The two results worth taking to a desk are the closed-form cumulant series and the polynomial-degree theorem. Both give independent checks on smile-expansion code. Their value exceeds the reported runtime figures.
The setting is a forward-variance model,
dS/S = sqrt(V)(ρ dW + sqrt(1-ρ²) dW⊥),
with the curve driven by an adapted functional f and deterministic kernel g. Its variance contract, M(T) = ∫ ξ_0(u) du, gives the zeroth-order total implied variance. Alòs, Gatheral and Radoičić (AGR) showed that the cumulant generating function (CGF) of the log price has a forest expansion. Diamond-product recursion generates binary trees, with leaves labelled by either the variance contract M or the log price X.
For a tree of order ℓ containing m leaves of the first kind and x of the second, ℓ = 2m + x - 2. Its symmetry weight is w_τ = 2^{-s(τ)}, where s counts nodes with identical subtrees. At the shifted Fourier argument u - i/2, the tree contributes to the CGF through λ_a^m a^x, multiplied by its weight, with λ_a = a(a-1)/2.
The source of k³ and k⁴
Smile recovery requires matching Fourier integrals order by order, following Gatheral's 2006 matching condition. A direct expansion of both exponentials puts formal moments B[λ_a^{n-1}] on the right-hand side. The tree products must match those moments, whose degree in log-strike is high. If j trees have total order ℓ, their product generates powers of k through degree ℓ + 2(j-1). Every degree above ℓ must disappear. AGR's Remark A.1 describes the vanishing k³ and k⁴ coefficients as a mysterious cancellation, a description repeated in this paper's introduction. Lewis and then Bergomi and Guyon encountered the same cancellation at second order.
Bourgey and Gatheral make the inversion explicit by defining
Bq = q(1/2 - ∂_k) I_0(k)/I_0(k),
where I_0 is the Gaussian kernel in the Fourier integral. Set ζ = 1/2 + k/M and θ = 1/M. Hadamard's lemma then reduces B to exp(-θ/2 ∂_ζ²), the backward heat semigroup. Consequently, B maps (a - 1/2)^j to θ^{j/2} He_j((ζ-1/2)/√θ). The prefactors for individual trees follow directly: B[a] = 1/2 + k/M, while B[a²] = k²/M² + k/M - 1/M + 1/4.
The matching condition's right-hand side becomes the integral of the formal moment generating function G(y) = B[e^{yλ_a}]. Its logarithm has the closed form
K(y;ζ,θ) = yκ²/(2(1+θy)) - (1/2)log(1+θy) - y/8,
with κ = k/M. Expansion yields K_1 = λ_ζ - θ/2 and K_m = (-θ)^{m-1}(κ²/2 - θ/(2m)) for m ≥ 2. Since λ_a is quadratic, each K_m remains quadratic in ζ and contains θ^{m-1}.
Differentiate the matching condition and apply the heat equation for B. The result is a nonlinear backward heat equation for the smile, with the simple initial condition Σ̃(ζ,0) = ψ̃(ζ)/λ_ζ. Sorting terms by the number of tree factors gives a recursion. Each source for the j-tree contribution uses only prefactors with fewer trees and the K_m.
Degree exactly ℓ
The second-order calculation contains the full mechanism. Moment matching assigns the square of the order-one tree the prefactor
(1/2)B[λ_a a²] - (1/2)B[a]²B[λ_a].
Its quartic and cubic terms cancel identically. What remains is
(1/(4M))(-5k²/M² - 2k/M + 3/M + 1/4).
Those higher powers never enter the recursion. From the start, its source is the quadratic S_2 = -(5/4)ζ² + (3/4)ζ + θ/4. Two θ-derivatives followed by a Taylor expansion recover the same prefactor. Proposition 3.8 extends the result: for a j-tree product of total order ℓ, the prefactor equals (-θ)^{j-1} times a polynomial of degree exactly ℓ in ζ, with a positive leading coefficient.
At order ε^ℓ, degree exactly ℓ is immediately useful.
Any independent implementation gets a free unit test: a k^{ℓ+1} term signals a bug, and each additional tree in a product must supply a factor (-1/M)^{j-1}. The θ-counting further implies that, at fixed k, multi-tree prefactors contain M^{-(j-1)}. This scaling applies only to the prefactors.
Does order six beat order two?
The analysis ends with that structural result. The authors clearly divide the work between a model-independent cumulant series and universal prefactors computed once. Coefficients through order four appear in the appendix, with those through order six in the repository. Computing the diamond-tree expectations remains a separate task for each kernel g and functional f_t(ξ). The paper leaves that model-dependent half untouched.
Reported symbolic generation takes 0.086 s and 19.3 s through orders six and ten. These are medians of five runs on an M5 MacBook Air, including imports and rendering. They are one-time costs. A user who downloads the repository pays zero.
We found no accuracy numbers in the paper. There is no Monte Carlo benchmark, exact-price comparison, truncation-error statement in T, |k| or vol-of-vol, or radius of convergence for the ε-series. Validation remains internal, with the authors acknowledging its scope through agreement with the AGR Bell-polynomial computation "as far as we have compared them, through order six". They also disclose extensive generative-AI assistance while exploring ideas and checking intermediate arguments, and assert that they verified the work themselves. Publication of both code and coefficients provides the appropriate response, while the AGR comparison supplies the audit trail.
Our use case and the missing evidence
We are building a strategy for US listed ETF options from end-of-day prices, Greeks and implied volatilities. We have not tested the paper's method. The paper provides an implied-variance expansion. We must choose and calibrate a specific forward-variance model, so every result will depend on that choice. Calibration and hedging also remain ours to define.
Because the expansion is perturbative, we must select a truncation order and examine whether calibration remains stable out of sample at long maturities and far strikes. Our option data are end-of-day and contain neither executable bid-ask quotes nor an intraday surface. Any apparent relative-value edge therefore cannot be validated after true option spreads or intraday hedging cost.
Evidence that would change my view is a table comparing implied-volatility errors with Monte Carlo at orders two, four and six. It should use one power-law-kernel parameter set, report a one-month and a one-year maturity, and extend to two standard deviations in k. A material reduction in near-money error at order four would turn the bookkeeping into a pricing tool. Otherwise, the degree theorem remains a clean account of an old cancellation and a strong test for smile-expansion code.