Duality and convergence for binomial markets with friction (Q354186)

The paper under review is to prove a limit theorem for the super-replication cost of European options in a binomial model with friction. It is a purely probabilistic proof that uses a Kusuoka construction [\textit{S. Kusuoka}, Ann. Appl. Probab. 5, No. 1, 198--221 (1995; Zbl 0834.90049)] on transaction costs in an essential way to provide a lower bound (existence of a liquidity premium) for the continuous-time limit of the super-replication costs, where the upper bound follows from compactness and two general lemmas. The authors extend homogenization techniques for viscosity solutions from [\textit{S. Gökay} and \textit{H. M. Soner}, Math. Finance 22, No. 2, 250--276 (2012; Zbl 1277.91170)] to the case of non-Markovian claims and to a more general liquidity function \(g\). The super-replication cost of European options behaves quite differently due to the various structures of the function \(g\): (i) the super-replication cost is very high if \(g\) is nondifferentiable at the origin and (ii) any continuous trading strategy has no liquidity cost if \(g\) is differentiable. The super-replication of a European claim is a European contingent claim with maturity \(T=1\) and payoff \(F_n = F(W_n(S^{(n)}))\), where \(S^{(n)}\) is the underlying stock price given by a binomial sequence. The super-replication price is defined by \[ \begin{aligned} V_n = V_n(g, F_n)& =\inf\{ x: \text{ there exists a self-financing portfolio \(\pi\) with initial capital \(x\)}\\ &\phantom{=}\;\text{ such that } Y^{\pi}(n, g) \geq F_n \, \text{\(Q\)-a.s.}\}, \end{aligned} \] where \(Y^{\pi}(n, g)\) is the portfolio value at the discrete time \(k/n\) before a transfer is made at this time. The Legendre transform (convex conjugate) \(G\) of the trading cost \(g\) plays a dual relation between \(G\) and \(g\). The penalty function is a binomial version of the linear liquidity model of \textit{U. Çetin} et al. [Finance Stoch. 8, No. 3, 311--341 (2004; Zbl 1064.60083)]. Three typical examples of the trading cost function \(g\) are given in Subsection 2.2. The truncation \(g^c_N(t, S, v)\) of \(g\) at level \(c\) is given in Definition 2.3 as the dual Legendre transform restricted on level \(c\). The first result of the paper is a characterization of the dual problem given in Theorem 3.1 in Section 3, proved in Section 4. Among all probability measures \(Q\) in the probability space, the duality between the trading cost \(g\) with the initial capital and its Legendre transform \(G\) with the payoff is \[ \begin{aligned} V_n &= \inf\{ x: \text{ there exists a self-financing portfolio \(\pi\) with initial capital \(x\)}\\ &\phantom{=}\;\text{ such that }Y^{\pi}(n, g) \geq F_n \, \text{\(Q\)-a.s.}\}\\ &= \sup_{Q} E^{Q}\left(F_n - \sum_{k=0}^{n-1} G(k/n, W_nS^{(n)}(n), E^{Q}[S^{(n)}(n)|{\mathcal F}_k] -S^{(n)}(k))\right).\end{aligned} \] For any admissible control level \(c\), \(V_n^c \leq V_n\), hence Corollary 3.6 shows that \[ \liminf_{n\to \infty} V_n \geq \sup_{a\in A}E^P\left(F(S_a) - \int_0^1 \hat{G}(t, S_a, \alpha(t, S_a)S_a(t))dt\right) \] among all bounded, nonnegative progressively measurable processes, where \(\alpha \) is related to the quartic variation density of \(\ln S_a\). The main result of the paper shows the asymptotic behavior of the truncated super-replication costs \(V_n^c\) in Theorem 3.5, which is proved in Section 5. The limiting convergence result is \[ \lim_{n\to \infty } V_n^c = \sup_{a\in A^c}E^P\left(F(S_a) - \int_0^1 \hat{G}(t, S_a, \alpha(t, S_a)S_a(t)dt\right), \] where \(A^c\) is the set of all admissible control levels \(c\) in a complete probability space \(P\) associated with the standard one-dimensional Brownian motion. Section 2 provides preliminaries and the basic model for the wealth dynamics, and super-replication and trading cost functions as well as the Legendre transform \(G\) of \(g\). Section 3 states main results of the paper and corresponding assumptions required in the theorems. Section 4 is devoted to a complete proof of the duality, and Section 5 gives a proof of the main result of limit convergence. Section 6 constructs a sequence of martingales on the discrete space that approximate a fixed martingale via Kusuoka's smooth volatility processes. Section 7 provides a technical lemma to prove the upper bound of the main result of limit convergence, and two other lemmas related to the optimal control. It is an interesting question to see if there is a nonzero gap between \(\liminf_{n\to \infty} V_n\) and \(\sup_{a\in A}E^P(F(S_a) - \int_0^1 \hat{G}(t, S_a, a(t, S_a)S_a(t))dt)\), as well to know under which assumption the gap is zero.