A commodity desk could use stochastic spot/convenience-yield correlation with memory. Günther and Overbeck deliver the model and prove that its joint Fourier-Laplace transform closes. They provide no values for its parameters.
Begin with Gibson-Schwartz. Log spot has a drift equal to a constant less the convenience yield and half the spot variance. The convenience yield mean-reverts at rate kappa toward theta. A two-dimensional Brownian motion Z drives both quantities. In the classical model, the instantaneous covariance of Z is a fixed 2x2 matrix. Spot volatility, convenience-yield volatility and their correlation are three calibrated constants.
Günther and Overbeck replace that matrix with a process V on the cone of positive semidefinite 2x2 matrices:
V_t = V_0 + integral of k(t-s)(b + M V_s + V_s M-transpose) ds + integral of k(t-s)(sqrt(V_s) dW_s Q + Q-transpose dW_s-transpose sqrt(V_s)).
A 2x2 matrix of Brownian motions W generates Z. Take W rho + sqrt(1 - rho-transpose rho) times an independent Brownian motion, with |rho| at most 1. Rho supplies the leverage channel, allowing shocks to the covariance matrix to feed back into spot and the convenience yield.
Memory enters through the scalar kernel k. Current covariance reflects a k-weighted integral of earlier drift and martingale increments, so k determines how quickly a variance shock fades. A Markovian Wishart specification instead forces exponential decay. V_11 is spot variance, V_22 is convenience-yield variance, and V_12 divided by the square root of their product gives the moving instantaneous correlation.
The paper's citations supply the economic motivation. Routledge, Seppi and Spatt (2000) argue that the spot/convenience-yield relationship changes through time. Omura and West (2015) report stronger co-movement when inventories are low. Pindyck (2004) finds that high-volatility episodes coincide with increases in both quantities.
Memory leaves the transform intact
The mathematics closes. Theorem 3.1 writes the conditional Fourier-Laplace transform in exponential form, with a matrix Riccati-Volterra equation determining the exponent. Its validity depends on the candidate exponential being a true martingale. Theorem 3.6 then represents that exponent as an affine functional of the past path of X.
Proposition 3.4 introduces the Brownian functional N_t, defined as the integral of the trace of H(s)-transpose sqrt(X_s) dW_s. The commodity model uses it to carry leverage. Within the Riccati equation, N appears only through C_H(psi) = HQ psi + psi Q-transpose H-transpose and a term in HH-transpose. The affine form remains in place.
Theorem 4.2 is what turns the formal transform into one that can be used. Set d = 2, require Q to be invertible, and let k be twice continuously differentiable on [0,T]. Under those conditions, the complex Riccati-Volterra equation has a unique global solution with a positive semidefinite real part.
With Re psi positive semidefinite, the nonnegativity of the resolvent L and Assumption 4.1 bound the candidate exponential by 1 in modulus for every t in [0,T]. The exponential is therefore a true martingale. Assumption 4.1 matters again later. At the cone boundary, the inward-pointing condition becomes a sum of two squares. The inequality rho-transpose rho is at most 1 makes it hold. The same invertibility and smoothness conditions also give uniqueness in law.
Existence comes through a shifted kernel
The existence argument requires b_0 to dominate k(0)(d-1)alpha, where alpha = Q-transpose Q. The paper describes this as a sufficient condition rather than a necessary one. It is the Volterra counterpart of Bru's classical requirement that beta minus (d-1)Q-transpose Q belong to the cone. Multiplying the auxiliary-SDE condition by lambda = k(0) produces it. Economically, the condition acts as a drift floor tied to the kernel's value at the origin. The Markovian Wishart model has no corresponding floor.
A fractional kernel t^(alpha-1)/Gamma(alpha) has k(0) infinite, which rules it out. The paper substitutes (t+epsilon)^(alpha-1)/Gamma(alpha), with alpha = H + 1/2 in (1/2,1). At zero, this shifted kernel equals epsilon^(alpha-1)/Gamma(alpha). Since Alpha - 1 is negative, taking epsilon toward zero to recover the singularity drives k(0) to infinity. The admissible drift set then contracts to nothing.
The abstract acknowledges the exclusion before saying that shifted fractional kernels "provide a tractable specification with power-law memory". Away from the origin, the power-law decay is genuine and gives the variances persistence. Roughness comes from small-time behavior, precisely where epsilon smooths the kernel. Solution paths are Hölder continuous of every order strictly below gamma/2, with gamma at most 1. The paper explains why gamma above 1 cannot arise for a continuous kernel with k(0) positive: the relevant integral behaves like k(0) squared times (t-s). Below-1/2 Hölder regularity remains Brownian grade.
The authors state the trade plainly. Their main focus, they write, is "not the construction of genuinely rough matrix-valued Volterra processes, but rather the associated matrix transform theory, the analysis of the corresponding Riccati-Volterra equations, and their application to commodity markets." The transform theory is complete on those terms. A desk seeking persistent stochastic correlation gets something useful. Epsilon remains unfitted.
The direct SDE construction and the quadratic construction do not nest, as the authors explain through Abi Jaber 2022 and Cuchiero and Teichmann 2019. The quadratic approach admits singular kernels and obtains positivity by construction. The direct approach permits a general affine drift while requiring k(0) to remain finite. The choice is between roughness and drift flexibility.
A second restriction concerns kernel shape. The proof takes K(t) = k(t) times the identity, with k nonnegative, bounded, continuous, nonincreasing, nonnegativity-preserving and satisfying k(0) positive. Genuinely matrix-valued kernels fall outside the result.
The obstruction is explicit. At a boundary point, v-transpose K(r) sqrt(X) A Q v equals (sqrt(X) K(r)-transpose v)-transpose A Q v. This expression need not vanish because K(r)-transpose v need not belong to the kernel of X. The authors observe that a sandwich action K(.)K-transpose preserves the cone by construction. The invariance question remains open because the martingale increments themselves are not positive semidefinite.
The convenience yield cannot decay more slowly than spot.
Assumption 4.1 adds another unresolved condition. It concerns nonnegative shifted resolvents, and the commodity theorem uses it to establish the true-martingale bound. We did not find a verification that the shifted fractional kernel satisfies it.
No empirical target
The paper contains no data, estimation, futures-curve fit or option-pricing exercise. It also gives no comparison with the Markovian Wishart extension of Schneider, Six and Tavin (2025), even though that model motivates the application. Any improvement from bounded power-law memory therefore remains untested.
The contribution is theoretical, and the theory is clean. We did not find a numerical scheme for the Riccati-Volterra equation. Every price requires solving that equation, making cost per quote the first number an implementer needs.
We could not run the model. It applies to commodity futures curves, and we have no futures price data. A substitute universe would test a different model. The paper also supplies no parameter estimates that we could use.
A calibration to a futures options surface would change the verdict. At some fitted epsilon, the shifted kernel would need to beat constant-correlation Gibson-Schwartz and the Markovian Wishart version on dates outside the fit. The authors would also need to report solve time.