What is "best" when you can't improve one goal without hurting another?
Topic 23 keeps one decision-maker but gives them several objectives that genuinely conflict, so “optimal” stops being a point and becomes a curve. Game theory (topics 11 to 15) drops the other assumption: when another party is also choosing in response to you, your best plan depends on theirs and there is no single answer to compute, only an equilibrium to find. It underwrites markets that clear through auctions, security proofs written as games between an adversary and a defender, and populations that settle into stable mixes of strategy.
Almost every real decision has more than one axis of “good”: cost and risk, speed and quality, profit and emissions, coverage and privacy. Most of this course assumed those could be folded into a single objective number. Topic 23 is what you do when they honestly cannot, when the axes conflict, so pushing one down pushes another up, and there is no exchange rate between them that everyone would agree on.
The organising idea is Pareto optimality (Vilfredo Pareto, 1906). A solution is Pareto-optimal, or efficient, if you cannot improve any one objective without making at least one other worse. The set of all such solutions is the Pareto front, and it is the honest output of a multi-objective problem: not an answer, but the complete menu of non-wasteful trade-offs. Anything not on the front is dominated (beaten outright by something on it) and can be discarded with no argument. Choosing a point on the front is a values judgement, and the method's job is to present the front clearly, not to pretend it can make that call for you.
This is the exact mirror image of topic 19. There, constraint satisfaction had no objective and the only question was feasibility. Here there are several objectives and feasibility is easy: the difficulty has moved entirely into what “best” even means. Between them, the two topics bracket the assumption every other topic on this page quietly relied on: that there is exactly one number to optimise.
With objectives f1, …, fk all to be minimised over a feasible set X, point x dominates point x′ if fi(x) ≤ fi(x′) for every i and strictly less for at least one. x* is Pareto-optimal if nothing in X dominates it. The front is the image of all Pareto-optimal points in objective space. Two standard methods turn this back into ordinary single-objective problems:
weighted sum: minimise Σ wᵢ fᵢ(x) , wᵢ ≥ 0 , Σ wᵢ = 1
ε-constraint: minimise f₁(x) s.t. fⱼ(x) ≤ εⱼ for j ≠ 1
Every weighted-sum optimum is Pareto-optimal (for strictly positive weights). The one-line proof: if something dominated it, that something would have a strictly smaller weighted sum. But the converse fails, and the figure is a worked counterexample: P5 = (8, 2) is Pareto-optimal, yet no weight vector makes it the weighted-sum winner, because it lies above the lower convex hull of the front. Beating P4 in the weighted sum needs w < 1/4; beating P6 needs w > 1/3; no w does both. Sweeping 100,001 weight values in the script confirms it: the winners are P1, P2, P3, P4, P6, never P5.
The ε-constraint method has no such blind spot. Fixing “cost ≤ 8” and minimising delay returns P5 directly, and varying the bound traces the entire front, convex dents included (Haimes, Lasdon & Wismer, 1971). Its price is that each front point costs a fresh constrained optimisation, where weighted-sum only re-weights. For problems where the front is large and irregular, the workhorses are evolutionary. NSGA-II (Deb, Pratap, Agarwal & Meyarivan, IEEE Transactions on Evolutionary Computation 6(2), 182–197 (2002)) evolves a whole population toward the front at once, keeping it spread out, which ties this topic straight back to topic 10's genetic algorithms, and more loosely to topic 15's replicator dynamics. Verification: a script recomputes this.
The classic instance is financial: Markowitz's efficient frontier (1952) is precisely the Pareto front of expected return against variance, and picking a point on it is the investor's risk appetite, not a theorem: work that shared the 1990 Nobel Memorial Prize in Economic Sciences.
Every engineered object you own An aircraft wing trades weight against drag against manufacturing cost; a chip layout trades speed against power against area; a delivery schedule trades fuel against on-time rate against driver hours. None has a natural exchange rate, so the standard practice is to compute the Pareto front (often with NSGA-II) hand it to the people who own the decision, and let them pick. The method's honesty is the point: it makes the trade-off explicit instead of smuggling a weighting in under the label “the optimal design.”
And on the quantum side of this page Designing a control pulse for a quantum gate is multi-objective in practice: gate speed against fidelity against leakage out of the computational subspace, with no fixed rate between them. A faster gate is a noisier gate; suppressing leakage costs time. Pulse-engineering tools compute the trade-off surface and a human picks the operating point: the same shape as the wing, on a chip cooled to fifteen millikelvin. This is also why topic 22's robust pulses and this topic's Pareto pulses are usually built together: you want a point on the trade-off front that also survives the uncertainty set.
Check yourself
Why can the weighted-sum method never return the Pareto-optimal plan P5?
Beating P4 needs a weight below 1/4 and beating P6 needs one above 1/3: no weight does both. The epsilon-constraint method finds P5 directly.