Option traders and risk modelers need price models that match observed option prices at a handful of maturities and keep prices strictly positive with volatility proportional to price. A geometric Bass martingale is such a model: among all martingales that match given distributions of the price at two dates, it is the one whose log-price fluctuates the least, so it sits as close as possible to geometric Brownian motion, the textbook log-normal price model. The theory is now in place, as Geometric martingale Benamou–Brenier transport and geometric Bass martingales gives the defining stochastic differential equation, a dual problem, and an explicit change-of-measure bijection with the older arithmetic Bass martingales. That paper also states that efficient numerical algorithms remain open. We propose to build one by pulling the problem back to the arithmetic side, where a fast quantile-space fixed-point calibration already exists. We then use a classical fact about optimal stopping to check the result.
That fact is the smooth fit principle. When an American option holder decides whether to exercise, the option's value function must meet the exercise payoff at the exercise boundary without a kink whenever the price process diffuses. Brownian-type noise makes any kink exploitable. The true geometric Bass martingale is a diffusion with nondegenerate volatility inside its support, so the exact American value under it has no kink. Any kink we measure in a computed American value is therefore numerical error. It can be measured without a reference solution, and it shows where in price and time the error sits. Our central claim is that this "kink certificate" ranks discretizations of the calibrated model correctly, including atomic quantile grids, piecewise-linear quantiles, and neural or tokenized representations. It also detects calibration error that price-matching checks at the calibration maturities cannot see, because the model can reproduce vanilla prices at those maturities while its path dynamics in between are wrong.
Core Claim
The three research directions play different roles here. Geometric Bass martingales are the object we compute. Smooth fit is the measuring instrument. Discrete numerical representation is the design space being measured: the finite object (atoms, grid nodes, network weights, number tokens) that stands in for a continuous quantile function or exercise boundary.
The certificate is defined in log-price coordinates $y=\log m$. For an American put with value $V(t,y)$, payoff $g(y)=(K-e^{y})^+$, and computed exercise boundary $b(t)$, the kink is
$$\kappa(t) = \big|\partial_y V(t,b(t)^+) - g'(b(t))\big|.$$
Here $\partial_y V(t,b(t)^+)$ is the one-sided derivative from the continuation region. Smooth fit says $\kappa\equiv 0$ for the exact model. We will test whether the integrated kink, $\int_0^T \kappa(t)\,dt$, predicts the American price error and the boundary error more reliably than conventional checks. Those checks are vanilla repricing error at the calibration maturities and the change in price under grid refinement. One caveat must be designed in. With finitely many exercise dates (Bermudan exercise), a kink of order the exercise spacing is legitimate, because the value is a maximum of two smooth functions. The certificate is therefore computed along a sequence of exercise frequencies. Only a kink that does not vanish as exercise dates are refined counts as model or discretization error. This is consistent with the higher-order regularity seen under periodic exercise in American options under periodic exercise opportunities.
Research Program
The solver. The first component is a geometric Bass solver that reuses arithmetic machinery. Backhoff-Veraguas et al. prove an explicit bijection between arithmetic and geometric Bass martingales through a change of measure. We expect it to act on the marginals as a numéraire change, reweighting each target law in proportion to price. Pinning down the exact form from that paper and verifying it numerically is the first deliverable.
On the arithmetic side we reuse the method of Calibration of the Bass Local Volatility model. That method recasts calibration as a fixed-point operator on quantile functions. The operator is non-expansive in the $\infty$-Wasserstein distance (the worst-case gap between quantile functions) and a strict contraction modulo translations, and it converges in about 15 iterations in the paper's synthetic tests. Mapping the solution back gives a geometric Bass martingale, which we then simulate. As an independent second solver we will implement the $L^2$-lifted gradient flow of The Gradient Flow of the Bass Functional in Martingale Optimal Transport. That paper lifts the Bass objective to a space of random variables, shows the gradient is driven by the classical optimal transport map, and proves exponential convergence in one dimension for compactly supported marginals.
Neither paper treats the geometric case. Both prove their guarantees under compact support, which excludes the log-normal targets that matter most here. Studying behavior on truncated versus untruncated log-normal mixtures is therefore part of the work, not an afterthought.
Stopping and the certificate. The second component solves the American problem under the calibrated model and extracts $\kappa$. In one dimension we use a fine finite-difference solver in log-price, with the calibrated local volatility read off the Bass map. We also train a neural stopping policy in the style of Deep optimal stopping, which learns sigmoid stop-or-continue decisions from simulated paths and gives a lower bound from the policy and an upper bound from a dual. The certificate is applied to both.
The theoretical part of the project is to bound $\kappa$ by the local-volatility error of the discretization. This rests on the observation in On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models that smooth fit is equivalent to the first-order optimality of the exercise threshold. Under that equivalence, a kink is a non-zero derivative of value with respect to the boundary, which yields a direct error bound. That paper works with time-homogeneous Lévy processes and one-sided thresholds. Extending the equivalence argument to a time-inhomogeneous diffusion with a moving boundary is the specific open step.
Optimal Stopping of Integral Functionals and a "No-Loss" Free Boundary Formulation constrains the design. It shows smooth fit can genuinely fail for irregular diffusions. Where target marginals have gaps or the calibrated volatility degenerates, a non-zero kink may be real, so the certificate is only claimed inside irreducible components of the marginals.
Representations. The third component compares discrete representations of the same calibrated model. These are the atomic "semidiscrete" quantile representation of Acciaio et al., a piecewise-linear quantile interpolation, and a small neural parameterization of the quantile map.
As a deliberately narrow stress test of the certificate, we also train a small sequence model to emit exercise-boundary values in three ways. The first uses ordinary digit tokens with cross-entropy. The second uses xVal: A Continuous Number Encoding for Large Language Models, which encodes each number as one embedding scaled by its value and decodes it through a regression head. The third uses digit tokens trained with the distance-aware loss of Teaching Metric Distance to Discrete Autoregressive Language Models, which replaces one-hot targets with discretized Gaussians around the true value.
Prior work reports that digit tokenization snaps continuous outputs onto grids and produces staircase artifacts (Golkar et al.; Alampara et al.). A staircase boundary should produce a characteristic non-vanishing kink pattern. The probe is not meant to propose sequence models as option pricers. It is meant to show that the certificate detects representation artifacts the price alone hides.
Baselines
Naive reference. The do-nothing baseline is the Black–Scholes geometric Brownian motion fitted to the at-the-money implied volatility, with no marginal calibration. It shows how much American prices move once the full marginals are respected. When the targets are exactly log-normal, the geometric Bass martingale must reduce to geometric Brownian motion, because geometric Brownian motion is the unique process that is both an arithmetic and a geometric Bass martingale (Backhoff-Veraguas et al.). This case doubles as an exact correctness check. There, a dense binomial tree supplies the reference American put price, and the solver must match it and show a vanishing kink.
Arithmetic Bass local volatility. The strongest competing model is arithmetic Bass local volatility calibrated with the fixed point of Acciaio et al.. It is the closest existing fast, exactly calibrated method. It uses the same target marginals, the same quantile grid sizes, and the same iteration cap. The reported reference figures are convergence in approximately 15 iterations and simulation up to 80 times faster than classical Dupire. Beating it does not mean faster calibration, since we inherit its speed. It means producing materially different American prices and exercise boundaries on non-log-normal targets, with the difference confirmed by a kink certificate that is smaller for the geometric model. If the two models agree within Monte Carlo error, that is a negative result we will report.
Dupire local volatility. Dupire local volatility on interpolated strikes is the industry default. We run it on the same discrete set of maturities and strikes, interpolated with a standard smooth scheme. No reported result was found on Lacuna for American pricing accuracy under Dupire versus Bass models, so the comparison is new.
Neural stopping reference. For the stopping layer, the reference is the Deep optimal stopping policy network, which on the Bermudan max-call benchmark reports a lower bound of 136.55 against an upper bound of 136.98 at 500 dimensions, in 500 seconds on one GPU. Our multi-asset runs use independent geometric Bass components in a handful of dimensions. No reported result was found on Lacuna for that benchmark at low dimension, so we will rerun the method ourselves under an identical simulation and training budget. The recurrent policy-gradient method of Deep Recurrent Optimal Stopping is included as a second neural baseline. The file quotes no exact numerical result for it.
Ablations. The key ablation removes the certificate and substitutes the two conventional error indicators: vanilla repricing error at the calibration maturities, and price change under one grid halving. The certificate earns its place only if its rank correlation with true American price error, across many discretizations, is clearly higher than both.
Why not just…?
Why not just use Dupire local volatility on log-prices? Dupire needs prices at a continuum of strikes and maturities. Acciaio et al. document that the required interpolation and extrapolation of discrete quotes is numerically unstable. Our Dupire baseline measures this directly. If Dupire matches the geometric Bass solver on prices and kinks for smooth synthetic targets, the case for the new solver rests only on sparse-maturity regimes, and we will say so.
Why not just use arithmetic Bass local volatility, since positive terminal prices already force a positive martingale? This is the real threat. When the target marginals sit on the positive half-line, the arithmetic Bass martingale also stays positive, so the textbook "negative prices" objection does not apply. The models differ instead in how volatility is spread along paths. One minimizes variation of the price, the other variation of the log-price. That difference matters only for path-dependent and early-exercise payoffs. Our head-to-head American pricing comparison on bimodal and heavy-tailed log-marginals is designed to settle whether the difference is economically material. A null result would redirect the project toward the certificate itself, which applies to both models.
Why not just refine the grid until prices stop moving? Refinement detects non-convergence but not convergence to the wrong model. A miscalibrated volatility converges stably to a wrong price. Refinement also gives no localization. The certificate costs one solve, and because the true model has $\kappa\equiv 0$, it flags exactly the boundary segment where volatility is wrong. The ablation above tests whether this advantage is real.
Why not just learn the exercise boundary with a neural net, or a sequence model, and skip the solver? In one dimension a finite-difference solver is cheaper and more accurate than any learned policy. The curse of dimensionality that motivates Deep optimal stopping does not bite. We use learned representations only as objects to be certified, not as competitors.
Proposed Experiments
The data are synthetic target marginals in log-price. These include log-normal (the exact-answer case), mixtures of log-normals with two modes, and log-Student-t. We also use log-transformed versions of the truncated normal, logistic, and Student-t targets from the Acciaio et al. calibration benchmark, each in truncated and untruncated form to probe the compact-support assumption. The file lists no market option datasets, so market quotes are out of scope here.
Calibration metrics are the $\infty$-Wasserstein distance between computed and target quantile functions, the number of iterations to tolerance, and agreement between the fixed-point solver and the gradient-flow solver. Stopping metrics are the American put price and its error against a high-resolution reference, the exercise boundary error, the gap between the neural lower and upper bounds, and the integrated kink $\int_0^T \kappa(t)\,dt$ across exercise frequencies.
Three outcomes would decide the project. In the log-normal case, the solver must reproduce the binomial American put and drive the kink to zero under exercise-date refinement; failure here falsifies the solver. Across roughly a hundred discretizations of non-log-normal targets, with varying atom counts, interpolation schemes, and network widths, the integrated kink must rank-correlate with true price error more strongly than repricing error and grid-halving change. Finally, on bimodal log-marginals, geometric and arithmetic Bass must differ in American price by more than Monte Carlo error, or we report that they do not.
The sequence-model probe is secondary. It succeeds if digit-token boundaries show a non-vanishing kink that xVal and distance-aware training remove, with the certificate catching this when the price error alone is too small to notice.
References
- Geometric martingale Benamou–Brenier transport and geometric Bass martingales
- American options under periodic exercise opportunities
- Calibration of the Bass Local Volatility Model
- The Gradient Flow of the Bass Functional in Martingale Optimal Transport
- Deep optimal stopping
- On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models
- Optimal Stopping of Integral Functionals and a “No-Loss” Free Boundary Formulation
- xVal: A Continuous Number Encoding for Large Language Models
- Teaching Metric Distance to Discrete Autoregressive Language Models
- MatText: Do Language Models Need More than Text & Scale for Materials Modeling?
- Deep Recurrent Optimal Stopping