Positivity as a Probability Test: Discrete Schemes for Uncertain-Volatility Risk That Are Positive and Time-Consistent by Construction

Pricing or risk-measuring a contract when the volatility of the market is not known exactly means taking the worst case over a whole family of possible probability models. That family has to be "pasting-stable": following one model up to some time and another afterwards must give a model that is still in the family. This property is what makes the risk number computed today agree with the numbers recomputed tomorrow. Separately, numerical analysts have spent years building schemes that keep intrinsically non-negative quantities such as interest rates, volatilities and concentrations from going negative. The two literatures have not been connected. This note argues they are describing the same property seen from two sides. For a numerical scheme that shifts values by cash in the obvious way, "every non-negative claim gets a non-negative value" holds exactly when "each backward step is a worst case over genuine probability weights rather than signed weights." If those genuine weights are chosen independently at each step, the scheme is pasting-stable by construction.

The practical consequence is a design rule and a diagnostic. The rule is to build each backward time step of an uncertain-volatility solver as a maximum over non-negative, mass-one quadrature weights placed on forward nodes that themselves stay positive. The diagnostic is a pair of quantities that can be measured on any existing solver. The first is the most negative effective weight the solver implicitly uses. The second is the gap between solving over a horizon in one pass and restarting the solver mid-horizon from its own intermediate values. We expect neural solvers for second-order backward equations to fail both tests on models with a singular zero boundary. We expect the constructed schemes to pass both tests exactly, at a cost and accuracy competitive with the best positive forward schemes.

Core Claim

Consider one backward step of a discrete scheme. It maps tomorrow's value function $V_{n+1}$ to today's value $V_n$, and we write it as $V_n = S_h[V_{n+1}]$. Suppose $S_h$ is cash-additive, meaning $S_h[V+c] = S_h[V]+c$ for any constant $c$, and positively homogeneous, as a sublinear expectation is. Then three statements are equivalent. The first is that $S_h$ maps non-negative functions to non-negative values, which is the positivity property the numerical literature wants. The second is that $S_h$ preserves order, $V\le V' \Rightarrow S_h[V]\le S_h[V']$, which is the discrete comparison property. The third is that $S_h$ admits the representation

$$S_h[V](x) = \sup_{a\in A}\ \sum_{k} w_k(x,a)\, V\big(x_k(x,a)\big), \qquad w_k\ge 0,\ \ \sum_k w_k = 1,$$

where $a$ ranges over admissible volatility values in a set $A$, $x_k(x,a)$ are the forward nodes reached from state $x$, and $w_k$ are their weights. The implication from the first statement to the second is the standard argument for monetary operators. The third statement is the finite-dimensional version of the dual representation theory in Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes and Coherent and convex monetary risk measures for unbounded càdlàg processes. Those papers characterize time-consistency through acceptance-set decomposition and through concatenation of the dual functionals.

When the scenario sets are chosen state by state and step by step, the composed scheme $S_h^{N}$ is itself a worst case over all concatenations of one-step kernels. That is exactly the discrete analogue of the pasting condition. Superhedging and Dynamic Risk Measures under Volatility Uncertainty shows that this condition is necessary and sufficient for the dynamic programming principle in continuous time. If the forward nodes are additionally required to stay strictly positive, the scheme also never evaluates a singular coefficient such as $c_{-1}/x$ at a negative or zero state.

The claim is not that monotone schemes are new. The claim is that, for risk measures under volatility uncertainty, positivity of the state, positivity of the value, comparison and discrete pasting stability are one constraint, namely non-negative one-step weights on positive nodes. Treating it as one constraint does two things. It produces positive, time-consistent schemes for singular rate models under volatility uncertainty, a combination that no existing method in the file provides. It also gives a quantitative test that exposes when popular solvers silently leave the class of genuine risk measures.

Research Program

The first part is the equivalence and its quantitative version. Beyond the exact equivalence, we will prove a stability statement for perturbed operators. Suppose a solver's effective one-step operator is within $\varepsilon$ of a monotone sublinear operator in sup norm, for example because a regression or neural network approximates the conditional expectation. Then its sign violations, comparison violations and pasting defect grow at most linearly in $N\varepsilon$ over $N$ steps. To make the pasting defect measurable, define it at an intermediate time $t_m$ as

$$D_m = \big\|\, V_0^{\text{one pass}} - V_0^{\text{restart}(m)} \big\|,$$

where $V_0^{\text{restart}(m)}$ is computed by rerunning the solver backward from time $t_m$, starting from the solver's own value $V_m$ at that time. For a recursive monotone scheme $D_m$ is zero by construction. For forward-shooting neural solvers $D_m$ need not be zero. In Machine Learning Approximation Algorithms for High-Dimensional Fully Nonlinear Partial Differential Equations and Second-order Backward Stochastic Differential Equations, the Deep 2BSDE method fits the initial value and the gradient and Hessian networks globally against a terminal loss. Its intermediate values are therefore not guaranteed to be conditional worst-case values.

The second part is the construction. Forward nodes $x_k(x,a)$ are generated with the positivity-preserving maps already proved for singular rate models. The first such map is the tamed Euler-type step of Unconditionally positivity-preserving explicit Euler-type schemes for a generalized Ait-Sahalia model. It treats the $1/x$ drift implicitly through the positive root of a quadratic and tames the remaining terms. The second is the double-implicit Milstein root of Convergence, Non-negativity and Stability of a New Milstein Scheme with Applications to Finance, which attains strong order one in one dimension. In both maps the noise increment enters as a discrete quadrature point instead of a sampled Brownian increment. Weights $w_k$ are chosen to match the first two conditional moments for each volatility value $a$, subject to $w_k\ge0$. This is a small linear feasibility problem per state and per $a$. Where it is infeasible, we widen the node stencil rather than accept a signed weight. The driver, for instance a discounting or default-intensity term, enters in the tamed fashion of Convergence and qualitative properties of modified explicit schemes for BSDEs with polynomial growth. We will check that taming preserves cash-additivity and monotonicity; where it does not, we will modify it so that it does. This answers the file's Gap 2 directly: the one-step weights tell us whether taming leaves the class of risk measures.

The third part is the high-dimensional variant. In 100 dimensions an explicit stencil is too expensive. We will use the one-step operator only to produce regression targets on simulated positive paths, and fit $V_n$ with a network whose output is constrained to be non-negative. The non-monotone part is then confined to the fit, and the stability result bounds its effect. Here the program is honest about what is guaranteed and what is only measured.

Baselines

The naive reference point is an explicit Euler–Maruyama forward simulation, with the backward step computed by the same moment-matching sup over $a$ but allowing signed weights, and no positivity treatment. It is what one gets from transcribing the continuous equation directly. It shows how often states cross zero and how often singular coefficients blow up or produce undefined values as the step size grows. It is run on identical nodes and step sizes.

The strongest existing method for the high-dimensional uncertain-volatility question is the Deep 2BSDE solver of Machine Learning Approximation Algorithms for High-Dimensional Fully Nonlinear Partial Differential Equations and Second-order Backward Stochastic Differential Equations. It is the only method in the file that directly targets second-order backward equations of the kind that represent sublinear expectations, and it reports solving the 100-dimensional Black-Scholes-Barenblatt equation "within minutes". The file gives no numeric error table for that benchmark, and for the G-Brownian motion benchmark there is no reported result found on Lacuna. We will therefore rerun it under its published settings and use our own measured error as the reference. For first-order problems we include the Deep BSDE solver of Deep learning-based numerical methods for high-dimensional parabolic partial differential equations and backward stochastic differential equations. It reports a relative $L^1$ error of 0.30% on 100-dimensional Allen-Cahn and 0.17% on 100-dimensional HJB in 330 seconds, which sets the accuracy bar our high-dimensional variant must stay within on those same equations. Both neural solvers receive the same number of time steps, the same number of simulated paths, and the same GPU wall-clock budget as our method, with learning-rate tuning over an equal grid for each.

For the low-dimensional singular-rate question, the strongest competitors are the positivity-preserving forward schemes. These are the backward Euler method of On the backward Euler method for a generalized Ait-Sahalia-type rate model with Poisson jumps, proved positive for any step size with mean-square order one-half, and the tamed Euler-type method, reported as order 0.5 and "nearly three times faster" than backward Euler on its Example 5.1. Neither handles volatility uncertainty. We therefore pair each forward map with a backward step that is a sup over $a$ using unconstrained least-squares weights. This isolates whether positive forward states alone suffice, or whether the non-negativity of the backward weights is what matters. That pairing is the key ablation. A second ablation keeps non-negative weights but uses Euler nodes, which isolates the role of positive nodes.

Beating the baselines means the following. Our scheme has zero sign violations, zero comparison violations and zero pasting defect at every step size tested. Its error against a fine-step reference is no worse than the strongest baseline at matched cost. In high dimension, its relative error stays within a factor of two of the Deep 2BSDE solver while showing a pasting defect at least an order of magnitude smaller.

Why not just…?

Why not just clip negative values to zero? Clipping the value, $\max(V,0)$, does make values non-negative. It does not restore order between two claims whose values were computed with signed weights. Clipping the state is fatal in singular models, because the tamed Euler-type paper's Aït-Sahalia drift contains $c_{-1}/x$ and a clipped state of exactly zero divides by zero. The tamed backward equation paper also documents that naive explicit steps with polynomial drivers explode toward plus infinity, which no lower clip prevents. Our experiments measure comparison violations between ordered claim pairs for a clipped Euler baseline. If clipping showed none, the value-level part of our claim would be weakened. We expect clipping to fail in exactly the regimes where signed weights appear.

Why not just use a fully implicit or standard monotone grid scheme? In one or two dimensions this is a real threat. A carefully built monotone grid scheme with an implicit positive forward step would also be positive and pasting-stable. By our equivalence, it is in fact an instance of the same class. The contribution in low dimension is therefore not to beat such schemes on accuracy. It is to supply positive nodes for singular drifts cheaply, by inheriting the roughly threefold speedup of tamed Euler over backward Euler reported in the tamed Euler-type paper, and to show that the weight test identifies which schemes are in the class. In high dimension, grid-based implicit solves are intractable, which is why the neural solvers exist.

Why not just give the Deep 2BSDE network a positive output layer? A softplus output makes $V$ positive, but positivity of the output alone does not give comparison or pasting stability. The global terminal-loss fit still lets intermediate values differ from the restarted worst case. The restart test of the pasting defect is designed to settle this. If softplus Deep 2BSDE shows a negligible pasting defect and no comparison violations on the singular-rate and Black-Scholes-Barenblatt problems, the high-dimensional part of this proposal would be unnecessary, and we would report that.

Why not just discretize into an MDP with augmented actions? The augmented-action method of Risk-Averse Reinforcement Learning via Dynamic Time-Consistent Risk Measures obtains time-consistent CVaR-type risk with a contracting Bellman operator. The risk-envelope policy gradient of Policy Gradient for Coherent Risk Measures handles Markov coherent risk similarly. Both assume a single transition kernel per action and finite or low-dimensional states, and the authors of the former state that scaling the auxiliary variable to high-dimensional continuous states is open. Our one-step operator is a Markov coherent risk operator whose envelope is the non-negative moment-matching weights. The MDP view is therefore compatible with ours, not an alternative to it. What it lacks is the positive-node construction for singular continuous dynamics.

Proposed Experiments

The first experiment is the decisive low-dimensional test. It uses the generalized Aït-Sahalia rate model from Examples 5.1 to 5.3 of the tamed Euler-type paper, in both its non-critical and critical parameter regimes. The diffusion coefficient is made uncertain, ranging over an interval scaled from 0.8 to 1.2 times its nominal value, and the claim is non-negative: a zero-coupon bond and a cap paying $(r-K)^+$. We sweep step sizes from coarse to fine. For each method we measure five quantities: the fraction of paths or nodes with negative or undefined state; sign violations of the value; comparison violations across forty ordered claim pairs; the pasting defect $D_m$ at every intermediate time; and the error against a fine-step run of our scheme, cross-checked against the backward Euler pairing at the finest step. The prediction that would falsify the core claim is that the unconstrained-weight ablation with positive nodes shows zero comparison violations at coarse steps. That would mean non-negative weights add nothing beyond positive forward maps.

The second experiment is the high-dimensional stress test. It reruns the 100-dimensional Black-Scholes-Barenblatt and G-Brownian motion benchmarks of the Deep 2BSDE paper, and adds a 100-dimensional product of independent positive rate factors under uncertain volatility, with a non-negative basket payoff. We compare Deep 2BSDE, its softplus variant, and our regression-target variant, reporting relative error, pasting defect, comparison violations and wall-clock time. We also run Allen-Cahn and HJB to confirm that the monotone construction does not degrade accuracy badly against the 0.30% and 0.17% figures.

The third experiment adds jumps. It applies the Poisson-jump rate model from the backward Euler jump paper, and includes a one-dimensional Lévy-driven risk process in the style of Dynamic Risk Measures for Processes via Backward Stochastic Differential Equations Associated with Lévy Processes, which builds time-consistent risk measures from jump-driven backward equations but provides no positive scheme. The test is whether non-negative jump weights can be found without the restrictive parameter conditions the backward Euler jump analysis requires. Success in any regime the backward Euler analysis excludes would be the clearest sign that the probability-weight view does more than reorganize known results.