acceptodds
Withdrawn from ICLR 2027

Local-Volatility Calibration with Hard-Constrained Differentiable Galerkin Layers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local-volatility calibration is an ill-posed inverse parabolic problem, and standard pipelines typically separate surface interpolation from the pricing operator used downstream. We instead learn only the variance-rate coefficient and obtain prices from the exact log-moneyness Dupire equation through differentiable weighted-Galerkin layers. A positive parameterisation makes coefficient admissibility architectural; the weighted discretisation makes every implicit state matrix symmetric positive-definite (SPD); and reverse-mode gradients reuse the same sparse solves as the forward pass. Smoothness remains a statistical regulariser rather than a surrogate for either positivity or the PDE. We prove that finite soft penalties cannot certify membership in the admissible ellipticity set and test the distinction through synthetic recovery, a mechanism-level hard-versus-soft comparison, and out-of-sample calibration on an S&P 500 option snapshot. The results show that structural guarantees can be retained without claiming universal coefficient accuracy or a complete static-arbitrage certificate.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
PINN, Galerkin method, ML for finance, option calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_otuaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates local volatility calibration as a differentiable weighted Galerkin inverse problem. The proposed framework learns only the variance rate, while option prices are computed through a discretization of the Dupire equation. A positive parameterization enforces a lower bound on the variance coefficient, and the weighted Galerkin formulation produces symmetric positive definite (SPD) implicit systems. Gradient computation can reuse the matrix factorizations obtained during the forward solves, improving computational efficiency.

The authors evaluate the method through synthetic experiments and a real-market S&P 500 option dataset. The synthetic experiments demonstrate successful coefficient recovery and option price fitting, although the neural parameterization does not outperform conventional Tikhonov regularization. In comparisons with soft physics-constrained methods, the hard-constrained approach avoids negative variance coefficients. However, differences in state representations make it difficult to attribute this improvement exclusively to the constraint mechanism.

On an S&P 500 option snapshot, the proposed method outperforms monotone SSVI when evaluated on held-out maturities but underperforms it on held-out strike blocks. Overall, the experimental results support the claimed structural guarantees and numerical properties, but do not establish uniformly superior calibration accuracy or a comprehensive guarantee of static no-arbitrage.

Critical strengths

1. The mathematical guarantees are clearly formulated and appropriately scoped.

Sections 3–6 derive the exact weighted Dupire formulation and establish that, under a positive variance coefficient, the implicit Galerkin matrices are symmetric positive definite (SPD) and the discrete coefficient-to-price mapping is differentiable. Importantly, the authors explicitly acknowledge the limitations of these results: they do not establish unique recovery of the local volatility coefficient or a complete static no-arbitrage guarantee for the resulting option surface. This distinction makes the theoretical contribution precise and avoids overstating the implications of the structural guarantees.

2. The empirical evaluation transparently reports cases where a classical baseline performs better.

In the synthetic experiment in Section 8.1, the neural coefficient parameterization achieves a mean coefficient error of 7.1% and a held-out relative price error of 0.23%, compared with 5.2% and 0.20%, respectively, for nodal Tikhonov regularization. By reporting these results, the paper provides a more balanced assessment of the proposed approach rather than claiming uniformly superior reconstruction accuracy. The contribution is therefore appropriately centered on maintaining coefficient positivity and positive definiteness throughout calibration, rather than on demonstrating that neural parameterization necessarily improves accuracy.

Critical weaknesses

1. The hard-versus-soft comparison does not isolate the contribution of hard constraints.

Section 8.2 compares the proposed constrained coefficient parameterization combined with a Galerkin solver against PINN, Deep Ritz, and VPINN baselines that independently learn the state solution through neural networks. Consequently, the experiments vary not only the constraint mechanism but also the state representation and training objectives. The negative coefficients, larger recovery errors, and occasional indefinite matrices reported in Figure 3 therefore cannot be attributed solely to the use of finite penalty terms.

Furthermore, Proposition 2 establishes that finite penalties cannot theoretically guarantee the prescribed coefficient lower bound, but it does not imply that appropriately tuned soft-constraint methods will generally perform worse in practice. A more controlled comparison that keeps the state representation and numerical solver fixed while varying only the constraint mechanism would be necessary to substantiate the claimed empirical advantage of hard constraints.

2. The real-market evaluation is limited to a single option snapshot.

The market experiments demonstrate different performance patterns depending on the holdout strategy. In the maturity-holdout experiment, the proposed method achieves an implied volatility RMSE of 2.29 volatility points, compared with 3.05 for monotone SSVI. However, in the strike-block holdout, the ranking reverses, with RMSE values of 2.44 and 1.84, respectively.

Although these results provide evidence of the method's applicability to real option data, they do not establish whether the observed performance differences generalize across trading days, volatility regimes, or market conditions. Evaluating multiple market snapshots and reporting the variability of calibration errors would strengthen the empirical evidence. The current results support a limited real-market validation rather than a claim of consistently superior calibration accuracy.

3. The principal comparative experiment does not directly evaluate the exact Dupire operator.

Positive coefficient parameterizations, symmetric positive definite finite-element systems, and adjoint-based differentiation are individually established numerical techniques. The paper's main methodological contribution is their integration within the exact weighted Dupire formulation.

However, the hard-versus-soft comparison in Section 8.2 is conducted using the auxiliary symmetric operator QSQ_{S}, whereas the synthetic coefficient recovery and real-market calibration experiments use the exact Dupire operator QDQ_D. This discrepancy limits the extent to which the comparative results directly support the proposed framework in its primary application.

A corresponding controlled comparison using QDQ_D would provide more direct evidence of whether the integrated formulation offers practical advantages over soft physics-constrained alternatives in the actual local volatility calibration problem.

Questions for authors

Q1. Novelty and positioning relative to recent work.

The proposed weighted Galerkin formulation integrates positive coefficient parameterization, symmetric positive definite (SPD) linear systems, and adjoint differentiation for the exact log-moneyness Dupire equation. However, each of these components has an established mathematical and numerical foundation.

Which specific aspect of this integration constitutes the main methodological advance, and what capabilities does it provide beyond existing inverse-PDE calibration approaches?

A more explicit comparison with recent local volatility calibration methods, particularly [Bias-Reduced Local Volatility — Cho et al. (2026)], would help clarify the methodological novelty. Additionally, [Derivative-Constrained PINNs — Hoshisashi et al. (2026)] investigate derivative-constrained PDE learning, including financial volatility applications, while their earlier work, Probability-Density-Consistent PINNs — Hoshisashi et al. (2025), addresses stochastic local volatility calibration. Although these approaches differ from the deterministic weighted-Galerkin formulation considered here, could the authors clarify the methodological distinctions, particularly regarding constraint enforcement and numerical optimization?

Q2. Interpretation and validity of the hard-versus-soft comparison.

The proposed calibration framework is based on the exact weighted Dupire operator QDQ_D, whereas the hard-versus-soft comparison in Section 8.2 uses the auxiliary symmetric operator QSQ_S. Moreover, the soft-constraint baselines employ separate neural state representations, introducing architectural differences beyond constraint enforcement.

Could the authors provide a controlled comparison using QDQ_D, keeping the Galerkin solver, network backbone, training data, boundary conditions, and optimization budget fixed while varying the coefficient constraint-enforcement mechanism? Such a comparison would help isolate the effect of hard constraints from other architectural differences.

The empirical conclusions also warrant clarification. Nodal Tikhonov regularization outperforms the neural parameterization in the synthetic recovery experiment. On the single S&P 500 option snapshot, the proposed method outperforms monotone SSVI on maturity holdouts but underperforms it on strike-block holdouts.

Could the authors evaluate additional trading dates or clarify the extent to which these results support claims of superior calibration performance?

Q3. Scope and interpretation of the theoretical guarantees.

The SPD and differentiability results establish useful structural properties of the discrete numerical solver. Proposition 2 demonstrates that finite penalty terms cannot guarantee the prescribed lower bound on the coefficient, but this does not establish that penalized optimization generally performs worse in practice.

Do the authors claim any new theoretical guarantee concerning coefficient identifiability, recovery accuracy, or calibration error, or is the primary theoretical contribution the exact weighted Dupire formulation and its structural guarantees?

Reviewer_Zk1rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a local-volatility calibration method that combines a positive coefficient parameterisation with differentiable Galerkin solvers for the Dupire equation. It evaluates the method on synthetic data and S&P 500 option prices.

Critical strengths
  • Personally, I liked the topic because I believe there is still room for interesting discoveries when AI is applied to classical quantitative finance problems, including pricing, hedging, and calibration.
Critical weaknesses
  • The paper is difficult to follow. The problem and methodology are poorly explained, making the main idea and contribution hard to identify.
  • The related work reads as a list of loosely connected topics rather than a focused comparison with the closest calibration methods.
  • Section 4 proves Propositions 2 and 3 before stating them in Section 6. It is difficult to find their theoretical value.
  • I don't understnad the role and meaning of Section 5 "Symmetric Benchmark Theory”. Its connection to the exact Dupire formulation and the relevance of its results are very unclear.
  • (Minor) Figure 1 is difficult to read.
Questions for authors

To save the authors time and effort, I would like to share my honest view. I think the paper needs substantial revision beyond what a rebuttal can address. I encourage the authors to improve the exposition and clarify the contribution.

Reviewer_361FOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies local-volatility calibration, namely, the recovery of the diffusion coefficient in the Dupire equation from observed option prices. The authors parameterize the volatility directly, compute prices by solving a discretized Dupire equation, and fit the prices with gradient-based training. They test the method on synthetic data, compare it with PINN-type methods, and run it on one S&P 500 dataset.

Critical strengths
  • The weighted reformulation of the Dupire equation is clean and handles the first-order term exactly. The resulting layers are symmetric positive definite, and the adjoint reuses the forward factorizations.
  • The paper separates what is guaranteed by construction from what is only checked empirically (Table 1), and it does not overclaim static no-arbitrage.
  • The authors report negative results openly, for instance that nodal Tikhonov regularization outperforms the neural field in Section 8.1.
Critical weaknesses

The paper is honest about the limits of its claims. However, the method seems to be standard PDE-constrained optimization applied to a classical inverse problem, the theoretical results are elementary, and the neural parameterization brings no benefit over classical regularization in the paper's own experiments. I am not an expert in this area, but it appears to me that the contribution from a machine learning perspective is therefore limited, and I lean towards rejection. The work may be better suited to a venue in computational finance or numerical PDEs.

The presentation is hard to follow. Several results are proved before they are stated (the proofs of Propositions 2 and 3 in Section 4 precede their statements in Section 6), and many symbols and terms are used before they are defined, including Q_D (Figure 1), Q_S (Section 8.4, defined only in the appendix), B_n, the interface coefficients, the admissible set, and the flux density. Some terms, such as the "σ-Q/SPD correction", are never defined. Domain-specific terms such as SVI/SSVI and Black–76 are also used without explanation, which makes the paper difficult for the ICLR audience.

Questions for authors

How does the method compare with existing PDE-based or neural local-volatility calibration methods on market data?

Section 6.3 refers to "Appendix 4"; this reference appears to be incorrect.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.