In a theoretical model of delegation, a decision-maker who knows only the average of an agent's private information is given a better worst-case guarantee by a randomized upper limit than by a fixed cap. The comparison ranges over every information distribution consistent with that known mean, so the result addresses uncertainty about the shape of the distribution, not just uncertainty about one value. It is a mathematical result, not evidence from workers, firms or markets.
The proposed mechanism draws a cap, called Z, independently of the hidden state and reveals it to the agent. The agent receives the delegation interval from zero to Z; under the model's incentive-compatible rule, the action is the smaller of the privately observed state and Z. The cap varies randomly, but the instruction remains a single upper-bound rule.
A rule built around uncertainty
The question is whether simple delegation survives when the principal lacks a full probability model. In the baseline setup, the state theta lies between 0 and 1, the agent observes it privately, and the principal commits to an incentive-compatible direct mechanism before nature chooses the distribution. Nature may choose any distribution on that interval with the known mean mu, where b denotes the constant bias, subject to mu being greater than b.
The sequence matters. The mechanism has to be chosen before the worst-case distribution is selected, so the objective is max-min: find the mechanism whose guarantee is highest after accounting for the least favorable distribution permitted by the mean. There are no sampled records or empirical participants in this exercise, and there is no control group to compare with the mechanism.
The model also narrows the problem to a specific loss function and incentive structure. It uses quadratic loss, meaning deviations are penalized through squared error, assumes a constant upward bias in the agent's preferred action, and includes no transfers or participation constraints. Those choices define the setting in which the reported optimum has been established.
A mathematical worst case
Under Assumption 1, the paper constructs a pair - a randomized mechanism and a least-favorable distribution - that satisfies saddle inequalities. In ordinary terms, the mechanism is robust against the constructed distribution, while that distribution is the one that makes the mechanism's guarantee hardest to improve. The max-min and min-max calculations meet at the same value, V-star, across the class of incentive-compatible mechanisms and the allowed mean-constrained distributions.
The reported robust value, called V-star, is the negative variance of the constructed worst-case distribution. Variance is a measure of spread, so the formula links the robust value to how dispersed the state is in that least-favorable distribution. It is not a payoff estimated from a survey, experiment or market record; it is the value delivered by the model's optimization.
That worst-case distribution has two recognizable pieces. Across its continuous region, its survival function - the share of probability remaining above a given state - follows an exponential pattern. At the upper endpoint, the distribution places an atom, or a discrete lump of probability. The result therefore does not rely on a single smooth curve across the whole state space.
One structural result explains how the random cap can still be recovered as a simple rule. Its expected-action schedule is nondecreasing and concave as the hidden state rises, and the slope of that schedule uniquely determines the distribution of caps. The paper uses this relationship, along with a double-indifference argument, to build the saddle point.
The key comparison is that randomization strictly improves the principal's robust guarantee relative to deterministic caps. The authors interpret the gain as a form of hedging: different cap values expose the mechanism to different worst-case distributions, while the implementation remains an upper-bound rule. The claim is about the guarantee under the model's ambiguity set, not a promise that randomness will improve decisions in every organization.
The fixed-cap benchmark has two regimes. When mu is at most 1 - 2b, every cap z between 1 - 2b and 1 is optimal. When mu is above 1 - 2b, the unique optimal fixed cap is z = (1 + mu) / 2 - b. These formulas provide the reference class for the paper's strict comparison with randomization.
The boundaries of the result
The paper's main ambiguity setup assumes the mean is above the bias. It separately examines a boundary regime in which mu is below b. There, the optimal mechanism needs an additional incentive-neutral outcome lottery, and pure random caps are strictly suboptimal. The cap idea remains part of the construction, but it is not by itself sufficient in that regime.
An extension considers mean-preserving-contraction, or MPC, ambiguity under Assumptions 1 and 2. The paper reports another saddle point for a random-cap mechanism in that setting, with a value equal to the negative variance of the relevant worst-case distribution. Because this is an extension under additional conditions, it does not remove the dependence on the model's assumptions.
In a separate signal-design benchmark, full pooling is the worst case. Two rules are robustly optimal there: a singleton rule that sets the action to mu - b for every signal, and a deterministic cap at mu - b. This benchmark addresses how information is signaled rather than changing the baseline claim about distributional ambiguity, so it remains a related theoretical comparison.
The assumptions leave several reasons for caution. The analysis notes that a state-dependent bias could make the candidate cap distribution infeasible or block a global saddle point. It also says a single upper bound may be too restrictive, with richer stochastic intervals or other delegation sets potentially needed. Nonquadratic preferences could make higher moments or the full action distribution matter, rather than the particular variance expression used here.
The financial-contracting interpretation is reduced-form: it does not model verification decisions, audit costs, default reporting, enforcement or investment incentives. Nor does the preprint provide empirical validation or show causal effects in people, organizations or markets. It is an arXiv version 1 document dated 20 August 2026, with no journal publication or DOI reported in the supplied metadata.
The supplied analysis therefore presents random caps as a conditional benchmark, not a universal prescription. Its open questions include state-dependent bias, nonquadratic preferences, randomizing both upper and lower bounds, a fuller financial foundation, and tests in real organizations or markets.
Paper data and sources
Original title: Random Cap: Optimal Informationally Robust Delegation
Authors: Zhiyuan Jia
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text