Context

Every full-solution structural estimator pays the same bill: each candidate parameter vector triggers a dynamic-programming solve, and the optimizer visits many candidates. NFXP pays it with a nested fixed point, MPEC moves the fixed point into the constraint set of a large optimizer. UFXP refuses to pay it at all.

Source Ideas

The construction comes from Bray. Instead of maximizing a likelihood, UFXP works with the first-order conditions of Bellman’s equation: under logit shocks, the log-odds of any two actions at a state must equal the difference of their choice-specific values. Those conditions involve the value function, which normally forces a solve per candidate. Bray’s observation is that the conditions only ever use the value function through fixed linear functionals, and each such functional has a dual representation as a fixed point in the empirical choice probabilities alone — independent of the parameters. Computing the duals once, before the search, removes the value function from the problem.

Oguz and Bray (2026) add the optimal weighting of those conditions, prove the weighted estimator is as asymptotically efficient as maximum likelihood, and use the construction to train neural-network utility functions inside dynamic discrete choice models. The econirl implementation covers the linear-utility case, where the optimally weighted estimator collapses to a single closed-form solve.

Where UFXP Fits

UFXP sits with NFXP, CCP, and MPEC in the structural family. It recovers the utility parameters in the same parameterization as the data-generating process, supports counterfactual re-solving, and reports standard errors. Like CCP it starts from inverted empirical choice probabilities, so it shares CCP’s sensitivity to thin state coverage; unlike CCP it scores all of Bellman’s restrictions with efficient weights rather than inverting once, and its optimal weighting restores maximum-likelihood efficiency. Against NFXP the trade is plain: NFXP is the exact finite-sample MLE at meaningful compute cost, UFXP is its asymptotic equal at almost no cost.