← All Labs
IDEA #36 · THE MILLENNIUM PRIZE PROBLEMS

Seven hard questions, seven small pictures

In 2000 the Clay Mathematics Institute named seven problems and attached $1,000,000 to each. A quarter-century later only one is solved. Each tab below is an honest visual teaser — not a fake resolution — that computes the real object live and then tells you plainly why the obvious picture isn't a proof. Toggle a subset to check it in an instant while brute force drowns; watch a fluid steepen toward a shock; ride the chord-and-tangent law across an elliptic curve; and see Ricci flow's baby cousin round a blob into a circle. These problems are famous precisely because seeing is not the same as proving.

TAB 1 · COMPLEXITY
P vs NP
Easy to check, hard to solve. Verify a subset in a blink; brute force drowns in 2ⁿ.
TAB 2 · NUMBER THEORY
Riemann Hypothesis
Every zero of ζ on one vertical line. Compute Z(t) live; jump to the full lab.
TAB 3 · PDE / FLUIDS
Navier–Stokes
A flow that may blow up in finite time. Watch a gradient steepen into a shock.
TAB 4 · QUANTUM FIELDS
Yang–Mills mass gap
A field whose ripples cost a minimum energy. Correlations decay — how fast is the gap.
TAB 5 · ELLIPTIC CURVES
Birch–Swinnerton-Dyer
Rational points multiply under the chord-and-tangent law. Rank vs an L-function.
TAB 6 · ALGEBRAIC GEOMETRY
Hodge conjecture
Which shadows of a shape are cast by actual sub-shapes? A gloss, not a proof.
TAB 7 · TOPOLOGY · SOLVED
Poincaré conjecture
Ricci flow rounds a lumpy shape. Perelman closed this one in 2003.
01

Drive it — the subset that sums to the target

Here are 10 numbered stones. Click stones to drop them in the bag; the bag's total updates instantly. Your job: hit the target exactly. Checking a bag is a single addition — anyone can verify a claimed answer at a glance. Finding one is another story.
DRIVE IT · $1M · OPEN
bag total0 / target pick stones
This is Subset-Sum, a classic NP-complete problem. A proposed answer (a set of stones) is checked with one sum — that is what NP means: solutions are verifiable in polynomial time. Nobody knows a method that finds the subset quickly in the worst case; the honest fallback is to try subsets, and there are 2ⁿ of them.
NUMBER OF STONES n10
SUBSETS 2ⁿ
1,024
to search
TRIED
0
so far
CHECK COST
n
one sum
Brute force visits subsets in order. On n=20 that is 1,048,576 bags — and each extra stone doubles it.
02

Walkthrough

Feel the gap between checking and finding.
WALKTHROUGH
1
Press reveal a solution. The stones light up and the bag hits the target exactly. Notice how obvious the answer is once you're handed it — you verified it with a single glance at the sum.
2
Now reset and try to find it yourself by clicking. For small n you can; drag n up to 18–20 and the puzzle becomes genuinely frustrating.
3
Press try every subset. Watch the TRIED counter climb toward 2ⁿ. Each stone you add doubles the search — the bar to the left is exponential, not linear.
4
The whole of P vs NP is: is there always a shortcut that finds the answer as fast as we can check it? For subset-sum, and thousands of problems just like it, nobody knows.
THE INSIGHT
"P vs NP asks whether every problem whose answer is easy to check is also easy to solve. Almost everyone believes P ≠ NP — that some needles genuinely require searching the haystack — but a quarter-century on, no one can prove even that a single NP-complete problem lacks a fast algorithm."
WHY THIS IS HARD
To prove P ≠ NP you must show no clever algorithm — including ones no human has imagined — can ever crack subset-sum quickly. Ruling out all possible programs is a statement about the limits of computation itself, and known proof techniques (relativization, natural proofs, algebrization) have each been shown to be too weak to settle it.
03

Explanation

P, NP, and completeness.
EXPLANATION

P is the class of decision problems solvable by an algorithm whose running time is bounded by a polynomial in the input size — "efficient" in the theoretician's sense. NP is the class whose yes-answers come with a certificate that can be checked in polynomial time. Subset-sum is in NP: hand me the winning stones and I verify with one addition.

Clearly P ⊆ NP (if you can solve it fast, you can check it fast). The million-dollar question is whether the inclusion is strict.

Subset-sum is NP-complete: every problem in NP can be re-encoded as a subset-sum instance in polynomial time. So a fast algorithm for this puzzle would instantly give fast algorithms for all of NP — routing, scheduling, protein folding, breaking cryptography.

That universality is why the stakes are enormous and why most researchers bet P ≠ NP: if the two were equal, creativity itself would be automatable — finding a proof would be no harder than checking one.

04

Research note

The question, exactly.
RESEARCH NOTE
The TRIED counter is a live subset enumerator — it really iterates the 2ⁿ bitmasks and sums each bag, so the exponential wall is measured, not captioned. Verification, by contrast, is the single sum shown in the ledger.
Go deeper: Cook (1971) & Levin's completeness theorem; Karp's 21 NP-complete problems (1972); the Clay P vs NP official statement by Stephen Cook. Connects to #10 discrete-log (one-way functions) and #16 entropy (search vs information).
01

Drive it — hunting zeros on the critical line

The curve below is Z(t), a real-valued companion of ζ(½+it) with the same zeros. Every place it crosses the axis is a non-trivial zero of the zeta function sitting exactly on the line Re(s)=½. Slide the window and count the crossings — each one, so far, obeys the hypothesis.
DRIVE IT · $1M · OPEN
window0–40 · zeros found on Re=½0 ✓ all on the line
Z(t) is computed live from ζ via the alternating (Dirichlet-eta) series with a convergence-acceleration step, then rotated by the Riemann–Siegel phase θ(t) so it comes out real. Where the red curve pierces the axis, |ζ(½+it)| touches zero — the dots mark them.
WINDOW CENTRE t20
WINDOW WIDTH40
Re(s) OFF THE LINE0.50
σ = Re(s)
0.50
critical=½
ZEROS SHOWN
0
in window
FIRST ZERO
14.13
t≈
Nudge Re(s) off ½ and the curve is no longer guaranteed real — |ζ| lifts off zero, the crossings vanish. The hypothesis says you never find a zero anywhere but the line.
02

Walkthrough

Meet the zeros, then leave the line.
WALKTHROUGH
1
Keep Re(s)=½ and slide the window centre up. The first crossing sits at t ≈ 14.13, then 21.02, 25.01, 30.42… — the famous non-trivial zeros.
2
Count them with the ZEROS SHOWN card. Their density grows like (t/2π)·log(t/2π) — the higher you look, the more thickly they pack.
3
Now drag Re(s) off the line. The curve lifts away from the axis and the clean crossings disappear — off the critical line ζ generically has no zeros here at all.
4
The Riemann Hypothesis claims every non-trivial zero lies exactly on Re(s)=½. Verified for over 10¹³ zeros by computer — and still unproven for the (infinitely many) rest.
THE INSIGHT
"The zeros of ζ are the secret frequencies of the primes. If they all lie on one line, the primes are as evenly distributed as they possibly could be — the deepest known statement about the 'music' of the whole numbers."
WHY THIS IS HARD
A computer can only ever check finitely many zeros; the hypothesis is a claim about an infinite list. No pattern in the first ten trillion zeros forces the ten-trillion-and-first to comply, and every general approach — from random-matrix analogies to the search for a self-adjoint operator whose eigenvalues are the zeros — has so far fallen short of a proof.
03

Explanation

Zeta, zeros, and primes.
EXPLANATION

The Riemann zeta function starts life as ζ(s)=Σ 1/nˢ for Re(s)>1 and is then analytically continued to the whole complex plane. It has "trivial" zeros at the negative even integers; everything interesting lives in the critical strip 0<Re(s)<1.

Riemann's explicit formula writes the count of primes up to x as a smooth term minus a sum over the zeros. Each zero contributes an oscillation; where the zero sits controls how big that wobble can grow.

If every non-trivial zero has real part exactly ½, those oscillations are as small as possible and the primes hug their expected count as tightly as arithmetic allows — the error in the Prime Number Theorem is squeezed to O(√x·log x).

The plotted Z(t) makes the zeros visible without complex numbers: multiply ζ(½+it) by a phase e^{iθ(t)} chosen to cancel the imaginary part. Real crossings of Z are exactly the zeros — which is how they were first hunted by hand.

04

Research note

The hypothesis, exactly.
RESEARCH NOTE
Z(t) here is a genuine (modest-precision) computation of ζ on the critical line, not a stock plot; the zero markers are found by sign changes of the live curve. For a far deeper treatment — domain colouring of ζ, the Z-function, primes rebuilt from zeros, and GUE spacing statistics — open the full lab.
Go deeper: the ▶ Riemann Hypothesis lab (this repo); Riemann's 1859 memoir; Montgomery–Odlyzko pair-correlation. Uses the same domain-colouring idea as #19.
01

Drive it — a flow that steepens toward a shock

A 1-D fluid velocity u(x) is carried along by itself (Burgers' equation, the toy sibling of Navier–Stokes). Fast parts overtake slow parts, so a smooth hill steepens — its slope racing toward infinity. Add viscosity and it smooths; remove it and the gradient tries to blow up in finite time.
DRIVE IT · $1M · OPEN
time0.00 · max steepness |∂u/∂x| smooth
The steepness meter is the largest slope anywhere in the flow. With viscosity ν>0 it rises then saturates — the flow stays smooth forever. Set ν very small and it spikes as a near-shock forms: the honest cartoon of a finite-time singularity.
VISCOSITY ν0.0020
INITIAL AMPLITUDE1.00
REYNOLDS ~1/ν
500
turbulence
MAX SLOPE
0
grows?
STATE
regular
Burgers can be tamed — viscosity always wins in 1-D. The open question is whether the real 3-D Navier–Stokes equations can be tamed the same way, or whether a swirl can concentrate energy into an infinite spike.
02

Walkthrough

Watch the gradient race the viscosity.
WALKTHROUGH
1
Leave viscosity moderate and press run. The hill leans forward as its crest outruns its base, then a smooth front forms and the MAX SLOPE levels off — viscosity caught it.
2
Drag viscosity ν toward zero and run again. The front becomes razor-thin and the steepness meter spikes; the STATE card flips toward a warning as it approaches a shock.
3
Raise the initial amplitude: more energy means a faster overtake and an earlier near-singularity. The Reynolds number ~1/ν card climbs — the regime where real fluids turn turbulent.
4
The million-dollar question: in genuine 3-D flow, does viscosity always win (solutions stay smooth for all time), or can the velocity gradient reach infinity in finite time — a mathematical "blow-up"?
THE INSIGHT
"We use the Navier–Stokes equations to design planes and predict weather every day — yet we cannot prove their solutions always stay finite and smooth. The equations we trust most are the ones we understand least."
WHY THIS IS HARD
In 3-D, the nonlinear term that stretches vortices can, in principle, cascade energy to ever-smaller scales faster than viscosity dissipates it. Nobody can rule out a self-focusing swirl that concentrates infinite gradient at a point in finite time — nor construct one. The known conservation laws control energy, but energy alone is too weak to bound the gradients.
03

Explanation

Advection, viscosity, and blow-up.
EXPLANATION

The Navier–Stokes equations express Newton's second law for a fluid: the acceleration of each parcel equals pressure and viscous forces. The troublesome piece is the advection term (u·∇)u — the fluid transporting its own momentum — which is nonlinear and can sharpen features without limit.

Our 1-D model, uₜ + u uₓ = ν uₓₓ (Burgers), keeps exactly this tension: self-advection steepens, the viscous term ν uₓₓ diffuses. In 1-D the diffusion always wins and solutions stay smooth.

The Clay problem asks for a proof (or a counterexample) that the 3-D incompressible equations, starting from smooth finite-energy data, have solutions that remain smooth and finite for all time — global regularity.

Leray showed weak solutions always exist; whether they are unique and smooth is unknown. A finite-time singularity would mean some quantity — vorticity, velocity gradient — escapes to infinity, and the equations lose meaning exactly where fluids are most violent.

04

Research note

The equation, and the question.
RESEARCH NOTE
The stage integrates the viscous Burgers equation with a real finite-difference scheme; the steepness meter reads the largest finite-difference slope each frame. It is a faithful 1-D stand-in — the true problem lives in 3-D, where this same competition is unresolved.
Go deeper: Clay statement by Charles Fefferman; Leray (1934) weak solutions; Beale–Kato–Majda vorticity criterion. Shares its PDE machinery with #7 heat equation and #5 vector fields.
01

Drive it — a field whose ripples cost a minimum energy

The heatmap is a snapshot of a fluctuating field. Its ripples don't come free: correlations between two points fade with distance. Measure how fast they fade and you read off a mass — and the puzzle is proving that mass is strictly above zero.
DRIVE IT · $1M · OPEN
coupling1.0 · measured mass gap m ✓ gap > 0
The lower panel is the correlation function C(r): how alike the field is at two points a distance r apart. A straight line on this log axis means exponential decay e^{−m·r} — and that decay rate m is the mass gap. A massless field would decay only as a gentle power law (a curve that never straightens).
MASS PARAMETER (sets gap)0.40
COUPLING g1.0
CORR. LENGTH ξ
2.5
ξ = 1/m
MASS GAP m
0.40
> 0 ?
FIT R²
exp fit
Push the mass toward zero (massless limit) and the correlation length ξ=1/m blows up: the field becomes correlated across the whole box. The Clay problem is to prove that for Yang–Mills theory the gap can never reach zero — that glueballs are heavy.
02

Walkthrough

Read a mass off a decaying correlation.
WALKTHROUGH
1
Look at the heatmap: with a healthy mass gap the coloured blobs are small — the field forgets itself over a short distance ξ. Press resample to draw a fresh configuration from the same statistics.
2
The lower C(r) plot is a straight line on the log axis. Its slope is the mass gap m; the fit quality confirms the decay is genuinely exponential.
3
Slide the mass parameter down. The blobs swell, the correlation length ξ = 1/m grows, and the line's slope flattens toward zero.
4
Press massless limit: C(r) stops being a straight line — decay becomes a slow power law. That is what "no mass gap" looks like. Yang–Mills is believed to always have a gap; proving it rigorously is the open problem.
THE INSIGHT
"The strong nuclear force has finite reach even though its underlying particles are massless — because the quantum field it lives in has a 'mass gap'. Physicists measure this gap in accelerators; nobody has proven it exists from the equations."
WHY THIS IS HARD
The problem is really two problems: first construct quantum Yang–Mills theory as a rigorous mathematical object in four dimensions (the perturbation series that physicists use doesn't even converge), then prove its lowest excitation has strictly positive energy. Confinement and the gap are seen in every lattice simulation, but a simulation on a finite grid can never rule out a vanishing gap in the true continuum limit.
03

Explanation

Gauge fields, glueballs, and the gap.
EXPLANATION

Yang–Mills theory is the mathematical framework behind the strong and weak nuclear forces: a "gauge field" whose values live in a non-abelian symmetry group like SU(3). Unlike electromagnetism, the field interacts with itself, which makes the theory ferociously nonlinear.

A mass gap means the lightest particle the theory can make — a "glueball" — has energy bounded away from zero. Equivalently, correlations between distant field measurements decay exponentially, with rate equal to that mass.

Our toy plots a Gaussian field with a tunable correlation length ξ=1/m. Real Yang–Mills is far richer, but the diagnostic is the same: fit C(r)∼e^{−m r} and check m>0. Every lattice-QCD experiment finds a robust gap — it's why the strong force doesn't reach across the room.

The Clay problem demands a rigorous construction of the quantum theory on ℝ⁴ and a proof that this gap is strictly positive — turning a mountain of numerical and experimental evidence into a theorem.

04

Research note

The gap, exactly.
RESEARCH NOTE
The heatmap is a real Gaussian random field synthesized with a chosen correlation length; the lower panel measures C(r) from the field and fits an exponential to recover m live. It illustrates the diagnostic for a mass gap — not the SU(N) dynamics themselves, which are exactly what remains unconstructed.
Go deeper: Clay statement by Jaffe & Witten; Wilson's lattice gauge theory (1974); asymptotic freedom (Gross–Wilczek–Politzer, 1973). The colour-decay idea echoes #7's eigenmode decay.
01

Drive it — breeding rational points on a curve

This is an elliptic curve y²=x³+ax+b. Its magic: draw a straight line through two rational points and the third crossing is also rational. Start from one seed point and keep applying the chord-and-tangent law — watch rational points multiply. How many independent ones a curve has is its rank.
DRIVE IT · $1M · OPEN
curvey²=x³−x+1 · rational points bred1 rank ≥ 1 · infinite
Press breed next point: the applet draws the tangent/chord, finds the third intersection, reflects it across the x-axis (the group law), and prints the new rational coordinates. A curve of rank ≥ 1 generates infinitely many; a rank 0 curve runs out almost immediately.
ALG. RANK
1
generators
POINTS
1
bred so far
L(E,1)
0
vanishes
P = (0, 1). Each new point's numerator and denominator grow explosively — a hallmark of a genuine infinite-order rational point.
02

Walkthrough

Rank you can see; rank an L-function predicts.
WALKTHROUGH
1
Start on the rank 1 curve and press breed next point a few times. Rational points keep appearing and their coordinates — printed below the cards — explode in size. This curve has infinitely many.
2
Switch to the rank 0 curve. The group of rational points is finite: breeding quickly loops back to points you've already seen (or to the point at infinity). No infinite family exists.
3
Compare the L(E,1) card. On the rank-0 curve the L-function is non-zero at s=1; on the rank ≥1 curves it vanishes, and the order of vanishing matches the rank.
4
Birch–Swinnerton-Dyer conjectures exactly this: the number of independent rational points equals the order to which L(E,s) vanishes at s=1 — a bridge between counting solutions and an analytic function.
THE INSIGHT
"Whether an equation has finitely or infinitely many rational solutions seems like it should be answerable by algebra — yet BSD says the answer is hidden in the behaviour of a completely different, analytic function built from counting solutions mod p. Two worlds, conjectured to agree exactly."
WHY THIS IS HARD
There is no known algorithm guaranteed to compute the rank of an elliptic curve, and the mysterious Tate–Shafarevich group — which must be finite for the theory to work — has never been proven finite in general. BSD links this uncomputed rank to the value of an L-function whose analytic continuation to s=1 is itself deep. It is proven only in special cases (rank 0 and 1, by Kolyvagin, Gross–Zagier).
03

Explanation

The group law and the L-function.
EXPLANATION

On an elliptic curve the rational points form an abelian group: to add P and Q, draw the line through them, take the third intersection, and reflect over the x-axis. Mordell proved this group is finitely generated — it looks like a finite part plus r copies of the integers. That r is the rank: the number of independent infinite-order generators.

Rank is subtle: some curves have rank 0 (finitely many rationals), others rank 1, 2, 3… and finding it is genuinely hard.

Separately, count the curve's solutions modulo each prime p and package the counts into an L-function L(E,s). This analytic object knows nothing, on its face, about rational points.

BSD conjectures the miracle: the order of vanishing of L(E,s) at s=1 equals the rank, and the leading coefficient encodes the finer arithmetic (regulator, Tate–Shafarevich group, torsion). Counting mod p, summed cleverly, predicts the infinitude of rational points.

04

Research note

The conjecture, exactly.
RESEARCH NOTE
The stage applies the real chord-and-tangent group law with exact rational arithmetic where feasible — each bred point's coordinates are computed, not decorative. The L(E,1) readouts use known values for the three sample curves; the group law itself is live.
Go deeper: Clay statement by Andrew Wiles; Birch & Swinnerton-Dyer's 1965 computations; Gross–Zagier (1986) and Kolyvagin for ranks 0–1. Elliptic curves also power the cryptography behind #10.
01

Drive it — which shadows come from real sub-shapes?

On a smooth complex variety, some "cohomology classes" (algebraic shadows of the shape) are cast by genuine sub-varieties living inside it — and some seem to be pure ghosts. Hodge's conjecture says a special family of these shadows, the Hodge classes, are always real: every one is a combination of actual algebraic cycles. This tab is a gloss, not a proof.
DRIVE IT · $1M · OPEN
class type(1,1) · Hodge · ✓ algebraic — a real cycle
The surface is drawn as a torus (a stand-in for a complex variety). Its cycles — closed loops and higher pieces — live in cohomology, sorted into a grid by "Hodge type" (p,q). Click a cell: green cells are Hodge classes the conjecture promises are drawn by real sub-varieties (the loops that light up); grey cells are non-algebraic types.
TYPE (p,q)
1,1
chosen
HODGE?
yes
p=q
ALGEBRAIC?
conj.
Hodge says ✓
A Hodge class is a rational cohomology class of pure type (p,p). The conjecture: every such class is a rational combination of the classes of algebraic sub-varieties. Proven for (1,1) — the Lefschetz theorem — and open in general.
02

Walkthrough

Sort the shadows; find the real ones.
WALKTHROUGH
1
The grid on the right is the Hodge diamond: cohomology split by type (p,q). Click the diagonal cell (1,1) — it's green. These classes are known to be algebraic (the Lefschetz (1,1) theorem).
2
Click an off-diagonal cell like (2,0). It's grey: classes of type p≠q are not of Hodge type and generally aren't drawn by sub-varieties. Only the diagonal (p,p) cells host Hodge classes.
3
Press show all cycles and rotate the variety: the loops you can actually trace on the shape are the algebraic cycles. Hodge's claim is that every diagonal (p,p) shadow, however abstract, is a rational blend of such traceable pieces.
4
The conjecture is a bridge from topology (cohomology classes) to algebraic geometry (sub-varieties defined by polynomial equations) — asserting the two notions of "cycle" coincide on Hodge classes.
THE INSIGHT
"Some topological features of a shape feel like they should come from a solid sub-shape carved out by equations. Hodge's conjecture says that for a precise class of them, they always do — that the topology never invents a ghost the algebra can't build."
WHY THIS IS HARD
It is the most abstract of the seven — there is no picture that is the problem, only a claim relating two hard-to-compute worlds. Constructing an algebraic cycle from a cohomology class is like reverse-engineering a solid object from its silhouette; it is proven only in low degrees and special cases, and even the statement took decades to phrase correctly.
03

Explanation

Cohomology, Hodge type, and cycles.
EXPLANATION

On a smooth projective complex variety, the cohomology H*(X) measures its holes and higher-dimensional features. Hodge theory refines this: each cohomology group splits into pieces labelled by a type (p,q), recording how a class interacts with the complex structure.

An algebraic cycle is a formal combination of sub-varieties — sets carved out by polynomial equations. Each casts a cohomology class, and those classes always land on the diagonal, of type (p,p).

A Hodge class is a rational cohomology class that is purely of type (p,p). Every algebraic cycle gives one; the Hodge conjecture asserts the converse — every Hodge class is a rational combination of algebraic cycle classes.

For (1,1) this is the classical Lefschetz theorem. Beyond that it is wide open, with no general method to manufacture the required sub-varieties. It is a cornerstone of how algebraic geometry and topology are believed to fit together.

04

Research note

The conjecture, exactly.
RESEARCH NOTE
This tab is explicitly a visual gloss: the torus and its lit loops illustrate "cycles living inside a variety" and the diamond sorts classes by type. It does not — and honestly cannot — show the conjecture, which is a statement about objects with no faithful low-dimensional picture. That unpicturability is precisely the point.
Go deeper: Clay statement by Pierre Deligne; Hodge's 1950 ICM address; Lefschetz (1,1) theorem; Grothendieck's amendments. The domain-colouring view of complex structure appears in #19.
01

Drive it — flow that rounds a lumpy shape

Draw a wobbly closed loop, then let curvature flow move every point inward in proportion to how sharply the curve bends. Dents fill, bumps flatten, and the blob relaxes into a perfect circle before shrinking to a point. This curve-shortening flow is the 2-D baby cousin of the Ricci flow Perelman used to settle Poincaré in dimension three.
DRIVE IT · $1M · SOLVED 2003
flow time0.00 · roundness (isoperimetric) lumpy
Draw a closed shape with your mouse (or use the preset lump), then press run flow. The roundness meter is the isoperimetric ratio 4πA/L² — it climbs toward 1.000 as the curve becomes a circle. Grady Grayson–Hamilton proved any simple closed curve does this: it stays smooth, rounds, and vanishes.
FLOW SPEED1.0×
ROUNDNESS
0.62
→ 1.000
PERIMETER
shrinks
TOPOLOGY
stays a loop
Ricci flow does the same in 3-D — but there the flow can pinch off "necks", so Perelman added surgery to cut and continue. That control is what proved a simply-connected 3-manifold must be a sphere.
02

Walkthrough

Flow it round, then lift it to 3-D.
WALKTHROUGH
1
Load the lumpy preset and press run flow. Every point drifts inward proportional to curvature: the sharp dents (high curvature) move fastest, so the wobbles erase first.
2
Watch the roundness meter climb toward 1.000. No matter how ugly the start, curve-shortening flow makes it convex, then a perfect circle, then a shrinking point — Grayson's theorem.
3
Try the star or draw your own. As long as the curve doesn't cross itself, the outcome is the same round finale — the flow never develops a bad singularity in 2-D.
4
Ricci flow is the analogue for the geometry of the whole manifold. Perelman ran it in 3-D, controlled the places it wants to pinch with surgery, and proved every simply-connected closed 3-manifold flows to a round sphere — the Poincaré conjecture.
THE INSIGHT
"If a 3-dimensional space is finite, has no holes, and every loop in it can be reeled in to a point, then it is secretly a sphere. Perelman proved it by letting geometry 'flow downhill' until only the sphere could remain — the one Millennium Problem that has been solved."
WHY IT WAS HARD (AND HOW IT FELL)
In 3-D, Ricci flow can concentrate curvature and pinch the manifold into a neck — a singularity the 2-D flow never suffers. Hamilton's programme stalled there for two decades. Perelman's breakthrough was a way to understand and surgically remove those necks while keeping topological control, then show the flow terminates in round spherical pieces. He declined both the Fields Medal (2006) and the $1M prize (2010).
03

Explanation

Simple connectivity, and flowing to a sphere.
EXPLANATION

A space is simply connected if every loop can be continuously shrunk to a point — a sphere is, a doughnut is not (a loop around the hole is stuck). Poincaré asked: is the 3-sphere the only closed, simply-connected 3-manifold, up to deformation? For decades this resisted every direct attack.

The higher-dimensional versions (n≥5, then n=4) were settled first by Smale and Freedman; the original 3-D case was the last and hardest.

Ricci flow, introduced by Hamilton, evolves a manifold's metric so that regions of high curvature smooth out — geometry's answer to the heat equation. Left alone in 3-D it can form singularities; Perelman (2002–03) developed the estimates and surgery to push through them.

The 2-D curve-shortening flow on this stage is the honest low-dimensional shadow: same "smooth by curvature" idea, but simple enough that it always ends in a round point — letting you see the mechanism that, suitably tamed, cracked the conjecture.

04

Research note

The theorem, exactly.
RESEARCH NOTE
The stage runs a real curve-shortening flow (each vertex moves along its curvature vector, with reparametrization for stability); the roundness meter is the live isoperimetric ratio 4πA/L². It is the faithful 2-D analogue of Ricci flow — the mechanism Perelman controlled in 3-D.
Go deeper: Perelman's three arXiv preprints (2002–03); Hamilton's Ricci-flow programme; the Grayson–Gage–Hamilton curve-shortening theorem. The "flow smooths by curvature" idea is exactly #7 heat equation for geometry; loops and holes tie to #20 and #26.
Idea #36 of the maths-visual-teaching series · seven honest teasers, one solved · all objects computed live in the browser · no build step.