Pith. sign in

REVIEW 3 major objections 5 minor 31 references

On Inverse Problems for Mean Field Games with Common Noise via Carleman Estimate

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper proves Lipschitz and Hölder stability for the stochastic mean-field-game system with common noise and a uniqueness theorem for an x1-independent source, all via two new Carleman estimates.

desk verdict A serious and mostly detailed extension of Carleman inverse problems to MFGs with common noise, but the theorems as stated overreach the regularity of U. read the letter →

arxiv 2412.08483 v1 pith:AI3CNRYM submitted 2024-12-11 math.AP math.OC

Qi Lü , Zhonghua Liao This is my paper · ORCID
classification math.APmath.OC MSC 35R3060H1549N80
keywords meanfieldgamescommonnoiseinverseproblemsCarlemanestimatesLipschitzstabilityHöldersourceproblemstochasticFokker-Planck-HJBsystem
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

This paper studies two inverse problems for mean-field games with common noise, described by a coupled forward-backward stochastic system of Fokker-Planck and Hamilton-Jacobi-Bellman equations. It proves that the solution pair is Lipschitz stable in the terminal and initial observations (Theorem 1.1), that it is Hölder stable when only the two terminal observations are used (Theorem 1.2), and that an unknown source term depending only on the transverse variables is uniquely determined by the terminal density together with lateral Cauchy data on two parallel boundaries (Theorem 1.3). The engine is a pair of new Carleman estimates for the forward and backward stochastic parabolic operators of the system. If correct, these results extend the deterministic mean-field-game inverse program to the common-noise setting that arises in financial markets, traffic flow, and other systems with a shared random factor.

What carries the argument

The load-bearing objects are the two Carleman estimates of Section 2 (Theorems 2.1 and 2.2, with the slab versions 2.3 and 2.4). A Carleman estimate is a weighted integral inequality for solutions of a parabolic equation, built from an exponential weight θ=$e^{{λ(t+2)^μ}}$; the proofs multiply the equation by a judiciously chosen test expression, apply the stochastic chain rule (Itô's formula), and integrate so that weighted norms of the unknown, its gradient, and its Laplacian are bounded by the source terms and by boundary, terminal, and initial data. The structural relation βhat=(1+$β^{2}$)/2 between the common-noise coefficient β and the diffusion coefficient βhat is what lets the undesirable stochastic cross terms be absorbed. For the inverse source problem, the additional devices are dividing the HJB equation by R, differentiating in x1 so the unknown r—independent of x1—drops out, inserting a time cut-off χ, and using the lateral Cauchy data to bring the estimate to a close.

What would settle it

On the one-dimensional slab G=(0,1), take β=0, R≡1, and a source difference r1-r2 supported in time; solve the linearized forward-backward system with U chosen in L2 but not in H1 and check whether the lateral Cauchy data plus terminal density still force r1=r2. If such a U makes the Carleman estimate (2.3) inapplicable, the missing H1 hypothesis is exposed; if the identity still holds, the H1 condition is unnecessary.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1.3: two solutions of the coupled stochastic FP-HJB system with the same terminal cost, the same terminal density, and equal lateral Cauchy data for ρ and u together with their first and second x1-derivatives on the two sides of the slab G=(0,1)×$R^{{n-1}}$ must have the same x1-independent source factors r1 and r2, almost surely, whenever the multiplier R is bounded away from zero. Theorems 1.1 and 1.2 are the companion stability statements: for two solutions satisfying the a priori bounds of Assumption 1.3, the H1 difference of the pair is controlled by the measurements—Lipschitz when terminal value, terminal density, and initial density are all observed, Hölder when only the two terminal measurements are available and the estimate is taken on [ε,T] with ε∈(0,T).

Load-bearing premise

The load-bearing premise is that the correction process U is spatially differentiable in a square-integrable sense, even though the stated solution class only assumes it is square-integrable without derivatives.

Editorial extensions

If this is right

  • Adding terminal density observation restores uniqueness for the coupled forward-backward system, which the paper notes is otherwise rare without a monotonicity assumption.
  • With the full set of three measurements, the solution map is Lipschitz; with only the two terminal measurements, the map is Hölder on any subinterval [ε,T].
  • The source uniqueness theorem implies that an x1-independent cost or interaction source can be recovered from terminal density plus boundary polls in a small neighbourhood of two parallel planes, because the Cauchy data in (1.7) can be approximated from such neighbourhood data.
  • Because the estimates hold for every β∈[0,1], the deterministic MFG inverse results appear as the limiting case β=0 of the common-noise framework.

Reading between the lines

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

  • The paper proves uniqueness but not stability for the inverse source problem; writing a quantitative version of the x1-differentiation argument would likely yield a conditional Hölder estimate for r, but that step is absent.
  • The constants in the Carleman estimates grow like e^{2λ(T+2)^μ}, so a numerical reconstruction from these measurements would be severely ill-conditioned without regularization; the paper does not address algorithms.
  • One could test whether Theorem 1.3 survives with fewer boundary observations, for instance only ρ and u on the lateral boundary rather than their first and second x1-derivatives; the proof as written uses all six conditions in (1.7).
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

3 major / 5 minor

Summary. The paper studies two inverse problems for mean field games with common noise, described by the coupled forward-backward stochastic parabolic system (1.1). The first problem (IP1) asks whether the terminal density observation, together with initial and terminal data, determines the solution pair (ρ,u) with Lipschitz or Hölder stability; the second problem (IP2) asks whether an x1-independent source term r in a factorized running cost can be uniquely recovered from terminal density and lateral Cauchy data. The main results are Theorem 1.1 (Lipschitz stability), Theorem 1.2 (Hölder stability), and Theorem 1.3 (unique determination of the source term). The proofs rely on two whole-space Carleman estimates (Theorems 2.1 and 2.2), proved in detail via Itô's formula, and two boundary Carleman estimates (Theorems 2.3 and 2.4), stated without proof. The paper linearizes the Hamiltonian under Assumptions 1.1–1.3 and reduces the inverse problems to weighted energy estimates.

Significance. If the results are valid, the paper provides the first stability and uniqueness results for inverse problems in MFGs with common noise, extending a substantial line of work on deterministic MFG inverse problems (e.g., [14,15]) to the stochastic setting. The whole-space Carleman estimates in Section 2 are proved carefully, and the treatment of the Hamiltonian nonlinearity through the coefficients B1–B5 and F1,F2 is systematic. The authors correctly identify the main stochastic difficulties, namely the gradient noise in the Fokker-Planck equation and the nonhomogeneous correction term U in the HJB equation. The use of the relation between β and β̂ to control the gradient drift is a notable technical contribution. However, the manuscript as written has a serious regularity gap: the Carleman estimates require the stochastic integrand and the solution to lie in higher Sobolev spaces than the stated solution classes provide. Since this gap affects the central applications of the estimates, the results cannot be considered fully established without repairs.

major comments (3)
  1. [§3, application of Theorem 2.1 to (3.9) and Theorem 2.2 to (3.8)] Theorem 2.1 requires the noise integrand f2 to lie in L^2_F(0,T;H^1(R^n)) and the solution w to lie in L^2_F(0,T;H^2(R^n)), while Theorem 1.1 only assumes U∈L^2_F(0,T;L^2(R^n)) and (ρ_i,u_i)∈L^2_F(0,T;H^1(R^n)). The proof applies (2.3) to the u-equation (3.9) with f2=U, which is not justified for U of class L^2 only. Moreover, the term -β div U in (3.9) is not well-defined as an L^2-valued dt-integrand when U is merely L^2. The same problem occurs for ρ: Theorem 2.2 is applied to (3.8), but that theorem requires p∈L^2_F(0,T;H^2(R^n)), whereas the stated solution class for ρ in Theorem 1.1 is only H^1(R^n). The authors should either strengthen the solution classes (e.g., U∈L^2_F(0,T;H^1(R^n)) and u,ρ∈L^2_F(0,T;H^2(R^n))) or prove and cite a regularity result showing that Assumption 1.3 upgrades U and ρ to the required classes.
  2. [§4, equations (4.6)-(4.16)] The proof of Theorem 1.3 applies the boundary Carleman estimate (2.21) to w=χ v_{x1} with noise integrand W=χ V_{x1}, and (2.23) to p=χ ρ_{x1}. For (2.21) to apply, one needs W∈L^2_F(0,T;H^1(G)) and w∈L^2_F(0,T;H^2(G)); for (2.23), one needs p∈L^2_F(0,T;H^2(G)). The hypotheses of Theorem 1.3 only assume U∈L^2_F(0,T;L^2(G)) and u∈L^2_F(0,T;H^1(G)). Even with Assumption 1.3 (u∈W^{2,∞}, ρ∈W^{1,∞}), the required H^2 regularity on the unbounded slab G=(0,1)×R^{n-1} does not follow, since bounded second derivatives do not imply square-integrability on unbounded domains. Consequently, the estimates (4.14)-(4.16) are not justified as written. The solution classes in Theorem 1.3 need to be strengthened, or a separate approximation/regularization argument must be supplied.
  3. [§2, Theorems 2.3 and 2.4] Theorems 2.3 and 2.4 are essential for the proof of Theorem 1.3, but they are stated without proof, with the remark that their proofs are 'very similar' to those of Theorems 2.1 and 2.2. This is not sufficient as it stands: (2.20) contains the additional drift coupling f3·f2, which is absent in the whole-space case and forces the parameters λ0(f3), µ0(f3) to be chosen in a way that is not quantified; the Dirichlet boundary condition w=0 on (0,T)×Γ also generates boundary terms in the weighted integration-by-parts argument that do not appear in the whole-space proof. The authors should provide complete proofs of these two estimates, or a precise reduction to Theorems 2.1 and 2.2 with full details of the boundary and f3 terms.
minor comments (5)
  1. [§4, equation before (4.4)] The line 'Put v = u/R and V = V/R' contains a typo; it should read 'Put v = u/R and V = U/R'.
  2. [§2, equations (2.3) and (2.21)] In the left-hand side of (2.3) and (2.21), the |f2|^2 term is written with weight θ, whereas the proof of Theorem 2.1 and the intermediate estimate (2.7) contain θ^2; the displayed inequalities appear to have missing squares.
  3. [§2, Theorem 2.4, equation (2.23)] The last integral in (2.23) is over R^n, while all other integrals are over G; this is presumably a typo and should be ∫_G.
  4. [§1, Theorem 1.2] In the statement of Theorem 1.2, the norm ‖u1-u2‖ is written as L^2(ε,T;H^1(R^n)) without the subscript F, whereas the ρ term uses L^2_F; the u term should also be L^2_F for consistency.
  5. [§4, equations (4.12)-(4.13)] In (4.12) and (4.13), the final bound is written as M^2∫_G p(T,y)^2 dy, but p=χρ_{x1} depends on (t,y); this should be ∫_G |χ(t)ρ(t,y)|^2 dy or a similar expression involving p(t,y), with the time variable made explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the Carleman estimates are derived in-paper from Ito's formula, and the stability and uniqueness theorems follow from them without assuming or importing their conclusions.

full rationale

The derivation chain is self-contained. The two Carleman estimates are proved inside the paper: Theorem 2.1 from the Ito identity (2.5) after multiplying by 2*beta_hat*Delta*v - 2*lambda*mu*(t+2)^(mu-1)*v and integrating, and Theorem 2.2 from the analogous identities (2.12)-(2.18); each estimate is an a priori inequality obtained by integration by parts and Young's inequality, and neither proof assumes the estimate it derives. The stability theorems then apply these internally proved estimates to the difference equations (3.8)-(3.9), with the measurement terms entering exactly through the boundary/terminal terms of the Carleman inequalities, so no fitted quantity is renamed as a prediction. Theorem 1.3 similarly derives r = 0 from the Carleman inequalities applied to (4.6)-(4.7) and the standard finite-variation-against-martingale identity (4.20); the conclusion is not assumed in the hypotheses. Self-citations [21] and [23]-[25] appear only in the literature survey and are not load-bearing. The flagged concerns are rigor gaps rather than circularity: Theorem 1.1 assumes U in L^2_F(0,T;L^2) while Theorem 2.1 requires f2 in L^2_F(0,T;H^1), and Section 4 differentiates V = U/R in x1 without proving the needed spatial regularity of U; additionally, the proofs of Theorems 2.3 and 2.4 are omitted as 'very similar'. These affect correctness and completeness, not the independence of the derivation chain.

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

The paper introduces no fitted parameters and no invented entities. The load-bearing inputs are structural assumptions on the MFG data, the unproved boundary Carleman estimates, and an implicit higher regularity of the correction process U. These are the only places where the conclusions could fail without one of the theorems being false.

assumptions (4)
  • domain assumption Assumptions 1.1 and 1.2: F is C^1, B has bounded derivatives up to third order, and K is bounded in L^2 with bound M2.
    Used to linearize the system and bound the coefficients B1 through B5 and F1, F2 in Section 3. This is a regularity and boundedness hypothesis on the game data, not derived from the system.
  • domain assumption Assumption 1.3: the a priori bound ||u||_{L^∞(0,T;W^{2,∞})} + ||rho||_{L^∞(0,T;W^{1,∞})} <= M3 holds.
    Standard a priori regularity used to control the nonlinear terms in (3.7) and to justify the H^2 norms needed by the Carleman estimates. Not proven from the equations.
  • ad hoc to paper The correction process U has enough spatial regularity to serve as the noise integrand in the Carleman estimates, at least H^1, and V = U/R can be differentiated in x1 in Section 4.
    Only U in L^2 is assumed, but Theorem 2.1 requires f2 in H^1 and the proof of Theorem 1.3 differentiates U/R. This extra regularity is not stated or justified.
  • ad hoc to paper The boundary Carleman estimates Theorems 2.3 and 2.4 are valid as stated.
    Theorem 1.3 relies on these inequalities; the paper omits their proofs with the comment that they are very similar to Theorems 2.1 and 2.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Inverse Problems for Mean Field Games with Common Noise via Carleman Estimate." pith.science (2026). https://pith.science/paper/AI3CNRYM

@misc{pith2026241208483,
  author       = {Pith},
  title        = {Pith review of: On Inverse Problems for Mean Field Games with Common Noise via Carleman Estimate},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AI3CNRYM}},
  note         = {Machine review of arXiv:2412.08483}
}
read the original abstract

In this paper, we study two kinds of inverse problems for Mean Field Games (MFGs) with common noise. Our focus is on MFGs described by a coupled system of stochastic Hamilton-Jacobi-Bellman and Fokker-Planck equations. Firstly, we establish the Lipschitz and H\"older stability for determining the solutions of a coupled system of stochastic Hamilton-Jacobi-Bellman and Fokker-Planck equations based on terminal observation of the density function. Secondly, we derive a uniqueness theorem for an inverse source problem related to the system under consideration. The main tools to establish those results are two new Carleman estimates.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages

  1. [1]

    Bardi and M

    M. Bardi and M. Fischer, On non-uniqueness and uniqueness of solutions in finite-hor izon mean field games. ESAIM Control Optim. Calc. V ar.,25(2019), Paper No. 44

    work page 2019

  2. [2]

    Bensoussan, J

    A. Bensoussan, J. Frehse and P . Y am, Mean field games and mean field type control theory. Springer, New Y ork, 2013

    work page 2013

  3. [3]

    Cardaliaguet and F

    P . Cardaliaguet and F. Delarue, Selected topics in mean field games. ICM–International Congress of Mathematicians. V ol. 5. 3660–3703, EMS Press, Berlin, 2023

    work page 2023

  4. [4]

    Cardaliaguet, F

    P . Cardaliaguet, F. Delarue, J. Lasry and P . Lions, The master equation and the convergence problem in mean field games. Ann. of Math. Stud., 201, Princeton University Press, 2019

    work page 2019

  5. [5]

    Carmona and F

    R. Carmona and F. Delarue, Probabilistic theory of mean field games with applications. I. mean field FBSDEs, control, and games. Springer, Cham, 2018. 18

    work page 2018

  6. [6]

    Carmona and F

    R. Carmona and F. Delarue, Probabilistic theory of mean field games with applications. II. mean field games with common noise and master equations. Springer, Cham, 2018

    work page 2018

  7. [7]

    L. Ding, W. Li, S. Osher and W. Yin, A mean field game inverse problem. J. Sci. Comput., 92(2022), Paper No. 7, 35 pp

    work page 2022

  8. [8]

    M. Ding, H. Liu and G. Zheng, Determining internal topological structures and running c ost of mean field games with partial boundary measurement. arXiv 2408.08911v1, 2024

    arXiv 2024

Show all 31 references
  1. [9]

    Huang, R

    M. Huang, R. P . Malhamé and P . E. Caines, Large population stochastic dynamic games: close-loop MCKean-Vlasov system and the nash certainty equ ivalence principle. Commun. Inf. Syst., 6(2006), 221–252

    2006

  2. [10]

    O. Y u. Imanuvilov, H. Liu and M. Y amamoto, Lipschitz stability for determination of states and inverse source problem for the mean field game equations. Inverse Probl. Imaging, 18 (2024), 824–859

    2024

  3. [11]

    Kachroo, S

    P . Kachroo, S. Agarwal and S. Sastry, Inverse problem for non-viscous mean field control: example from traffic. IEEE Trans. Autom. Control, 61(2016), 3412–3421

    2016

  4. [12]

    M. V . Klibanov, A coefficient inverse problem for the mean field games system, Appl. Math. Optim., 88(2023), Paper No. 54pp

    2023

  5. [13]

    M. V . Klibanov, The mean field games system: Carleman estimates, Lipschitz s tability and uniqueness. J. Inverse Ill-Posed Probl., 31(2023), 455–466

    2023

  6. [14]

    M. V . Klibanov and Y . Averboukh, Lipschitz stability estimate and uniqueness in the retro- spective analysis for the mean field games system via two Carl eman estimate. SIAM J. Math. Anal., 56(2024), 616–636

    2024

  7. [15]

    M. V . Klibanov, J. Li and H. Liu, Hölder stability and uniqueness for the mean field games system via Carleman estimates. Stud. Appl. Math., 151(2023), 1447–1470

    2023

  8. [16]

    M. V . Klibanov, J. Li and H. Liu, On the mean field games system with the lateral Cauchy data via Carleman estimates. J. Inverse Ill-Posed Probl., 32(2024), 277–295

    2024

  9. [17]

    M. V . Klibanov, J. Li and Z. Y ang, Convexification numerical method for the retrospective problem of mean field games. Appl. Math. Optim., 90(2024), Paper No. 24 pp

    2024

  10. [18]

    V . N. Kolokoltsov and O. A. Malafeyev, Many agent games in socio-economic systems: cor- ruption, inspection, coalition building, network growth, security. Springer, Cham, 2019

    2019

  11. [19]

    Lacker, A general characterization of the mean field limit for stocha stic differential games

    D. Lacker, A general characterization of the mean field limit for stocha stic differential games. Probab. Theory Relat. Fields, 165(2016), 581–648

    2016

  12. [20]

    J. M. Larsy and P . L. Lions, Mean field games. Jpn, J. Math., 2(2007), 229–260

    2007

  13. [21]

    Liao and Q

    Z. Liao and Q. Lü, Stability estimate for an inverse stochastic parabolic pro blem of deter- mining unknown time-varying boundary. Inverse Problems, 40 (2024), Paper No. 045032. 19

    2024

  14. [22]

    H. Liu, C. Mou and S. Zhang, Inverse Problems for mean field games. Inverse Problems, 39(2023), Paper No. 085003

    2023

  15. [23]

    Lü, Carleman estimate for stochastic parabolic equations and inverse stochastic parabolic problems

    Q. Lü, Carleman estimate for stochastic parabolic equations and inverse stochastic parabolic problems. Inverse Problems, 28(2012), 045008

    2012

  16. [24]

    Q. Lü, X. Zhang, Inverse problems for stochastic partial differential equa tions: some pro- gresses and open problems. Numer. Algebra Control Optim., 14(2024), 227–272

    2024

  17. [25]

    Lü and X

    Q. Lü and X. Zhang, Mathematical control theory for stochastic partial differential equations. Springer, Cham, 2022

    2022

  18. [26]

    K. Ren, N. Soedjak and K. Wang, Unique determination of cost functions in a multipopulation mean field game model. airXiv: 2312.01622, 2023

    2023 arXiv

  19. [27]

    K. Ren, N. Soedjak, K. Wang and H. Zhai, Reconstructing a state-independent cost function in a mean field game model. arXiv: 2402.09297, 2024

    2024 arXiv

  20. [28]

    Tang and X

    S. Tang and X. Zhang, Null controllability for forward and backward stochastic p arabolic equations. SIAM J. Control Optim., 48(2009), 2196–2216

    2009

  21. [29]

    B. Wu, Q. Chen and Z. Wang, Carleman estimates for a stochastic degenerate parabolic equation and applications to null controllability and an in verse random source problem. In- verse Problems 36 (2020), 075014

    2020

  22. [30]

    Wu and J

    B. Wu and J. Liu, On the stability of recovering two sources and initial statu s in a stochastic hyperbolic-parabolic system. Inverse Problems 38 (2022), Paper No. 025010

    2022

  23. [31]

    Y uan, Conditional stability in determination of initial data for stochastic parabolic equa- tions

    G. Y uan, Conditional stability in determination of initial data for stochastic parabolic equa- tions. Inverse Problems, 33(2017), Paper No. 035014. 20

    2017

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.