Price every name in a long-only book against its own worst-case expected payoff, approve each position, and the combined trade can still give money away. Zhang shows the failure in Example 2.1 with six numbers. I would pin this part of the paper above a risk desk. The infimum falls outside the sum, and that placement drives the result.
Separation produces a shadow price
The market has finitely many states and discrete time, with one riskless numéraire and M risky assets quoted two-sided. Purchases cross the ask and sales hit the bid. Establishing a position and liquidating it therefore use different sides of the spread, so a single linear functional cannot price both.
Finite-dimensional separation handles the single-period case. Subtract the positive orthant from attainable payoffs and the result is a closed polyhedral cone. No arbitrage separates that cone from the nonnegative quadrant. Strong separation then supplies a strictly positive vector f.
Test the separating inequality with one unit of the riskless asset and again with minus one unit. Its components f_k/f_0 must sum to 1. Normalising them produces the pricing measure Q, which assigns strictly positive mass to all K states.
For a purchase of one unit of asset m, the time-0 ask must be at least the Q-expected time-1 bid. For a sale of one unit, the time-0 bid cannot exceed the Q-expected ask. Those inequalities make the initial spread overlap the expected terminal spread. Any point in that overlap defines a shadow price S satisfying S(0) = E_Q[S*(1)].
Once short sales are banned, the sale strategy disappears from the test. The long inequality remains, while martingale equality gives way to S(0) >= E[S(1)].
Zhang carries the same construction onto a finite path tree E^{T+1}, allowing a possibly nontrivial initial sigma-field. Nodewise solvency cones and their duals describe the market. A Farkas argument is then applied separately to every initial subtree, producing an adapted dual process Z with Z_0 > 0, density Z_0(T)/Z_0(0), and shadow prices Z_m/Z_0.
The final step replaces a single pricing measure with a family Q. Zhang defines a lower condition through the infimum of conditional expectations over Q and an upper condition through the supremum.
The paper contains no data and claims none. Its only worked figures are the two-state example and the corresponding multi-period restatement. I checked the arithmetic of those six numbers.
What remains of uncertainty?
Every state must receive positive mass from at least one measure in the family P. Under this support assumption, Lemma 2.1 selects one such measure for each state and averages them. The resulting R_P = (1/K) sum of at most K members of P has full support. For nonnegative X, sup over P of E_P[X] > 0 holds exactly when E_{R_P}[X] > 0.
Zhang states the consequence in the introduction and repeats it in the conclusion. No arbitrage depends on the family only through the union of its supports, and "the model is not nondominated in the technical sense." He explains why the reduction is made explicit: the probability family could otherwise be confused with the auxiliary family of pricing measures introduced later. The abstract makes the same point in its second sentence.
The collapse is intentional and disclosed. Yet the title and abstract still sell model uncertainty, while the surviving uncertainty concerns support. After P is reduced to a full-support reference measure on a finite state space, bid-ask duality under a strategy cone supplies the remaining substance. The proof of Lemma 3.1 says so directly: "the same multiplier construction is the finite-state FTAP of Kabanov and Stricker (2001) (see also Schachermayer (2004))." Denis and Martini (2006), Bouchard and Nutz (2015, 2016) and Bayraktar and Zhang (2016) appear in the introduction as the nondominated literature. None of the quasi-sure machinery associated with those citations enters the proofs. An implementer arriving for model-uncertainty apparatus instead finds transaction-cost geometry.
Little guidance from constrained duality
The paper leaves an important selection problem unresolved. Without constraints, the two test inequalities force S(0) into the intersection of two intervals. A convex weight alpha_m in [0,1] then determines S(1) through the martingale property.
With the short-sale constraint, the converse proof chooses the ask for S(0) and the bid for S(1). The supermartingale inequality follows automatically. This works as an existence proof, yet offers little as a valuation rule: the construction places the shadow price at a corner of the spread at both dates. Anyone using that price for marks or hedges must provide a separate selection criterion. The paper gives none.
Corollary 2.1 and Corollary 3.1 are thin in the same way. A lower system exists because Q may be the singleton containing the supermartingale measure already obtained. The introduction calls this "an existence corollary, not a new equivalent characterization."
Six numbers
Consider two risky assets with identical initial spreads [0.45, 0.50] and no spread at time 1. Asset 1 pays 1.2 in state one and 0.1 in state two. Asset 2 reverses those payoffs, returning 0.1 and 1.2. Let Q contain the two Dirac measures, and assign both shadow prices 0.50 at time 0.
For either asset, the infimum over Q of the expected terminal payoff is 0.1. Each therefore clears the lower condition because 0.50 >= 0.1. Now hold minus one unit of cash alongside one unit of each risky asset. The cost, minus 1 plus 0.50 plus 0.50, is zero. In either state, terminal value is minus 1 plus 1.2 plus 0.1, leaving 0.30.
Viewed separately, both assets look expensive against their worst cases.
Together, they are free money.
The ordering of the infimum causes the failure. With nonnegative holdings, the infimum of the expected portfolio dominates the sum of the per-asset infima. That direction cannot provide an upper bound for the portfolio. Long-only sizing rules based on worst-case expectations for individual names therefore miss the trade. Worst-case analysis has to occur at portfolio level, with aggregation performed in the working order.
Zhang's multi-period restatement replaces the Diracs with full-support measures whose one-step conditional probabilities are (0.9, 0.1) and (0.1, 0.9). At every node, either payoff has a lower conditional expectation of 0.21 against a shadow price of 0.50. The portfolio continues to pay 0.30 on every terminal path.
The abstract acknowledges that the lower condition is not sufficient, and the conclusion says the same, so Zhang does not conceal the gap. My objection is narrower. The lower side supplies only an existence corollary, while the upper side is merely sufficient. The Q-family results consequently give no characterization. Remark 3.3 agrees that the pair should not be presented as a single necessary-and-sufficient statement unless Q satisfies additional stability assumptions, which the paper does not identify.
Strict spreads on the tree
The backward recursion over modified dual cones is laid out precisely. Remark 3.2 describes it as necessary only: every consistent price system belongs to those cones, while their nonemptiness is explicitly never substituted for the complete dual argument. Implemented directly, the recursion gives a screen.
The main construction applies the Farkas alternative to each subtree. Trades are separated into nonnegative purchase, sale and free-disposal variables. Node multipliers obey flow equalities for cash and inequalities for constrained risky holdings.
Efficient friction is required at every node, meaning each dual cone must have nonempty interior. This restricts the eligible assets. The dual cone is {lambda(1,s): s_m in [bid, ask]}, whose interior is nonempty only when bid < ask at every node.
A zero-spread asset falls outside the scope.
When the initial sigma-field is nontrivial, initial cost becomes a random variable. Both the no-arbitrage test and the valuation interval must then be compared node by node at time 0, rather than treated as scalars.
The paper also draws a clear boundary around its valuation claims. The bounds pi and pi-bar are "dual expectation bounds only". Interpreting them as arbitrage-free extension prices or superhedging prices would require a claim-extension result and a superhedging duality theorem, neither of which is assumed.
Our own examination ended after one pass with nothing to report. Testing the mechanism requires observed bid and ask quotes for each asset at each node. We have daily open-high-low-close bars, without bid-ask quotes. A flat cost assumption would substitute an assumed spread and test that assumption instead of the paper's inequality. There is nothing empirical to measure anyway: Example 2.1 consists of six numbers and one inequality.
Stability or pasting conditions on Q could change my view if they joined the lower and upper systems into one dynamic characterization. The computational question also remains open in the paper, which says only that backward recursion "may provide practical valuation bounds". Without either development, the lasting contribution is the 0.30 counterexample. That example alone earns the paper a read.