Stir a cup of coffee. As long as the liquid never tears and nothing spills, at least one point ends up exactly where it started. The reason is purely geometric: if every point had moved, the direction "from $f(x)$ toward $x$" would be defined and continuous everywhere; follow it to push each point out to the boundary and you get a continuous retraction of the disk onto its boundary circle. But a disk can be shrunk to a point and the circle cannot. Fixed points exist because the disk has no hole.
None of the three conditions can be dropped. Compact (closed and bounded) stops points from escaping to infinity; convex guarantees no hole — rotation of an annulus has no fixed point; continuous forbids the map from tearing neighbouring points apart. Kakutani extended $f$ to set-valued maps, because the "best choice" is often not unique — and that is the version game theory uses.
A topological fact about holes became the universal currency of existence in economics: Nash equilibrium, general equilibrium, and solutions of the Bellman equation all come out of this one mould. But the beauty carries a crack — Brouwer's proof is non-constructive: it tells you the solution is there, not where. That crack was later formalised as the complexity class PPAD.
By contrast, Banach's theorem demands that the map be a contraction ($d(f(x),f(y))\le c\,d(x,y)$, $c<1$). A far stronger hypothesis, and a far larger payoff: the fixed point is unique, iteration is guaranteed to converge, and you get a rate for free. Contraction means you can compute it; mere continuity means you can only know it exists.
The "invisible hand" was a metaphor until Arrow and Debreu turned it into a theorem in 1954. The idea is plain: fix a vector of prices, let everyone work out how much they want to buy and sell, and add it all up to get the excess demand for each good. Raise the price where demand exceeds supply, lower it where supply exceeds demand. That price-adjustment rule is itself a continuous map from prices to prices — and its fixed point is exactly the price vector at which every market clears at once. Equilibrium is not computed by anyone; it is the point at which the map reproduces itself.
$p$ is the price vector and $Z(p)$ total excess demand. The numerator is "raise the price in proportion to excess demand"; the denominator normalises onto the simplex — which both supplies the compact convex set Brouwer requires and encodes the fact that scaling all prices changes nobody's choice. A fixed point of $T$ means no price needs adjusting; Walras's law $p\cdot Z(p)=0$ (everyone spends what they earn) then implies markets really do clear.
First, Walras's law means $n$ markets carry only $n-1$ independent equations — exactly matching that scale invariance. One symmetry consumes precisely one degree of freedom, the same bookkeeping as Noether's theorem (Day 18).
Second, and more worth remembering: the Sonnenschein–Mantel–Debreu theorem proves that aggregate excess demand can be almost any shape at all, beyond being continuous, homogeneous of degree zero, and obeying Walras's law. Individual rationality places almost no constraint on the aggregate. Micro-level order does not propagate upward into macro-level order: existence is secured, uniqueness and stability are lost together.
Locational marginal pricing in electricity markets is a general equilibrium actually being solved every day: many goods clearing at once under network constraints, with prices read off as Lagrange multipliers; cloud spot markets and ad-slot allocation share the mechanism. But there is a hard computational wall — solving general equilibrium, like Nash equilibrium, is PPAD-complete. "The market will find the equilibrium" is a claim about computation, and that computation may admit no efficient algorithm.
What is an option worth? Intuition says: look at how much the stock rises on average. The answer is that you don't. An option can be replicated — hold a certain number of shares plus some cash, continually adjusting the share count as the price moves, and the portfolio matches the option's value in every possible state. Two paths that agree everywhere must agree in price now, or buying one and selling the other is riskless arbitrage. Price comes from replication, not prediction.
Everything hinges on "adjusting": hold more shares when the price rises, fewer when it falls. That ratio is exactly $\Delta=\partial V/\partial S$. Once directional risk is hedged away, the only risk left is the size of the wobble — so the expected return $\mu$ vanishes from the formula entirely, leaving only the volatility $\sigma$.
$V(S,t)$ is the option's value, $\sigma$ the volatility, $r$ the risk-free rate. The second term comes from Itô's lemma: along a random path the second-order term cannot be dropped ($(\mathrm{d}W)^2=\mathrm{d}t$), so convexity $\partial^2V/\partial S^2>0$ is itself value-creating — the wilder the wobble, the more the option is worth. The third and fourth terms are the financing cost of the hedge. Note that $\mu$ is absent.
After a change of variables it is the heat equation (Day 22): uncertainty diffusing through probability space and heat diffusing through a metal bar share one operator.
The second layer of beauty is a change of measure (Day 21). Eliminating $\mu$ is equivalent to swapping the real-world probability for the risk-neutral measure — a fictitious world in which every asset earns $r$, and pricing is a single expectation taken there. Change the measure and a PDE becomes an integral.
The everyday use is implied volatility: feed the traded price back into the formula and solve for $\sigma$. Black–Scholes earns its keep not by producing prices but by translating prices into a comparable quantity. And the recovered $\sigma$ traces a "volatility smile" across strikes, which is precisely a refutation of the model's lognormal assumption — practitioners knowingly use a wrong model as a coordinate system. LTCM in 1998 and the 2008 crisis point at the same place: the model assumes continuous trading and unlimited liquidity, and markets go illiquid exactly when the hedge is needed.
Three voters rank A, B, C as A>B>C, B>C>A, and C>A>B. Pairwise: A beats B 2:1, B beats C 2:1, C beats A 2:1. The "majority view" is therefore A>B>C>A — a cycle. Every individual preference is a well-formed order; the aggregate is not. That is Condorcet's paradox, and in 1951 Arrow pushed it to its limit: this is not a defect of majority rule in particular.
A social choice function aggregates everyone's ranking into one social ranking. Three requirements, all of which look mild:
Arrow's theorem: with three or more alternatives, no aggregation rule satisfies all three. It does not exist.
It converts "what counts as fair" from an endless moral argument into a decidable mathematical proposition — and the answer is that the kind of fairness you wanted is unavailable. This belongs to the family of Gödel, Turing, and Galois (Day 42): not that we have yet to find the method, but that the method cannot exist.
Watch IIA above all. It forbids the rule from using any information about the intensity of preference, permitting only order. Relax it and you escape the theorem — Borda counts and score voting both violate IIA, at the cost of opening room for strategic manipulation. An impossibility theorem is a list telling you what you must give up.
Fusing several ranking signals in search and recommendation is social choice in disguise, and optimal Kemeny aggregation is NP-hard. Closer to home is RLHF: the reward model compresses the pairwise preferences of many annotators into a single scalar reward function. Arrow's warning is that this is not a technical detail but a normative choice — flattening orders into a scalar quietly assumes that preference intensities are comparable and additive (that is, it abandons IIA). Mathematically, nothing guarantees that "human values" admit a consistent aggregate.