Every Law on the Rack

The final pile and the total number of topplings do not depend on the order the grains fall in. That commutativity turns the recurrent states into an abelian group — whose identity element is itself a striking fractal.

A market maker moves the price by excess demand. Fundamentalists buy what is cheap against the fundamental F; chartists buy the trend they read off the chart, r_t, with bounded positions. The shares n follow realised profit through the logit. At F the reflexive gain is μ·g·n_c; past one the fundamental cannot hold.

What a light wave in a medium at rest predicts: an interferometer moving through the aether at v sees its fringes slide by 2L(v/c)²/λ on turning a quarter — 0.4 of a fringe for Michelson and Morley’s 11 m. They found far less than a twentieth of it.

Anything above the limit folds back into (−f_s/2, f_s/2]. Divide by N and you get the spin you actually perceive — this is the line the instrument computes each frame.

Scale every length by L and a bone’s cross-section grows as L² while the weight it carries grows as L³. The stress — load per area — rises in proportion to size: a photocopied giant is a broken giant.

To hold the stress constant the diameter must grow faster than the length — the femur of Galileo’s giant is stouter at every size. It works, and the bill is the skeleton: its share of the body grows as M^1/3.

A warm body loses heat through its skin, so it should burn in proportion to its surface. The exponent Kleiber displaced — and that later re-analyses keep bringing back.

Basal metabolic rate grows slower than mass and faster than surface: a whale is a hundred million times a mouse’s mass and burns a million times its power. B₀ is Kleiber’s ≈ 70 kcal a day per kg^3/4.

Cardiac output follows metabolism (M^3/4); the stroke volume is a fixed share of blood volume (M). Their ratio is the heart rate: a mouse near 600 beats a minute, a whale near ten.

Power to fly grows as weight times speed, W^7/6; flight muscle, a fixed share of the body, delivers ∝ W^e with e ≈ 1. The ratio creeps up as W^1/6 and crosses one somewhere among the heaviest flapping birds.

Grow so that every size sits the same fraction below its buckling height and the trunk must thicken as h^3/2 — between the geometric twin (d ∝ h) that buckles and the stress rule (d ∝ h²) that never needs to.

Larger mammals stand straighter, raising the effective mechanical advantage of their limbs. With near-isometric bones (d ∝ M^0.36) that cancels almost all of the square–cube rise — until the legs are straight and EMA can grow no more.

Geometric similarity for flyers: weight is a volume, wing area a surface. Tennekes’s diagram runs one such line (N/m², W in N, approximate) from insects to airliners.

Branch n ways at every level; radii shrink by β, lengths by γ, down to a capillary that is the same in every mammal. Area-preserving (a = ½) and space-filling (b = ⅓) give 1/(2a+b) = 3/4: the number of capillaries — the power — grows as blood volume to the 3/4.

The Lemuel grammar: a quantity as a power of body size. On log–log axes every allometric law is a straight line whose slope is the exponent b — fitted, on every instrument of the edition, from the sizes you visit.

The lift equation with wing loading ∝ W^1/3: cruise speed grows as the sixth root of weight — a jumbo about a hundred times a fruit fly’s speed for 10¹¹ times its weight.

A wheel resting on a turning disk at distance y from its centre is dragged round at y times the disk’s speed, so its shaft turns by the running integral of y against the disk’s rotation x. The radius is the integrand, the disk is the variable, the wheel is the answer.

Feed the second integrator’s output back, reversed, into the first one’s lead screw, and the shafts can only turn in the ratio the equation allows: y and v trace a circle. The machine does not compute the solution — the solution is the only motion it has.

At equilibrium every route anyone uses costs the same, and no unused route is cheaper — so no driver can gain by switching alone. On Braess’s diamond, 4 000 drivers split evenly at 20 + 45 = 65 minutes; open a free connector and the route through it (40 + 0 + 40 at full use) undercuts both old roads for each driver, until all 4 000 take it and every commute is 80. Both states are stable; the second is worse for everyone.

The ratio of the total travel time at the selfish equilibrium to the least total any routing could achieve. With latencies linear in the flow it never exceeds 4/3, on any network. On the diamond it is 1 while the connector is used by everyone or by no one, rises with demand, and touches 4/3 exactly at 4 500 drivers (90 against 67.5 minutes); at 4 000 it is 80/64.69 = 1.237. A toll of 22.5 minutes on the connector alone — the extra delay a connector driver imposes on everyone else, beyond what an old-road driver does — moves the equilibrium onto the optimum (Pigou’s fix).

τ too small and the coordinates are near-identical — the cloud collapses onto the diagonal. τ too large and the fold tangles. The usable window is near the autocorrelation’s first zero (or the first minimum of mutual information): decorrelated, not yet folded.

How to get ∇L cheaply: the chain rule, run backwards. The error at the output is propagated layer by layer, and each weight’s gradient is its incoming activation times the error flowing back through it.

Throw n balls into n bins uniformly at random and the fullest bin holds about this many — about 9 for a million bins. The asymptote is slow: at a million it says 5.3, and Γ⁻¹ is what the bins actually do.

Let each ball look at d bins and join the least loaded and the fullest falls from logarithmic to doubly logarithmic: 4 for a million bins with d = 2. Almost all the gain is in the second look; the third only divides by ln 3 instead of ln 2.

Jobs arrive at rate λ per server, each joins the shortest of d sampled queues, service takes one unit on average. The fraction of queues with at least i jobs is s_i: geometric λⁱ for one choice, doubly exponential for two. At λ = 0.9 the mean time in system falls from 10 to about 2.61.

Keep a Bayesian belief (a Beta posterior) about each arm’s win rate, draw one sample from each, and play the arm whose sample is highest. Uncertainty makes wide posteriors likely to win a draw — exploration for free, from probability matching alone.

What you lose by not knowing the best arm from the start — the gap between always playing the true best (mean μ*) and what your policy actually earned. A good policy makes this grow like ln T, not T.

Optimism in the face of uncertainty: judge each arm by the top of its confidence interval, not its mean. A rarely-pulled arm gets a big bonus, so it is tried; a well-known poor arm is dropped. No randomness, and provably logarithmic regret.

ℓ is the share of deposits withdrawn early; the bank pays them from its liquid share m and then by selling loans worth θ on the unit. Under common knowledge every θ with 0 < ψ(θ) < 1 has two equilibria — nobody runs and it survives, everybody runs and it fails.

Why it’s optimal: for Gaussians, posterior ∝ prior × likelihood is again Gaussian, and its mean is precisely the Kalman update. The filter is Bayes’ rule, closed-form, once per tick.

The femur as a cortical tube (inner/outer ratio ρ) loaded axially and bent by an eccentric load arm εd. The load is the muscle force that balances the ground reaction kMg through the limb’s mechanical advantage — so posture enters the stress directly.

Everyone names a number; the winner is closest to p times the average. For p < 1 no number above 100p can win, so none above 100p² can, and so on down: the only equilibrium is that everyone says zero.

Start from a naive 50 and best-respond k times: 33, 22, 15 … for p = ⅔. A guess x implies a depth k = ln(x/50) / ln p. Nagel’s first rounds clustered near the first two steps.

Below T_c the thermal density saturates, so every further-cooled atom is forced into the ground state. That exact curve is what the inset dot rides.

In a harmonic trap the critical point follows the atom count and the trap stiffness — more atoms condense warmer. Slide N and the inset’s T_c line moves.

Condensation isn’t set by temperature alone but by phase-space density: when the de Broglie wave is as big as the spacing between atoms, nλ³ hits ζ(3/2).

The payoff: thousands of atoms share a single wavefunction — one amplitude, one phase. A quantum state grown to a size you can photograph.

Bell’s bound in testable form: four correlators at four angles. Every local-hidden-variable theory obeys S ≤ 2 — so one measured S above 2 rules them all out at once.

Quantum mechanics breaks the classical wall — but not infinitely. The most it reaches is 2√2, hit exactly at the optimal angles. Strange, yet still lawful.

Two particles, one state — neither has a spin of its own, yet along any shared axis they always come out opposite. The whole is defined exactly where the parts are not.

The average product of the two ±1 outcomes, fixed only by the angle between the detectors. It bows past every straight line a local model can draw — and that gap is the whole effect.

The value of a state is the best action’s immediate reward plus the discounted value of wherever it lands you. Optimal play is exactly the solution of this self-referential equation — the present worth of the whole future, folded into one line.

The one rule. Nudge the estimate toward the reward you just got plus the best value you now see ahead; the bracket is the TD error — the surprise. Learn the future from a better guess of the future, with no model of the world at all. This single line drives The Descent’s deep-RL descendants.

How good it is to take action a in state s and then play greedily forever after — the expected discounted reward. Learn this table and the policy is free: in every state, pick the action with the highest Q.

Birkhoff throws the flight between walls away: a billiard is fully told by two numbers per bounce — where it hit the wall and how steeply. The map preserves area, so what it draws in the (s, p) plane is an honest portrait of the dynamics: curves mean order, spray means chaos.

The one law on the table: the angle in is the angle out. Hero of Alexandria derived it around 60 AD from a least-path principle — light takes the shortest route via the wall — making this the oldest variational argument in physics. Everything else is the shape.

In an ellipse the product of a chord’s angular momenta about the two foci survives every bounce. A trajectory can therefore never cross the confocal conic its constant pins — it wraps around it forever, and that untouchable curve is the glowing caustic the chords weave.

The dark disc is bigger than the horizon itself: any ray aimed inside the critical impact parameter b_crit is swallowed. Lensing magnifies the r_s horizon into a √27 ≈ 5.2 r_s silhouette — the ring the Event Horizon Telescope photographed in 2019.

The attacker’s progress over the z confirmations taken as Poisson at the average pace, then the gambler’s ruin from wherever he is. 0.0009137 at q = 0.1, z = 5 — The Chain’s engine rebuilds the whole printed table. Run as an exact Poisson race the count is negative-binomial, and Rosenfeld’s form gives 0.00178 there.

A random walk that steps toward the attacker with chance q and away with chance p ever closes a gap of z with probability (q/p)^z — the gambler’s ruin. Below half the hash power the odds fall exponentially with every block; at half or above the attacker always catches up.

A pool of share α, winning γ of the honest miners in a tie, earns more than its share by withholding blocks inside this range: above ⅓ when ties go against it, ¼ when they split, any size when they all go its way. The Chain simulates the state machine and lands on the revenue formula to the third decimal.

Three steering urges — match their heading, move toward them, but keep your distance — summed with weights. That is the whole brain; nothing in it knows the flock exists.

In thermal equilibrium a state of energy E is occupied with weight e^−E/kT. Temperature sets how readily the system pays energy to leave a comfortable valley — the physics annealing borrows and aims.

The cloud is the square of the wave. Squaring discards the phase — which is exactly the half this instrument paints back in as colour.

With Δ = 2 the cross-section is conserved at every fork, so the branches of a tree squeezed together would make one trunk. Murray’s Δ = 3, the rule for vessels pumping fluid, thins the crown faster.

The share using predictor h follows its recent profit U — the rational predictor pays a cost C, the naive one is free. When the cobweb is unstable under naive expectations, raising β past β₁ = (2/C)·artanh(b_d/b_s) turns the steady price into a two-cycle, then a cascade of doublings into chaos: the crowd keeps leaving the expensive predictor while prices are calm, and coming back when they are not.

Einstein 1905 tied this walk to Avogadro’s number — the jiggle of a pollen grain as proof that matter is atoms. Five years earlier Bachelier had modelled prices as the same walk: the seed of quantitative finance and every backtest.

A slender column carries its load straight until P_cr, then bows sideways all at once. Stiffness E·I grows as d⁴, so doubling the length quarters the load a column of the same girth can bear.

A column standing under its own weight w = ρgπd²/4 per metre buckles when w·h³/EI reaches 7.837. Tall and thin fails first: a trunk twice as thick may stand only 2^2/3 ≈ 1.6 times as tall.

Anchor on the demand you expect, then adjust toward the stock you want (S*) and the supply line you should have (SL*). β is the weight a manager gives what is already on order: Sterman’s players gave it about a third. At β = 0 The Whip’s step game is still swinging at week 104; at β = 1 it settles.

An order-up-to stage that forecasts by a p-period moving average and covers a lead time of L periods, under AR(1) demand of correlation ρ: equality when it keeps no safety stock. Long windows and short lead times damp the whip. In The Whip’s engine (L = order + shipping delay) the retailer measures 3.47 against 3.46 at L = 3, p = 4, ρ = 0.5.

With oral messages, agreement among the loyal survives f traitors exactly when they are fewer than a third of the camp. The margin is a count: The Generals’ worst-case search breaks OM at n = 3, 4, 5, 6 for f ≤ 2 and finds no break at n ≥ 3f + 1. With unforgeable seals the third disappears: SM(m) handles any number of traitors up to m.

Round r of OM(m) relays every value along every path of r distinct generals, so the courier count is a falling product per round: 3 + 6 = 9 for OM(1) at n = 4, 6 + 30 + 120 = 156 for OM(2) at n = 7, 3 609 for OM(3) at n = 10 — all counted by the engine. Signed SM(m) relays each order once per lieutenant instead.

The same eight bits sort into four behaviours — uniform, periodic, chaotic, complex. Rule 110 sits in the last and is Turing-complete: a universal computer hiding in three bits.

Each cell sees only itself and its two neighbours, and a single local function decides its next state. No cell knows the whole world — all behaviour is local.

Three binary inputs make eight possible neighbourhoods, and a rule just picks an output bit for each — so the entire space of laws is countable and small.

Carnot’s waterfall: caloric C, conserved, falls from the hot body to the cold one, and the motive power is the caloric times the height of its fall, as water turns a mill. False as physics — heat is not conserved — yet exact as bookkeeping, because the quantity his axioms define is Q/T.

A wetting liquid climbs a tube until the column’s weight matches the pull of its meniscus: halve the bore, double the rise (slope −1). Non-wetting mercury is pushed down instead.

The three surface tensions meeting at the contact line balance along the plate, fixing the angle the liquid meets it at — near 160° on a lotus leaf, near 10° on clean glass.

A curved liquid surface holds a pressure jump proportional to its surface tension and its total curvature. Add gravity’s hydrostatic pressure and integrate from the apex, and it is the whole shape of a drop, a meniscus or a soap film.

The size at which surface tension and gravity are equal. Below it a drop is a spherical cap; above it, a puddle whose height stops at 2ℓ_c sin(θ/2).

Gravity on a volume against tension on a surface — the square–cube race for liquids. The crossover the drop’s fit may not cross is Bo = 1.

The most any engine working between two temperatures can turn to work. Carnot showed in 1824 that it depends on the two temperatures alone; written on the absolute scale it is the second law’s first number, and it survived the caloric it was derived from.

One integer matrix on the torus: shear, shear, wrap. The determinant is 1, so no information is ever lost — every pixel of the picture survives every step. It only stops looking like a picture.

On an N×N pixel grid the map is a permutation, so the scrambled picture must come back exactly. Dyson & Falk bounded the wait at 3N steps — but the schedule is wild: 101 pixels return in 25 steps, 100 pixels take 150.

The map’s eigenvalues are set by the golden ratio: along one irrational direction every gap is multiplied by φ² ≈ 2.618 per step, along the other it shrinks by the same factor. That one number is the Lyapunov exponent — the shredding rate — of the whole system.

A chain of weight w per unit length, pulled horizontally by H, hangs in a hyperbolic cosine. Its slope grows with the weight of chain below each point — a parabola would need the weight spread along the ground instead.

Standardise the sum of any finite-variance step law and it converges to the same standard Gaussian. The bell is the universal attractor of sums — the reason it is everywhere.

The bits per use you can send with vanishing error. Noise eats capacity; at p = ½ the channel carries nothing at all — input and output share no information.

A measure-zero set of exact periodic orbits survives the chaos — Euler’s collinear line, Lagrange’s equilateral triangle (the origin of the L-points), and the figure-eight.

Send at T₁ by your clock, the server stamps T₂ and T₃, the reply lands at T₄. The midpoint estimate of the offset is wrong by at most half the round trip — whatever the split between the two directions — and Cristian’s version (T₂ = T₃) is the same bound.

The spiral tightens when demand answers price more strongly than supply does, widens when supply is the steeper, and repeats a fixed square when they match. Nothing outside the market is needed for it to cycle for ever.

Demand a − b_dp clears against supply c + b_sp^e planted for the price farmers expect. With naive expectations — last season’s price — each deviation from the equilibrium p* comes back multiplied by −b_s/b_d: a glut, a shortage, a glut.

Thinkers of level k believe everyone else thinks less, spread in proportion to the Poisson shares below k, and best-respond to that mix; level 0 picks at random. One parameter, τ, the mean number of steps. With τ = 1.5 and p = ⅔ the model crowd’s average is 33.5 and its winning number 22.3.

The draws needed until every coupon has appeared m times — N ln N + (m − 1)·N ln ln N + O(N). The second complete set costs only the ln ln term, which is why two collectors who swap duplicates each finish at about two thirds of the solo cost.

Completion times are not bell-shaped: centred at N ln N with width N, they follow the double exponential of extreme values, with a long right tail. A thousand albums finish over a spread of several N, and one is always out near twice the mean.

Drawing uniformly from N equally likely coupons, the k-th new one costs N/(N − k + 1) draws on average — one at the start, N at the end. The total is N times the harmonic number: the last coupon alone costs as much as the first N/2 together.

Two conservation laws — momentum and energy — pin the outcome of a head-on elastic collision completely: no fitting, no friction, no choice. Huygens settled the rule in 1669; every clack in this instrument is this one line, applied exactly.

Each clack turns the state around the circle by the same angle 2θ, so the total count is how many fit in a half-turn: π/θ. Set the mass ratio to 100^(d−1) and θ ≈ 10^(1−d) radians — and the count spells the first d digits of π.

Rescale each velocity by the square root of its mass and energy conservation draws a circle; momentum conservation draws a line. Every collision is then a reflection — the messy clatter of blocks becomes a polygon of equal chords inscribed in one circle.

Discount a light by scaling each cone signal by the ratio of the target white to the assumed light’s white. Assume a different light and the same pixels become a different paint — the whole of the dress, in three ratios.

The chromaticities of typical daylight, from warm late sun to blue north sky, with x_D a cubic in 1/T (4000–25 000 K). The visual system’s guesses about the light live along this curve; D65 sits on it at 6504 K.

With fewer measurements than unknowns, infinitely many images fit y. Ask for the one whose coefficients in a basis Ψ have the smallest ℓ₁ norm: the corners of the ℓ₁ ball pick out sparse solutions, and if the scene really is sparse in Ψ, it is the scene. The prior is the basis — name it.

A k-sparse signal of length N is recovered from on the order of k log(N/k) random measurements — the scaling, not the constant. The Donoho–Tanner phase diagram draws where the bet actually pays.

A gradient step on the misfit, then soft-thresholding: every coefficient shrinks toward zero by λ/L, and the small ones land on it exactly. FISTA adds a momentum term and converges as 1/k² rather than 1/k (the plain iteration is Daubechies, Defrise & De Mol, 2004).

A read/write head on an unbounded tape, steered by a finite table of rules — and that is the whole of what is computable. The universal machine reads another machine’s table off the tape and runs it: the first proof that one device can do anything any device can.

Will program p, run on input x, ever stop? Turing’s diagonal argument proves no single program can answer for all p — the first problem shown to have no algorithm. The busy beaver is its shadow: knowing S(n) would solve halting.

Square tiles with coloured edges that may sit side by side only when the touching edges match. Wang asked the innocent question — can a given set tile the plane? — and conjectured that any set that can, can do so periodically. He was wrong, and the wrongness runs deep.

Among all n-state machines that do halt, how long can the longest run? S(n) grows faster than any computable function — so no program can compute it. S(5) = 47,176,870 was settled only in 2024; S(6) already dwarfs every number in physics.

Berger proved no algorithm can decide it — by encoding a Turing machine into tiles, so “tiles forever” means “never halts”. The proof spits out the first aperiodic set: tiles that cover the plane but never repeat. Penrose shrank it to two; the hat, in 2023, to one.

On a white square turn right, on a black square turn left; either way, flip the square’s colour and step forward. That is the entire program — and from it comes ten thousand steps of apparent chaos, then, with nobody asking, an ordered “highway” that runs forever.

For many simple programs there is no formula for the answer that is faster than just running them. The ant proves it cleanly: nothing in the rule foretells the highway — you can only watch. Wolfram’s claim is that this is the rule, not the exception.

Every cell starts as a superposition of all tiles. Collapse the most-constrained cell (lowest Shannon entropy) to one tile, propagate the adjacency rules to its neighbours, and repeat — a wave of certainty spreading by constraint, borrowing its name loosely from quantum measurement.

Generalise the ant — more internal states, more colours — and you have a Turing machine whose tape is the plane. The Tape’s head crawling a line and the Ant crawling a grid are the same creature; some turmites are themselves universal computers.

Bayes’ rule on the host’s protocol, not on the empty door: if your door hides the prize she opens either of the other two (½); if the other closed door hides it she has exactly one legal door (1). That forced hand carries the closed door to 2/3. A host who opened at random and happened to show an empty door would leave ½ · ½ — a different game.

The swap argument needs “twice or half, equally likely” at every amount you could hold. That is a uniform weight on the doubling ladder, and its total is infinite: no probability distribution can give it. Under any real prior the odds of doubling move with the amount — and the 1.25× disappears.

A sender keeps W packets in flight and grows W by one packet each round trip until the network drops one, then halves it. Chiu & Jain proved it the one linear rule that converges to a share both efficient and fair; the sawtooth is what it looks like.

A long-lived AIMD flow that loses a fraction p of its packets settles to this throughput: the window’s sawtooth between W/2 and W carries 1/p packets per cycle. Longer round trips and more loss both cost, whatever the wire’s speed.

Emmy Noether’s theorem: every continuous symmetry of the laws of motion yields a conserved quantity — and the symmetry of time (the laws are the same today as tomorrow) is exactly energy conservation. A wheel cannot gain energy from nothing because its physics does not know what time it is; run the same release a century later and the curves overlay. This is why the perpetual-motion wall stands, not merely that it does.

Integrated by velocity-Verlet (leapfrog) — symplectic, so energy stays bounded over thousands of periods. The drift readout proves the orbits are the law’s, not the solver’s.

Every curve of constant width w — circle, Reuleaux polygon or anything stranger — has perimeter exactly π times its width. The perimeter is the integral of the support function, and only its constant term, w/2, survives a full turn.

Of all curves with the same constant width, the Reuleaux triangle encloses the least area; the circle, π w²/4, encloses the most. Same width, same perimeter — about an 11% spread in area.

h(θ) is the distance from the centre to the tangent line facing direction θ; the caliper width is h(θ) + h(θ + π). It is the same in every direction exactly when h is w/2 plus odd harmonics only. The radius of curvature h + h″ must stay non-negative, or the curve folds over itself and is no roller at all.

A Reuleaux triangle of width w can turn inside a square of side w touching all four sides at once. One full turn sweeps everything but four thin slivers at the corners — a hole that is square to within 1.23% of its area.

Every compass-and-straightedge step intersects lines and circles already drawn — at worst a quadratic — so each new point at most doubles the degree of the field it lives in. Every constructible number therefore sits at some finite floor 2ᵏ over ℚ; nothing built this way can ever reach degree infinity.

Temperature falls a little each sweep. Cool slowly (α → 1, anneal) and the search settles in the global minimum; cool fast (quench) and it freezes wherever it happens to be — a defect.

The share of the grid that cooperates — the population’s order parameter, exactly as magnetization is for The Threshold. On the spatial dilemma it settles near 0.318 across a band of b: neither all-C nor all-D, but a living mixture.

Carry a direction round a closed loop without ever turning it relative to the surface — parallel transport — and it comes home rotated by the total curvature inside the loop: zero on a plane, the enclosed area on a unit sphere, negative on a saddle.

A polygon of great-circle arcs on a sphere has angles that add up to more than a flat polygon’s; the excess, times R², is its area. The octant has three right angles: an excess of π/2, an eighth of the sphere.

A prisoner who opens their own number’s box and then follows each slip walks exactly the cycle of the shuffle that contains them, and finds their slip within fifty opens if that cycle is no longer than fifty. So all hundred live or die together, on one property of one permutation: no cycle longer than half. A uniform shuffle of n has a cycle of length k > n/2 with probability exactly 1/k — at most one such cycle can exist — so the team fails with probability H₁₀₀ − H₅₀ and survives 31.18 % of the time, against (½)¹⁰⁰ ≈ 8 × 10⁻³¹ for a hundred independent guesses.

Matter has a wavelength set by its momentum — the reason a particle can interfere at all. Heavier or faster ⇒ shorter λ ⇒ finer, harder-to-see fringes.

Cold atoms have longer matter waves. Cool far enough and each atom’s wave grows until it overlaps its neighbours — the gas stops behaving like particles.

Divide by the blur where the signal is stronger than the noise and give up where it is not; the signal-to-noise per frequency is the prior, here estimated from the data’s own spectrum. Beyond the cutoff H = 0 and the filter returns exactly nothing.

The maximum-likelihood iteration for photon counts: multiplicative, so it stays positive. Its misfit falls without end — past the noise, into fitting it — so it is stopped where χ²/N first reaches 1 (Morozov’s discrepancy principle). Positivity can extrapolate a little past the cutoff for point stars: that is the prior, not the sky.

The image is the sky convolved with the point-spread function — the squared Fourier transform of the pupil P, with the wavefront error W as its phase — plus photon noise. For Hubble in 1990, W was a spherical aberration from a mirror ground to the wrong shape.

The transfer function is the pupil’s autocorrelation, and a pupil of diameter D has none beyond D/λ: finer detail reaches the detector multiplied by exactly zero. Deconvolution can divide by a small number; it cannot divide by zero.

Relativistic electrons push back only as hard as gravity pulls, at every radius alike: above this mass a cold star has no size at which it can stand. With the full prefactor and two nucleons per electron, ≈ 1.44 solar masses.

Atomic matter keeps its density, so a ball of N proton masses grows as N^1/3; squeezed by its own gravity, its electrons become degenerate and it shrinks as N^−1/3. The two meet near the mass of Jupiter — no cold ball can be wider than about that.

The hottest a ball of N protons can get before degeneracy stops the squeeze is kT ≈ α_G²N^4/3m_ec²; hydrogen burns near α²m_pc². Lighter balls stay brown dwarfs. As an order of magnitude ≈ 0.3 M_☉, against a measured ≈ 0.08.

Stack a signal against delayed copies of itself to build an m-dimensional vector. As t runs, the point v(t) traces a curve — and that curve is a faithful copy of the hidden attractor you never measured directly.

Forget absolute density; what matters is the contrast — how much denser than average each patch is. The early universe was nearly smooth, δ ~ 10⁻⁵ everywhere. Gravity does the rest.

The same descent, pushed too hard, is the route to chaos. Raise η and the step overshoots the minimum, then overshoots back — a period-2 oscillation that doubles into divergence, exactly the bifurcation The Cascade draws.

The same diagonal, aimed at proof. Arithmetize the syntax so that a formula can talk about provability; the fixed-point lemma then builds a sentence G that asserts “I have no proof.” If the system proved G it would prove a falsehood — so, if consistent, G is true and unprovable. The flipped diagonal, one register up.

Cantor’s trick, aimed at machines. Suppose H decides halting for every program. Build D that asks H about itself and does the opposite. Then D(D) halts iff H says it doesn’t — H(D,D) = ¬H(D,D), a flat contradiction. No total H exists; every decider ships with its own nemesis.

When the cart turns, its outer wheel rolls further than its inner; the differential turns the figure back by the difference over the track W. On a flat floor that exactly cancels the cart’s turn, so the arm keeps its bearing.

Melt the data: drift each sample gently toward the origin while injecting Gaussian noise, step after step, until after enough time every trace of structure is gone and the cloud is indistinguishable standard noise. It is The Walk’s Brownian motion with a leash.

Run time backward and the same diffusion becomes generative: the score term steers the noise uphill in probability, condensing formless static back into samples. Anderson proved a reversed diffusion is itself a diffusion — the theorem the whole field stands on.

The gradient of the log-density — at every point, the direction toward where data is denser. A real model learns it by denoising (gradient descent — The Descent); here we use the exact score of a known target, so the reverse flow is honest with nothing trained.

Drop the noise term and a plain ODE carries the same changing density: one fixed noise seed maps to one sample along a smooth path — the streamlines you watch the particles ride here. Same marginals, no randomness.

Diffusion on the grid is a weighted blur of the eight neighbours minus the centre — the only spatial coupling in the whole system.

Numbers made by multiplicative, scale-invariant growth spread evenly in the logarithm, so the leading digit d turns up with probability log₁₀(1 + 1/d) — a 1 thirty per cent of the time, a 9 under five. The mean absolute deviation of a ledger’s first-digit shares from the curve is Nigrini’s conformity measure; invented numbers fail it, because people make numbers by adding and spread them evenly.

The mean distance from each point to its nearest neighbour, over what complete spatial randomness of the same density predicts — ½√(A/n) for n points in area A, plus Donnelly’s 1978 term for the points a boundary of perimeter P robs of neighbours. Chance sits at one; clumping pulls R below it, and anything that holds points apart — a rake, a planting, territory — pushes it above.

The share of neighbouring pairs that switch. Independent fair bits switch half the time, with an exact binomial spread; people asked to be random switch more — they under-repeat and avoid runs — and the bias survives being told about it. It is the simplest case of a Markov transition table, and the one a penny-matching machine needs.

A fluent hand slows in its bends by a fixed law: speed v goes as the path’s curvature κ to the −⅓ (angular velocity A as curvature C to the ⅔), the gain γ changing only from stroke to stroke. Fit the slope of log speed against log curvature and a written signature lands near −⅓; a tracer, creeping along the same line at one speed, lands near zero. The picture is the same — the binding of speed to shape is not.

Take every step between horizontally neighbouring pixels and ask how heavy its tails are: the fourth moment over the square of the second, less the Gaussian’s three. A photograph steps gently almost everywhere and leaps at its edges, so the steps pile into a sharp peak with long tails (κ̂ ≫ 0). A field with the same spectrum but shuffled phases has nowhere to keep an edge; its steps fall into a bell.

If steps are uncorrelated their variances add, and a q-step move has exactly q times the variance of one step — a ratio of one at every horizon. Persistence (positive autocorrelations ρ_k) compounds the ratio above one; mean reversion pulls it below. Lo and MacKinlay’s overlapping, bias-corrected estimator made it a test of the random walk.

The longest streak of either face in n fair flips grows only like log₂ n — about eight in two hundred — and its law is exact by dynamic programming over the length j of the current run: each flip repeats the face (½, the run grows; past m the sequence is lost) or switches (½, back to one). A tape of two hundred that never reaches five happens to a fair coin less than once in a thousand.

For each node, the share of its pairs of neighbours that are linked — T_i the triangles through it, k_i its degree — averaged over the network. Growth through introductions leaves triangles behind; a degree-preserving shuffle keeps every degree and sinks C to the floor its hubs allow by accident, about (⟨k²⟩ − ⟨k⟩)² / (N⟨k⟩³).

Sort every run of three consecutive samples into one of the six ways three numbers can be ordered, and take the entropy of the census. An i.i.d. sequence speaks all six equally and scores one, whatever its values; a deterministic one has a grammar — the logistic map at r = 4 never falls twice running, a word it cannot say.

Slide a line against itself and find the longest stretch where it lands on its own notes, at a distance of six or more. A composed melody returns to its phrases — the repeat is architecture; a Markov chain, however well it knows each step, forgets within a few and cannot walk back. Dependence in real music decays slowly with distance; in a chain it dies off exponentially.

A solid fractal has no formula for “distance to the surface”, but the escape-time potential gives a safe lower bound. The ray leaps by that bound each step — far away it strides, near the surface it creeps — so the whole bulb resolves in a few dozen steps per pixel.

Release one particle far away. It random-walks — Brownian, undirected — until it touches the cluster, then freezes there forever. Repeat a million times. A seed becomes a branching tree.

A bump that pokes out intercepts more wandering walkers, so it grows faster and pokes out more — a runaway that sharpens tips and starves the bays between them. That instability is why the branches never fill in, and why they look the same at every magnification.

A falling body speeds up until drag equals weight. Weight grows as L³, Newton’s drag as L²v², so the speed at which they balance grows only as √L — the mouse lands slowly, the horse does not. (g′ is gravity less buoyancy.)

Below a Reynolds number of one, drag is viscous and linear in speed. Weight L³ against drag L·v gives v ∝ L²: halve a droplet and it falls four times slower — which is why fog hangs and bacteria never land.

One empirical curve joins Stokes (C_d = 24/Re) to Newton (C_d ≈ const). On the log–log plot of terminal speed it bends the slope from 2 down to ½ across a decade or two around Re = 1.

Solve the endgames first: two left, the first shooter at accuracy a wins a / (1 − (1 − a)(1 − b)). With three standing, each turn either ends in one of those duels (hit) or passes the turn (a miss, or the air), so the three-standing states form a single cycle with an exact sum. At 0.3 · 0.8 · 1.0, firing at the Marksman leaves you 0.258, at the Gunner 0.237 — and into the air 729/2150 = 0.339, because the two strong shooters are each other’s largest threat and you enter every endgame holding the first shot.

The exact central interval for K, the count of wins in n independent plays at win probability p — computed from the rules alone, before a single play. A record inside it is the rules working; a record outside it happens one time in twenty, and the band says so in advance. When the chance of each play is fixed before it is played but differs from play to play, the same band comes from the Poisson-binomial sum.

Matter tells spacetime how to curve; the curvature tells matter (and light) how to move. Gravity isn’t a force in the picture at all.

Agent i goes to the bar when the best predictor in its own bag — the one with the smallest running error e — forecasts attendance below the threshold L. Every predictor is rescored each week against what the week brought. A forecast everyone shares cannot survive the rule: believed below L it fills the bar, believed above it empties it.

Ω is the number of microstates that look like the current macrostate. The corner has vanishingly few; "spread out" has almost all of them — so that is where the gas goes. Engraved on Boltzmann’s tombstone.

Surprise, in bits — the average doubt in one channel use. Maximal (one full bit) when p = ½, the channel a pure coin flip; zero when the outcome is certain. This is the thing a good code never wastes.

Rolf Landauer’s 1961 result: erasing one bit of information in a thermal environment costs at least k_BT ln 2 joules — irreversibly dissipated as heat. This is the bridge between information theory and thermodynamics: information is not just abstract, it is physical.

The same sum The Arrow carves as heat (S = k·ln Ω): there it counts microstates and points time forward, here it counts bits and bounds what a channel can carry. Entropy and information are one measure, read in two worlds.

What the instrument plots: coarse-grain the box into cells, count the occupancy pᵢ, sum. It is maximal at a uniform spread — the Gibbs/Shannon form of the same quantity.

Ptolemy’s speed fix: the epicycle’s centre P rides a circle about C, e off the Earth, but turns uniformly about Q, twice as far — the motion is uniform about a point that is not the centre. On a sky of Mars it takes the worst error from six degrees to about one.

The deep identity: an outbreak is the giant connected component of “occupied” edges, each present with probability T. Vaccination is site percolation; transmissibility is bond percolation. The epidemic threshold is a percolation threshold.

Most of the time take the best action you know; a fraction ε of the time gamble on a random one — because the only way to find a better path is to risk a worse one. Too little ε and the agent locks onto the first route it finds; too much and it never commits.

The crowd’s average compounds the arithmetic mean of the factor; any one life compounds the geometric mean, which is always smaller. When ⟨r⟩ > 1 > e^{⟨ln r⟩} the same coin enriches the ensemble and ruins almost every member of it — both verdicts exact, both forever.

Multiply your stake by a random factor, repeat — the whole law. No strategy, no memory, no choice. The line never says who takes the average of the outcomes — over copies of you, or over your own tosses — and that silence is worth a sign flip.

Both top bidders pay. With fixed steps s and bankroll B, solve from the top rung down: the last affordable raise wins, so the verdict alternates rung by rung all the way to the first — and opening is worth −s. Allow jump bids and a single committed opening of x* (in units of s, with budget b and prize v) takes the note while the other rationally passes: O’Neill’s solution, commitment over reaction.

A crystal embryo pays surface tension σ on its skin and earns Δg per volume; their sum is a hill in size whose crest — the critical nucleus — has height ΔG*. Undercooling raises Δg, so the crest falls as 1/ΔT² and the waiting time, e^+ΔG*/kT, collapses exponentially. A dust grain lowers the hill by a contact factor f.

Five constants, one line. The Square’s hidden room runs it as a contradiction: if π sat on some floor of the tower it would be algebraic, so iπ would be algebraic too, so Lindemann–Weierstrass would force e^iπ to be transcendental — but e^iπ = −1, an integer, ground floor. Something has to give, and it is π.

For any connected drawing in the plane with no crossings, vertices minus edges plus faces is 2 — however the drawing writhes. The surface keeps this number, not the drawer.

Houses never wire to houses, so no face can be a triangle — every face spends at least four edge-sides and each edge has only two to give. Six vertices buy eight edges; the ninth has no face to live in.

Every handle a surface grows lowers its Euler number by two and raises the wire budget by four. One handle: the cap moves from 8 to 12, and K_3,3’s nine wires land flat with room to spare.

A walk using every bridge exactly once may hold odd degree only at its two ends — every land mass it merely passes through spends its bridge-ends in one-in, one-out pairs. Zero or two odd shores, never four.

Every edge lifts the degree at both of its ends by one, so the degrees sum to exactly twice the edge count — and an even sum can never be built from an odd number of odd terms. Odd-degree shores always come in pairs, so a count of four can never fall to two by routing.

Inside a wall too tall (E < V₀) the wave doesn’t oscillate — it decays. Not to zero, just smaller: thin enough, and something survives to the far side.

A fast excitable variable v with a cubic nullcline, coupled to a slow recovery w. A small kick decays; a kick past threshold fires a full spike, then a refractory pause — the engine of every wave here. (Simulated via Barkley’s numerically-robust cousin.)

One stable rest state and a threshold. Below it perturbations die; above it the response is all-or-none and identical however hard you pushed, followed by a recovery window in which it cannot fire again. Threshold plus refractoriness is the whole secret.

Diffusion D and reaction rate r set how fast an excitation front marches — faster coupling, faster wave. Multiply by the recovery period T and you get the wavelength: the spacing between crests in a train, or the arms of a spiral.

A curved wavefront travels slower than a flat one by an amount set by its curvature κ. A broken end lags, the front pivots about the gap, and the lag winds it into a spiral that turns forever about its tip — why spirals are inevitable, not accidental.

The expected price moves a fraction w of the way to the last price. The cobweb then converges whenever |1 − w(1 + b_s/b_d)| < 1 — enough caution tames a market that naive expectations (w = 1) would throw apart.

One number for how evenly n users share: 1 when all get the same, 1/n when one gets everything, and k/n when k share equally and the rest starve. It ignores how much there is — utilisation is a separate count.

Raise every flow’s rate together; freeze a flow when its demand is met and share what is left among the rest. No flow can gain without taking from one that has less. It needs every demand and the capacity in one place — an oracle, not a protocol.

Fix c and iterate every starting point instead and you get its Julia set. It is one connected piece exactly when c lies in M — otherwise it shatters into a Cantor dust. M is the index of them all.

An N-spoke wheel looks identical every 1/N turn, so the rate that matters isn’t its spin — it’s how often it returns to an identical pose: N × the spin.

The doubling intervals shrink by a fixed ratio that converges to δ ≈ 4.6692 — and Feigenbaum found the same number for any smooth one-hump map. The route into chaos is universal; it doesn’t care what the map is.

On the real axis z² + c is the logistic map in disguise: the period-1 cardioid, the period-2 disc, then 4, 8, … bulbs accumulating at Feigenbaum — the same cascade The Cascade (INST·02) walks, here as a map.

The flash must be placed ahead of the moving ring by θ_PSE to be seen aligned with it; divided by the ring’s angular speed ω, that offset is a time — how far behind the present the flash is placed by the moving thing’s percept.

A single vertex folds flat only if the sector angles between its creases, taken alternately, add to 180° each — walk round the folded vertex and you must come back to where you started.

At every vertex of a flat fold the mountain creases outnumber the valleys, or the valleys the mountains, by exactly two — the turns of the paper round the vertex must close.

In steady level flight the wing’s lift carries the weight: air density, the square of airspeed, wing area and a lift coefficient of order one. Solve it for V and size enters only through W/S.

Each agent reads only the neighbours inside a radius r — never the whole flock. Everything below is computed over that handful, and that locality is the entire point.

One number for "how much one mind": 0 when headings are random, 1 when unanimous. Raise the noise and φ collapses past a critical η_c — order is a phase, not a dial.

Fourier claimed something outrageous in 1822: any periodic signal — even one with corners and jumps — is a sum of smooth sine and cosine waves at integer multiples of one fundamental frequency. The harmonics slider adds these terms one by one.

Euler’s formula turns each harmonic into a rotating vector of radius |cₙ| spinning at speed n. Stacked tip-to-tip they are the epicycle chain on the left; the pen’s height is the signal.

Run it backwards: the discrete Fourier transform reads the amplitude of each harmonic out of the samples — the spectrum bars. The same operation lives inside MP3, JPEG and every spectrum analyser.

The whole signal folded onto the frequency axis: perfect pitch resolution, and zero sense of when. The wavelet trades a sliver of this for all of the missing time.

Any shape is a sum of pure modes — a Fourier sine series. A triangular pluck at position p has these exact weights, so a centre pluck (p=½) cancels every even harmonic.

Count how many pairs of reconstructed points fall within distance r; it grows as a power of r whose exponent is the attractor’s fractal dimension. The deep payoff: an invariant of a system you only ever sampled through one wire — kin to The Set and The Dendrite.

The number of particles within radius R scales as a power of R, and the exponent D ≈ 1.71 is fractional — denser than a curve, sparser than the plane. The cluster is mostly empty space, at every scale.

How far apart the bright bands sit. Wider wavelength or closer slits ⇒ wider fringes — the two spatial knobs on the panel move exactly this.

A cable carries its load w(x) per horizontal length in pure tension, its horizontal pull H the same everywhere. A uniform deck gives a parabola; the cable’s own weight, spread along its length, gives the catenary. Inverted, the same line carries the same load in pure compression: the arch.

The only place the goal ever enters: the fitter a genome, the more likely it breeds the next generation. Everything else is blind — selection alone carries the information about what “better” means.

Holland’s “fundamental theorem”: short, above-average building blocks (schemata) get exponentially more trials each generation. It is the reason blind variation plus selection actually climbs instead of wandering.

Two surviving parents cut their gene strings at a random point and swap tails. Recombination is what lets evolution mix two good half-solutions into one whole — building-block by building-block, faster than mutation could alone.

Each gene has a small chance μ of flipping to a fresh random value — the one source of genuinely new material. Too little and the population stagnates; too much and selection can’t hold its gains.

Each G′ is the commutator subgroup of the one before it — the swaps a radical built on the previous layer can undo. A polynomial is solvable by radicals exactly when this ladder reaches the trivial group in finitely many rungs. S₂, S₃ and S₄ all ground out; S₅ does not.

A₅ — the even permutations of five roots, order 60 — is simple and equal to its own commutator subgroup. Nest the ladder as deep as you like: it never shrinks, so no radical formula can ever name the roots of a general quintic. This is the rung that refuses to break.

For a fat or soft wall, the leak is set by the area of forbidden region the wave must cross — Gamow’s factor, the same integral that dates a rock by its alpha decay.

The price earns a steady drift μ and is kicked by Brownian noise σ, both scaled by the price itself — so returns, not absolute moves, are what’s random. The multiplicative twin of The Walk’s additive diffusion.

The skew made precise: the mean rides above the median by e^σ²t/2. A symmetric shock to returns is an asymmetric shock to price — upside unbounded, downside floored at zero.

Itô’s lemma integrates the SDE exactly: log-price is a drifting Gaussian, so the price is log-normal. This is the increment each sample path multiplies by — no discretisation error.

At a jump, no finite sum of smooth waves ever closes it — a ripple overshoots the edge and, as you add harmonics, only narrows, never flattening. Its height homes in on a fixed ≈ 8.95% of the jump: the square hits it almost at once, the sawtooth climbs toward it from below.

How a glass forgets: not one exponential but a lazy shoulder on a log-time axis — the sum of many exponentials with no single τ. The smaller β, the wider the spread of relaxation times. It breaks, on purpose, the one-τ decay a single well gives.

The empirical fit to a cooling liquid’s relaxation time: an Arrhenius law whose effective barrier blows up at a temperature T₀ above absolute zero. Fragile liquids follow it steeply, strong ones stay nearly Arrhenius (Angell’s map).

Cages hop among traps whose depths are exponentially distributed; escape from a trap is Arrhenius. Below T_g the mean trapping time diverges, so the system keeps finding deeper traps: its relaxation time grows with its own age, and equilibrium recedes as it is approached.

Each depositor sees θ through private noise and withdraws when the chance the bank fails beats the cost ratio c. At the marginal depositor’s signal the share of others withdrawing is, to her, uniformly uncertain — so at the threshold the run is exactly 1 − c, whatever the noise. With ψ(θ) = θ this is Morris and Shin’s θ* = 1 − c.

α is the precision of a public signal, β of each private one. While the inequality holds the threshold is unique; beyond it the public signal becomes common enough to coordinate, and a precise enough headline can start a run on a bank the private evidence calls sound. With the bank’s resilience slope ψ′ the bound is α ≤ ψ′√(2πβ).

The delay between a satellite’s time stamp and the receiver’s clock, times c: the true range, plus the receiver clock’s error b — the same for every satellite — plus noise. Four of them fix x, y, z and b; a nanosecond of b is thirty centimetres of every range.

H holds, for each satellite, minus the unit vector toward it and a 1 for the clock. Range noise σ becomes position-and-time error σ·GDOP: satellites spread over the sky keep it near 2; bunched together, the rows of H nearly agree, HᵀH nearly loses its inverse, and the error ellipse swells.

A proxy U is the value V plus an independent error X, both normal. Pick by U and you get back only this fraction of the pick in V — the rest of the selection was error. Galton’s regression toward the mean, read as a law of metrics.

Gao, Schulman & Hilton’s fitted form for the gold reward against the distance d a best-of-n policy moves from the initial one (d is the square root of its KL divergence, which for best of n is known in closed form). It rises, peaks at d = α/2β, and falls.

The expected largest of n standard normals, exactly by the order-statistic density, and its slow asymptote. For a Gaussian proxy the true value of the best of n is σ_V²/σ_U times it — it keeps rising, ever more slowly.

M is the link matrix (a column per node, its 1s split evenly down its out-links). The teleport term — jump to a random node with probability 1−d — makes G positive and connected, so a single, well-defined rank is guaranteed to exist.

A teller who finds the listener already knew stops with chance 1/k (Demers et al.’s feedback-and-coin rule). In the continuous-time limit the fraction never told solves this equation: ≈ 0.203 at k = 1, Daley & Kendall’s number, and falling exponentially in k.

Each informed node tells one random node a round. The informed set doubles until most know (log₂ n rounds), then the last stragglers are found like coupons — each round catches a fixed fraction of the rest (another ln n).

An uninformed node asks a random node and stays ignorant only if it asked another ignorant one: the ignorant share squares every round. Push spends its last rounds looking for the few who have not heard; pull lets them come looking.

The flywheel speeds up when the steam torque k cos φ beats the load F; the balls swing out when the spindle’s centrifugal pull beats gravity, and the sleeve’s friction b resists them. The throttle closes as they rise. Linearised about the working speed, the pair becomes one cubic.

The whole of learning. The gradient ∇L points uphill, so step the opposite way, scaled by the learning rate η. Repeat. There is no cleverer secret underneath modern AI than this line.

A mountain’s base flows when the weight of one molecule’s column, A m_pgh, matches the energy that melts it — a small fraction ξ of a Rydberg. Weight ∝ h³ on a base ∝ h²: the square–cube law of a bone, written on a planet. On a rocky body g ∝ M^1/3; where h_max reaches the body’s own radius, the potato radius, gravity stops shaping it.

B autocatalyses through the AB² term — it eats A to make more of itself; feed F tops up A, kill k removes B.

The area inside a closed curve is a sum taken only along the curve. Walked counter-clockwise it is positive, clockwise negative; a figure-eight returns the difference of its lobes. On a polygon the sum is the shoelace formula.

The wheel on the tracer arm rolls only with the part of its motion perpendicular to the arm. Summed round a closed trace, that sideways roll ℓ times the arm length b is the area the tracer enclosed — so long as the pole arm swung out and came back without going round.

With the pole inside the outline both arms make a full turn, and the wheel misses the area of one special circle — the one the tracer could ride forever without the wheel turning. Here a is the pole arm, b the tracer arm and c the wheel’s place along the tracer arm from the hinge; a wheel set behind the hinge makes c negative.

Two reflections per round: the oracle O_w flips the sign of the marked answer, then the diffusion operator reflects every amplitude about their mean — and the marked one grows while the rest shrink.

The whole algorithm is one rotation in the plane spanned by the answer |w⟩ and everything-else |s′⟩: each step turns the state by 2θ toward |w⟩. Run it too long and it sails past the answer.

About π4√N rounds land the amplitude almost entirely on the answer — a quadratic speedup over the N/2 a classical scan averages, and provably the best any quantum search can do.

The velocity distribution relaxes to Maxwell–Boltzmann (the line in the histogram), and H falls monotonically: the H-theorem. The temperature never changes — only the shape does.

Codewords sit at least three flips apart, so any single error still lands nearest its original — sphere-packing in bit-space. A 3-bit syndrome names the flipped position outright. Hamming, 1950.

To move n disks, move n − 1 out of the way, move the largest once, and move the n − 1 back on top: the count doubles and adds one, so it is 2ⁿ − 1 — and no solution is shorter. Each disk doubles the wait; sixty-four of them at one move a second outlast the stars.

With a fourth peg, park the top k disks using all four, move the other n − k with three, and bring the k back: choose k to minimise. The wait stops doubling per disk — 64 disks fall from 1.8×10¹⁹ moves to 18,433. Optimal for four pegs (Bousch 2014); open beyond.

The walker-arrival probability obeys Laplace’s equation — the same discrete Laplacian The Skin diffuses with — and the cluster grows fastest where that field is steepest. Drift biases the walkers and the tree leans into the wind.

A new server on the ring takes only the keys between its predecessor and itself — on average one share in S + 1, the least any map could move. Under key mod N a key stays only if its remainder is the same for S and S + 1, about one key in S + 1; all the rest move.

The S·v positions are uniform points on the circle, so the arcs between them are nearly independent exponentials and a server’s share is a sum of v of them. With one position each, the busiest holds about ln S times the mean; with v, the spread shrinks like 1/√v.

Each node keeps a finger to the node succeeding it at every power of two around the ring. Forwarding a lookup to the closest finger before the key at least halves the remaining distance, so a lookup takes at most log₂ S hops and about half that on average.

The size of one immigrant’s family when every member has Poisson(n) children. Below n = 1 its tail is cut off exponentially; at n = 1 the cut-off vanishes and P(s) falls only as s^−3/2.

The rate of events right now: an exogenous base μ, plus a decaying kick from every past event. Each kick integrates to n, the branching ratio — the mean number of events one event sets off — and fades on the time scale 1/β.

Each immigrant brings 1 + n + n² + … events in all, so the long-run rate is the exogenous rate divided by 1 − n, and the endogenous share of activity is n. At n = 1 the series diverges: there is no finite rate, and no amount of news is needed to keep the market busy.

The walker density obeys the heat equation; its point-source solution is the spreading Gaussian — the silver curve, of width σ = √(2Dt). One blind walker becomes a smooth, deterministic law.

The curve where a mode never moves. Sand bounced off the violently vibrating antinodes drifts until it lands on these still lines and stays — so the silence, not the sound, is what you see. Chladni’s 1787 figures are these zero sets made visible.

Count the modes below frequency k: they grow as the drum’s area times k², with a correction in its perimeter. So the spectrum betrays the area and the edge exactly — you can hear how big the drum is. Just not, Kac asked, its shape; in 1992 the answer came back: no.

A drumhead’s standing waves are the eigenfunctions of the Laplacian on its domain Ω, pinned to zero at the rim. Each solution φ is a pure mode shape; its eigenvalue k² fixes the tone. The shape of Ω chooses the whole ladder of modes — that is the question.

On a rectangle the modes are products of sines, frequencies following √((m/a)² + (n/b)²); on a disk they become Bessel functions J_l with circular and radial nodes. The sand settles wherever φ = 0.

Immunise a fraction p and you thin the network of usable paths. Past p_c the infected component can no longer span the graph and the outbreak stops — even though a majority remain susceptible.

Lamarck’s engine, made quantitative: the young inherit a share h of what their parents acquired, Δ. Under a standing use the trait climbs every generation with no selection at all.

The germ line is set apart from the body: nothing acquired in a life (Δ) writes to what is passed on. Dock tails for five generations and every pup is born whole.

A decision, in closed form: the chance of finishing above a target K is Φ(d₂). The fraction of sample endpoints above the line converges to it — the histogram checking the formula.

Replace the forward sum with a max and keep a back-pointer: δ carries the score of the best path reaching each state, and the trace-back yields the single most-likely sequence of regimes — one global decode, not a string of per-step guesses.

A hidden chain of states you never see — transitions a — each state emitting an observation x through its own distribution b. Three things define it: the start π, the transition matrix, the emissions. Everything else is inference.

Sweep forward accumulating α (the evidence up to t) and backward accumulating β (the evidence after it); their product, normalized, is the probability of each hidden state at every step — the soft regime bands painted under the price.

When the transitions and emissions are unknown, Expectation–Maximization re-estimates them from the very posteriors forward–backward computes, alternating until they stop moving. The chain learns its own parameters — the unsupervised cousin of The Lens’s given model.

Any vector field splits (Helmholtz) into a divergence-free flow plus the gradient of a scalar. We solve a Poisson equation ∇²p = ∇·u* for that scalar and subtract its gradient, leaving a mass-conserving field. This is the projection step every frame.

A plate records intensity only, but the cross term R*O carries the object wave’s phase, written into the spacing of the fringes. Its conjugate RO* — the twin — is recorded too. In-line the two share every spatial frequency; tilt the reference and they part.

A field is a sum of plane waves — its 2-D spectrum — and each plane wave advances by a known phase. Multiply, transform back, and the field is anywhere you like: run z negative and a plate refocuses to any depth of the scene that lit it.

The near-axis limit of the same propagation: every point of the plate sends a spherical wavelet, approximated as a parabola. It is why one point’s hologram is a zone plate — rings that crowd outward — and why a small piece of plate still holds the point.

Pick any two distinct fibres: Gauss’s integral, which counts how many times one loop threads the other, always returns exactly one. No two are unlinked — the whole weave is chain-mail with no free rings, kin to the trefoil of The Knot.

Every point of the 3-sphere (a,b,c,d) is sent to a point of the ordinary 2-sphere. The map forgets one circle’s worth of information — and the whole circle it forgets lands on the same spot below. That circle is a fibre.

Fix a target (θ, φ) on the base sphere and let the shared phase t run once around: the pre-image is a full circle in the 3-sphere. Slide (θ, φ) and the circle sweeps out a torus — the whole weave is these circles, one per direction.

The 3-sphere lives in 4-D, so to see it we cast it into ordinary space from one pole — the same trick that flattens a globe to a map. It sends circles to circles, so every fibre stays a clean ring you can orbit.

Carve each pattern ξ^μ into the wiring: neurons that agree in a memory get a positive link, those that disagree a negative one. Memory is a property of the connections, never one address.

The Ising energy of The Threshold, but with learned couplings w_ij. Every neuron flip can only lower it, so the state slides downhill and must stop — in a memory.

Store more than ≈0.138N patterns into N neurons and the basins overlap and collapse into spurious blends — the net forgets everything at once. The spin-glass price of a distributed memory.

Show a corrupted cue, then let each neuron align with the majority vote of the rest. The state rolls down E into the nearest stored pattern — content-addressable memory: the cue is the address.

Any consistent system strong enough to do arithmetic leaves sentences it can neither prove nor refute — and one of them, G, is true. Add it as an axiom and the new system grows its own G′. The shoreline moves; the sea never empties.

A consistent system cannot prove its own consistency — Con(T) is itself one of its dark stars. The auditor cannot audit herself. Gentzen (1936) proves it anyway, but only from a stronger vantage: induction up to ε₀.

λ is the public record’s log-odds, s = ±1 the private signal. Once the record leans by more than one signal’s worth, every arrival takes the same door whatever she saw; her action then tells the record nothing and λ stops moving. With binary signals that happens after the first two agreeing choices (a tie is a coin).

The chance the crowd ends up locked on the better door, and on the worse one, for signals right with probability p. At p = 0.6 one crowd in three cascades the wrong way — with a hundred people still to come, each holding a signal that would have helped.

The correlation of the signals at two points — two dishes — is the Fourier transform of the source’s brightness, read at their separation in wavelengths. An interferometer never sees the sky; it hears one coefficient of its transform per baseline, per moment.

The dirty image is the sky convolved with the dirty beam B. Find its brightest point, subtract a fraction γ of the beam there, record a delta-function component; repeat. Then convolve the components with a clean Gaussian beam and add back the residual. It is a prior — the sky is a sum of points — and it fills the unsampled uv plane with it.

The finest detail a telescope separates is about the wavelength over its size — or, for an array, over its longest baseline. At 1.3 mm across the Earth that is some 25 micro-arcseconds: a grapefruit on the Moon. The shortest baselines set the largest scale it can see at all.

In the cell, P and Q stay on one ray from the fixed pivot O and the product of their distances never changes: Q is P inverted in the circle of radius k. Inversion sends every circle through O to a straight line — so drive P round a circle through O, and Q walks a line, exactly.

No fitting, no simulation — Onsager solved it on paper. Above it, noise wins; below it, order.

Each spin σ = ±1 only wants to match its four neighbours (coupling J > 0); the field h tilts the whole sheet one way.

Temperature is the only referee: cheap moves are always taken, costly ones only a fraction e^−ΔE/kT of the time. Run it and the sheet is a draw from the thermal ensemble.

The magnetisation leaves zero with a vertical tangent at T_c — that exact curve is what the inset dot rides.

At T_c the correlation length blows up and the lattice looks the same at every zoom. The kicker: these exponents are universal — the very same numbers describe water’s critical point and a ferromagnet.

With every signal acted on and counted, a majority of N independent voters, each right with probability p > ½, is right with a probability that tends to one. The sealed street is a Condorcet jury; the published one is not.

The heart of it: the gain is the prediction’s share of the total uncertainty. Confident prediction (small P⁻) or noisy sensor (large R) → K small, trust the model. Loose model (large Q) → K → 1, trust the blip.

The model pushes the belief forward — here a constant-velocity guess — and inflates the covariance by the process noise Q. With no correction this step alone is a forecast: the ellipse only grows.

How wrong the prediction was, and how much that surprise could be noise. S is the innovation’s covariance — the prediction’s uncertainty plus the sensor’s.

The new estimate is the prediction nudged toward the blip by K of the innovation, and the covariance shrinks — the ellipse closing. Information only ever sharpens the belief.

Run predict without correct, over and over, and P grows without bound: the covariance ellipse becomes a forecast cone. The Lens, blinded, is The Oracle.

The attractor has a fractal (Hausdorff) dimension of ≈ 2.06 — more than a surface but less than a volume. Trajectories are confined to this set forever, yet are locally unstable in every direction. This is what "strange" means: attracting yet chaotic.

The exact single-soliton solution to KdV. Amplitude = c/2, speed = c — taller solitons travel faster. The width narrows as amplitude rises, so a soliton’s energy is concentrated into ever-tighter pulses. When two solitons meet, they pass through each other, emerging with the same shape — only their positions shift (the phase shift).

After two KdV solitons collide and separate, each soliton is displaced from where it would have been without the collision. This phase shift is the only trace of the interaction — their shapes are entirely preserved. Zabusky and Kruskal named these objects “solitons” in 1965, because they behave like particles.

The KdV equation governs shallow water waves, plasma waves, and ion-acoustic waves. The nonlinear term 6u∂u/∂x steepens tall waves (making them faster); the dispersive term ∂³u/∂x³ spreads them out. Their balance produces the soliton: a stable, shape-preserving pulse that travels at constant speed forever.

Bernoulli’s idea, made a law by Breiman: value log-wealth, not wealth. Maximising expected log-wealth is exactly maximising the geometric growth rate — and it almost surely beats every other strategy over the long run, while a bigger expected dollar return can march the median to zero.

Bet the fraction f* of your bankroll set by your edge over the odds — win probability p, loss q = 1 − p, payoff b to 1. No edge, no bet; a big edge at short odds, bet big. This one fraction maximises long-run growth.

The exponential rate your wealth compounds at when you stake fraction f. It is a hill: zero at f = 0, peaking at f*, and back to zero at the no-growth point f = 2f* — past which a favourable game loses money to volatility drag.

The same criterion for a lognormal asset (The Oracle’s GBM): stake the excess return μ − r over the variance σ². Thorp carried this from blackjack to the market — the growth-optimal portfolio is the Sharpe ratio, squared and scaled.

Put a two-bar arm’s pen into the curve’s polynomial and products of cosines become sums of cosines. Each term is a rod at angle mα + nβ + γ, a point at radius c on it; add the points and the sum’s x-coordinate is f − C. Hold it on x = −C and the pen is on the curve.

Two similar contraparallelograms share a rod, the first’s corner pinned to the second’s long bar: the third rod is the mirror of the first in the second. Chained from a fixed rod it multiplies an angle; two of them through a bisecting rod add two.

The orbit is an ellipse with the Sun at a focus, and the line to the planet sweeps equal areas in equal times. Two curves fit the sky that a chain of circles only approaches — and the area law is a reason, the one Newton would derive.

A polynomial fingerprint built by switching one crossing at a time. It tells the trefoil from its mirror image — something no amount of wiggling can do by eye — and it fell out of the algebra of quantum observables, not geometry.

A real number that measures how much the loop coils around itself, summed over every pair of points along the curve — the same double integral Gauss wrote for linking. Averaged over all viewing angles it is the signed crossing count you actually see.

Treat the curve as charged and let it push itself apart. Flowing downhill in this energy untangles every loop to its tidiest form — a plain circle for the unknot — yet it can never force a real knot open. That obstinacy is the invariant.

The rate at which thermal fluctuations knock a particle over a potential barrier ΔU, with D = σ²/2 the noise diffusion. In the threshold model of stochastic resonance, this sets how often the noisy signal crosses the threshold — and therefore the output firing rate that tracks the input signal.

Above threshold the coherence rises continuously from zero — a second-order phase transition. Exact for a Lorentzian spread; the silver curve the live point tracks.

The ignition point. Widen the frequency spread g and you raise the bar; sharpen it and the chorus locks for almost any nudge. The blue line on the bifurcation inset.

Only oscillators slow enough to be captured join the cluster; |ωᵢ| > Kr keep drifting around the rim. The locked fraction is on the readout — and a bimodal spread can refuse to pick one rhythm at all.

Because the coupling is global, every oscillator feels a single pull toward the mean phase ψ, of strength Kr. The pull grows with the order it has already created — the feedback that makes the transition sharp.

Each oscillator runs at its own frequency ωᵢ and is pulled toward every other by the coupling K. With global coupling this is the whole model — Kuramoto’s solvable distillation of Winfree.

The population’s mean phasor. r measures coherence: 0 when the phases are scattered, 1 when they march in lockstep. It is the arrow drawn on the circle and the number on the readout.

Every graph that refuses to lie flat hides one of exactly two minimal culprits — the five-clique or the utilities board. This instrument is not an example of the wall; it is half of the wall’s entire statement.

Overdamped particles in a sawtooth potential that switches on and off, s(t) ∈ {0, 1}. On: they slide into the teeth. Off: they spread as pure diffusion. On again: the asymmetry harvests the spread — net transport uphill against the load F, paid for entirely by the switching.

The bend at an interface — not a separate law but the price of least time: minimise T over the crossing point and the derivative vanishes exactly when n sin θ is conserved across the boundary. Past the critical angle there is no real θ₂ and the light reflects entirely.

Of all the paths light could take between two points, it takes the one whose travel time is stationary — almost always the least. The speed in a medium is c/n, so the ray bends to spend less length where it would crawl. Every law of ray optics falls out of this one line.

The fastest descent is a cycloid — the curve a chalk mark on a rolling wheel traces, run upside down. Bernoulli found it by a trick: a bead falling through ever-faster layers is a light ray refracting through ever-thinner glass, so sin θ / v = const — Snell’s law — gives the curve. Least time, twice.

Demand that the action be stationary against every small wiggle of the path and this is what survives — the equation the extremal curve must obey. Feed it L = T − V and out comes Newton’s F = ma; feed it the descent-time integrand and out comes the cycloid. One machine, every path.

Of all the ways a system could move between two fixed events, it takes the one whose action is stationary. This single sentence contains Newton, optics, relativity and — through Feynman — quantum mechanics. The rack’s deepest answer to what is the law under the laws?

Why does the classical path win? A quantum particle takes every path, each carrying a phase e^(iS/ℏ). Away from the stationary path the phases churn and cancel; near it they march in step and add. “Least action” is the place where neighbouring histories stop interfering — classical mechanics as the loud silence of a quantum sum.

Geometric length weighted by the index — the distance light “feels”. Fermat’s path is the one of stationary OPL; a perfect lens works by making every route to the focus take equal optical path, so they arrive in phase.

The time for a bead to slide a curve y(x): arc length over speed, with v = √(2gy) from energy conservation. The brachistochrone is the y(x) that makes this integral least — a problem about a whole function, not a number, which is what the calculus of variations was invented to solve.

Not energy (that is T + V) but its mirror — kinetic minus potential. Nature’s ledger: a path is cheap when it banks potential early and spends kinetic sparingly. Integrate L over the trip and you get the action, the thing the true motion makes least.

Every cell reads its neighbourhood through a kernel K, runs the total through a growth map G, and edges up or down by a small Δt. Take Δt → 1 and the kernel to Life’s 3×3 count and this is The Garden — Lenia just lets the grid, the clock and the state all go smooth.

A bell curve in [−1, 1]: a neighbourhood total near the sweet spot μ makes a cell grow, anything too sparse or too crowded makes it decay. Life’s hard “born on 3, survive on 2–3” became a smooth dial — and the width σ is how forgiving it is.

The kernel is a soft ring — each cell weighs an annulus of neighbours, not just the eight touching it. That single ring, normalised to sum to one, is the whole difference between a pixel grid and a medium an organism can swim through.

Solve for the apparent position θ given the true source β. When the source sits dead behind the mass, every solution is a circle — a full ring of radius θ_E. (This is the line the shader inverts per pixel.)

Remove the finite-variance condition (Cauchy / heavy tails) and the classical CLT is void: the sum stays a fat-tailed Lévy α-stable spike, no Gaussian ever forms — the wildness real markets live in (Mandelbrot vs Bachelier).

One law for every cell: a dead cell with exactly three live neighbours is born, a live cell with two or three survives, anything else is death by loneliness or crowding. No parameters at all.

Each cell counts only the eight around it (the Moore neighbourhood) and cares only about the total — not which neighbours are alive. A local, totalistic rule, exactly like The Rule one dimension down.

A ray grazing a mass bends by this angle — exactly twice the Newtonian guess. Eddington’s eclipse confirmed it and made Einstein famous overnight.

M is the visitor’s match, S what a photometer would set (the luminance), R what full constancy would set (the paint). Zero is a luminance meter, one is perfect albedo recovery; the log form treats equal ratios as equal steps.

Within each framework of illumination the highest luminance is anchored to white (reflectance 0.9) and everything else scales from it; the reported lightness blends the local framework (the shadow) with the global one (the whole scene).

What reaches the eye from a matte surface is its reflectance (the paint, R) times the light falling on it (E). One number, two causes: a dark paint in sun and a light paint in shade can send the same L, and the eye has to decide which product it is looking at.

For any nonzero algebraic number α, e^α is transcendental — never the root of any polynomial with rational coefficients. Setting α = iπ forces π itself off every floor of the constructible tower: degree infinity, and no compass-and-straightedge square can ever equal the circle.

A denser-than-average region pulls in more matter and grows denser still — runaway under its own gravity, slowed only by the cosmic expansion. While the contrast is small every patch grows by the same factor D(t): the rich get richer.

Walk round a four-bar linkage and you come home: crank and coupler, as complex vectors, add up to ground and rocker. Give the crank angle and the loop fixes the other two — one circle–circle intersection. Watt’s pen is the coupler’s midpoint; at his symmetric proportions its error grows only as the fifth power of the stroke.

No friction: energy is conserved and phase-space volume is preserved. The system is deterministic and reversible — and still unpredictable. Determinism ≠ predictability.

The only order a distributed system can know: a before b if a could have caused b — earlier on one machine, or linked by a chain of messages. Two events with no such chain either way are concurrent, a ∥ b, and no observer-independent order exists between them.

One integer per process, ticked on every event and pushed past every stamp it receives. It never contradicts causality — but the converse fails: C(a) < C(b) does not mean a could have caused b.

One counter per process instead of one in all. Each process ticks its own entry and takes the component-wise maximum of every stamp it receives; then a happened before b exactly when every entry of V(a) is at most V(b)’s — causality, captured both ways.

Robert May’s 1976 toy ecology: a population is the previous one grown by r and damped by its own crowding (1 − x). One quadratic, iterated — and it breeds the full route to chaos.

The map between frames moving at relative speed β (units of c). It mixes space into time and time into space — which is why a boost shears the diagram instead of just sliding it.

Einstein’s second postulate: light’s speed is the same for every observer in uniform motion. No medium, no wind, no fringe to slide — the null the interferometer kept returning, made into a law.

The one number every observer measures the same. It is the metric of flat spacetime, and the reason the 45° cone is absolute while the axes are not.

The Lorentz factor. At β = 0 it is 1; as β → c it runs to infinity — moving clocks slow and moving rulers shrink without bound. The ember dot on the inset rides this curve.

Two events at the same time in one frame are not at the same time in another. "Now" is a frame-dependent slice — the root of every relativity paradox, drawn live in the Simultaneity scene.

The time a carried clock actually reads. The straight (inertial) path between two events racks up the most proper time, so the twin who turns around comes home younger — by exactly 1/γ.

Inside the cone (timelike, s² < 0) cause precedes effect for everyone — the order is absolute. Outside it (spacelike, s² > 0) no signal connects the events and their order is up for grabs.

The whole instrument in one line. E pushes a charge straight; the v × B term pushes it sideways, always perpendicular to motion — so it does no work and can only turn. Magnetism curves; electricity accelerates.

In a pure magnetic field the charge circles forever at the Larmor radius r_L, turning at the cyclotron frequency ω_c — which depends on the field and the charge-to-mass ratio, not on speed. Faster particles just trace bigger circles in the same time.

Cross an electric field with a magnetic one and the gyrating charge slides sideways at a steady drift — and the drift velocity is the same for every charge, positive or negative, light or heavy. The plasma moves as one; that’s the “aha”.

Below ρ_c ≈ 24.74 the only attractors are the two fixed points C±. Above it they lose stability and the strange attractor is born — a fractal of dimension ≈ 2.06, never a surface, never a volume. This bifurcation is the instrument’s phase knob.

σ and ρ come from meteorology (Prandtl and Rayleigh numbers). Lorenz derived these equations in 1963 while modelling atmospheric convection rolls — and discovered chaos by accident when he re-ran a simulation from a rounded printout.

Prey x breed freely (α) and are eaten on encounter (βxy); predators y grow only by eating (δxy) and starve otherwise (γ). Two coupled rates — the minimal ecology, and the first answer to “if conditions change, what happens to the populations?”

The one nonzero state where both rates cancel. But it is a center, not an attractor: nudge the system and it does not return — it circles. Counter-intuitively, x* depends only on the predator’s constants and y* only on the prey’s.

Every trajectory rides a level set of V — Lotka–Volterra is conservative, like a frictionless pendulum or The Orbit’s two-body ellipse. The orbit you start on is the orbit you keep; the swing’s size is a memory of the initial state, not the parameters.

The whole organism is one short rule that rewrites a symbol into a longer string. Here every F becomes a stem that sprouts two branches and carries on — the entire plant is folded into this one line, the way a seed folds the tree.

A turtle reads the finished string like sheet music: F draws a step, +/− turn it by the angle δ, and the brackets [ ] save and restore its position — which is exactly how one stem remembers where to grow a side branch and then return.

Lindenmayer’s twist on a grammar: each generation rewrites every symbol at once, not one at a time — modelling cells that all divide together. That single change turns a dry formal grammar into a growing, living thing.

Average stretching per step. λ < 0 → nearby states converge (a predictable cycle); λ > 0 → they separate exponentially (chaos). In the period-3 window near r ≈ 3.83 it dips back below zero — order hiding inside the disorder.

Any gap, however small, grows exponentially. λ is the slope of the climbing line in the inset — measured live by Benettin’s method.

Every extra Lyapunov time τ costs another factor of e in the precision you’d need. You can’t buy much future: the cost grows exponentially, the reach only logarithmically.

Two nearby trajectories separate exponentially: δ(t) ≈ δ₀ e^λ₁t. If λ₁ > 0 the system is chaotic — any uncertainty in the initial state doubles in 1/λ₁ ≈ 1.1 s. We measure λ₁ live by Benettin’s method: track a shadow orbit, renormalise, accumulate.

The boundary is endlessly detailed — Hausdorff dimension 2 (Shishikura) — with whole mini-Mandelbrots buried at every depth, and the same shape governs the chaos onset of any smooth family. Infinite structure, five symbols of law.

The black region: the set of c whose orbit of 0 never escapes. Outside, colour is the smooth escape rate — how many steps the point survives before it flies away.

No 3D multiplication behaves like the complex plane’s, so White & Nylander defined one by analogy: write the point in spherical coordinates (r, θ, φ), multiply the two angles by n and raise the radius to the n. The escape test |z| > 2 carries over unchanged.

A ticket that pays 2^k with probability 2^−k adds one coin to its expectation for every depth, forever. No single price is fair; Feller showed that n tickets together are fair at about n·log₂ n, so the running mean climbs like log₂ n with no ceiling — the more you buy, the more each is worth.

Stop a fair game by any rule that cannot see the next coin and its expected value does not move; give the house an edge and the expected loss is exactly the edge times the expected stake (Wald’s identity). A betting system chooses only when to stop and how much to stake — so it can change the shape of the outcome, never its mean.

Stone is strong in compression, has no strength in tension and does not slide. Then if any line of thrust in equilibrium with the loads lies inside the masonry, the arch stands; it fails only when no line fits and hinges open where the line touches the edge.

In a closed vessel nothing is made and nothing lost: weigh everything and the books close. Lavoisier’s sealed retort made it a measurement — the calx gains exactly what the air inside loses.

Maxwell’s 1867 thought experiment: a tiny intelligent being controls a trap-door and sorts fast from slow molecules, creating a temperature difference without doing work — apparently defeating the second law. The resolution, completed by Landauer and Bennett, is that the demon’s memory must be erased, and that erasure generates entropy that restores the ledger.

The letters available d ms after the flash — the cued row’s score times the number of rows — falling from A₀ towards what whole report gives. Sperling plotted the curve; the exponential is the standard summary of it, fitted here by least squares.

For a change-detection test that probes one item of N: if K of them are held, a change to a held item is always caught and the rest are guessed, so the hit rate H exceeds the false-alarm rate FA by K/N. Below capacity K follows N; above it, K levels off — at about four.

Always take a downhill move; take an uphill move of cost ΔE only with probability e^−ΔE/T. High T accepts almost anything; as T → 0 only improvements pass. This one rule lets a search climb out of a trap.

The film between two rings of radius R a distance h apart is a catenary spun round the axis. Its waist c solves R = c cosh(h/2c), which has a root only while h/2R < 0.6627 — then the film must snap.

A film with the same air on both sides has no pressure across it, so by Young–Laplace its mean curvature is zero: curved one way exactly as much as the other, at every point — the surface of least area for its boundary.

The degree-n polynomial closest to f in the worst case is the one whose error E = ‖f − p*‖_∞ is reached n + 2 times with alternating sign. Born from trying to straighten Watt’s linkage: spread the error evenly across the stroke instead of making it tiny in the middle.

N agents each pick a side, aᵢ = ±1, with the best of S strategies that read the last m minority sides; the minority wins. σ² is the variance of A. Coin tosses give σ²/N = 1. Measured for any N and m, σ²/N falls on one curve of α: far above 1 when α is small (the crowd herds), a minimum near α_c, back toward 1 when memory outgrows the crowd.

How wrong the network is, averaged over every point. Squared so that big misses dominate and the surface is smooth — a landscape in weight space the optimiser can roll down.

Monte Carlo is the other half: roll N independent futures and the empirical distribution converges to the closed form at rate 1/√N. Watch the ember cloud tighten onto the cone as you raise N.

Two neighbouring inputs; each is delayed and multiplied with the other, and the two products subtracted. A pattern drifting from 1 to 2 makes the first product win, the opposite drift the second — an opponent motion signal from nothing but timing.

At sea level the muon intensity falls with zenith angle roughly as cos²θ — the slant path through the atmosphere is longer, and more of the parent pions and the muons themselves decay or stop on the way. A horizontal detector adds one more cos θ for its projected area.

Ionisation costs a muon about a = 2 MeV per g/cm²; above a few hundred GeV radiation (bremsstrahlung, pair production) adds bE with b ≈ 4 × 10⁻⁶ per g/cm². To cross a column density X it must start with E_min — about 46 GeV through 100 m of stone — and few do.

N counts observed where the model expects μ: Poisson noise is √μ, so a fixed fractional excess grows in significance as √t. A claim waits for 5σ — and for the excess to agree from more than one side, because the model of the stone is never perfect.

A profile where no player can do better by changing strategy alone. Every finite game has one. In the one-shot dilemma it is mutual defection — individually unbeatable, collectively the worst stable outcome.

The whole of classical fluid motion in two lines: a parcel accelerates from pressure gradients, viscous drag (ν∇²u) and any body force f. The constraint ∇·u = 0 says an incompressible fluid can swirl freely but never pile up.

Each neuron takes a weighted sum of the layer below plus a bias, then bends it through a non-linearity. Stack two such layers and the network can carve curved, disconnected regions — not just a single straight cut.

Stand on the curve, slide down the tangent to where it crosses zero, repeat. Near a root the error is squared every step — correct digits double — which is why this three-century-old update still sits inside every solver and every GPS fix.

Colour every starting point by the root it falls to and the plane divides into basins. Two roots: a straight border, as Cayley proved. Three or more: the borders coincide — every boundary point touches all the basins at once, and the map of destiny turns fractal.

Take only a fraction a of each tangent step and the same roots keep their basins — but the borders between them swell, curl and bloom. The fractal isn’t decoration on the method; it is the method, rendered at the places where the next step is undecided.

Every mass pulls every other. ε is Plummer softening — it rounds off the 1/r² singularity so a near-collision integrates cleanly instead of exploding.

The headline of 1948: codes exist that drive error to zero for any rate below capacity — and none above it. The wall p* is where R meets C. Our codes give out earlier; better ones reach all the way to it.

“Beats more often than not” is not an order: it can run in a circle. Efron’s four dice each beat the next two times in three, and the last beats the first — so whatever is chosen first, there is an answer that wins. The arrow refuses to be a straight line.

XY reads as a binary number: its bit k is 1 when the last k symbols of X equal the first k of Y. Four such overlaps give the exact odds that B appears before A in a fair-coin stream — HHH loses to THH 7 : 1 — and the same self-overlap AA sets the mean wait (HHT 8 flips, HTH 10). Against b₁b₂b₃ the beater is b̄₂b₁b₂.

A photon travels perfectly straight; spacetime is curved, so its path curves with it. This is the exact orbit equation the instrument integrates per ray — not a thin-lens shortcut like The Well. The extra u² term is general relativity; drop it and you get Newton’s half-strength, straight-ish light.

Each numerosity n is represented as a noisy magnitude whose spread grows with n (scalar variability, Weber fraction w). The chance of picking the larger of two depends only on their ratio; w is fitted to the visitor’s answers by maximum likelihood.

Sample slower than twice the feature rate and the motion is lost — it can no longer be told apart from a slower one. This single inequality is the whole limit.

Scale-free in time as well as size: the activity’s power spectrum runs as 1/f, the ubiquitous flicker noise found in resistors, rivers, heartbeats and starlight — the puzzle self-organized criticality was invented to explain.

Parallelograms of sides a, b and angle γ, each tilted θ out of the plane: every vertex of the sheet follows from θ alone. Folding shrinks both the run S and the rise L at once, so the in-plane Poisson’s ratio is negative.

A node’s rank is the rank poured into it along incoming links, each source splitting its own rank evenly across its out-links. Importance is recursive — you matter if things that matter point at you — plus a small flat share (1−d)/N so nothing starves.

Count the pairs standing out of order and take minus-one to that power. Every arrangement is even or odd — a single swap flips the sign, so no chain of swaps can ever change a shuffle’s parity by stealth.

Both factors flip on every slide, so their product never does. Loyd’s deal carries I = −1; the solved board carries +1. The two values are two islands, and sliding never builds a bridge.

A legal slide is exactly one transposition — the tile trades cells with the blank — and exactly one unit step of the blank. Two coins, flipped together on every move, with no third option on the board.

Game A is a plain biased coin; game B picks its coin by whether your capital is divisible by 3 — a vicious coin on multiples of three, a generous one elsewhere. Played alone, each drifts your capital down. The law never mentions alternation.

The drift per round, exactly: π₀ is how often the capital sits on a multiple of 3 — the stationary state of a three-state Markov chain. Game B digs its own trap (π₀ high, the bad coin often); mixing in A shuffles π₀ down, and the same formula changes sign.

Let sin θ ≈ θ and the amplitude vanishes from the period — every swing, wide or narrow, takes the same time. This is Galileo’s isochrony, the founding approximation of timekeeping, and it is only true in the limit of swings too small to see.

Straight from the Lagrangian L = T − V. No approximation, no randomness — every pendulum obeys this to the letter (Δ = θ₁ − θ₂).