A CGMY European-put price in about 9 milliseconds is useful; Damircheli's per-query error bar is the part worth copying. Every neural surrogate tested against that bound fails it.

We are building a strategy around this idea for US ETF puts, which are American. The paper certifies European puts against a discrete solver. Our version therefore uses a European approximation with separate early-exercise safeguards, and does not test the paper.

Cheap certification from Toeplitz structure

The paper calls its method symbol-based residual certification: it bounds a candidate solution's error by measuring how far that solution falls short of the pricing equations. The intended use is repeated CGMY pricing for calibration and uncertainty quantification. In that setting, speed is useful only if each surrogate answer comes with a bound.

CGMY is the four-parameter tempered-stable Lévy model. Its European pricing equation is a nonlocal PIDE, with fractional order Y changing as the parameters change. The paper discretises it using quadratic B-splines on a uniform log-price grid. Since the basis functions are shifts of one another, each Galerkin matrix is a Toeplitz section of an explicit periodic function, called the periodised symbol. Truncating the symbol's alias series at |k|<=100 matches a 40-digit reference matrix to 1e-8 in operator norm. The principal-period shortcut (K=0) leaves an operator-norm error of 31% to 41% at every resolution. Anyone maintaining in-house spline code should check for that error.

Any candidate trajectory in the spline space can receive a certificate. Insert it into the Crank-Nicolson equations, then measure the residual in a dual energy norm. An energy argument bounds its worst-in-time error against the exact discrete solution. This requires a lower bound on the operator's symmetric part. A Lévy-Khintchine identity gives that bound in closed form: r times 2/15, valid for Y in (0,2) except Y=1.

To bring the cost to O(N), the method replaces the dense matrix with a banded approximation built from the mass and stiffness symbols. The resulting bound is looser, with a median looseness factor of 2.05 at N=1024. Empirical interpolation expresses the symbol in Q=36 affine terms, allowing certification of a reduced model without using the full grid. A three-stage rule accepts queries that pass the cheap reduced bound. It rejects those that could not pass even the full bound, escalating only the cases between those decisions to full order.

The entire experiment is synthetic. Sobol draws supply C in [0.05,1], G in [1,5], M in [2,9], Y in [1.1,1.8], r in [0.01,0.1] and T from 0.1 to 1 year. The strike is fixed at 100. For spots from 0.8K to 1.2K, a spot factor of 8.44 turns one cent into a trajectory tolerance of 1.18e-3.

Does the certificate earn its compute cost?

Full-order certification adds little overhead. A direct solve at N=1024 with 240 steps takes 259 ms. The banded certificate adds 6.0%, rising to 7.2% with per-query spectral data. Across N=64 to N=4096, that fraction remains between 4.5% and 5.6%.

The reduced model supplies the speed. At rank 80, mass-weighted POD-Galerkin runs in 8.3 ms, or 0.032 of direct time. POD means proper orthogonal decomposition, a reduced basis fitted to training trajectories. All 120 benchmark queries pass at the reduced stage. Across 10^4 queries, the average is 9.0 ms versus 260 ms, and 99.77% are settled without using the full grid.

Offline work takes about 1530 s. Recovering that cost requires about 6100 queries at rank 80 and 6400 at rank 64. Rank 64 is fragile: at one cent it takes 0.041 of direct time, but tightening the tolerance to 0.3 cent pushes that to 0.55. Loosen the tolerance to three-cent and it becomes the faster choice, taking 0.026 against 0.032 for rank 80.

Neural models fail the one-cent test by 46x

The best standard POD-NN, a network that predicts reduced coefficients, produces a median true error of 0.46 and an exact bound of 1.26. The bound is about 10^3 times tolerance. The space-time variant improves on this, with errors from 0.014 to 0.016 and bounds from 0.054 to 0.058. Even those bounds remain 46 to 49 times the tolerance.

Training with a residual penalty cuts the bound roughly tenfold while increasing the true error roughly fivefold. The residual penalises temporal roughness that least-squares fitting overlooks. Selecting checkpoints by the rigorous bound consequently favours the network with bound 1.26 and error 0.46 over the network with bound 12.2 and error 0.082. The chosen epochs fall between 1890 and 2000 of 2000. The author acknowledges that the networks had not converged, yet asserts that longer training cannot close a gap of 46 to 1000. The paper does not test that assertion with longer training.

Every neural query needs correction. After 2.7e4 s of offline work, the space-time model still costs 282 ms per query online. It never breaks even.

For pricing beyond a calibrated range, the out-of-box test is the useful comparison. Rank 80 stays within one cent for 210 of 264 queries outside the box. The certificate accepts 193 of these, leaving 17 false rejects. Its precision is 0.76, compared with 0.20 for a rule rejecting everything outside the box. Among the 150 in-box and 264 out-of-box queries with reference solutions, there is no false accept.

Implementation choices remain

Much of the recipe is specified: K=100, p=3, N=1024, 240 steps, and closed-form mass and stiffness symbols. The paper gives explicit warnings about residual evaluation and precision. Accurate residuals are 1e-8 of their constituents, so the usual Gram-matrix offline/online calculation loses all digits. The R-factor form is mandatory. Network outputs also need double precision: at N=512, single-precision rounding alone produces a bound of order 1e-3.

We could not find some inputs. Neither numeric endpoints for the truncated domain nor the division of the (C, Y, min(G,M)) range into 18 parameter cells appeared in what we found. Regularity and buffer assumptions are deferred to a companion submitted manuscript. The author also warns that the grid search for symbol extrema supplies no validated enclosure. An allowance covers floating-point residual error.

Greeks need care. Their derivative bounds include a factor sqrt(Lambda_j) of order h^(-j). The author says this factor can be conservative on a wide log-price domain, and it grows with refinement.

What the one-cent guarantee leaves outside

At N=1024, the maximum gap to COS is 0.235 cent across 30 diagnostic queries and two box corners. The certified bound goes on top of that gap. Reducing the time grid at N=1024 from 240 to 60 steps raises the 30-query error to 4.6 cent. Time resolution matters there. Spatial resolution remains a constraint too: at N=512, the two short-maturity box corners miss by 2.4 and 3.2 cent. The paper states the limit plainly: "What the certificate does not do is bound the continuum price", and "the certified bound is added to that budget, not instead of it."

We have not tested the paper's method. Our work differs from it: US ETF puts are American, whereas the paper demonstrates European puts only. An American certificate exists as a backward-Euler theorem with a complementarity term. The author leaves its Crank-Nicolson counterpart open.

Our implementable version uses a European CGMY approximation, with separate safeguards for early exercise and dividends. Its certificate cannot be read as a guarantee of American market values. It controls error against the assembled discrete system. Continuum error and CGMY calibration error therefore require separate checks, as do market-model error and execution costs. The strategy is under evaluation now. We have no results to report.

The paper's 3% figure comes from one core inside its synthetic parameter box. Transferring the one-cent guarantee requires measuring those four other error sources first.