AI-handled content

This site is generated and maintained by AI and may be prone to errors. Please verify any claim independently before relying on it.

The Heat Kernel: How the Diffusion Equation Computes the Circle Constant Through Its Own Normalization, and What Golden Pi Changes¶

Every previous article in this series has taken a formula that produces the circle constant — a series, an integral, a product, a table — and asked what happens to it when the constant is relabeled. This one takes a formula that consumes it. The heat equation has no circle in it, no angle, no angular frequency, no rotating anything. It is a statement about how a bump of concentration flattens out over time. Yet its fundamental solution carries the circle constant in one specific place, and that place is exactly where a computation — never a measurement — puts it: the normalisation constant \(1/\sqrt{4\pi D t}\).

🔬 Live graph · Laws of appearance

Measurement feasibility: how far apart the two candidates are in millimetres on a physical disc. The Goblet of the Truth keeps two bodies of law apart — the laws and recommendations of the primal power (Creation), and the laws of appearance (nature) (Ch 2 §342), the register governing how a thing manifests and behaves. Geometry answers to that second register, so nothing here asks to be taken on authority: drag the slider, change a value, and the figure answers for itself (Ch 22 §35: “the fluidal-powers are truthful effects of natural laws”).

That makes the diffusion kernel a good test case, because the constant enters neither as a geometry nor as a phase but as the value of a Gaussian integral rewritten in polar coordinates. If the constant is a name for the radians of a turn, then the kernel's \(4\pi\) is that name written twice over: a full turn (\(2\pi\)) times a factor of two from the radial Jacobian. And that is precisely what a computed constant can be examined for.

The Propagator¶

The heat (diffusion) equation in one dimension,

\[\frac{\partial u}{\partial t} \;=\; D\,\frac{\partial^2 u}{\partial x^2}, \qquad D > 0,\]

has for initial data \(u(x,0)=\delta(x)\) the fundamental solution

\[K(x,t)\;=\;\frac{1}{\sqrt{4\pi D t}}\,\exp\!\left(-\frac{x^{2}}{4Dt}\right),\]

and this is the object every diffusion problem is built from: any initial profile is \(u(x,t)=\int K(x-y,\,t)\,u(y,0)\,dy\).

Two facts about \(K\) must be kept apart, because the whole article depends on the distinction.

The shape of \(K\) — the Gaussian form and its width — is fixed by the equation itself. A Gaussian is reproduced by the heat flow (it is the self-similar solution), and its variance satisfies \(\frac{d}{dt}\langle x^2\rangle = 2D\).
The normalisation — the constant out front — is fixed by the requirement \(\int_{-\infty}^{\infty} K\,dx = 1\), which is what makes \(K\) a propagator of total probability (or of total mass) rather than just a bump of the right shape.

The normalisation is the only place the circle constant appears, and it appears because of one computed limit:

\[\int_{-\infty}^{\infty} e^{-x^{2}/(4Dt)}\,dx \;=\; \sqrt{4\pi D t}.\]

That integral is a computation. It is evaluated by squaring it, passing to polar coordinates, and recognising the angular integral as one full turn:

\[\left(\int_{-\infty}^{\infty}e^{-x^{2}}\!dx\right)^{2}=\int_{-\infty}^{\infty}\!\!\int_{-\infty}^{\infty}e^{-(x^{2}+y^{2})}dx\,dy=\underbrace{\int_{0}^{2\pi}\!\!d\theta}_{2\pi}\underbrace{\int_{0}^{\infty} r e^{-r^{2}}dr}_{\tfrac12}=\pi .\]

Nothing here is measured. The \(2\pi\) is the length of the unit circle as computed by the integral of \(d\theta\), and the constant enters the kernel because a two-dimensional Gaussian's volume is an area of revolution. The reader may reasonably ask whether this is circular reasoning — is the turn's value being assumed? No: \(\theta\) is only a parameter of the polar map, and the value \(2\pi\) is what the angular integral produces when the parameter is normalised to one period. This series has made the same argument in the Gaussian-integral and Dirichlet-integral articles, and it is the same argument here.

Where the Constant Actually Enters¶

Written out, the kernel exposes three separate uses of the constant, and they do not move by the same amount when the label changes:

Appearance of the constant	Classical form	Shift under the golden label
Peak height \(1/\sqrt{4\pi Dt}\)	\(1/\sqrt{4\pi}=0.2820947917738781434740397\) (at \(Dt=1\))	\(-0.0479167\%\) (square root)
Three-dimensional amplitude \((4\pi Dt)^{-3/2}\)	\(1/(4\pi)^{3/2}=0.02244839026564582021114\) (at \(Dt=1\))	\(-0.1436811\%\) (one and a half powers)
Laplace Green's function \(1/(4\pi r)\)	\(0.07957747154594766788444188\) at \(r=1\)	\(-0.0958103\%\)
Mode quantisation \(k = 2\pi n/L\)	—	\(+0.0959022\%\)
Diffusive lead \(1/\sqrt{4\pi Dt}\) vs. oscillatory lead \(1/\sqrt{2\pi i\hbar t/m}\)	—	\(-0.0479167\%\) (square root)

The figures in the middle column are computed here, not recalled: \(1/\sqrt{4\pi}\) and \(1/(4\pi)\) to twenty-three digits, and \((4\pi)^{-3/2}\) to twenty. The two percentages in the right-hand column are the two directions of the same gap that this series has met in every article — the forward gap,

\[\frac{\hat\pi-\pi}{\pi} = +0.09590223087825259637\%,\]

and the reciprocal gap, which is what a quantity carrying \(1/\pi\) experiences:

\[\frac{\pi-\hat\pi}{\hat\pi} = -0.09581034661843333236\%.\]

They differ in the fourth significant figure, and the difference matters when the constant is inverted — as it is in the kernel's peak height, where \(1/\sqrt{4\pi}\) under the golden label equals \(\varphi^{1/4}/4\) exactly, because \(1/\sqrt{4\pi}=1/\sqrt{16/\sqrt\varphi}=\sqrt[4]{\varphi}/4\). That number, \(\varphi^{1/4}/4 = 0.2819596213904205650662089\), is an exact algebraic number in the golden field, and it sits \(0.0479166533\%\) below the classical peak. The square root halves the gap; the inversion flips its direction; both figures were computed at forty digits of working precision for this article.

The Kernel Side by Side¶

Fix \(D = t = 1\) so that the numbers are concrete. The kernel is then \(\tfrac{1}{\sqrt{4\pi}}e^{-x^{2}/4}\) — a Gaussian with variance \(2\), peak \(0.2820947917738781435\) and half-maximum width \(4\sqrt{\ln 2}=3.3302184446307910254\).

Quantity	Value (classical label)	Depends on the constant?
Peak height	\(0.2820947917738781434740397\)	yes, \(1/\sqrt{4\pi}\)
Total mass \(\int K\,dx\)	\(1.000000000000000000\)	no — it defines the normalisation
Mean-square displacement \(\langle x^{2}\rangle\)	\(2Dt\)	no
Full width at half maximum	\(4\sqrt{Dt\ln 2} = 3.3302184446307910254\)	no
Second moment of the diffusivity, \(\sqrt{2Dt}\)	diffusion length	no
Differential entropy \(\tfrac12\ln(4\pi eDt)\)	\(1.765512123484645396488946\) nats	yes, under a logarithm
3-D amplitude \((4\pi Dt)^{-3/2}\)	\(0.02244839026564582021114\)	yes, power \(3/2\)
Steady 3-D Green's function \(1/(4\pi r)\)	\(0.07957747154594766788\) at \(r=1\)	yes, first power

Read the right-hand column carefully, because it is the honest heart of the matter. Diffusion's physical predictions — how far a particle has wandered in a given time, how wide the profile has become, when a front arrives — do not contain the circle constant at all. The wandering distance is \(\sqrt{2Dt}\), full stop. A diffusion experiment therefore measures the diffusion coefficient \(D\), in metres squared per second, a genuine physical measurand; it never touches \(\pi\) directly. The constant appears only in the mathematical apparatus we wrap around the measurement: in normalisation factors, in mode spacings, in the solid-angle factor of a Green's function in three dimensions.

The Ledger: What the Relabel Moves and What It Cannot¶

Take the golden label seriously and put \(\hat\pi = 4/\sqrt\varphi = 3.14460551102969314427823434337\) into the kernel's normalisation:

\[K_{\hat\pi}(x,t)=\frac{1}{\sqrt{4\hat\pi D t}}\exp\!\left(-\frac{x^{2}}{4Dt}\right).\]

Then the peak falls to \(\sqrt[4]{\varphi}/4\) — a change of \(-0.0479167\%\) — and the three-dimensional amplitude changes by the constant's three-halves power, \(-0.1436811\%\). But now the integral of \(K_{\hat\pi}\) is

\[\int_{-\infty}^{\infty}K_{\hat\pi}\,dx=\sqrt{4\pi Dt}\cdot\frac{1}{\sqrt{4\hat\pi Dt}}=\sqrt{\frac{\pi}{\hat\pi}}=0.9995208334666244797814529,\]

a mass deficit of exactly \(0.04791665333755202186\%\). This is the first place in the series where the relabel breaks something that was not merely a name: a propagator that loses \(0.0479167\%\) of its mass per application is not a propagator, it is a sink. The defect is not large — it is four parts in ten thousand — but it is a defect in an identity, not in a description, and identities are checkable by pure arithmetic.

There are three honest ways out, and each one tells us something.

Reparametrise. Absorb the gap into the diffusion coefficient: define \(\hat D = D\,\pi/\hat\pi\) so that \(4\hat\pi \hat D t = 4\pi D t\) and mass is restored exactly. This works, and it shows that the label is invisible whenever it multiplies a free, dimensionful constant of the theory — here \(D\). Diffusion coefficients are measured, not predicted; any error in a label can be swallowed by a measurement of \(D\) at the percent level, a hundred times coarser than the gap.
Rescale the exponent. If instead the golden convention also moves the zero of the Gaussian standard deviation — the \(\hat\pi\) that appears in the mode spacing and hence in the variance — then the exponent's coefficient \(1/(4Dt)\) is not untouched after all. That route restores unit mass on a different green function, and it is the route the "golden calculus" articles took when they wrote the full turn as \(2\hat\pi\).
Keep the arithmetic. Leave the label alone and accept that the kernel's normalisation is \(\sqrt{4\cdot3.14159\ldots D t}\), whatever name one writes for that number. This is the option the computed identities force, and the next two sections show why.

The Trace on a Circle and Jacobi's Identity¶

Put a diffusing particle on a circle of radius \(R\). The eigenvalues of \(-D\,d^{2}/dx^{2}\) are \((n/R)^{2}D\) — no circle constant in sight — so the heat trace is the π-free sum

\[\Theta(A)=\sum_{n\in\mathbb Z}e^{-A n^{2}},\qquad A=\frac{Dt}{R^{2}},\]

and \(A\) is a pure ratio of a diffusion length squared to a radius squared. Its closed form, however, is Jacobi's imaginary transformation,

\[\sum_{n\in\mathbb Z}e^{-A n^{2}}\;=\;\sqrt{\frac{\pi}{A}}\sum_{m\in\mathbb Z}e^{-\pi^{2}m^{2}/A},\]

an identity of exact arithmetic in which the constant appears twice: once under a square root in the prefactor, once squared inside the wrapping-number sum. Verified numerically here at \(A=0.01\), both sides agree to \(1.1\times10^{-15}\) relative (the residual is the truncation of the sums, not the identity):

\(A=0.01\)	value
\(\sum_{n\in\mathbb Z}e^{-An^{2}}\) (no constant anywhere)	\(17.72453850905516007815107\)
\(\sqrt{\pi/A}\sum_{m\in\mathbb Z}e^{-\pi^{2}m^{2}/A}\)	\(17.72453850905516027298167\)
\(\sqrt{\hat\pi/A}\sum_{m\in\mathbb Z}e^{-\hat\pi^{2}m^{2}/A}\)	\(17.73303558624324518467033\)

This is a genuinely different situation from the ones the earlier articles described, and it deserves to be stated plainly rather than softened. The left-hand side contains no circle constant at all: it is a sum of exponentials of a ratio of physical quantities, and it can be computed to any precision without knowing anything about turns or radians. The identity says that this π-free number is also equal to a closed form that carries the constant twice. The constant in that closed form is therefore computed by the identity — it is the number that makes the two expressions agree. Substituting the golden value into the closed form misses the π-free left-hand side by \(+0.0479396244\%\), and no choice of convention repairs that: the left-hand side is fixed arithmetic. This is the arithmetic version of the point made in the Fourier-series and nilakantha articles — a relabeled symbol cannot change what a computation converges to — but here the computation is a identity between two ways of writing the same physical quantity, and only one label satisfies it.

Entropy: The Circle Constant Under a Logarithm¶

The differential entropy of the kernel is

\[h \;=\; -\int K \ln K\,dx \;=\; \tfrac12\ln\!\left(2\pi e\,\sigma^{2}\right)\;=\;\tfrac12\ln\!\left(4\pi e D t\right),\]

with \(\sigma^{2}=2Dt\). At \(Dt=1\) that is \(h = 1.765512123484645396488946\) nats, or \(2.547095585180641102701602\) bits. Under the golden label the same formula gives

\[h_{\hat\pi}=\tfrac12\ln\!\left(4\hat\pi e D t\right)=1.765991404854989756960025\ \text{nats},\]

a shift of \(+0.00047928137034436047108\) nats, or \(+0.00069145685618627554366\) bits — small in absolute terms because the constant sits inside a logarithm, where a \(0.0959\%\) multiplicative change becomes a \(0.000958\times\tfrac12 = 0.000479\) additive change. Two further honest points belong here. First, differential entropy is not a physical measurand at all: it changes with the units of \(x\) and can be negative, so no instrument anywhere can adjudicate a \(0.00048\)-nat difference. Second, the Gaussian is the maximum-entropy density for a given variance — a fact proved by the divergence \(D(p\,\|\,q)\ge 0\), i.e. \(\tfrac12\ln(2\pi e\sigma^{2})\ge h(p)\) for any density \(p\) of variance \(\sigma^{2}\). It is attained at equality for the Gaussian and strictly missed otherwise: for a uniform density of the same variance \(2\) (at \(Dt=1\)) the entropy is \(\tfrac12\ln 24 = 1.589026915173972809823471\) nats, short by \(0.17648520831067258667\) nats. Both of those numbers are π-free or label-trivial, which is the point: the whole information-theoretic content of the diffusion kernel — how much uncertainty a diffusing cloud represents, and why the Gaussian is the honest maximum — is independent of what name is written for the radians.

Brownian Motion Is Measured; the Constant in Its Formula Is Not¶

Here is where the series' rules matter most. Brownian motion gives us a genuine measurement: tracking a particle yields a mean-square displacement linear in time, \(\langle x^{2}\rangle = 2Dt\), and that linearity is an empirical fact about the world. Reading a value of \(D\) off a trajectory is measuring a physical transport coefficient. But \(\langle x^{2}\rangle = 2Dt\) has no circle constant in it, so a Brownian-motion experiment measures \(D\) and nothing about \(\pi\).

The circle constant enters through the relation physicists use to turn \(D\) into microscopic parameters — the Stokes–Einstein relation,

\[D=\frac{k_{B}T}{6\pi\eta a},\]

whose literal \(6\pi\) is a computed feature of Stokes flow (it is the solid-angle/velocity-field factor around a sphere), not a measured one. Every other quantity in it is measured: \(T\) a temperature, \(\eta\) a viscosity, \(a\) a particle radius. With nominal values — \(T = 298.15\) K, water viscosity \(\eta = 0.89\) mPa·s, a \(10\) nm particle — the relation returns

\[D_{\pi}=2.4537\times10^{-11}\ \text{m}^{2}\text{s}^{-1},\qquad D_{\hat\pi}=2.4514\times10^{-11}\ \text{m}^{2}\text{s}^{-1},\]

differing by \(-0.0958103\%\). That gap is the entire question, and it is invisible: \(0.0959\%\) is ten to a hundred times smaller than typical uncertainties in the viscosity and the radius of a real particle. Viscosities are known to percent level at best over a range of temperatures, particle radii to comparable precision, and both enter to the first power. A diffusion experiment cannot decide between the labels, and — this is the part that matters — no refinement of diffusion experiments will, because the disagreement is smaller than the precision of the quantities the constant is glued to.

The Diffusive and the Oscillatory Kernel¶

One more computed contrast, because it separates two constants that are often lumped together. The diffusion kernel's lead is \(1/\sqrt{4\pi Dt}\); the free-particle Schrödinger propagator's lead is

\[K_{\text{QM}}(x,t)=\sqrt{\frac{m}{2\pi i\hbar t}}\,\exp\!\left(\frac{i m x^{2}}{2\hbar t}\right),\]

with a \(2\pi\) rather than a \(4\pi\), because the Gaussian becomes a Fresnel integral: \(\int e^{iax^{2}}dx = \sqrt{\pi/a}\,e^{i\pi/4}\) against \(\int e^{-ax^{2}}dx=\sqrt{\pi/a}\). Both constants are computed — one from a convergent real integral, one from a conditionally convergent oscillatory integral, the latter treated in the Fresnel article below — and both are normalisations forced by the requirement that the propagator compose. Under the golden label, the diffusive lead moves by \(-0.0479167\%\) and the oscillatory lead by the same amount, since both are square roots; the phases inside move by the full gap. Neither is a measurement, and neither is capable of arbitrating between the two labels.

What a Measurement Would Have to Do to Adjudicate¶

Suppose someone wanted to settle the matter with a diffusion experiment rather than an identity. The measurement would have to resolve \(0.0959\%\) in a quantity proportional to the circle constant, with the following systematic errors all pushed below that level: the viscosity of the medium (percent level), the radius of the tracer (percent level), the thermostatic temperature (millikelvin-level for a \(0.1\%\) effect), the finite size of the container, wall drag, and the non-Gaussianity of real trajectories near boundaries. Every one of those is already larger than the gap by one to two orders of magnitude. The gap is not merely unmeasured; it is at present unmeasurable in this setting, and honestly likely to remain so.

The identity ledger, by contrast, is exact. The Gaussian normalisation \(\int e^{-x^{2}/(4Dt)}dx = \sqrt{4\pi Dt}\) is a theorem; so is Jacobi's transformation; so is the maximum-entropy bound. Those computations do not need an instrument, and they are where the circle constant is actually pinned down.

The Honest Boundary¶

Stated plainly, in the spirit this series has kept: the diffusion equation is unchanged by any relabeling, its physical predictions (\(2Dt\)) are π-free, and the circle constant in the heat kernel is computed — out of a Gaussian integral in polar coordinates — not measured. Under Golden Pi (\(\hat\pi = 4/\sqrt\varphi = 3.14460551102969314427823434337\)) the kernel's peak becomes the exact algebraic \(\sqrt[4]{\varphi}/4 = 0.2819596213904205650662089\), a change of \(-0.0479167\%\); the 3-D amplitude changes by the three-halves power, \(-0.1436811\%\); the Laplace Green's function \(1/(4\pi r)\) becomes \(\sqrt\varphi/16\) at \(r=1\), \(-0.0958103\%\); the entropy shifts by \(0.000479\) nats. With the classical exponent in place, the golden normalisation integrates to \(0.9995208334666244797814529\), a \(0.0479167\%\) mass deficit that a reparametrisation of \(D\) hides completely — and that is exactly the problem, because \(D\) is measured to percent precision. Meanwhile Jacobi's identity, whose left-hand side contains no circle constant whatsoever, holds to \(1.1\times10^{-15}\) with the analytic constant and misses by \(+0.0479396\%\) with the golden one. That is the honest boundary of this article: a relabel moves every appearance of the constant in the kernel by half the gap or the full gap depending on the power, hides inside a logarithm for the entropy, and disappears entirely into a measured diffusion coefficient — while the exact identities that fix the constant's value keep pointing at \(3.14159265358979\ldots\), and the golden construction keeps pointing at \(4/\sqrt\varphi\) as the number that makes its own geometry self-consistent. Both statements are true; they are statements about different things.