How can a posteriori error bounds be computed for the Uncertain Volatility Model?

The bounds bracket the UVM price from one candidate

In the Uncertain Volatility Model, volatility and correlation are not fixed by a single model; they are chosen dynamically and adversarially inside a prescribed uncertainty set. The robust price is therefore a stochastic control problem whose value function solves a fully nonlinear Hamilton-Jacobi-Bellman equation, with a supremum over admissible volatility controls inside the nonlinearity. The paper's goal is not another point estimate of that price but an interval attached to a single numerical candidate, quantifying how far the candidate sits from the true UVM value.

The construction is primal-dual. On the primal side, a smooth candidate approximation of the value function induces a feedback control through its Hessian, and that admissible control yields a lower bound. On the dual side, the authors derive a representation of the UVM value in terms of a matrix-valued Gamma field, which produces an upper bound through a nonnegative Hamiltonian penalty. Both representations are exact when the candidate coincides with the true value function.

The penalty has a concrete reading: it is the gap between the Hamiltonian and its value under a given admissible control, so it measures how badly that control fails to maximize the Hamiltonian for the prescribed Gamma field. A candidate can therefore be accurate in value and still leave a wide interval if its derivatives are poor.

Second-order information decides how tight the interval is

The dual constructions depend explicitly on derivatives of the candidate, and the paper states that an accurate approximation of the option price does not by itself guarantee a tight a posteriori interval. Inaccuracies in those derivatives may lead to a substantial duality gap. This is the paper's central practical claim, and the numerical experiments are designed around it.

The mechanism is visible in the dimension-one case, where the extremal volatility is selected according to the sign of the Gamma of the value function, producing a bang-bang type control. In higher dimension, analogous regime changes are driven by the matrix of second-order derivatives. Because the dual quantities exploit the candidate Gamma directly through the UVM Hamiltonian, they make explicit use of second-order information, unlike a martingale correction built only from the candidate gradient.

The authors also note that the optimal feedback control need not depend smoothly on the state, which is part of why approximating the UVM price is hard even in moderate dimension: in dimension d the control involves d marginal volatilities and d(d-1)/2 pairwise correlations, and the correlation matrix must stay positive semidefinite while respecting componentwise bounds. Optimizing the Hamiltonian is thus a constrained matrix optimization problem at every time and state.

Discrete-time versions come with quantified discretization error

The continuous-time bounds are not directly computable, so the paper introduces discrete-time approximations of the primal and dual quantities and quantifies the corresponding discretization errors using existing results from the stochastic-control literature. This step matters because the interval a reader actually sees is produced by these discrete approximations, not by the exact continuous objects.

The discretization also interacts with how the bounds are evaluated. On the primal side, the exact expectation obtained under the admissible control is a lower bound for the value, so a resimulated value retains that interpretation up to Monte Carlo sampling error. On the dual side, Monte Carlo simulation only approximates the expectation associated with the chosen control; even if that expectation were evaluated exactly, it would be no larger than the supremum defining the theoretical dual quantity.

The paper is explicit that the loss of the upper-bound property comes from the optimization error rather than from the Monte Carlo approximation. Replacing the suprema by values obtained under fixed resimulated controls does not preserve the theoretical upper bound, and the same conclusion applies to the time-grid formulation.

The upper bound can be read as a BSB residual

When the dual field is generated by the candidate itself, the upper bound admits an equivalent representation in terms of the residual of the associated Black-Scholes-Barenblatt equation. In that form, the dual quantity is the candidate value plus an a posteriori error term, and evaluating that error takes the form of a stochastic control problem.

This gives two numerically distinct routes to an upper estimate from the same candidate: a direct Gamma-based dual quantity and a residual-based one. The paper reports that for the backward stochastic policy-gradient candidates and the finite-difference benchmark, the direct estimate is numerically tighter than the corresponding residual-based estimate, with the physics-informed neural network the only exception, where the residual estimate is slightly tighter.

The residual route also has a natural affinity with how some candidates are trained. The paper observes that the PINN's strong residual estimate is consistent with its training objective, which directly penalizes the BSB residual, while its direct estimate is also tight even though that quantity is not targeted by the PINN loss.

Sobolev training on Hessian targets tightens the duals

The numerical study uses candidates from stochastic policy-gradient methods and physics-informed neural networks, and compares the resulting estimates with a martingale dual approach from the stochastic-control literature. The SPG candidates follow a backward architecture in which a separate actor and critic are trained at each time level while the already learned critic at the next time level is frozen and supplies the continuation value.

Value regression alone does not control the spatial derivatives of the critic, so the authors add first-order Sobolev training, and then extend it with second-order information because the a posteriori residual depends on the matrix of second-order derivatives. To avoid constructing and storing a matrix-valued target, they use random directional second derivatives: for a direction drawn from a standard normal, the scalar second-order target is the corresponding directional derivative of the frozen-action map.

Directional probing retains information about the complete Hessian, including mixed derivatives. The paper shows that the expected squared directional error is equivalent to the squared Frobenius norm of the full Hessian error, bounded between two and d+2 times that norm. At the last backward step, where the terminal function need not be differentiable, the derivative targets are approximated by centered finite differences, with the same random realizations and the same frozen action used at all perturbed states.

Derivative-informed training tightens the dual estimates

In the two-dimensional experiments the primal estimates and candidate values remain close to the reference across all four constructions, whereas the dual estimates discriminate more strongly between the quality of the learned derivatives. The standard SPG critic gives an accurate value approximation but noticeably looser dual estimates.

Derivative-informed training substantially tightens the dual quantities. First-order Sobolev training produces a large improvement over the standard critic, and the additional second-order supervision further improves both Hessian-based estimates. That extra improvement is not reflected systematically in the martingale estimate, which depends on the candidate gradient rather than its Hessian. The paper reads this as support for the interpretation that the second-order supervision primarily improves the second-order information relevant to the Hessian-based duals.

The finite-difference candidate provides the natural benchmark: at the continuous-time level the exact value function makes the dual quantity equal the value, while its BSB residual and terminal discrepancy vanish identically, and the corresponding numerical direct and residual estimates sit close to the reference. The PINN also performs well for both Hessian-based constructions. The martingale dual behaves differently, with the finite-difference candidate again tightest while the PINN and the two Sobolev critics are closer to one another; the paper cautions that some of these differences are comparable with their Monte Carlo uncertainty and should not be over-interpreted.

The same ordering persists in five dimensions and nonconvex payoffs

The five-dimensional experiment confirms the conclusions obtained in dimension two. Primal estimates and candidate values stay close to the reference across all four constructions, while the dual estimates again discriminate more strongly between the quality of the learned derivatives, and the standard SPG critic again gives an accurate value approximation with noticeably looser dual estimates.

The numerical ordering for the two Hessian-based constructions is reported as finite difference, then PINN, then Sobolev-2, then Sobolev-1, then standard. The improvement from Sobolev-1 to Sobolev-2 is singled out as particularly informative, because the second-order loss directly targets curvature information and improves both Hessian-based estimates, whereas the martingale dual does not improve from Sobolev-1 to Sobolev-2. The PINN gives the tightest dual estimates among the learned candidates, and the direct estimate is again tighter than the residual-based estimate for all SPG candidates.

A two-dimensional best-of butterfly with a nonconvex payoff serves as a confirmation experiment, using the same methodology and candidate constructions. Its results confirm the qualitative conclusions of the outperformer spread: derivative-informed SPG training substantially improves the dual estimates relative to the standard critic, second-order Sobolev supervision further improves the two Hessian-based constructions, and this improvement is specific to the Hessian-sensitive estimates rather than a systematic improvement of the martingale dual.

Second-order supervision costs more but buys the tighter interval

The computational times show that the additional training cost remains moderate relative to the subsequent dual evaluations. First-order Sobolev training takes approximately 1.5 times the standard SPG training time, while second-order Sobolev training takes approximately twice as long, and the costs of the subsequent dual evaluations are broadly comparable across candidates.

The paper's own framing of the comparison is deliberately limited: the observation that the resulting upper estimates are closer to the reference UVM value for most examples is an empirical observation rather than a general ordering between the different dual formulations, and the tightness of these estimates depends strongly on the accuracy of the candidate derivatives. The martingale construction carries additional discretization and finite control-search errors, with the relevant time-discretization contribution of order (T/N)^(1/4) under the assumptions considered.

For a reader deciding what to take from this work, the operative priority is derivative accuracy, and in particular second-order information, when the goal is a tight a posteriori interval rather than a good point value. The paper's evidence supports choosing second-order Sobolev supervision when the Hessian-based duals are the ones being reported, and treating the martingale dual as a diagnostic that responds to gradient quality rather than to curvature.

Sources used in this answer