Marini's Bass-type calibration accepts quoted option prices without first reconstructing a risk-neutral density. Whether it can replace an existing Bass local volatility implementation remains a question for a timed comparison. The paper calls itself "a parallel computing approach," but we found no timed run anywhere in it. Its guarantees depend on mathematical assumptions. Read it as a design document.
Building the q_x-Bass model
Marini starts from the Bass local volatility model of Conze and Henry-Labordère, which uses Brownian motion as its reference process. That model fits full marginals exactly at finitely many maturities. Marini instead uses q_x-Bass martingales, replacing the single reference distribution with kernels q_x indexed by starting point. The coupling between two dates minimizes the average squared 2-Wasserstein distance from the model's conditional law to the reference kernel. Those kernels are transition laws for a reference diffusion dM = σ̂(t,M) dB. Any diffusion coefficient σ̂(t,x) is admissible if it is Lipschitz in x and has square-integrable solutions, as required by the paper's (A1) and (A2).
The continuous-time objective is to minimize E∫‖σ_t − σ̂(t, M_t)‖²_HS dt over Brownian martingales satisfying the market constraints. Evaluating the reference coefficient along the candidate path M gives the greatest weight to mismatches in regions the calibrated model visits. Computation begins with a partition T that includes every quoted maturity. Marini chooses n+1 marginals increasing in convex order to minimize the sum of n one-step transport costs. Conditional expectations along the reference filtration extend each step to continuous time. The paper calls the joined process a Bass interpolation.
Theorem 3.3 bounds the gap between this interpolation's cost and the true continuous optimum by C√|T|. Here |T| denotes the largest step. The result relies on Lemma A.1: over a step of length h, transport cost is bounded above by (1 + C√h) times the volatility cost of any martingale. For the Bass extension itself, the reverse inequality holds.
What fills the gaps between quotes?
The practical attraction is that "calibration constraints can be imposed directly in terms of observed derivative prices, without requiring the reconstruction of entire risk-neutral marginal distributions." A constraint can have the form E[(M_{s_i} − K_j)^+] = c_{i,j}. More generally, it can use any payoff affine on finitely many convex regions, covering calls, puts, baskets and spreads. This avoids both a Breeden-Litzenberger step and strike interpolation.
The space between strikes still has to be filled. Each date in the numerical formulation has a grid of m atoms, with an (n+1)×m row-stochastic weight matrix as the unknown. Proposition 4.4 establishes that taking the infimum over all grids recovers the undiscretized multimarginal problem's value on the same partition. Condition (11) places every atom within the convex hull of the following date's grid. Beyond that condition, we found no guidance on selecting m or locating the atoms.
The reference model also fills the gaps. As the paper observes, partial marginal information can lead its volatility criterion and a maximum-covariance criterion to choose different optimal chains. Fix the marginals, however, and both criteria give the same optimal coupling. Choosing the objective therefore expresses a view about unquoted prices.
For a vol desk, the latent-factor extension is the most interesting idea. It turns a Heston-style variance V into a martingale through Y_t = e^{κt}(V_t − θ), while imposing calibration constraints only on S. The optimization treats Y's dynamics jointly with the rest of the model; those dynamics need not survive. The paper contrasts this construction with the stochastic Knothe-Rosenblatt model of Beiglböck, Hasenbichler and Pammer. Their model prescribes the price marginals and retains the latent dynamics.
Parallel work, coupled projection
The decomposition does provide parallel work. Given a fixed chain of n+1 marginals, computing the subgradient requires n independent two-marginal dual problems. The Generalized Martingale Sinkhorn Algorithm solves each problem, and the solves can run concurrently. The outer projection remains coupled. All (n+1)×m weights are projected onto a convex set containing every price and convex-order constraint. This requires a constrained quadratic program over all (n+1)×m weights.
There are costs beyond those parallel solves. Every step needs reference kernels q^{(i)}_x; the paper itself cautions that a nontrivial σ̂ may require PDE solves. The kernels concentrate as |T| shrinks. Remark 4.6 proposes rescaling them to improve conditioning in the Sinkhorn step. Nor does that step necessarily reach an optimal dual potential within finitely many iterations. Marini describes the proposed step size, τ_k = 1/(k+1), as "not necessarily the most effective." When the partition is large, an additional odd/even split of the chain may help. The paper acknowledges that "convergence to a global optimizer is not ensured in general."
The conditions behind the error bound
The square-root-of-mesh guarantee is quite conditional. Theorem 3.3 requires a discrete optimizer that exists and admits a Bass interpolation. Even its constant depends on more than structural parameters: the proof folds a constant times MOT, the unknown optimal cost, into C. In Lemma A.1, the constant also contains C′ = 2exp(2L²(t₁ − t₀)). It therefore grows exponentially with the squared Lipschitz constant times the horizon. The mesh rate is real. Its multiplier could be large.
Convergence of the projected subgradient method in Theorem 4.8 assumes an optimizer in K^ε, a set that holds every weight strictly away from the feasible region's boundary. That margin and the assumed existence of an interior chain with irreducible steps guarantee a Bass martingale at every step. An optimal chain seeking zero weight on an atom falls outside this assumption. The analysis also assumes exact duals. Marini states the restriction explicitly: "For simplicity, we restrict our convergence analysis to the idealized setting in which the dual potentials are computed exactly." Approximate subgradients receive treatment only through citation.
We could not run this ourselves.
The natural test would use the paper's relaxed formulation, whose penalty λΨ allows inconsistent quotes. An honest choice of tolerance requires bid-ask quotes for the option panel, and we have no bid-ask price data. The method also supplies a model without a trading signal.
Wall-clock time against the Bass local volatility fixed point, using the same n maturities, is the comparison that matters most. Until someone measures it, the C√|T| bound establishes consistency and says nothing about speed.