Optimization

Transfers

Optimization is the mathematical discipline of finding the best element from a set of alternatives, given an objective function and constraints. As a mental model, it has colonized decision-making language so thoroughly that we barely notice: we “optimize for” outcomes, seek “the best trade-off,” worry about “local maxima,” and evaluate solutions against “constraints.” The model’s power is its precision — it makes the structure of choice legible as landscape navigation. Its risk is that precision about the wrong objective is worse than vagueness about the right one.

• The objective function — optimization requires a single scalar measure of goodness. This is the model’s most consequential import: it forces you to define what “better” means, numerically. In engineering, this is often natural (minimize weight, maximize throughput). In human domains, choosing the objective function is the decision. “Optimize for shareholder value” and “optimize for employee well-being” produce different strategies from identical constraints. The model cannot help you choose the objective; it can only tell you how to pursue one.

• Constraints define the feasible region — you cannot simply maximize the objective; you must do so within limits. Budget constraints, physical laws, ethical boundaries, and regulatory requirements all carve out the space of permissible solutions. The model’s structural insight is that constraints are not obstacles to optimization but part of its definition: the optimal solution is always optimal given its constraints. Removing a constraint changes the optimum.

• Local vs. global optima — the most widely borrowed concept. A local optimum is a solution better than all nearby alternatives but not the best overall. Hill-climbing algorithms get stuck on local peaks because every small step downward looks like a mistake. The metaphor transfers to careers (staying in a comfortable role that prevents reaching a better one), organizations (incremental improvement that forecloses radical redesign), and strategy (optimizing an existing business model instead of searching for a superior one). Escaping local optima requires accepting temporary degradation — stepping downhill to reach a higher peak.

• Trade-offs and Pareto frontiers — when multiple objectives conflict, no single solution can optimize all of them simultaneously. The Pareto frontier is the set of solutions where improving one objective requires worsening another. The model imports the idea that trade-offs are not failures of analysis but structural features of multi-objective problems. Any solution on the frontier is “optimal” in the Pareto sense; choosing among them requires a value judgment the mathematics cannot provide.

Limits

• Goodhart’s problem — “when a measure becomes a target, it ceases to be a good measure.” Optimization amplifies whatever the objective function rewards, including unintended consequences. Teaching to the test, gaming metrics, and reward hacking in AI systems all follow the same pattern: the optimizer finds solutions that score well on the formal objective while violating the spirit of what was intended. The model assumes the objective function captures what you actually want; in practice, all objective functions are imperfect proxies.

• The landscape is not given — mathematical optimization assumes a fixed search space. In creative, social, and entrepreneurial contexts, the landscape itself is being constructed. A startup does not search a pre-existing space of products; it creates new points in the space. A scientist does not optimize within known theories; she invents new frameworks. The optimization model cannot account for landscape-altering moves because they are, by definition, outside the current search space.

• Incommensurable values — the model requires reducing all considerations to a common currency (the objective function). But many important decisions involve values that resist comparison: how much safety is worth how much convenience, how much freedom is worth how much equality, how much present pleasure is worth how much future security. Forcing these into a single scalar produces a number, but the number conceals the moral choices embedded in the weighting.

• Computational intractability — even in domains where the objective function is well-defined, many optimization problems are NP-hard: the search space grows so fast that finding the true optimum is infeasible. The model’s prescriptive advice (“find the best solution”) becomes vacuous when the best solution is computationally inaccessible. In practice, we satisfice — find a good-enough solution — but the optimization frame makes satisficing feel like failure rather than rationality.

• Temporal myopia — optimization typically evaluates solutions at a point in time. But many systems operate over time, and the best solution now may be suboptimal dynamically. Over-optimizing for current conditions produces brittle systems that fail when conditions change. Robust systems often appear suboptimal under any single set of conditions because they maintain slack, redundancy, and adaptability that static optimization would eliminate.

Expressions

• “Optimize for X” — the generic form, treating any goal as a mathematical objective function
• “We’re stuck in a local maximum” — the diagnostic for incremental improvement that cannot reach a better solution without disruption
• “That’s a constrained optimization problem” — reframing a messy situation as a formal structure, often to justify a particular trade-off
• “There’s no silver bullet” — folk version of the no-free-lunch theorem, acknowledging that no single solution optimizes all objectives
• “Perfect is the enemy of good” — the satisficing counter to optimization’s perfectionism
• “Hill climbing” — algorithmic metaphor for incremental improvement, borrowed into strategy and career advice
• “Pareto optimal” — technical term that has leaked into management language to describe solutions where nothing can improve without something else getting worse

Origin Story

Mathematical optimization has roots in the calculus of variations (Euler, 1744; Lagrange, 1788), which sought curves and surfaces that minimized or maximized quantities. Linear programming, developed by George Dantzig in 1947, made optimization practical for military logistics and industrial planning. The simplex method could solve problems with thousands of variables, and operations research became a discipline.

The metaphorical expansion accelerated in the late 20th century as computing made optimization algorithms ubiquitous. Machine learning reframed intelligence as optimization (minimize a loss function), and Silicon Valley adopted optimization as a worldview: A/B test everything, measure everything, optimize everything. The language migrated from engineering into personal productivity (“optimize your morning routine”), relationships (“optimize for compatibility”), and governance (“optimize policy outcomes”). The expansion has been so thorough that Herbert Simon’s alternative — satisficing, finding solutions that are good enough — reads as a contrarian position rather than the common sense it originally was.

References

• Dantzig, G. “Maximization of a linear function of variables subject to linear inequalities” (1947) — the birth of linear programming
• Simon, H. “A Behavioral Model of Rational Choice” (1955) — satisficing as the realistic alternative to optimization
• Wolpert, D. & Macready, W. “No Free Lunch Theorems for Optimization.” IEEE Transactions on Evolutionary Computation 1.1 (1997) — formal proof that no optimizer dominates across all problems
• Goodhart, C. “Problems of Monetary Management” (1975) — the law that metrics degrade when targeted