chronochromatic.org
Who got here first

Inspirations

Maths is discovery, not invention. Nearly every piece of this site was found before — often decades before, sometimes by engineers solving an unrelated problem. This page names them. Arriving somewhere already furnished does not make the walk less real.

The digits

Red −1, green 0, blue +1 in base two — a redundant representation, where the same number has many spellings.

digits −1, 0, +1 in base 2 Signed-digit representation · Balanced ternary is its base-3 cousin
Booth, 1951
adding two grids in one pass, no carries Redundant binary representation — the redundancy is what buys carry-free addition
Avizienis, 1961. The 20,000-pair test was rediscovering his theorem.
push — spreading colour as far as it goes Booth’s multiplication algorithm recodes runs the same way
NAF — the fewest lit cells Non-adjacent form · used to make elliptic-curve multiplication a third faster

The stalks

Blue and red edges read as a number, and the +/− address that names it.

blue–red stalks with values Hackenbush · Surreal numbers
Conway, On Numbers and Games, 1976; Berlekamp–Conway–Guy, Winning Ways, 1982
the + − + − address The sign expansion — a sign expansion of a surreal number is exactly a Hackenbush chain. Combinatorial game theory
pinning 2⁰ so everything is a fraction Dyadic rationals — exactly the numbers with finite birthdays
The standard Hackenbush recipe already puts the binary point at the first colour change.

The anatomy of 0

Where the digits come from — and why they cannot start any earlier.

0 is not a point. Between −0 and +0 sits an entire structure of games, and the numbers on the rest of this site only begin once you are out of it. The order, Red to Blue, with every step ascending:

−0 < tiny-on < tiny < up < over < +ε < [ every real ] < ±ω < [ every real ] < −ε < under < down < miny < miny-on < +0

It is not linearly ordered, and that is the point. The game ∗ = {0 | 0} is not less than 0, not greater than 0, and not equal to it — it is confused with it. Trichotomy fails, so there is no side to be on. Left and right are total-order words and cannot be used in here. Red and Blue can, because they name whose advantage rather than which direction — which is also the Winning Ways convention: bLue is Left, Red is Right. The vocabulary did not happen to fit. It is the only vocabulary that can.

The rim is made of the centre. tiny-on is not tiny stacked on tiny. on = {on |} is a game, pronounced omega, and it is above everything; tiny-on is tiny parameterised by it — the smallest thing you can build, made out of the biggest thing there is. So the structure closes on itself. Stretch the pinhole at 0 open and look down, and on is what is at the bottom, because it was the material all along. That is inversion in a circle: the map that swaps a disc’s centre with its boundary. Every positive infinitesimal is greater than 1/on, so the floor of the whole zoo is the reciprocal of what sits at its middle.

Zero is doubled and ω is unified, and that asymmetry is a choice. Floating point doubles both; the Riemann sphere unifies both. Doubling only zero is what makes this a disc with a centre rather than a sphere, and what turns the chain above into a closed loop rather than a line.

The bits cannot start until you are out. A sign expansion is a number’s address, and ∗, ↑ and tiny are not numbers — they have no position for a bit string to point at. Everything else on this site begins exactly where this zoo ends, and that is forced by the theory rather than chosen. Meanwhile everything ordinarily called a number — the counting numbers, the reals, the irrationals, all of analysis — is the single greyed band above, between ε and ω. One step of thirteen. All the structure is at the two ends.

a game confused with 0 — neither side of it Star · Nimbers, its family
Which is why the ordering is a partial order and trichotomy fails.
up, down, and the infinitesimals below every number Up · Infinitesimals
↑ is positive and smaller than every positive number — below 1/ω, below 1/ω², below all of them.
the order of the low side, corroborated Owen Maitzen, Hackenbush: a window to a new world of math — gives the descending chain as the reals, ε, the powers of ε, over, up, tiny-1, tiny-2, and calls tiny-on “the tiniest number there is”
Which is exactly the left half of the chain above, from a second source. It also settles a direction by etymology rather than proof: over is named for sitting over up — “because it’s bigger than up, we’ll call it over” — and names under as its negative. An expository video is corroboration, not a citation; the strict form over > n·↑ is still open.
tiny, miny, over, under, on Berlekamp–Conway–Guy, Winning Ways vol. 1 p. 126 for tiny and miny; Conway, On Numbers and Games ch. 11 for the loopy games over, under and on
The book writes tiny-x as +x and miny-x as −x — the two halves are called plus and minus at the source. miny is exactly −tiny, so the halves of the disc are one negation apart.
up sitting just above the tinies It is one: +0 = {0 | {0 | 0}} = {0 | ∗} = ↑, Winning Ways vol. 1 p. 126
A tiny shrinks as its parameter grows, faster than any multiple can climb back. So up, with parameter 0, is the largest tiny, and tiny-on — parameter on, the largest there is — is the smallest. The chain is a single family, not a stack of species.
Red and Blue instead of left and right Combinatorial game theory names the players bLue and Red
The colour convention here is theirs, arrived at for the same reason: a partial order has no directions, only players.
−0 and +0 as two places, outside the disc Signed zero · IEEE 754 has both, and 1/+0 = +∞ while 1/−0 = −∞
The reciprocal sends the two outer rings to the centre. Floating point got there for engineering reasons.
one ω in the middle, the line closed into a loop Real projective line · Riemann sphere · ω and the ordinals
looking through the pinhole to see the centre Inversion in a circle · a conformal map exchanging the origin with the boundary

The atlas

Rings of integers, ordered by value — which turned out to be an FFT, a machine part, and a hyperbolic plane.

bits reversed over 2L Van der Corput sequence, 1935 · the base of every low-discrepancy sequence
why counting scrambles the arc Bit-reversal permutation — the input reordering of the Cooley–Tukey FFT, 1965
ring r is 2r+1 even directions Roots of unity — the FFT’s twiddle factors, in polar form
concentric rings, each doubling its sectors An absolute rotary encoder disc · often in Gray code so only one track changes at a time
Engineers machined this geometry into shafts because it is the efficient way to read an angle.
the outermost tile of every ring is all ones Those are the Mersenne numbers, 2n−1 — and all-ones in a base is a repunit, so ring r’s edge is the binary repunit of length r+1
Forced, not chosen: ring r spans [2r, 2r+1−1], so its last tile is 2r+1−1. Which means the edge tile’s value is 1 − 2−L — the horizon tile, one place value short of closing the ring.
so some rings have a prime on their edge Ring r’s edge is prime exactly when r+1 is a Mersenne exponent. In the default eight-ring Atlas that is rings 1, 2, 4 and 6 — 3, 7, 31, 127. Next are r = 12, 16, 18, 30, 60, 88, 106, 126.
A free filter comes with it: 2n−1 can only be prime if n is, so rings 3, 5, 7, 8 and 9 are ruled out without factoring anything. 52 Mersenne primes are known; the largest is 2136279841−1, 41,024,320 digits, found by GIMPS in October 2024. Each is confirmed by the Lucas–Lehmer test, which exists because the numbers have this shape. See largest known prime.
big in the middle, small at the rim, a boundary never reached The Poincaré disc model of hyperbolic geometry · the rim is an ideal point at infinite distance
Linear radius against a population that doubles each ring is the signature: circumference growing like e^(r ln 2) puts the curvature at −(ln 2)² ≈ −0.48. That is why it is a horizon, not a wall.

Shifting

Move the point, or drop a digit — two different operations, both classical.

dropping a digit; orbits and periods The doubling map · Symbolic dynamics
bar notation reaching thirds Repeating expansions · block length divides 2m−1 by the multiplicative order of 2
a binary tree of every rational Stern–Brocot tree

Multiplication and division

One is a rectangle you read; the other is the same rectangle, searched for.

every cell of A against every cell of B, no carries The grid method — the box taught in primary school · Outer product · Multiplication algorithms
It works because a product of sums expands. Division has no such expansion, which is the whole difference.
squashing the rectangle: sum the anti-diagonals, then carry Lattice multiplication — gelosia, the method Napier’s bones mechanise
The rectangle is the lattice; the squash is its diagonal sum. Which is also where the carries the grid deferred come due — about two grids in three owe one.
rows of B, shifted and subtracted, one yes-or-no each Long division — the tableau is the multiplication rectangle sheared, so its anti-diagonals become columns you can subtract one at a time
stopping early and keeping the odd chunk out Euclidean division — quotient and remainder, exact at every width, instead of an expansion that never ends
The oldest idea on this page, and still the one doing the most work.
a stalk in (−1,1) with a multiplier on the boundary Floating point · the significand normalised into a band, the exponent naming which band
Arrived at from the other direction, which suggests the layout is forced. The difference: floating point rounds inside the band and forgets. This keeps the remainder.

The cascade

The anti-diagonal sums this site has drawn since its first commit turn out to be one step of a cascade that generates Pascal’s triangle. We had been printing row two the whole time without knowing what it was.

summing a grid’s anti-diagonals That is convolution — a Cauchy product of the two digit strings
Measured, not assumed: squashDiagonals(hexProduct(A,B)) equals the convolution of A and B on 81 of 81 shapes. Weight depending only on r+c is what makes it one.
arcs(n) — where the commas fall The two-fold convolution of all-ones, which is Pascal’s second diagonal
arcs(n) equals the squash of an n×n all-ones grid, 12/12. arcs(8) = 1 2 3 4 5 6 7 8 7 6 5 4 3 2 1 is the 255 × 255 pyramid in the spec. Convolving all-ones k times gives C(i+k−1, k−1) exactly — checked for k = 2 to 6.
dividing by a repunit, over and over Division by B−1 in base B is the suffix-sum operator — a prefix sum from the other end — and iterating it on all-ones builds the same triangle
The same cascade as the row above, run backwards. We multiply by repunits and convolve; dividing by them sums. Casting out the remainder is casting out nines in base 2p.
a ring that comes back to itself under doubling Multiplication by 2m modulo 2p−1 is a circular shift of the p-bit register
Checked on 25,644 of 25,644 cases. The algebraic form of the rotary encoder disc two sections up — and D − x is the bitwise complement, 2,344/2,344, which is the colour flip for 0/1 digits.
all-ones rows behave unlike every other row Lucas’s theorem — every binomial in row 2k−1 is odd, because the row index is all ones in binary
Those rows are the repunits, which are the Mersenne numbers, which are the horizon tiles of the Atlas rings. The same integers keep turning out to be the exceptional ones.
signed digits, chosen so nothing carries Kummer’s theorem says the power of 2 dividing a binomial is the number of carries when you add in base 2. So carries are not noise in this cascade — they are what decides its structure.
Which makes our digit set a trade rather than a free win. One step through the grid is one convolution, but the result leaves {−1, 0, +1} and cannot be fed back in as digits; reconciling first destroys the counts. We killed the carries and lost the iteration. Held open: what the cascade looks like in signed digits, where the borrows Kummer counts have nowhere to live.
where this section came from Vladimir Cepeda, The Straggler Sets of Greedy Mersenne Decomposition are Pascal’s Triangle in Base 2, draft note, 30 August 2026
His result, not ours: he found the triangle inside repeated division by 2p−1, and reading it sent us looking for the same cascade here. An unrefereed draft — the suffix-sum slope is derived, the bridge from his model to the true chain is measured, and he labels which is which.

The wub ±

Phasors summed tip to tail, riding surfaces, tying knots.

phasors summed tip to tail tracing a closed curve Epicycles · Fourier series · Phasors
Ptolemy had the picture; Fourier had the theorem.
(p,q) from two rates Torus knots on a torus · Knot theory
counting crossings in a projection Crossing number — the minimum over all diagrams, which is why ours is only an upper bound

The fold

Anti-diagonals, mirror symmetry, inversion in a circle, and the basin boundary they turn out to be.

constant along anti-diagonals Hankel matrices
the fold-mirror J Aᵀ J Persymmetric matrices — symmetric about the anti-diagonal
fixes the Fold, swaps Inner and Outer, undoes itself Inversion in a circle — those three properties are its definition
laying a digit string along the anti-diagonals and cutting it into Inner / Fold / Outer, so the three sum back to the number exactly Placed, in 2026, as a Julia set and its two Fatou basins. Give a cell the coordinate ρ = 2^((r+c)−(n−1)) and Inner is |ρ|<1, Outer |ρ|>1, and the Fold the unit circle — which is the Julia set of z ↦ z², whose two basins are exactly those two hemispheres, with the anti-transpose as the inversion ρ ↦ 1/ρ between them. 89,439 cells checked, no exceptions. Cayley asked the matching question in 1879 — where Newton’s method lands from a guess — and for two roots the answer is this picture. For three he could state it and not solve it.
The geometry, not a new object: z ↦ z² is the oldest example there is, and the doubling map credited under Shifting is what it does on the circle — named here for years without the circle ever being named as its invariant set. Still held loosely on one point: the coordinate is an exact place value on the product rectangle and a magnitude ordering on a single folded stalk.
thirteen attempts to reach prior art from inside this geometry, and what each of them found instead A separate series read these pages as coding theory and built what each reading implied. Every round landed on something already named — the square on product codes, place value on residue arithmetic codes, the Atlas ordering on interleavers, the Wub on Reed–Solomon, the file-scale work on PAQ. None of it is claimed here.
The negative result is the useful half: across all thirteen, the fold's own partition was never the mechanism. Inner and Outer appear nowhere in that work as regions, and nor does the anti-transpose that swaps them. Every derivation walked past the one construction with no name. That is not proof of anything — only a record.
the fourteenth attempt, from the inside: one residue check per region, the fold's own I + F + O = V doing the accounting It works, and it has a name: an interleaved residue (AN) code, with the fold as the interleaving pattern.
And as a pattern the fold is legitimate and sub-optimal. Its 496 / 32 / 496 split separates two random errors 53% of the time; a plain three-way split manages 67%. Corrected in 2026: that comparison used the wrong family — against a fair two-way split the fold wins at every width — and the statistic itself moves no error channel at all. What the geometry costs is burst spread.
the fifteenth and sixteenth, from further inside: the anti-transpose as the code, three ways, and the green as the code, two ways The anti-transpose as one extra residue, V − σV, gives each hemisphere a second equation and names any same-region pair with no search: a two-syndrome arithmetic (AN) code, the residue cousin of a two-locator BCH code, blind to the Fold it turns on. As a mirror it is a repetition code that reads a burst off the partner at twice the size; as an interleaver it is a no-op. The green's forced slack is the 2-adic valuation of the value — mean one, and no random square carries its own checks — and a green cannot be a three-valued erasure on a run, because the signed-digit alphabet spells adjacent cells the same integer two ways: the identity push is built on.
The Wub's two-valued coin is not an approximation of the erasure model; it is the only one that decodes. Three constructions with names, one theorem about the greens, and then a name for the fold itself — see the partition entry at the top of this section.
the seventeenth, and the first to leave the square: the fold's anti-transpose applied to a whole file rather than to one grid A contiguous wound in the stored file becomes a scattered set of source addresses, so a block sees thin erasures instead of total loss — an interleaver again, but one that is its own inverse and costs not a single byte. It multiplied the survivable damage by up to 974×.
It buys spreading, not equations: surviving a 4 KB wound needs one equation per damaged byte per region, which is Reed–Solomon, a room this project has already walked into from another door. The same round found that the free variable in four rounds of work was the one nobody had touched: how much the cell holds.
the eighteenth: if burst spread is what the geometry costs, then which three-class partition of the grid actually minimises a burst? Asked properly — the worst cells landing in one class over every placement of a length-L run, along a row, a column, an anti-diagonal, and the row-major tape that wraps at row boundaries — it has a floor of ⌈L/3⌉ by the pigeonhole principle, and the partitions that reach it are block interleavers: class(j) = g(j mod L) for a balanced g. Named again, and not by this project.
The useful half is the obstruction. When 3 divides L a window at the floor has no slack, which forces the tape to be L-periodic, and then a coset count decides everything: for a run that fits across the grid, a partition at the floor exists exactly when 3 divides both L/gcd(n, L) and L/gcd(n−1, L). So it is arithmetic in the grid’s width that rules most widths out, not the shape of the partition — while the fold’s own (r+c) mod 3 is ruled out at every width for a simpler reason: an anti-diagonal is its level set, so a burst along one lands entirely in a single class. Nothing here is a claim about the instrument; it is what the instrument’s geometry looks like once the right question is asked of it.