Pith. sign in

REVIEW 2 major objections 4 minor 25 references

Lipschitz stability for Bayesian inference in porous medium tissue growth models

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Mild solutions of a tissue-growth porous medium equation depend Lipschitz-continuously on the diffusion exponent, with explicit constants.

desk verdict Solid explicit L1-Lipschitz stability in the exponent for the porous-medium tumor model, but the Bayesian posterior-convergence lemma rests on a missing D(A) hypothesis. read the letter →

arxiv 2506.04769 v1 pith:FDF5VX4Q submitted 2025-06-05 math.AP

Tomasz Dębiec , Piotr Gwiazda , B{l}ażej Miasojedow , Katarzyna Ryszewska , Zuzanna Szymańska , Aneta Wróblewska-Kamińska This is my paper
classification math.AP MSC 35B3035B3535B4535K5735K6535Q92
keywords PorousmediumequationTumourgrowthStabilityContinuousdependenceParameterestimationBayesianinversionLipschitzmildsolutions
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

Tissue-growth models of porous-medium type write the cell density $u(t,x)$ as a solution of $\partial_t u = \Delta(u^\gamma) + uG(p)$, where the pressure $p = \frac{\gamma}{\gamma-1} u^{\gamma-1}$. The paper proves that mild solutions depend Lipschitz-continuously on the exponent $\gamma$ in the $L^1$ norm, with an explicit bound in terms of time, the growth rate $G$, and the $L^1$/$L^\infty$/$BV$ norms of the initial data. The motivation is Bayesian inversion: gradient-based samplers such as the Metropolis-Adjusted Langevin Algorithm need the likelihood to be at least Lipschitz in the fitted parameter, and this estimate supplies exactly that regularity. A companion result shows that the posterior obtained from the implicit Euler discretisation converges to the true posterior in total variation at rate $n^{-1/2}$.

What carries the argument

The key machinery is a localised stability estimate for the resolvent operator $(I+\tau A)^{-1}$, where $A(u) = -\Delta\varphi(u) - (u)_+G(p(u))$. Proposition 2 proves, by the doubling-of-variables technique, that for two different triples $(\varphi_i,G_i,p_i)$ the difference of resolvents is controlled by the $L^1$ difference of the inputs, a term proportional to $\|f_1\|_{BV}$ times the square-root difference of the diffusion coefficients, and terms proportional to differences in $G$ and $p$. Iterating this estimate along the implicit Euler steps $(I+\frac{t}{n}A)^{-n}$ and passing to the limit through the Crandall–Liggett semigroup converts the resolvent bound into the Lipschitz estimate for mild solutions. The explicit constants are obtained at the end by inserting the power-law $\varphi$ and $p$ and carefully bounding the functions $\sqrt{\varphi'_1}-\sqrt{\varphi'_2}$ and $p_2-p_1$ on the relevant intervals.

What would settle it

Take $d=1$, $G\equiv 0$, and the same nonnegative initial datum $u_0$ equal to the indicator function of an interval (which lies in $L^\infty\cap BV$). Solve $\partial_t u = \partial_{xx}(u^{\gamma})$ for two close exponents $\gamma_1<\gamma_2$, measure $\|u_1(t)-u_2(t)\|_{L^1}$ at a fixed time $t$, and compare it with the right-hand side of Corollary 1: the constant is finite and explicit, so exceeding it for any such pair would disprove the claimed bound.

Watch

Extended reading notes

Core claim

Corollary 1 is the paper's main result: for exponents $\gamma_2 > \gamma_1 > 1$, nonnegative initial data with $u_0^1 \in L^\infty(\mathbb{R}^d) \cap BV(\mathbb{R}^d)$, and growth functions $G_i \in W^{1,\infty}$, the two mild solutions satisfy an $L^1$ estimate of the form $\|u_1(t)-u_2(t)\|_{L^1} \le e^{tG(0)}\|u_0^1-u_0^2\|_{L^1} + t e^{tG(0)} \|G_2-G_1\|_{L^\infty}\|u_0^1\|_{L^1} + C |\gamma_2-\gamma_1|$, where $C$ is fully explicit, involving $M = e^{tG(0)}\|u_0^1\|_{L^\infty}$, $\|u_0^1\|_{L^1}$, $\|u_0^1\|_{TV}$, and $\|G'_2\|_{L^\infty}$. This follows from the more general Theorem 3, which establishes the same continuous-dependence statement for the abstract filtration equation $\partial_t u = \Delta\varphi(u) + (u)_+ G(p(u))$, and then specialises to $\varphi_i(t)=\operatorname{sign}(t)|t|^{\gamma_i}$ and $p_i(t)=\frac{\gamma_i}{\gamma_i-1}|t|^{\gamma_i-1}$. The result is therefore a stability statement about the model's pressure-density law as much as about the exponent $\gamma$ itself.

Load-bearing premise

The proof requires one of the two initial data, $u_0^1$, to be in $L^\infty(\mathbb{R}^d) \cap BV(\mathbb{R}^d)$; without bounded variation the cross-diffusion term in the doubling-of-variables estimate is not controlled, so no Lipschitz bound is shown for merely $L^1$ initial data.

Editorial extensions

If this is right

  • The likelihood in the Bayesian inverse problem for $\gamma$ is Lipschitz, so gradient-based samplers such as MALA can be applied to the tissue-growth model without paying the extra error of a smoothed approximation.
  • The implicit Euler discretisation inherits the same stability estimate, and Theorem 2's error bound shows the resulting posterior converges in total variation at rate $n^{-1/2}$, so the discretised inference problem shares the well-posedness of the continuous one.
  • Theorem 3's abstract statement applies to any $\varphi$ and $p$ satisfying (5)–(7), so other tumour models with different pressure-density relations automatically get a continuous-dependence estimate of the same type.
  • The bound exposes how the regularity of the initial cell-density profile ($L^\infty$ and $BV$) controls the conditioning of the inverse problem: rough initial data enlarge the Lipschitz constant and make $\gamma$ harder to infer.

Reading between the lines

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

  • The Lipschitz estimate paves the way to prove differentiability of the map $\gamma \mapsto u(t)$ in $L^1$ for BV data; the paper itself does not establish a derivative formula, but if such a formula holds, MALA could use exact gradients instead of finite-difference approximations.
  • Because the proof only needs one datum in $BV$, an approximation argument with truncated mollifiers might extend the estimate to all $L^1\cap L^\infty$ initial data; the paper does not settle whether the Lipschitz constant remains finite in that larger class.
  • The constants contain factors like $|\ln M|$ and $M^{\gamma-1}$ with $M=e^{tG(0)}\|u_0\|_{L^\infty}$, suggesting the estimate degrades for large initial densities and long times; testing numerically whether these factors are sharp would indicate how practical the stability guarantee is for real tumour data.
  • Lemma 2's posterior-stability statement concerns the implicit Euler discretisation only; extending it to finite-volume or finite-element solvers actually used in practice would require an analogous stability bound for those discretisations, which the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper studies the porous-medium-type tissue growth equation ∂tu = Δ(u^γ) + uG(p) with pressure p = γ/(γ−1)u^{γ−1}, and proves a quantitative L1 stability estimate for mild solutions with respect to the exponent γ. The main result is Corollary 1, obtained as a specialization of the abstract stability theorem Theorem 3 for ∂tu = Δφ(u) + (u)_+G(p(u)) under assumptions (5)–(7). The proof uses regularized resolvent operators, the doubling-of-variables technique, BV estimates, and passage to the limit through the Crandall–Liggett semigroup. The paper also claims in Lemma 2 an O(1/√n) total-variation convergence of the posterior distribution based on the implicit time discretization.

Significance. The main PDE stability estimate is a valuable quantitative result: the Lipschitz constant is derived explicitly rather than assumed, the proof is largely self-contained, and the regularized-to-unregularized transfer in Sections 3–4 is carefully executed. If the paper limited its claims to the PDE stability theorem, it would be a clear contribution to the literature on continuous dependence for degenerate parabolic equations. However, the advertised Bayesian bridge in Lemma 2 is not established as written, because the cited Crandall–Liggett rate requires an initial datum in D(A) that is not assumed there. The significance of the paper therefore depends on repair or reframing of the Bayesian claims.

major comments (2)
  1. [§1.1, Lemma 2 and Eq. (10)] The inequality ||u − u^n||_{L1} ≤ C/√n in Eq. (10) is justified by citing Theorem 2, but the displayed Crandall–Liggett rate in Theorem 2 is stated only for u0 ∈ D(A), with D(A) defined in Proposition 5, Eq. (36). Lemma 2 imposes no such condition on the initial datum, and the cited rate does not apply to, for example, a BV ∩ L∞ initial datum for which Δφ(u0) is a measure rather than an L1 function. Thus the total-variation convergence of the numerical posterior to the exact posterior is not established as written. Please add the missing D(A) hypothesis to Lemma 2 and to the posterior-convergence claim, or supply a rate that is valid for all L1 initial data.
  2. [§4, Corollary 1] Corollary 1 applies Theorem 3 with p_i(t) = γ_i/(γ_i−1)|t|^{γ_i−1}, but these functions are not nondecreasing on all of R: for negative arguments |t|^{γ−1} decreases as t increases. Hence the formal application of Theorem 3, whose assumption (7) requires p nondecreasing on R, is outside the stated hypotheses. For nonnegative initial data the proof only evaluates p_i on the interval [0, M], and on that interval monotonicity holds, but this restriction should be stated explicitly in the proof of Corollary 1 before invoking Theorem 3.
minor comments (4)
  1. [Throughout] There are several typographical and notation issues, including 'covvarriance' in Eq. (3), 'Banch space' in Theorem 2, and inconsistent notation for the observation map (m_kj, m_ij, m(u)); a careful copy-edit is needed.
  2. [§4, Theorem 3 and Corollary 1] Theorem 3 states its estimate for almost all t ∈ (0,T), while Corollary 1 states the estimate for all t. The uniform convergence supplied by Theorem 2 likely justifies the stronger statement, but the two formulations should be reconciled explicitly.
  3. [§1.1, Lemma 2] The proof of Lemma 2 asserts that the likelihood potential is bounded both from below and from above; this requires a uniform bound on the forward map M(γ) over the parameter space Γ, which is not stated or proved in the lemma. The boundedness of Γ alone does not imply such a bound.
  4. [§4, Corollary 1] In the proof of Corollary 1, M = e^{tG(0)}||u0_1||_{L∞} may be less than 1, in which case |ln M| appears; the case u0_1 = 0 should be handled separately, and a short comment would remove ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the Lipschitz estimate is derived from explicit resolvent estimates and external semigroup theory; the Bayesian-convergence lemma has hypothesis gaps but no circular reduction.

full rationale

The central claim, Corollary 1, does not reduce to its inputs. Theorem 3 is proved by iterating the resolvent-level estimate (16), which is derived by the doubling-of-variables method from the two resolvent equations (18); every constant in the final bound comes from explicit quantities (||u0_1||_{BV}, ||u0_1||_{L∞}, ||G'_2||_∞, sup of sqrt(phi'_1)-sqrt(phi'_2), sup|p2-p1|) and none is fitted to the target difference. Corollary 1 then only supplies explicit elementary bounds (43) and (44) for these suprema in the power-law case; the factor |gamma_2-gamma_1| appears because sqrt(phi'_i)-sqrt(phi'_j) and p_i-p_j are Lipschitz in gamma, not because such dependence was assumed. The Crandall-Liggett theorem and the Benilan-Brezis-Crandall resolvent theory are external classical results, and the cited works by the authors are contextual Bayesian tumour-growth references rather than load-bearing. The only concerns are non-circular gaps: Lemma 2 applies the Crandall-Liggett rate ||u(t)-(I+t/n A)^{-n}u0|| <= (t/sqrt(n))||Au0|| although Lemma 2 does not assume u0 in D(A) (with D(A) defined in Proposition 5 by -Delta phi(u) in L1), so the stated TV convergence of the numerical posterior is not fully justified for general L1 initial data; and Corollary 1 applies Theorem 3 with p_i(t)=gamma_i/(gamma_i-1)|t|^{gamma_i-1}, which is nondecreasing only on [0,infty), so the formal monotonicity hypothesis (7) should be read together with the nonnegativity assumption on the initial data. These are correctness/completeness issues, not circularity: no equation is defined in terms of the quantity it is used to prove, and no prediction is a renamed fit.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted; all constants in the stability estimates are explicit functions of initial data and γ. The proof rests on standard nonlinear semigroup theory and the stated monotonicity assumptions.

assumptions (4)
  • standard math Crandall-Liggett semigroup generation theorem (Theorem 2 of the paper, [6])
    Used to construct mild solutions and to obtain the O(1/√n) discretization rate for the implicit Euler scheme.
  • standard math Lions pseudomonotone operator theorem ([16])
    Used in Proposition 1 to prove surjectivity of the regularized resolvent operator.
  • standard math Bénilan-Brezis-Crandall L1 theory for semilinear elliptic equations ([1])
    Used in Proposition 3 for unique solvability of the resolvent equation in L1.
  • domain assumption Model assumptions (5)-(7): φ increasing with φ(0)=0; G continuous with G(0)>0 and G′≤0; p nonnegative and nondecreasing
    These monotonicity and positivity conditions are needed for the accretivity of the operator and for the comparison estimates in the doubling-of-variables proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lipschitz stability for Bayesian inference in porous medium tissue growth models." pith.science (2026). https://pith.science/paper/FDF5VX4Q

@misc{pith2026250604769,
  author       = {Pith},
  title        = {Pith review of: Lipschitz stability for Bayesian inference in porous medium tissue growth models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDF5VX4Q}},
  note         = {Machine review of arXiv:2506.04769}
}
abstract

We consider a macroscopic model for the dynamics of living tissues incorporating pressure-driven dispersal and pressure-modulated proliferation. Given a power-law constitutive relation between the pressure and cell density, the model can be written as a porous medium equation with a growth term. We prove Lipschitz continuity of the mild solutions of the model with respect to the diffusion parameter (the exponent $\gamma$ in the pressure-density law) in the $L_1$ norm. While of independent analytical interest, our motivation for this result is to provide a vital step towards using Bayesian inverse problem methodology for parameter estimation based on experimental data -- such stability estimates are indispensable for applying sampling algorithms which rely on the gradient of the likelihood function.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 24 canonical work pages

  1. [1]

    Bénilan, H

    Ph. Bénilan, H. Brezis, and M. G. Crandall. Semilinear elliptic equation inL1(RN ). Annali Scienze, 2(4):523–555, 1975

    work page 1975

  2. [2]

    Byrne and D

    H. Byrne and D. Drasdo. Individual-based and continuum models of growing cell populations: a comparison. J. Math. Biol., 58(4–5):657–687, 2009

    work page 2009

  3. [3]

    J. A. Carrillo, K. Craig, and F. S. Patacchini. A blob method for diffusion.Calc. Var. Partial Differ., 58(2):53, 2019

    work page 2019

  4. [4]

    J. A. Carrillo and J. Skrzeczkowski. Convergence and stability results for the particle system in the Stein gradient descent method.Math. Comp., 94:1793–1814, 2025

    work page 2025

  5. [5]

    Cockburn and G

    B. Cockburn and G. Gripenberg. Continuous dependence on the nonlinearities of solutions of degenerate parabolic equations.J. Differ. Equ., 151(2):231–251, 1999

    work page 1999

  6. [6]

    M. G. Crandall and T. M. Liggett. Generation of semi-groups of nonlinear transformations on general Banach spaces.Am. J. Math., 93(2):265–298, 1971

    work page 1971

  7. [7]

    Falćo, D

    C. Falćo, D. J. Cohen, J. A. Carrillo, and R. E. Baker. Quantifying tissue growth, shape and collision via continuum models and Bayesian inference.J. R. Soc. Interface, 20:20230184, 2023. 19

    work page 2023

  8. [8]

    Y. Feng, L. Liu, and Z. Zhou. A unified Bayesian inversion approach for a class of tumor growth models with different pressure laws.ESAIM Math. Model. Numer. Anal., 58(2):613–638, 2024

    work page 2024

Show all 25 references
  1. [9]

    Friedman

    A. Friedman. Mathematical analysis and challenges arising from models of tumour growth. Math. Models Methods Appl. Sci., 17:1751–1772, 2007

    2007

  2. [10]

    A. E. Gelfand and A. F. M. Smith. Sampling-Based Approaches to Calculating Marginal Den- sities. J. Am. Stat. Assoc., 85(410):398–409, 1990

    1990

  3. [11]

    A. Goriely. The Mathematics and Mechanics of Biological Growth. Interdisciplinary Applied Mathematics, vol. 45,Springer New York, NY, 2017

    2017

  4. [12]

    Gwiazda, B

    P. Gwiazda, B. Miasojedow, J. Skrzeczkowski, and Z. Szymańska. Convergence of the EBT method for a non-local model of cell proliferation with discontinuous interaction kernel.IMA J. Numer. Anal., 43:590–626, 2023

    2023

  5. [13]

    W. K. Hastings. Monte Carlo sampling methods using Markov chains and their applications. Biometrika, 57(1):97–109, 1970

    1970

  6. [14]

    Kahle, K

    C. Kahle, K. F. Lam, J. Latz, and E. Ullmann. Bayesian parameter identification in Cahn– Hilliard models for biological growth.SIAM/ASA J. Uncertain. Quantif., 7:526–552, 2019

    2019

  7. [15]

    Korba, A

    A. Korba, A. Salim, M. Arbel, G. Luise, and A. Gretton. A non-asymptotic analysis for Stein variational gradient descent.Adv. Neural Inf. Process. Syst., 33:4672–4682, 2020

    2020

  8. [16]

    J. L. Lions. Quelques méthodes de résolution des problèmes aux limites non linéaires.Dunod, Paris, 1969

    1969

  9. [17]

    Lipkoá, P

    J. Lipkoá, P. Angelikopoulos, S. Wu, E. Alberts, B. Wiestler, C. Diehl, C. Preibisch, T. Pyka, S.E.Combs, P.Hadjidoukas, andK.VanLeemput. Personalizedradiotherapydesignforglioblas- toma: integrating mathematical tumor models, multimodal scans, and Bayesian inference.IEEE Trans...

    2019

  10. [18]

    Liu and D

    Q. Liu and D. Wang. Stein variational gradient descent: A general purpose Bayesian inference algorithm. Adv. Neural Inf. Process. Syst., 29:1–9, 2016

    2016

  11. [19]

    Metropolis, A

    N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equation of state calculations by fast computing machines.J. Chem. Phys., 21(6):1087–1092, 1953

    1953

  12. [20]

    N. Pavel. Nonlinear Evolution Operators and Semigroups: Applications to Partial Differential Equations. Lecture Notes in Mathematics, vol. 1260,Springer-Verlag, Berlin, 1987

    1987

  13. [21]

    Perthame, F

    B. Perthame, F. Quirós, and J. L. Vázquez. The Hele–Shaw asymptotics for mechanical models of tumor growth.Arch. Ration. Mech. Anal., 212(1):93–127, 2014

    2014

  14. [22]

    G. O. Roberts and R. L. Tweedie. Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli, 2(4):341–363, 1996

    1996

  15. [23]

    A. M. Stuart. Inverse problems: A Bayesian perspective.Acta Numer., 19:451–559, 2010

    2010

  16. [24]

    Szymańska, J

    Z. Szymańska, J. Skrzeczkowski, B. Miasojedow, and P. Gwiazda. Bayesian inference of a non- local proliferation model.R. Soc. Open Sci., 8(11):211279, 2021

    2021

  17. [25]

    J. L. Vázquez. The porous medium equation: Mathematical theory.Oxford University Press, Oxford, 2007. 20

    2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.