Automated analysis of the paper

A global SPX-VIX calibration cannot fit a smile beyond the reach of monthly stitching when both methods share the same prescribed SPX marginals. Acharya, Sun, Augustino, Chakrabarti, Sureshbabu and Che prove the equivalence of exact local and global feasibility.

Bid-ask bands require a narrower claim than fixed marginals. A joint global optimization can work when a prescribed sequential rule fails, since that rule may choose an interpolated shared marginal outside the compatibility region. The authors draw the boundary clearly. Their stronger practical result concerns joint selection and minimum relaxation, rather than a broader exact-feasibility set. Global coupling instead creates a strictly larger family of path laws. Every member still matches each monthly calibration, leaving measurable cross-period model risk invisible to the calibration instruments.

The construction

The paper extends Guyon's joint SPX-VIX problem to m SPX maturities and m-1 VIX maturities, spaced thirty days apart, under zero rates. Any calibrated law for the full state vector must match the SPX and VIX smiles. It must also satisfy a conditional martingale row for the next SPX level and a conditional dispersion row. The latter equates the forward-starting log-contract L(x) = -(2/tau) ln x, tau = 30/365, with VIX squared.

Stitching calibrates each three-variable block (S_i, V_i, S_{i+1}) on its own, then joins adjacent blocks through their shared SPX marginal. The global program applies the same martingale and dispersion identities conditional on the full history F_i, rather than only on (S_i, V_i).

The first theorem makes the feasibility result exact. Begin with any globally feasible law and replace each continuation kernel by its S_i-conditional version. This block-preserving SPX-Markovization, called Mmu in the paper, preserves every monthly triple law (S_i, V_i, S_{i+1}). Feasibility survives, and the resulting law is stitched. Nonemptiness of the global set, the stitched set and every monthly set therefore amounts to the same condition.

The classes separate once m is at least 3. The authors construct finite-tree marginals where V_1 equals 0.15 or 0.55 with equal probability. Either branch may arrive at S_2 = 100, while V_2 discloses the originating branch. Conditional on S_2, V_2 remains dependent on V_1. The pair (V_2, S_3) consequently retains information discarded by the stitched kernel.

This law meets every condition available to a monthly calibration while remaining outside the stitched class. Since the laws coincide on every payoff measurable within a single block, such payoffs cannot detect the Markov restriction.

For the numerical work, the authors discretize a tensor in solver coordinates (S_1, V_1, Z_1, V_2, Z_2). They divide the rows into hard constraints, handled by cyclic correction, and soft constraints, handled by penalties. Market inputs come from P-spline-smoothed surfaces observed in a single 2026-04-21 snapshot. SPX expiries are 23, 57 and 87 days; VIX expiries are 27 and 56 days. Each SPX surface has 20 strike targets and each VIX surface has 15, laid onto a grid of shape (30,25,8,25,8), totaling 1.2 million cells.

How much dependence goes unidentified?

On the finite tree, Markovization reduces Cov(V_1, V_2) from 0.0400 to 0.0110. The attenuation is 72.4%, while the global feasibility residual is 1.42e-14. Both laws still assign identical prices to every monthly instrument.

The difference appears in the VIX-spread call (V_2 - V_1 - K)+. Under the global law, its values are (0.050, 0, 0, 0) for K = (0, 0.1, 0.2, 0.3). After Markovization they become (0.113, 0.063, 0.045, 0.027). The at-the-money value more than doubles, and each out-of-the-money strike moves from exactly zero to a positive price.

This is a constructed tree, rather than a market portfolio. The paper leaves computation of the price interval across all admissible gluings for future work. It includes no P&L and no hedging test. The supported conclusion is narrower: the interval is non-degenerate.

Entropy returns to stitching

One result determines whether the larger global class changes an actual desk process. Under the paper's reference measure, which combines lognormal SPX transitions with independent draws from the VIX marginals, relative entropy splits exactly:

D_KL(mu || mubar) = D_KL(Mmu || mubar) + D_KL(mu || Mmu).

Markovization weakly lowers the objective. Every finite-entropy minimizer over the global set is therefore stitched. The authors state the implication directly, calling stitching "the minimum-information exact completion of the local calibrations."

The abstract says non-Markov dependence "requires cross-period information, an appropriate objective, or a history-dependent prior," while the conclusion says global coupling "becomes economically operative when cross-period information or objectives are supplied." The concession is accurate, and it carries the practical cost. The paper supplies or calibrates none of those inputs.

Information loss does have an exact expression: a sum of conditional mutual informations I((V_i, S_{i+1}); H_{i-1} | S_i). The missing quantity is therefore well specified. Until one of the listed inputs becomes available, the entropy objective selects the stitched law and the larger global state space contributes no non-Markov memory.

What the solver shows

For the synthetic infeasible instance, cyclic row projection ends with a marginal residual of 1.9e-2. The marginal-priority method reaches 7.4e-4, about 25 times tighter, while revealing a conditional residual of 1.1e-3 against 1.3e-5 for cyclic projection.

The instance consists of two discretized Gaussians on a 41-point grid, giving 1681 variables and 107 rows, with infeasibility controlled by a variance-ratio knob. At a ratio of about 0.7, a conditional-Jensen bound of 0.022043550052 against a target variance of 0.015749999517 confirms infeasibility. On the feasible instance, cyclic projection performs better: 5.1e-6 marginal and 8.8e-14 conditional, compared with 2.8e-5 and 2.5e-4. The authors caution that the 25x ratio depends on this row order and budget.

The headline sweep matters less than two details: the conflict between the constraint groups and the amount of tail mass excluded from the reported statistic. A conditional-only Newton post-projection lowers the displayed conditional errors to 1.4e-7 and 2.9e-10. At the same time, the worst smile error jumps from 0.531 to 10.786 vol points.

Tail treatment changes the reported fit sharply. At the preferred penalty of 1e4, the 10%-of-peak mass filter produces a dispersion statistic of 6.3e-2 and a martingale statistic of 7.6e-4. The filter keeps 78.1% of conditioning mass for the first transition and 44.3% for the second. Once every positive-mass cell is included, dispersion rises to 1.94 and martingality to 5.0e-3.

The authors disclose both versions, describe the headline figure as a thresholded bulk residual, and decline to use the table as a feasibility test. Fair. With less than half of the conditioning mass retained in one transition, however, the quoted conditional quality applies to the distribution's center rather than its wings.

The penalty path also lacks monotonicity. A setting of 3e4 performs worse than 1e4 on both conditional summaries, producing 8.7e-2 and 8.7e-4. Each row is a separate fixed-budget solve, and the authors explicitly disclaim asymptotic convergence. The 1e4 choice is consequently a tuned operating point for one snapshot. Across the sweep, the worst fitted-smile error remains between 0.54 and 0.61 vol points.

Dispersion is the binding soft family in this experiment. Holding its penalty at 100, the martingale residual shifts only from 1.7e-3 to 1.5e-3 as its own weight spans four orders of magnitude. The authors limit that finding to "this implementation and dataset."

A fixed 720-sweep schedule, tested on grids ranging from 78k to 1.2M cells, gives runtime proportional to N^1.04 with R-squared 0.9998. Single-threaded runs take roughly 80 seconds to 22 minutes. The measurement is candid and narrow. Since the tensor dimension is d = n_S n_V^{m-1} n_Z^{m-1}, linear-in-d work per sweep offers little relief as m increases. The demonstrations end with the 5-axis and merged 7-axis cases.

We ran nothing against this: there is no strategy and no return series, only two dated snapshots.

The practical opening

The misaligned-date extension comes closest to a usable improvement. On the 2026-05-13 snapshot, containing three SPX surfaces and two VIX surfaces, the 7-axis merged-timeline model has 1.12e6 cells. Its 5-axis interpolated baseline has 1.20e6.

For the first VIX surface (VIX_1), max IV error declines from 1.228 to 0.711 vol points. For the first SPX surface (SPX_T1), it falls from 0.221 to 0.059. The table shows improvement across all five surfaces, so these two columns were not selected after the fact. Max conditional-row residual is 9.4e-4.

There are limits to the comparison. The two VIX spans overlap by two days, which the paper says places the run outside its own non-overlap reference proposition. It is also a single instance drawn from a different snapshot than the main experiments.

We made a related point about a rough-Heston result in which an exact structural identity came with an unpriced calibration cost (/articles/in-rough-heston-the-conditional-density-is-exactly-lognormal). The identification theorem here is correct in the same sense. Its stitching cost is written exactly as a sum of conditional mutual informations.

Whether that cost justifies a tensor growing as n_V^{m-1} n_Z^{m-1} depends on a cross-period target that the paper assigns to future work instead of providing. A calibrated history-dependent prior, or one liquid instrument pricing the I((V_i, S_{i+1}); H_{i-1} | S_i) term, would change my view.