Mostovyi and Zariphopoulou reduce the full preference sensitivity of optimal terminal wealth to a convergent power series, available to arbitrary order with recursively specified coefficients. Access requires a single random variable, Ŷ_T. No price history I know of supplies it.
The paper is theory throughout. It contains no data, no sample period, no universe and no performance figure of any kind: no return, no Sharpe, no alpha. Its contribution is a sensitivity calculus for utility maximization in a general semimartingale market.
Two consequences arrive immediately for anyone trying to build from the result. First, the generic semimartingale market contains no named tradable asset class. Any implementation must supply one, and our substitute would be liquid US ETFs. Second, the results depend on continuous-time semimartingale assumptions and a stochastic dominance condition that produces a distinguished dual optimizer. Price history reveals neither. A discrete-time, finite-scenario utility optimization would therefore extend their idea rather than test their theorem. For a practitioner, the live question is which objects in the paper can actually be assigned numbers.
Where does the linearity come from?
The preference class is CMIM. Its defining property is that the inverse marginal I = (U')^{-1} is completely monotonic. Bernstein's theorem represents I as a Laplace transform, I(y) = ∫ e^{-yz} mu(dz), where mu is a nonnegative measure on (0, ∞). With Log utility, mu(dz) = dz and I(y) = 1/y. For Power utility satisfying 1/q + 1/p = 1, mu(dz) = z^{-q} dz / Γ(1-q). A finite sum of these measures produces I(y) = Σ y^{1/p_i - 1}, a mixture of CRRA branches. Preference becomes a measure, and the perturbation is affine in that measure: mu_eps = mu_0 + eps·nu.
Assumption 1 supplies the market side. Under stochastic dominance, one dual element Ŷ dominates every Y in the dual domain. Through the authors' earlier work with Sîrbu, this is equivalent to infinite-order stochastic dominance. The same Ŷ then attains the dual infimum for every utility in the class.
The dual value function consequently becomes exactly affine in the perturbation parameter: v(y, eps) = v_0(y) + eps·v_1(y). Its representation has two terms. Preference enters the primal problem through severe nonlinearity, while the dual traces a straight line.
From there, the calculation reduces to a scalar equation. The budget condition is g_0(y) + eps·g_1(y) = x, with g_0(y) = E[Ŷ_T I_0(y Ŷ_T)]. Complete monotonicity gives (I_0)' < 0 pointwise, and therefore g_0'(y_0) = E[Ŷ_T² (I_0)'(y_0 Ŷ_T)] < 0. The analytic implicit function theorem yields an analytic Lagrange multiplier. Its coefficients satisfy y_m = -A_m / g_0'(y_0) and y_1 = -g_1(y_0)/g_0'(y_0).
Repeated differentiation of the composition brings in the partial Bell polynomials through Faà di Bruno. The appendix lists the first six, B_{1,1} through B_{3,3}. According to the authors, Bell polynomials have, to the best of their knowledge, never previously been used systematically in portfolio optimization.
One gain deserves full credit. Because the negativity of g_0' follows directly from complete monotonicity, the proof avoids the bounded relative risk aversion condition 0 < c ≤ -xU''/U' ≤ C used in Kramkov and Sîrbu's second-order theory. The authors observe that CMIM does not in general imply bounded relative risk aversion. CMIM is therefore a genuinely different class.
The coefficients in practice
For the value function, u_1(x) = v_1(y_0) and u_2(x) = -(v_1'(y_0))² / v_0''(y_0). Strict convexity of v_0 makes the second coefficient nonpositive for every admissible perturbation direction. The paper makes this statement specifically about the coefficients in the primal value function's expansion around eps = 0, with no broader claim. Among the conclusions, the sign of u_2 alone survives every choice of nu. Near eps = 0 and to second order, linear extrapolation overstates value in either direction.
The optimizer has a parallel expansion. Write Z_m = y_m Ŷ_T. Then X_0 = I_0(Z_0), X_1 = (I_0)'(Z_0)Z_1 + J(Z_0), and X_2 = (I_0)''(Z_0)Z_1² + (I_0)'(Z_0)Z_2 + 2J'(Z_0)Z_1.
The CRRA example carries the clearest economic content. Set a = p_0/(p_0-1), b = 1 + 1/(p_1-1) and M(α) = E[Ŷ_T^α]. The reference solution is y_0 = (x/M(a))^{p_0-1} and X̂_T = (x/M(a))·Ŷ_T^{1/(p_0-1)}. Its first correction is y_1 = (1-p_0) y_0^{1/(p_1-1)-1/(p_0-1)+1} · M(b)/M(a). Two moments of Ŷ_T determine the entire first order.
The measure mu_eps must remain nonnegative, which constrains the range of eps. In this example, the expansions around eps = 0 are right-sided only. A second CRRA branch can be blended in. Removing one creates a different problem.
Truncation comes before application
The authors establish analyticity in a neighbourhood of eps = 0. I found no quantitative radius and no error bound. They also provide no numerical experiment that compares a truncated expansion with a directly solved perturbed problem. Anyone cutting the series at second or third order needs that comparison.
The abstract promises "convergent analytic expansions of arbitrary order". The paper never answers the practical charge: convergent over what range, and accurate to what? Nor does eps have scale-free units. In the expression for y_1, the wealth-dependent factor y_0^{1/(p_1-1)-1/(p_0-1)+1} means that an unchanged nominal eps produces different corrections at different initial wealth.
The difference between the paper's two halves is larger still. Preference perturbations extend to arbitrary order. Once beliefs are added, the result stops at exactly two first-order coefficients, y_{1,0} and y_{0,1}, under G_y(y_0,0,0) < 0. Remark 6 gives the reason. Higher order requires analyticity of delta ↦ Ŷ_T^delta in a topology compatible with the budget equation, which the authors describe as a substantially stronger assumption. They leave it for future work, reasonably.
Their one explicit market model is complete Black-Scholes-Merton. It has one riskless and one risky asset, no transaction costs and no portfolio constraints, with H_mu = -(1/σ) H_T (W_T + θT). On my reading, completeness collapses the dual domain and makes dominance free. The assumption carrying the general theorem therefore costs nothing in the only market model developed on the page.
The perturbation direction nu also lacks identification from observed information. The authors say this in the introduction, where statistical elicitation and calibration of Bernstein measures appear among the open problems, alongside approximation of general preferences by CMIM utilities. Their framing is open work, and I take it as such. The user must supply both unobservables: the direction of preference change and the dual element whose moments M(a) and M(b) determine its price.
We could not test this. Ŷ_T is the maximal element of the dual domain under infinite-order stochastic dominance, and no price history I know of provides that object. The paper makes a narrower concession that belongs beside ours: statistical elicitation and calibration of Bernstein measures is listed as an open problem.
One result would change my view: a worked incomplete-market example that constructs Ŷ explicitly, followed by a count of how many orders are needed for the series to match the exactly solved perturbed problem at an eps large enough to alter an allocation. Until then, the tradeable content consists of two items, the sign of u_2 and the reduction of a preference change to one scalar equation. Both advance the incomplete-market literature. They still fall short of an allocation rule.