Robust maximization of asymptotic growth (Q453248)

The present paper addresses the question of how to invest optimally in a market when the financial planning horizon is long and the dynamics of the underlying assets are uncertain. Here, determining a robust way to invest across potential misidentifications is desirable (if not indispensable). Let \(X\) be the coordinate mapping defined on the canonical space of continuous functions from \([0,\infty)\) to \(\mathbb{R}^d\). Here, \(n\) might be thought of as representing the (relative) price of certain underlying assets, discounted by some baseline wealth process. It is assumed that there exists a probability measure \(Q\) under which \(X\) has a dynamics of the form \(dX_t= \sigma(X_t)\,dW^Q_t\). Here, \(c= \sigma\sigma'\) is representing the instantaneous covariance matrix, and \(W^Q\) is a standard Brownian motion under \(Q\), where \(Q\) acts as a ``dominating'' measure used to form a class of probabilities \(\Pi\) containing an unknown representative which capctures the true dynamics of the process. Apart from a technical assumption, it is required that, for \(t\geq 0\), each \(P\in\Pi\) is absolutely continuous with respect to \(Q\) on \(\sigma(X_s, 0\leq s\leq t)\) implying that the volatility process of \(X\) under each \(P\in\Pi\) is the same. Hence, the allowable models are not permitted to be wildly inconsistent with one another. Given the set \({\mathcal V}\) of all possible positive stochastic integrals against \(X\) (starting from unit initial capital), the asymptotic growth rate of \(V\in{\mathcal V}\) under \(P\in\Pi\) is the largest \(\gamma\geq 0\) such that \(\lim_{t\to\infty} P((1/t)\log V_t\geq\gamma)= 1\). The solution of this problem is given in terms of a generalized version of the principal eigenvalue-eigenvector pair of the eigenvalue equation \[ {1\over 2} \sum^d_{i,j=1} c_{ij}(x) {\partial^2\eta\over\partial x_i\partial x_j}(x)= -\lambda\eta(x),\quad x\in E \] (\(E\) being specified in defining \(\Pi\)). The optimal strategy is only based on the current level of \(X\). This is in contrast to the situation considered, e.g., by \textit{P. H. Algoet} and \textit{T. M. Cover} [Ann. Probab. 16, No. 2, 876--898 (1988; Zbl 0642.90016)]. The concept of robust growth optimality is related to that of robust utility optimization; see, e.g., \textit{I. Gilboa} and \textit{D. Schmeidler} [J. Math. Econ. 18, No. 2, 141--153 (1989; Zbl 0675.90012)], where, in contrast to the present paper, penalty functions are considered. In the paper by \textit{T. Knispel} [Ann. Appl. Probab. 22, No. 1, 172--212 (2012; Zbl 1237.91244)], it is assumed that the class of models considered there, can only be identified up to a (stochastic) affine perturbation away from a fixed model, whereas in the present paper there is no such a (a priori) fixed model. Finally, robust growth-optimal trading strategies are related to optimal strategies as considered by \textit{D. Fernholz} and \textit{I. Karatzas} [Ann. Appl. Probab. 20, No. 4, 1179--1204 (2010; Zbl 1206.60055)]. In fact, robust growth-optimal wealth processes can be regarded as long-term limits of optimal arbitrages.