Straightedge and compass obey exactly two rules: draw a line through two points, draw a circle about a point. The key that opened a two-thousand-year-old problem was translating the act of drawing into arithmetic. Line meets line and you solve a linear equation; circle meets line or circle and you solve a quadratic. So each new intersection's coordinates are, at worst, the old numbers combined by $+,-,\times,\div$ and one square root. The compass is a machine that can only take square roots.
Constructible numbers are therefore locked inside a tower: starting from the rationals, each step can only double the "dimension," so the reachable levels are forever $1,2,4,8,\ldots$ — while doubling the cube demands $\sqrt[3]{2}$, of dimension 3. The trouble was never a lack of ingenuity; 3 simply never fits inside a power of 2.
$\mathbb{Q}(\alpha)$ is the number system you get by adjoining $\alpha$ to the rationals; the bracket is its dimension as a vector space over $\mathbb{Q}$ — intuitively, how many independent rational coordinates it takes to describe $\alpha$. The minimal polynomial of $\sqrt[3]{2}$ is the irreducible $x^3-2$, dimension 3: excluded. Trisecting $60^\circ$ means constructing $\cos 20^\circ$, a root of the irreducible $8x^3-6x-1=0$: also 3. Squaring the circle needs $\sqrt{\pi}$, and Lindemann proved in 1882 that $\pi$ is transcendental — not even finite-dimensional.
For two millennia people searched inside geometry; the answer lay in algebra. Wantzel's 1837 proof examines no particular construction — it builds a model of the tool itself and then shows the target lies outside its range. That is the universal recipe for impossibility: find a quantity every legal move preserves, then show the goal doesn't have it.
The same theory yields a startling positive result: a regular $n$-gon is constructible $\iff$ $n$ is a power of 2 times distinct Fermat primes (Gauss constructed the 17-gon at 19 on this basis). The cyclotomic rings $\mathbb{Z}[x]/(x^n+1)$ on that same thread are today the arithmetic stage of Ring-LWE lattice cryptography (Day 32). And "characterize what a given machine can generate" is precisely the prototype of computability (Day 24) and complexity reduction (Day 36).
Everyone memorizes the quadratic formula; cubics and quartics fell to Cardano's circle in the sixteenth century. The quintic resisted for over two hundred years, and people assumed it was merely harder. Abel and Galois answered: not harder — nonexistent.
A radical formula is nothing but "$+,-,\times,\div$ plus repeated root extraction." Now, the roots of an equation carry a symmetry: permute them however you like and the coefficients are unchanged (coefficients are symmetric functions of the roots), and these permutations form a group. Each root extraction breaks a little of that symmetry, peeling one layer off the group. Solvable by radicals means you can peel layer by layer down to nothing, each peel being "gentle" (abelian). But $S_5$ contains $A_5$: non-abelian, with no proper normal subgroup at all — an atom that will not peel. The chain breaks there.
On the left, every number system sitting between the coefficient field $F$ and the field $K$ containing the roots; on the right, subgroups of the group of symmetries (automorphisms) fixing every element of $F$. The correspondence is bijective and order-reversing: the bigger the field, the fewer symmetries preserve it. The core theorem — an equation is solvable by radicals $\iff$ its Galois group is solvable, i.e. admits a chain of subgroups with abelian successive quotients; and $S_n$ is not solvable for $n\ge 5$.
This was the first time in mathematical history that a problem was moved wholesale into another category to be solved: "can it be solved?" became "what does this group look like?" — and the latter is finitely checkable. That manoeuvre became the main engine of modern mathematics. There is a bonus beauty here: the simplicity of $A_5$ both kills the quintic formula and is exactly the source of icosahedral symmetry — the same unpeelable atom is a prohibition on one side and a Platonic solid on the other.
Galois theory over finite fields underlies Reed-Solomon codes (Day 33) and the $\mathrm{GF}(2^8)$ arithmetic inside AES. Differential Galois theory powers the Risch algorithm: when SymPy reports that $\int e^{-x^2}dx$ "cannot be integrated," that is not a search timing out but a proof that no elementary antiderivative exists. Galois representations, meanwhile, are the language in which Wiles proved Fermat's Last Theorem.
Euclid's first four postulates read like plain common sense; the fifth reads like a theorem: through a point off a line there is exactly one parallel. For two thousand years people tried to derive it from the other four. Saccheri (1733) assumed parallels were not unique and pushed for a contradiction — but only ever hit strangeness, such as triangles whose angles sum to less than $180^\circ$. He declared these results "repugnant to the nature of the straight line." In fact he was already writing the theorems of hyperbolic geometry.
The breakthrough was not to push harder on the proof but to change the question: build a world. The Poincaré disk packs the entire hyperbolic plane inside one circle; "lines" are arcs meeting the boundary at right angles, and the closer you get to the rim the shorter your ruler becomes, so the boundary is never reached. There, infinitely many parallels pass through a point — and all four other postulates hold. The fifth cannot be derived, because a world exists in which the first four are true and it is false.
A proposition $P$ is independent of an axiom set $A$ when some model of $A$ makes $P$ true and another makes it false. A "model" is a concrete interpretation under which every axiom becomes a true statement — the Poincaré disk simply reinterprets "point" and "line." Sharper still is relative consistency: since the model is built from the points and circles of the Euclidean plane, any contradiction in hyperbolic geometry could be translated word for word back into Euclidean geometry. Consistency can never be established from nothing; it can only be transferred.
This is the birth of mathematics' meta-level: before, axioms were truths about space; after, they became a set of rules admitting many models. For the first time, "cannot be proved" turned from a failure into a property that can itself be proved. The path runs straight to Hilbert's programme, Gödel's incompleteness theorems (Day 12), and Cohen's forcing (Day 11) — while this toy world became, sixty years later, the language of general relativity (Day 10).
Hyperbolic volume grows exponentially with radius, matching the branching growth of trees, so hyperbolic embeddings preserve hierarchical structure in far fewer dimensions — knowledge-graph representation learning relies on it. On the engineering side, formal verification asks "is this property entailed by the spec?", and an SMT solver answers by finding a counter-model — logically identical to building the disk two millennia later. When no proof is forthcoming, construct a world in which the claim fails.
Pluck a guitar string and listen for a second: you can name the pitch. Listen for a millisecond and you hear only a click. The shorter the sound, the wider its spectrum. This is not an instrument's shortcoming — frequency is defined as repetitions per second, and if you deny it time to repeat, it has no definite frequency.
Squeeze a signal by a factor $a$ in time and its Fourier transform necessarily spreads by $a$ in frequency, and the product of the two widths has a floor that nothing can break. Heisenberg's principle is the physical instance: position and momentum wavefunctions are exactly a Fourier pair. The limits of physics are the shadow of a mathematical structure.
$f$ is the signal and $\hat f$ its spectrum; the two integrals measure how spread out the energy is along the time axis and the frequency axis (an energy-weighted variance), and $\|f\|_2$ is the total energy. The inequality says the two variances cannot both be small, and the right-hand side is a constant independent of $f$ — the wall treats every signal alike. Equality holds only for Gaussians, which is exactly why the Gabor transform favours a Gaussian window.
Of the four impossibilities here, this one is the most structural: it comes not from logic, not from a tool, but from scaling symmetry itself — the spectrum of $f(ax)$ is $\hat f(\xi/a)/|a|$, so compression and dilation are tied to opposite ends of one transform, and the bound merely quantifies that tie. Uncertainty is thereby stripped of its physical mystique and reduced to an inequality about functions: an unlocatable particle and an unpitchable click are the same theorem appearing twice.
The entire design tension of short-time Fourier transforms and wavelets (Day 40) lives here: a short window localizes time but smears frequency, and vice versa — the "metallic" artifacts of audio pitch-correction software are a side effect of that choice. Radar range resolution and Doppler resolution are constrained by the same ambiguity function and cannot both be optimal. One common misconception: what compressed sensing evades with a sparsity prior is the Nyquist rate, not this inequality — no prior can evade this one.