When Herbert Simon introduced satisficing in 1956, he challenged a foundational assumption of economic theory: that rational agents maximize expected utility. His proposal seemed almost heretical—agents who set an aspiration level and accept the first alternative meeting it, abandoning the search for optimum. For decades, this was treated as a concession to bounded rationality, a descriptive patch on normative theory.

Recent formal work reveals something more provocative. Under realistic conditions—costly information, bounded computation, non-stationary environments—satisficing is not merely tractable but provably optimal. The mathematics is unforgiving: when search costs exceed marginal gains from continued exploration, aspiration-level strategies dominate exhaustive optimization in both expected utility and computational efficiency.

This article develops the formal architecture underlying these results. We examine how aspiration thresholds should be calibrated as a function of information costs, derive optimal stopping rules from sequential search theory, and identify ecological conditions under which satisficing strictly outperforms maximization. The conclusion is not that maximization is wrong, but that the choice between strategies is itself a decision problem—one whose solution often points away from the maximizing prescription that decision theory has historically privileged.

Consider an agent sampling from a distribution F over alternatives, paying cost c per observation. Let A denote the aspiration level—the threshold at which the agent stops and accepts. The expected payoff is V(A) = E[X | X ≥ A] − c·E[N(A)], where N(A) is the random number of samples until acceptance, geometrically distributed with parameter 1 − F(A).

Differentiating yields the canonical first-order condition: the optimal aspiration A* equates marginal expected gain with marginal information cost. Formally, A* solves E[X − A* | X ≥ A*] = c / (1 − F(A*)). This is structurally identical to the reservation wage equation in McCall's labor search model, but the interpretation generalizes far beyond labor economics.

The comparative statics are illuminating. As c increases, A* decreases monotonically—costly search rationalizes lower thresholds. As the tail of F thickens, A* rises, because the option value of continued search grows. Crucially, when the agent is uncertain about F itself, A* incorporates a robustness penalty derived from the worst-case distribution within an ambiguity set.

This framework dissolves the apparent tension between satisficing and optimization. The satisficing agent is not abandoning rationality but executing a second-order optimization: choosing a threshold rather than choosing an alternative. The threshold itself encodes the agent's beliefs about the environment and the structure of search costs.

Empirical neuroeconomic studies—particularly those using sequential sampling paradigms with fMRI—reveal that ventromedial prefrontal cortex tracks something close to A* rather than the value of the currently considered option. The brain appears to implement aspiration-based decision rules, computing acceptance thresholds rather than performing exhaustive value comparisons.

Takeaway

Satisficing is not a failure of optimization but optimization performed at a higher level of abstraction—the agent optimizes the threshold, not the choice.

When alternatives arrive sequentially and recall is imperfect, the problem becomes one of optimal stopping. The mathematical structure was formalized by Wald and refined by DeGroot: at each stage t, the agent computes the continuation value W_t = E[max(X_{t+1}, W_{t+1})] − c and stops when the current observation exceeds W_t.

For stationary infinite-horizon problems with i.i.d. draws, W_t collapses to a constant W* satisfying W* = E[max(X, W*)] − c. This reservation value W* is mathematically equivalent to the satisficing aspiration A*—a result that unifies what appeared to be distinct theoretical traditions. Satisficing under sequential search is optimal stopping.

Finite horizons introduce non-stationarity. The reservation value declines as remaining opportunities diminish, producing the well-known secretary problem dynamics. Here the optimal policy is explicitly time-dependent: A*_t = g(t, T, F, c), where T is the horizon. Decision-makers facing deadlines should rationally lower their standards as time runs out.

Extensions to correlated draws, learning about F, and multi-armed bandit structures preserve the threshold form but complicate its computation. Gittins indices provide an elegant solution for certain bandit problems, demonstrating that even highly complex sequential decisions reduce to comparing scalar reservation values across alternatives.

What emerges from this analysis is a deep isomorphism: heuristic satisficing, normative optimal stopping, and observed neural threshold-crossing dynamics describe the same underlying mathematical object. The dichotomy between heuristic and optimal evaporates under sufficient theoretical scrutiny.

Takeaway

Optimal stopping reveals that lowering one's standards as opportunities dwindle is not weakness but mathematical necessity—the reservation value must decay with the horizon.

Gigerenzer and Brighton's ecological rationality program asks not whether a strategy is optimal in the abstract, but whether it succeeds in the environments where organisms actually operate. Formal analysis reveals environments where satisficing strictly dominates expected utility maximization, even in expected-value terms.

Consider environments with high-dimensional alternatives and small samples. Maximization requires estimating a value function over the full alternative space, incurring variance that scales with dimensionality. Satisficing requires only estimating whether each alternative exceeds a scalar threshold—a problem of vastly lower statistical complexity. The bias-variance tradeoff favors satisficing as dimensionality grows.

Robustness theorems strengthen this conclusion. Under Knightian uncertainty about F, the minimax-regret optimal policy is often a threshold rule with aspiration set to guarantee a floor on payoffs. Maximization is fragile to misspecification; satisficing is structurally robust. This explains the prevalence of threshold-based decision rules in domains characterized by deep uncertainty—medicine, military operations, ecological management.

Computational complexity provides another lens. Many real-world decision problems are NP-hard under maximization formulations but admit polynomial-time satisficing approximations with provable performance bounds. The PAC-learning framework formalizes this: aspiration-based agents can guarantee approximate optimality with sample complexity exponentially smaller than full optimization.

These results carry a striking implication. The historical privileging of expected utility maximization as the normative standard reflects assumptions—costless computation, known distributions, stationary environments—that rarely hold. When the assumptions are relaxed to match the conditions of actual choice, the normative ranking inverts.

Takeaway

Rationality is not a property of strategies but of strategy-environment pairs—a heuristic that fails in one ecology may be provably optimal in another.

The mathematics of satisficing dissolves an old dichotomy. What appeared as a pragmatic compromise—accepting good enough because best is unattainable—reveals itself as a coherent optimization framework operating at the level of decision rules rather than individual choices. The aspiration threshold is not a heuristic shortcut but a sufficient statistic.

This reframing has consequences beyond decision theory. If the brain implements threshold-based rules because they are optimal under realistic constraints, then descriptive and normative theories converge rather than diverge. The persistent empirical violations of expected utility theory may reflect not human irrationality but the inadequacy of expected utility as a normative benchmark for boundedly informed agents.

The deeper lesson is methodological. Asking whether agents are rational is incomplete; we must specify rational with respect to what objective, under what constraints, in what environment. The mathematics of satisficing suggests that much of what we call irrationality is rationality misidentified by an inappropriate yardstick.