Physics (Rio de Janeiro, 1980), Soc. Brasil. Mat., Rio de Janeiro, 1980, pp. 65–73

    1980

  8. [15]

    S. Cen, B. Jin, Q. Quan, and Z. Zhou , Numerical analysis of unsupervised learning approaches for parameter identification in pdes , Handbook of Numerical Analysis, Vol. 27, (2026), p. in press

    2026

  9. [16]

    S. Cen, B. Jin, K. Shin, and Z. Zhou , Electrical impedance tomography with deep Calderón method , J. Comput. Phys., 493 (2023), pp. 112427, 14 pp

    2023

  10. [17]

    Z. Chen, J. Xiang, P.-O. Bagnaninchi, and Y. Yang , MMV-Net: A multiple measurement vector network for multifrequency electrical impedance tomography , IEEE Trans. Neural Networks Learn. System, 34 (2023), pp. 8938–8949

    2023

  11. [18]

    Z. Chen, Y. Yang, J. Jia, and P. Bagnaninchi , Deep learning based cell imaging with electrical impedance tomography, in 2020 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), IEEE, 2020, pp. 1–6

    2020

  12. [19]

    Cheney, D

    M. Cheney, D. Isaacson, and J. Newell. , Electrical impedance tomography , SIAM Rev., 41 (1999), pp. 85–101

    1999

  13. [21]

    Colibazzi, D

    F. Colibazzi, D. Lazzaro, S. Morigi, and A. Samoré , Deep-plug-and-play proximal Gauss-Newton method with applications to nonlinear, ill-posed inverse problems , Inverse Probl. Imaging, 17 (2023), pp. 1226–1248

    2023

  14. [22]

    de Hoop, N

    M. de Hoop, N. Kovachki, M. Lassas, and N. Nelsen , Extension and neural operator approximation of the electrical impedance tomography inverse map . Preprint, arXiv:2511.20361, 2025

    2025

  15. [23]

    Dimas, V

    C. Dimas, V. Alimisis, N. Uzunoglu, and P. Sotiriadis , Advances in electrical impedance tomography inverse problem solution methods: from traditional regularization to deep learning , IEEE Acess, 12 (2024), pp. 47797–47829

    2024

  16. [24]

    Dos Santos Ferreira, C

    D. Dos Santos Ferreira, C. E. Kenig, M. Salo, and G. Uhlmann , Limiting Carleman weights and anisotropic inverse problems , Invent. Math., 178 (2009), pp. 119–171

    2009

  17. [25]

    H. W. Engl, M. Hanke, and A. Neubauer , Regularization of Inverse Problems , Kluwer Academic, Dor- drecht, 1996

    1996

  18. [26]

    F an and L

    Y. F an and L. Ying , Solving electrical impedance tomography with deep learning , J. Comput. Phys., 404 (2020), pp. 109119, 19 pp

    2020

  19. [27]

    Feldman, M

    J. Feldman, M. Salo, and G. Uhlmann , The Calderón Problem - An Introduction to Inverse Problems , AMS, Providence, RI, 2025

    2025

  20. [28]

    Foschiatti, R

    S. Foschiatti, R. Gaburro, and E. Sincich , The local complex calderón problem: Stability in a layered medium for a special type of anisotropic admittivity , SIAM J. Math. Anal., 57 (2025), pp. 4396–4424

    2025

  21. [29]

    K. R. Foster and H. P. Schwan , Dielectric properties of tissues and biological materials: A critical review , Crit. Rev. Biomed. Eng., 17 (1989), pp. 25–104. 20

    1989

  22. [30]

    Gaburro and E

    R. Gaburro and E. Sincich , Lipschitz stability for the inverse conductivity problem for a conformal class of anisotropic conductivities , Inverse Problems, 31 (2015), pp. 015008, 26 pp

    2015

  23. [31]

    Garde, M

    H. Garde, M. Hirvensalo, and N. Hyvönen” , Infinite-dimensional Lipschitz stability in the Calderón problem and general zernike bases , Calc. Var., 65 (2026), pp. 89, 18 pp

    2026

  24. [32]

    Garde and N

    H. Garde and N. Hyvönen , Linearized Calderón problem: reconstruction and Lipschitz stability for infinite- dimensional spaces of unbounded perturbations , SIAM J. Math. Anal., 56 (2024), pp. 3588–3604

    2024

  25. [33]

    Gilbarg and N

    D. Gilbarg and N. S. Trudinger , Elliptic Partial Differential Equations of Second Order , Springer-Verlag, Berlin-New York, 1977

    1977

  26. [34]

    N. M. Gottschling, V. Antun, A. C. Hansen, and B. Adcock , The troublesome kernel: on hallu- cinations, no free lunches, and the accuracy-stability tradeoff in inverse problems , SIAM Rev., 67 (2025), pp. 73–104

    2025

  27. [35]

    Guo and J

    R. Guo and J. Jiang , Construct deep neural networks based on direct sampling methods for solving electrical impedance tomography, SIAM J. Sci. Comput., 43 (2021), pp. B678–B711

    2021

  28. [36]

    Hamilton and A

    S. Hamilton and A. Hauptmann , Deep D-bar: real-time electrical impedance tomography imaging with deep neural networks , IEEE Trans. Med. Imag., 37 (2018), pp. 2367–2377

    2018

  29. [37]

    S. J. Hamilton, M. Lassas, and S. Siltanen , A direct reconstruction method for anisotropic electrical impedance tomography, Inverse Problems, 30 (2014), pp. 075007, 33

    2014

  30. [38]

    Harrach, Uniqueness and lipschitz stability in electrical impedance tomography with finitely many electrodes, Inverse Problems, 35 (2019), pp

    B. Harrach, Uniqueness and lipschitz stability in electrical impedance tomography with finitely many electrodes, Inverse Problems, 35 (2019), pp. 024005, 19 pp

    2019

  31. [39]

    K. He, X. Zhang, S. Ren, and J. Sun , Deep residual learning for image recognition , in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  32. [40]

    Henkin and M

    G. Henkin and M. Santacesaria , On an inverse problem for anisotropic conductivity in the plane , Inverse Problems, 26 (2010), pp. 095011, 18 pp

    2010

  33. [41]

    Hörmander , The Analysis of Linear Partial Differential Operators

    L. Hörmander , The Analysis of Linear Partial Differential Operators. I , Springer-Verlag, Berlin, second ed.,

  34. [42]

    Huang, H

    Y. Huang, H. Zhang, S. Y, J. Kolter, and A. Anandkumar , Training certifiably robust neural networks with efficient local Lipschitz bounds , Advances in Neural Information Processing Systems, 34 (2021), pp. 22745– 22757

    2021

  35. [43]

    Ito and B

    K. Ito and B. Jin , Inverse Problems: Tikhonov Theory and Algorithms , World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2015

    2015

  36. [44]

    D. P. Kingma and J. Ba , Adam: A method for stochastic optimization , in 3rd International Conference on Learning Representations, San Diego, 2015

    2015

  37. [45]

    Knudsen, M

    K. Knudsen, M. Lassas, J. L. Mueller, and S. Siltanen , Regularized D-bar method for the inverse conductivity problem, Inverse Probl. Imaging, 3 (2009), pp. 599–624

    2009

  38. [46]

    Knudsen and J

    K. Knudsen and J. L. Mueller , The Born approximation and Calderón ’s method for reconstruction of conductivities in 3-D , Discrete Contin. Dyn. Syst., (2011), pp. 844–853

    2011

  39. [47]

    Knudsen and A

    K. Knudsen and A. K. Rasmussen , Direct regularized reconstruction for the three-dimensional Calderón problem, Inverse Probl. Imaging, 16 (2022), pp. 871–894

    2022

  40. [48]

    J. M. Lee and G. Uhlmann , Determining anisotropic real-analytic conductivities by boundary measurements , Comm. Pure Appl. Math., 42 (1989), pp. 1097–1112

    1989

  41. [49]

    K. Li, K. Shin, and Z. Zhou , Learning-enhanced variational regularization for electrical impedance tomog- raphy via Calderón ’s method, J. Comput. Phys., 541 (2025), pp. 114309, 21 pp

    2025

  42. [50]

    W. R. B. Lionheart , Conformal uniqueness results in anisotropic electrical impedance imaging , Inverse Problems, 13 (1997), pp. 125–134. 21

    1997

  43. [51]

    Liu , Stability estimates for the two-dimensional inverse conductivity problem , PhD thesis, University of Rochester, New York, 1997

    L. Liu , Stability estimates for the two-dimensional inverse conductivity problem , PhD thesis, University of Rochester, New York, 1997

    1997

  44. [52]

    S. Ma, S. K. Sahoo, and M. Salo , The anisotropic Calderón problem at large fixed frequency on manifolds with invertible ray transform , J. Lond. Math. Soc. (2), 110 (2024), pp. e13006, 35 pp

    2024

  45. [53]

    Mandache , Exponential instability in an inverse problem for the Schrödinger equation , Inverse Problems, 17 (2001), pp

    N. Mandache , Exponential instability in an inverse problem for the Schrödinger equation , Inverse Problems, 17 (2001), pp. 1435–1444

    2001

  46. [54]

    P. A. Muller, J. L. Mueller, and M. M. Mellenthin , Real-time implementation of calderón ’s method on subject-specific domains , IEEE Trans. Med. Imag., 36 (2017), pp. 1868–1875

    2017

  47. [55]

    Murthy, Y.-H

    R. Murthy, Y.-H. Lin, K. Shin, and J. L. Mueller , A direct reconstruction algorithm for the anisotropic inverse conductivity problem based on Calderón ’s method in the plane , Inverse Problems, 36 (2020), pp. 125008, 21 pp

    2020

  48. [56]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox , U-net: Convolutional networks for biomedical image segmen- tation, in MICCAI 2015, Springer, Cham, 2015

    2015

  49. [57]

    Scarlett, R

    J. Scarlett, R. Heckel, M. Rodrigues, P. Hand, and Y. Eldar , Theoretical perspectives on deep learning methods in inverse problems , IEEE J. Sel. Areas Inf. Theory, 3 (2022), pp. 433–453

    2022

  50. [58]

    Shin and J

    K. Shin and J. L. Mueller , A second order Calderón ’s method with a correction term and a priori infor- mation, Inverse Problems, 36 (2020), pp. 124005, 22 pp

    2020

  51. [59]

    B. Sun, H. Zhong, Y. Zhao, L. Ma, and H. W ang , Calderón ’s method-guided deep neural network for electrical impedance tomography, IEEE Trans. Instrum. Meas., 72 (2023), p. 4507911

    2023

  52. [60]

    Sylvester and G

    J. Sylvester and G. Uhlmann , A global uniqueness theorem for an inverse boundary value problem , Ann. of Math. (2), 125 (1987), pp. 153–169

    1987

  53. [61]

    Szegedy, W

    C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus , Intriguing properties of neural networks , in International Conference on Learning Representations, 2014

    2014

  54. [62]

    D. N. Tanyu, J. Ning, A. Hauptmann, B. Jin, and P. Maass , A fair comparative study on deep learning and analytic-based approaches for electrical impedance tomography, in Data-Driven Models in Inverse Problems, T. Bubba, ed., de Gruyter, Berlin, 2024, pp. 437–470

    2024

  55. [63]

    Uhlmann and Y

    G. Uhlmann and Y. W ang , The anisotropic Calderón problem for high fixed frequency , SIAM J. Math. Anal., 56 (2024), pp. 4084–4103

    2024

  56. [64]

    W ang and Y

    J.-N. W ang and Y. W ang, A Lipschitz stability for the Calderón problem with small structure-free anisotropy . Preprint, 2025, https://www.math.ntu.edu.tw/~jnwang/pub/resources/draft-03-28-2025.pdf

    2025

  57. [65]

    W ang and Y

    Y. W ang and Y. Zhong, Robustness of data-driven approaches in limited angle tomography , SIAM J. Imaging Sci., 18 (2025), pp. 345–358

    2025

  58. [66]

    Zhang, D

    B. Zhang, D. Jiang, D. He, and L. W ang , Rethinking Lipschitz neural networks and certified robustness: A Boolean function perspective, Advances in Neural Information Processing Systems, 35 (2022), pp. 19398–19413

    2022

  59. [67]

    Zhou , Universality of deep convolutional neural networks , Appl

    D.-X. Zhou , Universality of deep convolutional neural networks , Appl. Comput. Harm. Anal., 48 (2020), pp. 787–794. 22

    2020

  60. [1990]

    Distribution theory and Fourier analysis

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.