Symbolic systems evolve to preserve what matters — and break symmetry only where meaning requires it.
Your visual system treats a rotated object as the same object. It throws away orientation almost for free, because most of the time orientation doesn't change what a thing is. Invariance is the default; the work is deciding where it should stop.
Turn both glyphs by the same angle. X has a 180° rotational symmetry, so it lands back on itself — invariant. b does not: at 180° it becomes q, a different letter with a different meaning. The boundary isn't in the glyph — it's learned, and meaning decides where it sits.
Groups act on symbols — rotate, reflect, mirror. On the complex plane the rotation states live on the unit circle: z → eiθz. Drag θ and watch the point sweep its orbit; the four cardinal states are exactly 1, i, −1, −i. Crank n and De Moivre's theorem turns one turn into n.
D₄ — four rotations and four reflections — is the symmetry group of the square, and of many glyphs. Apply an element to the symbol x; the set of everything you can reach is its orbit, 𝒪(x) = { g·x : g ∈ G }. A high-symmetry glyph barely moves; a low-symmetry one (try F) shows all eight.
D₄ is finite. Loosen what counts as an allowable transformation and the group grows: rigid motions ⊂ similarities ⊂ affine ⊂ projective. Felix Klein's Erlangen program says a geometry simply is the study of what stays invariant under its group. Climb the ladder — each rung adds freedom and sheds an invariant. The survivors get deeper.
D₄ has eight elements. The orbit on the circle, eiθ, is the continuous group SO(2) — every planar rotation on one smooth dial, and it commutes. Step into 3D and you get SO(3): three degrees of freedom, a whole sphere of orientations — and it does not commute. The order of rotation becomes part of the answer.
The same shape, moved through the same group, lands on different letters. Literacy is largely the work of un-learning rotational invariance exactly where it would cost you meaning.
One bowl-and-stem, four placements: b↔d is a mirror, b↔p is a flip, b↔q is a 180° turn. Same object; different orientation; different meaning.
Engineered to sit on the identity boundary: designed so a 180° turn reads the same. Deliberate symmetry where convention usually forbids it.
Four orientation states — 0°, 90°, 180°, 270° = 1, i, −1, −i. A single rotating mark is enough to see phase, rotation, and state transitions at a glance.
An invariant is whatever stays fixed across every allowable transformation. These are what a symbol can be recognized by — and what a good representation should store instead of raw pixels.
Back to D₄ from §04, now with its deepest accounting identity. Act on a glyph with all eight elements: the ones that leave it unchanged form its stabilizer (a subgroup); the distinct results it can reach form its orbit. The orbit–stabilizer theorem says their sizes always multiply to |G| = 8 — the symmetry you keep and the symmetry that moves you are exact complements. The eight tiles below split into orbit colour-groups, each of stabilizer size.
Put a value on each of the square's four corners — that is a function on the cyclic group C₄, the rotations of §03–§04. Representation theory says it decomposes uniquely into four irreducible pieces, one per character: under a 90° turn r, each piece is simply multiplied by 1, i, −1, or −i — the very numbers this whole site is coloured by. The decomposition is the discrete Fourier transform; the piece-magnitudes are exactly the rotation invariants of the pattern.
The §05 ladder — rigid, similarity, affine, projective — is nothing next to the full freedom of topology, where every continuous stretch and bend is allowed. Almost no quantity survives that. But V − E + F does. Refine this mesh however you like — subdivide edges, add chords, drop in interior points — and the Euler characteristic χ = V − E + F = 2 never budges. It is the deepest invariant on the site: it doesn't even need straight lines.
The deepest answer to why anything is conserved at all. Emmy Noether proved that each continuous symmetry of a system's dynamics forces something to stay constant: rotation-invariance ⇒ angular momentum, time-invariance ⇒ energy, shift-invariance ⇒ momentum. A planet orbits a star below — pick a symmetry and watch its invariant hold, or break exactly where the symmetry is absent.
Primary visual cortex (V1) simple cells behave like Gabor filters — a grating under a Gaussian window — sharply tuned to orientation but sensitive to phase (where the light/dark bands fall). Pool two of them in quadrature and take their energy and you get a complex cell: same orientation tuning, but invariant to phase. Drag the phase below — the simple cell flips through zero; the complex cell barely moves. Invariance, built from biology. Modern equivariant CNNs rediscover the very same trick.
A vowel is fixed by its first two formants F1, F2 (resonances of the vocal tract). A child's tract is short and a grown man's is long, so their formants sit at wildly different absolute frequencies — yet we hear the same vowel. The phoneme is invariant under speaker scaling because identity lives in the pattern of ratios, not the raw hertz. Slide the vocal-tract length: in raw Hz the whole vowel cloud slides and spreads; switch to normalized and every speaker collapses onto one template. (The same scaling move as the similarity group in §05.)
Perception seeks invariance; learning sets the boundaries — where orientation starts to matter.
Type design exploits symmetry and transformation: few shapes, reused, many meanings.
Groups, representations, and orbits formalize it — invariants are the fixed points.
Symmetries govern conserved properties. (Noether: every continuous symmetry yields a conservation law.)
Equivariant models preserve structure across transforms; latent spaces encode orbits, not pixels.
Every section so far read a glyph. But a game is the purest case of the whole idea: its rules are the allowable transformations, and the game is whatever they leave invariant. In these three, the invariant is not a metaphor — it literally decides what's reachable, what's solvable, and who wins. The full gaming reading →
Every legal slide swaps the blank with a neighbour — one transposition, and one step of the blank. Each flips a parity; together they leave permutation parity ⊕ blank parity invariant. Scramble with legal moves all day: it stays solvable. Swap two tiles by hand and you flip the invariant — into the unreachable half the puzzle can never enter.
Each press flips a cell and its neighbours — addition mod 2. The boards you can clear are exactly the column space of the 25×25 toggle matrix over GF(2), which has rank 23: only one board in four is solvable. Press mode plays the game (always stays solvable); switch to Paint to flip a single light and build a board by hand — three times out of four you'll land outside the subspace, and a hidden invariant (a “quiet pattern”) refuses it.
Take any number from one heap; last to take a token wins. The whole game collapses to a single invariant — the nim-sum, the bitwise XOR of the heaps. Nim-sum zero is a lost position for whoever must move; the Sprague–Grundy theorem says every impartial game hides exactly such a value. The computer plays it perfectly: if you ever leave a nonzero nim-sum, it drives it back to zero.