This is exactly the kind of answer that should carry its geometry in figures rather than index gymnastics — I'll activate the visualization capability and check its rules before drawing.

That mismatch between "start" and "return" is not an error of transport — it is a fact about the surface, and it comes with a precise exchange rate. Shrink the loop and the rotation shrinks in proportion to the enclosed area, not the perimeter: on a sphere of radius $r$, a tiny loop of area $A$ rotates vectors by $A/r^2$. So the sensible definition of curvature at a point is a density:

$$\text{curvature} \;=\; \lim_{A \to 0}\; \frac{\text{rotation picked up around the loop}}{\text{area enclosed by the loop}}.$$

This is intrinsic (no outside view needed), local (defined point by point), and operational (a transport experiment measures it). Note what it is not: it is not about how the space sits in some bigger space. A cylinder has zero curvature in this sense — unroll it and every loop transports trivially — even though it "looks bent." Curvature is exactly the part of shape that an inhabitant can detect, and no more.

Why the Riemann tensor is the right object

In two dimensions one number per point suffices, because there is only one plane to make loops in and rotations of that plane are measured by one angle. Spacetime has four dimensions, and the definition above forces the bookkeeping to grow in two ways — and only two ways:

• A loop needs an orientation. A small loop is specified by the 2-plane it spans — give two vectors $u, v$ and take the little parallelogram along them. Different planes can carry different curvature (a space can focus east–west loops and not north–south ones).
• The output is not one angle but a rotation. In more than two dimensions, "how transported vectors come back changed" is a linear map: each vector $w$ acquires a small change $\delta w$, linearly in $w$, and metric-compatibility of transport makes this map an infinitesimal rotation.

Put the pieces together and curvature at a point must be a machine that eats a loop plane $(u,v)$ and a vector $w$, and returns the per-unit-area change of $w$:

$$R(u,v)\,w \;=\; \lim_{A\to 0}\frac{1}{A}\Big(w_{\text{after loop}} - w\Big), \qquad \text{equivalently} \qquad R(u,v) = [\nabla_u, \nabla_v] - \nabla_{[u,v]},$$

where the commutator form says the same thing infinitesimally: go a step along $u$ then $v$, versus $v$ then $u$, and $R$ measures the failure of the two orders to agree. It is linear in every slot — double the loop, double the rotation — and that linearity is what makes it a tensor: a pointwise, coordinate-free object rather than a property of any coordinate grid.

Its famous symmetries are not algebraic accidents; each is a sentence about loops:

• $R(u,v) = -R(v,u)$: traversing the loop backwards undoes the rotation.
• $\langle R(u,v)w, z\rangle = -\langle R(u,v)z, w\rangle$: the output is a genuine rotation — transport never stretches, because it preserves the metric.
• The first Bianchi identity $R(u,v)w + R(v,w)u + R(w,u)v = 0$: the three parallelograms you can build from three directions have holonomies that cancel in a cyclic sum — a closure condition inherited from the connection having no torsion.

These symmetries cut the naive $4^4 = 256$ components down to $20$ independent ones in four dimensions. And here is the theorem that certifies Riemann as the right object, exactly — neither too small nor too large:

$R \equiv 0$ on a region if and only if the region is flat — there exist coordinates in which the metric is exactly the flat metric, and parallel transport is path-independent.

Nothing less would do: every nonzero component of $R$ is detectable by some small loop acting on some vector, so you cannot throw any of it away. Nothing more is needed: any big loop can be tiled by small parallelograms, and its holonomy is stitched together from theirs (a curved-space Stokes argument) — so the infinitesimal loops already generate all path dependence there is. The Riemann tensor is the complete local answer to "how does this space fail to be flat," with zero redundancy.

Curvature you can feel: the tide

Now move from surfaces to spacetime, where the "paths" that matter are worldlines and the straightest ones — geodesics, $\nabla_u u = 0$ — are the trajectories of free fall. Einstein's equivalence principle sharpens what curvature has to mean physically: inside a small falling laboratory, gravity is undetectable — you float, light travels straight, physics is special relativity. A single geodesic carries no local trace of gravity at all. What cannot be transformed away is the comparison of two nearby free-falls: your head and your feet, two floating dust grains, the near and far sides of the ocean. Release two test particles, mutually at rest, and watch the separation vector $\xi$ between them. The answer is the loop machine again, fed one timelike direction:

$$\frac{D^2 \xi}{d\tau^2} \;=\; -\,R(\xi, u)\,u ,$$

the geodesic deviation (Jacobi) equation: the relative acceleration of straight lines is a linear function of their separation, and the linear map doing it is the Riemann tensor with the 4-velocity $u$ in its loop slots. Curvature is the tidal field — and its sign is something you can watch:

The division of labor is the key to the whole theory: matter will get to dictate the Ricci half, and only the Ricci half. The Weyl half is the gravitational field's own degrees of freedom — it is why there is gravity in empty space at all. Outside the Earth, spacetime is vacuum: Ricci vanishes identically there, yet the Moon still orbits and the oceans still bulge, because Weyl curvature — sourced by the Earth's matter elsewhere, propagated through the vacuum — fills the region. A gravitational wave is the same thing untethered: a ripple of pure Weyl curvature, shearing LIGO's ring of mirrors at constant volume, exactly the right-hand panel above, oscillating.

The Einstein equation: matter focuses volume

Now we can build the field equation the way Einstein had to: by asking what equation could stand between geometry and matter. Newton's law in local form is $\nabla^2 \Phi = 4\pi G \rho$ — and $\nabla^2\Phi$ is precisely the trace of the tidal matrix. Newton's gravity constrains only the volume-changing part of the tide and ties it to density. The relativistic translation writes itself: set the trace of the tidal map — Ricci — proportional to the energy–momentum tensor $T$ (the object that packages density, pressure, and momentum flux, with $T(u,u) =$ energy density seen by observer $u$):

$$R_{\mu\nu} \overset{?}{=} 8\pi G\, T_{\mu\nu}.$$

This was Einstein's actual first guess, and it fails for a geometric reason worth savoring. Energy–momentum is conserved — $\nabla \cdot T = 0$, always, as an identity of matter dynamics. So the geometric side must be identically divergence-free too, or the equation would impose extra, bogus constraints on matter. Ricci alone is not: the contracted Bianchi identity gives $\nabla\cdot\mathrm{Ric} = \tfrac{1}{2}\nabla R$, generally nonzero. But the same identity hands you the unique repair — subtract half the trace:

$$G_{\mu\nu} \;=\; R_{\mu\nu} - \tfrac{1}{2}R\, g_{\mu\nu}, \qquad \nabla^\mu G_{\mu\nu} = 0 \;\;\text{identically},$$

and the field equation becomes

$$\boxed{\,G_{\mu\nu} + \Lambda g_{\mu\nu} = 8\pi G\, T_{\mu\nu}.}$$

The second Bianchi identity behind this is itself a loop statement — Wheeler's slogan is "the boundary of a boundary is zero": transport around the six faces of an infinitesimal cube, and the face holonomies cancel in pairs. So energy–momentum conservation is not an assumption bolted onto general relativity; it is enforced by the geometry of loops. The same cancellation is why matter must move on geodesics — the equations of motion are already contained in the field equation. And this repair is not merely one option: Lovelock's theorem says $G_{\mu\nu} + \Lambda g_{\mu\nu}$ is the only symmetric, divergence-free tensor built from the metric and its first two derivatives in four dimensions. Given the geometric setting, Einstein's equation is essentially forced.

What it actually says, geometrically. Take a small ball of test particles, initially at rest relative to each other, and let it fall. Written for the volume of that ball, the full content of Einstein's equation is one sentence (the Baez–Bunn formulation):

$$\left.\frac{\ddot V}{V}\right|_{t=0} \;=\; -\,4\pi G\,\big(\rho + p_x + p_y + p_z\big)$$

— the volume of a free-falling ball of test particles accelerates inward in proportion to the energy density plus the pressure in each direction at its center. That's it. That single statement, demanded to hold for every observer at every point, is equivalent to the full tensor equation. Everything else in the theory is this plus the Bianchi bookkeeping. And its fine print carries most of modern gravity:

• Ordinary matter: pressure is tiny compared to $\rho$ ($p \sim \rho v^2/c^2$), the statement reduces to Newton's $\nabla^2\Phi = 4\pi G\rho$ — the correspondence check.
• Pressure gravitates. In a massive star, the central pressure that resists collapse also sources focusing — squeezing harder makes gravity stronger. That feedback is why sufficiently massive stars cannot find equilibrium: collapse to a black hole is a prediction of the $+3p$, not an exotic add-on.
• $\rho + 3p < 0$ means repulsion. A medium with large negative pressure — the cosmological constant has $p = -\rho$ — makes the ball's volume accelerate outward: dark energy and the accelerating universe live entirely in the sign of that trace.
• Light bends twice as much as Newton says because light's "pressure" terms are as large as its energy density — the factor of 2 that made Einstein famous in 1919 is the pressure contribution to focusing.
• Vacuum is not trivial: the equation only pins Ricci to zero there. Weyl remains free, constrained across regions by Bianchi — which is exactly the propagation law that lets tides reach across empty space and lets gravitational waves carry curvature to a detector a billion light-years from the collision that made it.

The whole chain, once

$$\text{metric} \;\Rightarrow\; \text{parallel transport} \;\Rightarrow\; \text{holonomy of small loops} \;=\; \text{Riemann} \;\Rightarrow\; \underbrace{\text{Ricci}}_{\text{volume of the tide}} + \underbrace{\text{Weyl}}_{\text{shape of the tide}},$$

with Einstein's equation pinning the volume part to the local energy–momentum, and the Bianchi identity — "loops around a cube cancel" — simultaneously guaranteeing conservation, forcing geodesic motion, and propagating the free Weyl part through vacuum. No step in that chain is index manipulation; each is a statement about carrying directions around loops and watching families of free falls. The indices are just the ledger.

Before we go further — a quick check on whether the load-bearing pictures actually landed, so I know what to reinforce next:

All three "not sure"s are the honest result, and genuinely the most useful one: there's no wrong model to dismantle, just ground that didn't get poured before the essay built on it. So let me rebuild the three foundations slowly, each anchored to something you can do with your hands — and then check the floor again.

1. Walk the loop yourself: why area

Forget formulas; take one concrete walk on the Earth (treat it as a perfect sphere). Start on the equator carrying an arrow that points north. The transport rule is just: never twist the arrow relative to the ground as you walk.

1. Walk north along your meridian to the North Pole. Your arrow points along your direction of travel the whole way — it arrives at the pole pointing "ahead."
2. Turn right $90^\circ$ and walk south along a new meridian. You turned; the arrow didn't — that's the whole discipline. So it now points $90^\circ$ to your left: east, all the way down to the equator.
3. Walk west along the equator back to your starting point. Again the arrow keeps its bearing: still east.

You are home, and the arrow that left pointing north has returned pointing east: rotated by exactly $90^\circ$, though you never once twisted it. Now the diagnostic question — what did that $90^\circ$ measure? Check against the two candidates:

• Area: your loop enclosed one octant, $\tfrac{1}{8}$ of the sphere. On a unit sphere that's an area of $\tfrac{4\pi}{8} = \tfrac{\pi}{2}$ — and $90^\circ = \tfrac{\pi}{2}$ radians. Exact match.
• Perimeter: now squeeze the loop into a skinny wedge — go up the meridian, come down another meridian only $1^\circ$ of longitude away, walk the short equator arc home. The perimeter is essentially unchanged (still two full pole-to-equator legs, ~10,000 km each), but the enclosed area has collapsed — and the arrow now returns rotated by only $1^\circ$. Turn tracks the wedge angle, i.e. the area; the perimeter barely moved.

So rotation is bought with enclosed area, at a rate — degrees per square meter, so to speak — and that rate is what curvature is. But there's a deeper reason it had to be area, and it's the single picture that also underwrites the Einstein equation later, so it's worth one figure:

This cancellation is why "rotation per unit area" is a well-posed local quantity: any big loop's rotation is just the sum of tiny-cell contributions, each cell contributing its own local rate times its own little area. A long skinny loop rotates almost nothing because it contains almost nothing — perimeter never enters. And notice for later: "everything interior cancels in pairs, only the boundary survives" is a purely combinatorial fact about walking edges — we will meet its 3D version at the very end, where it becomes the reason energy is conserved.

2. What the Riemann tensor is when you stand inside it

Now the second foundation, and let me put it in the first person. You are floating in a space station — a sealed can, no windows. Can you tell whether you're in deep space or falling around the Earth? Not by any experiment at a single point: in free fall, gravity is locally erased; that's the equivalence principle. But let out two ball bearings, one near the floor and one near the ceiling, both perfectly at rest relative to you, and wait. In orbit, they slowly drift apart (the lower one is on a slightly faster orbit, the geometry stretches the pair along the up–down line) while two balls released side by side drift together. In deep space, nothing happens.

That drift — relative acceleration between neighboring free-fall paths, per unit separation — is the Riemann tensor, measured. Three things follow, and they were exactly the three true statements in the checkpoint:

• It's locally detectable in free fall. The equivalence principle erases the pull (you can't feel which way is down), but it cannot erase the stretch between two nearby points — that differential effect is the invariant, twist-proof content, the same thing the loop rotation measures.
• It's nonzero in vacuum. The space station's interior, and the whole region around the Earth, is empty — $T_{\mu\nu}=0$ right where you float. Yet the ball bearings drift. The curvature you're detecting was sourced by the Earth's matter, which is somewhere else, and the field carries it across the empty region — the same way the electric field fills the vacuum around a charge. Vacuum means no source here, never no field here.
• Zero Riemann means genuinely flat. If every such experiment — every loop transported, every pair of ball bearings, everywhere in a region — comes back with exactly nothing, there is no residual "hidden" curvature: coordinates exist making the region exactly the spacetime of special relativity. Riemann is the complete inventory of detectable shape; when it's zero there is nothing left to detect.

The refinement from last time, now easy to place: only part of the ball-bearing drift pattern requires matter at your location. The part that changes the volume of a whole cloud of bearings (Ricci) exists only inside matter; the stretch-one-way-squeeze-the-other pattern at constant volume (Weyl) is what fills the vacuum around the Earth — it's what you were feeling in the station.

3. Why the equation needed fixing: a ledger that must balance

Last foundation. The energy–momentum tensor $T$ comes with a non-negotiable bookkeeping law, $\nabla \cdot T = 0$. Concretely it is a continuity equation, like charge conservation: energy and momentum in any small box change only by flowing through the box's walls — never created, never destroyed in place. Every kind of matter obeys this automatically, as part of its own equations of motion.

Now suppose you propose any equation of the form $$\big(\text{something built from geometry}\big) = 8\pi G\, T.$$ Take the divergence of both sides. The right side gives zero — always, for any matter doing anything legal. So the left side must give zero identically: as a mathematical identity true for every geometry, not as an extra condition. If it doesn't, your equation quietly legislates new laws of physics for matter that matter never agreed to. That's precisely what kills Einstein's first guess $R_{\mu\nu} = 8\pi G\,T_{\mu\nu}$: Ricci's divergence is $\tfrac12 \nabla R$, not identically zero, and chasing the consequences forces absurdities — for instance that the trace of $T$ must be the same constant everywhere, so a star could not sit in empty space. The universe as observed is illegal under the guessed equation.

The rescue is that geometry happens to own one identity of exactly the needed kind — the Bianchi identity, which is the figure above promoted one dimension: walk loops around all six faces of a tiny cube, and every edge is shared by two faces and walked in opposite directions, so everything cancels. That cube cancellation, traced through, says the particular combination $$G_{\mu\nu} = R_{\mu\nu} - \tfrac12 R\, g_{\mu\nu}$$ has identically zero divergence — for every conceivable metric, before any physics is imposed. It is the only such combination available (Lovelock's theorem), so the field equation $G_{\mu\nu} = 8\pi G\, T_{\mu\nu}$ isn't one candidate among many; it's the only equation whose geometry side balances the same ledger matter is already bound to. Conservation of energy isn't an input to general relativity — it falls out of "interior edges cancel."

That's the floor rebuilt: (1) rotation is bought with area because holonomy adds over tiles, (2) Riemann is the ball-bearing drift you can measure inside a windowless falling can, nonzero even in vacuum, and (3) the Einstein tensor is the unique geometric quantity whose books balance identically, matching matter's conservation ledger. Since this was a substantial repour, one short check that it set — same ideas, no tricks:

Give these three a try — this time you've walked every idea they ask about.

The last checkpoint left Ricci and Weyl as names you were not ready to bet on. That is the right place to start a black hole. Schwarzschild spacetime is the existence proof that those two halves of Riemann are not the same thing: the vacuum around the hole is Ricci-flat — Einstein’s equation with no matter — and yet it is vividly curved. Every tidal stretch in the story is Weyl. The event horizon, for all its causal drama, is not where that curvature runs away.

A vacuum solution of Einstein’s equation is the statement

$$R_{\mu\nu}=0\qquad\text{for all }r>0.$$

Ricci is the trace of Riemann, the piece that changes the volume of a small cloud of freely falling particles. The Einstein equation sets that trace equal to the local energy-momentum. No matter, no Ricci. What remains of Riemann is the traceless piece — the Weyl tensor $C_{\rho\sigma\mu\nu}$ — which shears the cloud, stretching it on one axis and squeezing it on the others, at leading order without changing its volume. In Schwarzschild geometry Weyl is the whole of the curvature, and its strength is captured by a single invariant, the Kretschmann scalar

$$K=R_{\alpha\beta\gamma\delta}R^{\alpha\beta\gamma\delta}=\frac{48M^{2}}{r^{6}}.$$

At the horizon $r=2M$ this is $K=3/(4M^{4})$ — finite. As $r\to 0$ it diverges as $1/r^{6}$. The metric component $g_{rr}=(1-2M/r)^{-1}$ in Schwarzschild coordinates does the opposite: it blows up at $r=2M$ and is perfectly finite as you approach the singularity. That pole is a bad choice of coordinates, not a geometric wall. The next picture plots those two behaviours on a shared radius, then converts the invariant at the marked point into a tidal acceleration you could actually feel.

That is the whole local story of the horizon: every scalar you can build from the Riemann tensor is finite there, and in coordinates adapted to an infalling observer (Eddington–Finkelstein, Kruskal–Szekeres, Gullstrand–Painlevé) the metric itself is smooth. What is special at $r=2M$ is causal structure, not curvature. The surface is null. Outgoing light rays sit on it forever; everything inside has $r$ as a time coordinate and is condemned to smaller $r$. A distant observer using Schwarzschild $t$ never sees you cross, because $t$ diverges at the horizon — that is their bookkeeping, not your proper time. A radial “raindrop” geodesic (dropped from rest at infinity) reaches the horizon in finite proper time and then has only $\tau=4M/3$ left until $r=0$. For the Sun that remainder is about seven microseconds; for Sagittarius A* it is about half a minute.

The equivalence principle is what makes the crossing uneventful. A sufficiently small freely falling lab is Minkowski to first order everywhere the curvature is finite. Second order is the tidal field, the geodesic deviation

$$\frac{D^{2}\xi^{\mu}}{d\tau^{2}}=-R^{\mu}{}_{\nu\rho\sigma}u^{\nu}\xi^{\rho}u^{\sigma}.$$

In an orthonormal frame that is the Weyl pattern of eigenvalues $(2,-1,-1)$ times $M/r^{3}$: stretch along the radial direction, squeeze in the two angular directions, volume unchanged at this order because $R_{\mu\nu}u^{\mu}u^{\nu}=0$. The physical acceleration across a body of size $L$ therefore scales as $c^{4}L/(G^{2}M^{2})$ at the horizon — large holes are gentler. Below roughly $10^{5}$ solar masses a two-metre body already feels a crushing tide at $r=2M$; above that, and certainly at Sgr A* or M87, the horizon is a quiet place and the Weyl shear only becomes a death warrant deeper in. In every case the field is continuous. There is no bump, no infinite force, no local invariant that jumps. You do not “hit” the horizon. You pass a one-way causal membrane whose location cannot even be detected by a local curvature measurement.

That volume being pinned at $1$ is a statement about a moment, not about a lifetime. Ricci is the instantaneous trace of geodesic deviation — the second derivative of the volume of a small ball. Weyl is the shear. Once shear has been acting for a while it feeds the Raychaudhuri equation

$$\frac{d\theta}{d\tau}=-\frac13\theta^{2}-\sigma_{\mu\nu}\sigma^{\mu\nu}-R_{\mu\nu}u^{\mu}u^{\nu},$$

and even with $R_{\mu\nu}=0$ the $\sigma^{2}$ term is nonnegative, so a sheared congruence still focuses. Near $r=0$ the cloud is drawn into a line: the volume eventually vanishes because Weyl has been stretching it, not because Ricci ever turned on. Spaghettification is empty-space curvature doing what the Einstein equation allows empty space to do.

So where is Ricci in a real collapse? Inside the star. The Oppenheimer–Snyder dust ball is a piece of closed Friedmann geometry glued onto Schwarzschild. Homogeneous isotropic dust has Weyl $=0$ by symmetry — the only curvature is Ricci, sourced by $T_{\mu\nu}=\rho u_{\mu}u_{\nu}$ — while the vacuum exterior has Ricci $=0$ and all the Weyl. The two halves of Riemann live in two different regions, and they meet at the star’s surface. As the star falls through its gravitational radius the Ricci region is swallowed whole; after the crunch the manifold $r>0$ is pure Weyl, and the matter that sourced Ricci has been crushed into the singularity where the geometry itself ends.

The horizon’s growth from the centre is the same moral in a different tense. An event horizon is defined by who can still send light to infinity, which depends on the whole future. It is allowed to sweep through ordinary matter without any local curvature invariant noticing. That is why “where is the horizon right now in my lab?” is not a question a Riemann measurement can answer.

A rotating hole does not change the split. Kerr is vacuum, so Ricci vanishes identically outside the ring singularity; Weyl does all the work again, now with frame-dragging. The ring is where Kretschmann diverges. The inner (Cauchy) horizon of the exact Kerr solution looks smooth, but that smoothness is unstable: infalling radiation is infinitely blueshifted there, and the generic result is a curvature blow-up — again Weyl, sourced by the shear and energy of that radiation rather than by a local $T_{\mu\nu}$ sitting on the horizon. Charged holes (Reissner–Nordström) are the one common case where Ricci is not zero: the Maxwell field has stress-energy, so $R_{\mu\nu}\neq 0$ even though the Ricci scalar still vanishes. Both halves then diverge at $r=0$. The astrophysical case is Kerr-like vacuum, which is to say: Weyl.

What an infalling observer actually meets, in one list:

• At the horizon. Ricci is zero (vacuum). Weyl is finite, of order $1/M^{2}$. The metric is smooth in infalling coordinates. Proper time is finite. No local experiment detects a wall. A hovering observer does feel something — they are accelerating, with surface gravity $\kappa=1/(4M)$ — but that is Unruh/acceleration physics in a static frame, not a property of the geometry the geodesic sees. Semiclassically the horizon is special for quantum fields in the exterior vacuum; classically, and for a free-faller, it is not a curvature event.
• Between horizon and singularity. Still Ricci-flat. Weyl grows as $M/r^{3}$. Shear accumulates. For a supermassive hole this is where the story stops being gentle.
• At the singularity. The manifold ends. Kretschmann $\to\infty$. Ricci never “diverges on the manifold” in Schwarzschild because Ricci is identically zero wherever the manifold exists; the singularity is a Weyl singularity, reached in finite proper time, after which there is no further spacetime to extend into.

The slogan from the parent thread now has a spacetime to live in. Riemann vanishes throughout a region if and only if that region is flat. Ricci can vanish without flatness — the entire exterior and interior of a vacuum black hole is the example. Weyl is what empty space uses to be curved, what you feel as a tide, and what runs away at $r=0$. The horizon is a one-way causal surface drawn on that empty curved geometry, and it is allowed to be geometrically boring.

Give these three a try — they are the horizon/singularity split this turn was built to make load-bearing.

This is the derivation you asked for, run end to end: mode bases and the Bogoliubov transformation, the ray-tracing map that collapse imposes on the outgoing modes, the two Gamma-function integrals that produce $|\alpha_{\omega\omega'}| = e^{\pi\omega/\kappa}|\beta_{\omega\omega'}|$, and the wave-packet step that turns divergent coefficients into a steady rate. The pivot is section 04: thermality is not put in anywhere, it is the ratio of two analytic continuations of one function around a branch point at the last escaping ray — and the only thing the geometry contributes is that the map from affine parameter to Killing time on the horizon is exponential, at the constant rate $\kappa$. Sections 06–07 then restore the curvature barrier and show exactly where $\Gamma_{s\ell}(\omega)$ multiplies into $|\beta|^2$, what it does to the infrared, and why the high-frequency hole radiates from $27\pi M^2$ rather than the horizon area.

Three figures carry the argument: the exponential peeling and its logarithm, the $p$-plane arc where the Boltzmann factor is generated, and an interactive greybody panel built on the Boonserm–Visser rigorous bound (the curves there are a bound, not a numerical Regge–Wheeler solution — the caption says so, and section 07 uses that honestly to squeeze the true answer between two verified results). Two collapsible checkpoints sit after sections 04 and 07 if you want to test the load-bearing steps rather than read past them.

If you want, I can extend this to Kerr, where the same map runs with $\kappa=(r_+-M)/(r_+^2+a^2)$ and $\Gamma$ turns negative for co-rotating modes.

Why 1/3 and 2/3 are everyone's exponents

Your question has two halves that are often blurred together, and it's worth keeping them apart from the start. The first half — why a lattice deposition model and a stochastic PDE share the exponents $1/3$ and $2/3$ — is a question about a universality class: the claim is that these numbers are fixed by two structural facts (a symmetry and an invariant measure) that survive coarse-graining, so any model possessing them lands on the same large-scale statistics no matter what its microscopic rules are. The second half — Hairer versus Bertini–Giacomin — is a question about the KPZ equation itself, which is a genuinely ill-posed object: Bertini–Giacomin found a way to name a solution without ever making sense of the equation, and Hairer made the equation itself meaningful, with all the robustness that buys. Let me take them in order.

What the two numbers measure

Consider a one-dimensional growing interface $h(x,t)$ — the top of a deposition cluster, a burning front, the edge of a bacterial colony. Started flat, it develops fluctuations, and the KPZ class makes two scaling statements:

$$\delta h(t) \;\sim\; t^{1/3}, \qquad \ell(t) \;\sim\; t^{2/3}.$$

The first says the size of the height fluctuations around the deterministically growing mean is of order $t^{1/3}$ — much slower than the $t^{1/2}$ a sum of independent contributions would give, which is already a signal that something nontrivial (strong correlation buildup) is going on. The second says the lateral distance over which the interface knows about itself — the correlation length — grows like $t^{2/3}$. Equivalently, the dynamic exponent is $z = 3/2$: information propagates superdiffusively along the interface. The two are tied together by a third exponent, the roughness $\alpha = 1/2$ (a saturated interface of length $L$ has width $L^{1/2}$, i.e. it looks like a Brownian path), through the identity $\beta = \alpha/z$: fluctuations at time $t$ are those of a Brownian path viewed over the correlated stretch, $\delta h \sim \ell(t)^{1/2} \sim t^{2/3 \cdot 1/2} = t^{1/3}$.

So the question "why are $1/3$ and $2/3$ universal" is really the question "why are $\alpha = 1/2$ and $z = 3/2$ universal" — and the answer comes in two steps: first, why every model in the class is described by one equation; second, why that equation's exponents are pinned by symmetry rather than by any coefficient you could tune.

Step one: coarse-graining forces the equation on you

Take any local growth model — ballistic deposition, an asymmetric exclusion process viewed as a height function, Eden growth, polynuclear growth. Zoom out, and the evolution of the smoothed height must take the form

$$\partial_t h \;=\; F(\partial_x h,\, \partial_x^2 h,\, \dots) \;+\; \text{noise},$$

by locality alone. Now expand $F$ and throw away everything you're allowed to:

• A constant (overall growth speed): removed by moving to the co-moving frame.
• A term $\propto \partial_x h$: removed by a tilt of coordinates.
• $\nu\, \partial_x^2 h$: surface tension / smoothing — generically present.
• $\tfrac{\lambda}{2} (\partial_x h)^2$: the first term that cannot be transformed away whenever the growth speed depends on the local slope. And it generically does: a surface growing at speed $v$ normal to itself rises vertically at $v\sqrt{1+(\partial_x h)^2} \approx v + \tfrac{v}{2}(\partial_x h)^2$. That's exactly where lattice deposition models get their $\lambda$ — an overhanging tilted region catches more incoming particles.
• The noise: microscopically it's whatever the model does, but it is local and has short-range correlations, so by the central limit theorem it coarse-grains to space–time white noise $\xi$.

Everything higher order — $(\partial_x h)^3$, $\partial_x^4 h$, gradient–curvature couplings — is irrelevant in the renormalization-group sense: under the rescaling $h \mapsto b^{-\alpha} h(b x, b^z t)$ that keeps the large-scale interface of order one, those terms flow to zero. What survives is precisely

$$\partial_t h \;=\; \nu\, \partial_x^2 h \;+\; \tfrac{\lambda}{2} (\partial_x h)^2 \;+\; \sqrt{D}\,\xi,$$

the KPZ equation (Kardar–Parisi–Zhang, 1986). This is the sense in which the equation is not "one more model" but the normal form of the entire class: microscopic details survive only as the three numbers $(\nu, \lambda, D)$ — and, as we'll see, not even those matter for the exponents.

Step two: two rigid facts pin the exponents

Here is the part that makes $1/3$ and $2/3$ feel inevitable rather than accidental. In one dimension, the KPZ equation has two nonperturbative properties, and each one nails an exponent relation exactly:

Fact 1: the stationary state is Brownian, so $\alpha = 1/2$. In $d=1$ there is a small miracle: white noise in the slope variable $u = \partial_x h$ is an invariant measure not just for the linear (Edwards–Wilkinson) equation but for the full nonlinear one — the contribution of the $(\partial_x h)^2$ term to the evolution of that measure cancels identically. So the stationary interface is, statistically, a Brownian path, and Brownian paths have $\alpha = 1/2$, full stop. No coupling constant can change this. (Microscopic counterpart: the stationary measure of ASEP is product Bernoulli, whose height function is a random walk — the same statement one level down. That parallel is not a coincidence; it's the invariant measure surviving the scaling limit.)

Fact 2: tilt (Galilean) invariance, so $\alpha + z = 2$. The transformation $h(x,t) \mapsto h(x - \lambda\varepsilon t,\, t) + \varepsilon x$ maps solutions to solutions: tilting the interface just makes it drift sideways, with the same coefficient $\lambda$ coupling the two. Because the symmetry locks $\lambda$ to the space–time reparametrization, $\lambda$ cannot be renormalized — and demanding that the tilt symmetry survive the rescaling $h \to b^{-\alpha}h(bx, b^z t)$ forces the exponent identity

$$\alpha + z = 2.$$

Put the two together: $z = 2 - \tfrac12 = \tfrac32$, hence $\ell(t) \sim t^{1/z} = t^{2/3}$, and $\beta = \alpha/z = \tfrac{1/2}{3/2} = \tfrac13$.

This is the honest answer to "why should a lattice model and an SPDE agree": both exponents are consequences of a symmetry and an invariant measure, and those are exactly the kind of structure that coarse-graining preserves. A microscopic model doesn't need to resemble the equation; it needs a slope-dependent growth rate (any $\lambda \neq 0$), local smoothing, and local noise. Everything else washes out. Contrast $\lambda = 0$: you lose Fact 2, and Edwards–Wilkinson scaling takes over ($z = 2$, $\beta = 1/4$) — so the exponents jump discontinuously the moment the nonlinearity appears, which is what "different universality class" means operationally.

Two caveats worth having straight. First, this RG-style argument is a physicist's derivation; turning "irrelevant terms vanish" into theorems is precisely the hard mathematics. The rigorous confirmations came from a different direction — exactly solvable models: Baik–Deift–Johansson (1999) for the longest increasing subsequence and Johansson (2000) for TASEP extracted the $t^{1/3}$ fluctuations and more: the full limiting distribution is Tracy–Widom (GUE or GOE depending on geometry — curved vs. flat initial data), the same law as the largest eigenvalue of a random matrix. Universality in this class goes beyond two exponents to entire distributions and processes, and it has even been measured in the lab, in turbulent liquid-crystal interfaces (Takeuchi–Sano, 2010), with Tracy–Widom fits to the data. Second, the full "strong universality" statement — every model in the class converges under $1{:}2{:}3$ scaling to a single limit object, the KPZ fixed point (constructed by Matetski–Quastel–Remenik, 2018; equivalently the directed landscape of Dauvergne–Ortmann–Virág) — is proven for a growing but still special family of models. The exponents are the settled part; the sweeping process-level statement is a live frontier.

Bertini–Giacomin: naming a solution without an equation

Now the second half. The KPZ equation, taken literally, is nonsense. Its solution should locally look like the stationary state — Brownian in $x$, so Hölder continuous of exponent just below $1/2$. Then $\partial_x h$ is not a function but a distribution, and $(\partial_x h)^2$ is the square of a distribution: undefined. If you regularize the noise at scale $\varepsilon$ and compute, the nonlinear term blows up like $\varepsilon^{-1}$ — the equation only has a chance to mean something as

$$\partial_t h = \partial_x^2 h + \tfrac12\left[(\partial_x h)^2 - \infty\right] + \xi,$$

with an infinite subtraction. In 1997 that was far beyond any solution theory.

Bertini and Giacomin sidestepped it. The Cole–Hopf transform $Z = e^{(\lambda/2\nu) h}$ formally converts KPZ into the multiplicative stochastic heat equation

$$\partial_t Z = \partial_x^2 Z + Z\,\xi,$$

which is linear in $Z$ and perfectly well-posed via Itô calculus. So they defined the solution of KPZ to be $h := \tfrac{2\nu}{\lambda}\log Z$ — the "Cole–Hopf solution" — and, crucially, proved a universality statement: the height function of the weakly asymmetric simple exclusion process converges to exactly this object. That result used a second, independent miracle: Gärtner's observation that ASEP admits a discrete Cole–Hopf transform — an exponential of the microscopic height that satisfies a discrete SHE.

This was a landmark, but look at what it dodges. The object $h$ is never shown to satisfy any equation; there is no sense in which you can plug it back in. The construction is rigid in two ways at once: it works only for the exact KPZ nonlinearity with coefficients precisely matched to the Laplacian (the transform linearizes nothing else), and the convergence proof works only for microscopic models that happen to have their own exponential linearization (ASEP essentially uniquely, at the time). If you perturb either side — a slightly different nonlinearity, a deposition model without Gärtner's identity, a different regularization of the noise — the method gives you nothing. The "solution" is a name attached to a formula, and the physicists' derivation of the equation (the counterterm, the mollification, the claim that details don't matter) remains uninterpreted mathematics.

Hairer: making the equation mean something

What Hairer did (2013, "Solving the KPZ equation," and then the general theory of regularity structures, 2014) is give the equation an intrinsic, robust solution theory:

The well-posedness statement. Mollify the noise at scale $\varepsilon$, and solve the classical PDE

$$\partial_t h_\varepsilon = \partial_x^2 h_\varepsilon + \tfrac12 (\partial_x h_\varepsilon)^2 - C_\varepsilon + \xi_\varepsilon,$$

with an explicit diverging constant $C_\varepsilon \sim \varepsilon^{-1}$. Then $h_\varepsilon$ converges as $\varepsilon \to 0$, the limit does not depend on the choice of mollifier, and it coincides with the Cole–Hopf solution. The infinite subtraction the physicists wrote down informally is now a theorem, with the divergence structure made exact — and the agreement with Cole–Hopf certifies, after the fact, that Bertini–Giacomin had named the right object.

The mechanism. The obstruction was that classical analysis describes functions locally by Taylor polynomials, and no polynomial description survives at Brownian regularity. Hairer's move: describe the solution locally in terms of a small, explicit basket of stochastic objects built from the noise itself (the heat-kernel integral of $\xi$, its derivative squared, a few iterated combinations) — a generalized Taylor expansion whose "monomials" are random distributions. Given this enhanced noise (the model), the equation becomes a deterministic fixed-point problem in a space of locally-described functions, solved by ordinary contraction arguments; all the probabilistic difficulty is quarantined in constructing finitely many objects, and renormalization becomes a finite-dimensional group acting on them — which is where $C_\varepsilon$ comes from, canonically rather than by guesswork.

Why it matters beyond aesthetics. Three concrete payoffs. First, robustness: the solution map is continuous in the model, so approximation-independence is automatic — any reasonable regularization converges to the same limit, which is itself a universality statement Cole–Hopf could never express. Second, it unlocked universality proofs outside the Cole–Hopf world: Hairer–Quastel (2018) showed that weakly asymmetric growth models with an arbitrary even-polynomial slope-dependence — no quadratic structure, no Gärtner identity, no linearization anywhere — still converge to KPZ, with a renormalized effective $\lambda$ that picks up contributions from all the higher terms. That is the coarse-graining story of step one, finally as a theorem. Third, the machinery was never KPZ-specific: the same framework resolves a whole family of previously ill-posed SPDEs ($\Phi^4_3$, the parabolic Anderson model in 2d/3d, and more), which is why it earned the 2014 Fields Medal — KPZ was the flagship application, not the boundary. (In parallel, Gubinelli–Imkeller–Perkowski's paracontrolled calculus gives an alternative route to the same well-posedness, and the Gonçalves–Jara "energy solution" framework — uniqueness by Gubinelli–Perkowski, 2018 — handles stationary initial data and yet more microscopic models; the equation now has several interlocking solution theories where in 2010 it had none.)

The clean summary

One final distinction ties both halves together. The KPZ equation and the KPZ fixed point are different objects. The exponents $1/3$ and $2/3$ belong to the fixed point — the universal large-time limit, where Tracy–Widom lives. The equation is itself a member of the class, exhibiting those exponents only asymptotically (it crosses over from Edwards–Wilkinson behavior at short times); its distinguished role is as the weak-universality limit — the one canonical object every weakly asymmetric model passes through on its way to the fixed point. Bertini–Giacomin proved the first instance of that weak universality by transform tricks specific to one model and one nonlinearity; Hairer built the solution theory that makes the equation a well-defined mathematical object with a continuous, approximation-independent solution map — turning "the details don't matter" from the physicists' article of faith into the actual content of the theorem. And the reason a lattice gas and an SPDE share $1/3$ and $2/3$ in the first place is that neither number was ever theirs to choose: one is the roughness of the Brownian invariant measure, the other is forced by tilt invariance through $\alpha + z = 2$, and both structures are exactly what survives when you zoom out.

Two last checks — the elliptic-curve side of Shor's cost, and a first-party source on the signature-size problem.

This traces one causal chain end to end. SIKE was the small-key candidate — 676 bytes on the wire, less than a third of what we actually deploy today — and it was advanced to a fourth NIST evaluation round on 5 July 2022 and killed by a Magma script twenty-five days later. The report covers the break itself (Kani's 1997 reducibility criterion turned into a decision oracle, then Robert's dimension-8 generalisation that closed every escape hatch), what NIST ended up standardising instead, the exact byte cost of that choice at the TLS layer, and what the episode should change about how you read a security argument.

Load-bearing throughout are the byte counts, which come from the designers' own parameter tables and RFC 10024, and the arithmetic built on them — every derived sum is labelled as derived. Sections 03 and 04 carry the mathematics and are the densest part; the two collapsible checkpoints test whether the structural argument landed rather than whether you remember the numbers. Section 09 is the part that generalises: the failure was not a proof error, it was a bespoke assumption with published auxiliary data, and that shape is still deployed elsewhere.

The two checkpoints are collapsed by default — open them if you want to pressure-test the structural argument rather than just the byte counts.