A little π between belief and relativity

Place a standard probability textbook right next to a special relativity textbook.

The probability book asks how an observer should express prior ignorance about a coin toss before gathering any data. The relativity book asks how much energy a particle carries as it accelerates toward the cosmic speed limit.

At first glance, these two questions could not seem further apart: one belongs to the epistemology of beliefs, the other to the physics of spacetime.

Yet if you turn to the right pages, change a single coordinate, and set the two mathematical curves side by side, an astonishing thing happens. The two curves are identical, separated only by a factor of $\pi$.

Why on earth should a Bayesian rule for measuring human uncertainty know anything about the energy of a relativistic electron? And what is $\pi$, the universal constant of circles, doing right in the middle of a probability update?

When I first noticed the connection between light cones, Beta evidence, probability, and physics on August 12, 2026, this coincidence was one of the earliest signs that something deep and beautiful was happening. The first steps fit on the back of an envelope. By the end of this excursion, that envelope will hold a vibrating mechanical spring, an information-theoretic ruler, and an odds-combining formula that turns out to be Einstein’s velocity addition in disguise.

Meet the curve in a physics classroom

In special relativity, the total energy $E$ of a particle with positive rest mass $m$ is given by Einstein’s mass-energy relation:

$$E=\gamma m{c}^2.$$

Here $c$ is the speed of light in vacuum ($c\approx 3\times {10}^8\text{ m/s}$), and $m$ is the particle’s invariant rest mass, which measures its intrinsic inertia when stationary.

Write the particle’s velocity along a single spatial axis as $v$. To make velocity dimensionless, we divide by $c$ to define the speed ratio $u=v/c$. Because no massive body can reach or exceed the speed of light, $u$ is strictly confined to the open interval $(-1,1)$.

The quantity $\gamma (u)$, pronounced “gamma,” is the famous Lorentz factor:

Look at how this function behaves as a particle speeds up:

When the particle is at rest ($u=0$), the Lorentz factor is $\gamma (0)=1$, returning Einstein’s most famous equation for rest energy, $E=m{c}^2$.

If the particle reaches $60\mathrm{\%}$ of the speed of light ($u=3/5=0.6$), the Lorentz factor climbs to $\gamma =1/\sqrt{1-0.36}=1/0.8=5/4=1.25$.

If it reaches $80\mathrm{\%}$ of the speed of light ($u=4/5=0.8$), $\gamma$ rises to $1/\sqrt{1-0.64}=1/0.6=5/3\approx 1.667$.

As $u$ approaches the cosmic boundary ($u\to \pm 1$), the denominator $\sqrt{1-u^2}$ drops toward zero, and the curve shoots upward toward infinity. Accelerating a particle all the way to the speed of light would require infinite energy.

The mathematical silhouette of the Lorentz factor is unmistakable: a gentle, rounded basin near $u=0$ that curves upward into towering, vertical asymptotic walls at $u=\pm 1$.

Meet it again in a probability classroom

Now close the relativity textbook and open the probability textbook.

Suppose you are studying a Bernoulli trial, an elementary random experiment with two possible outcomes: a coin lands heads or tails, a detector records a photon or darkness, a medical test reads positive or negative.

Let $p\in (0,1)$ denote the unknown probability of success.

Before you observe a single trial, what prior probability distribution should you assign to $p$?

In Bayesian statistics, if you want a prior that represents complete neutrality and remains invariant under reparameterization, you choose the Jeffreys prior (named after geophysicist and statistician Sir Harold Jeffreys). For a Bernoulli model, this prior is the Beta distribution $\mathrm{Beta}(1/2,1/2)$, whose probability density function is:

A probability density measures probability mass per unit length of the horizontal axis. Notice the shape: it has a symmetric U-shape that dips to its lowest point at an even split ($p=1/2$) and curves upward toward infinity as $p\to 0$ and $p\to 1$.

Now let us bring probability into the coordinate system of physics.

In standard probability, $p$ lives on the unit interval $[0,1]$, where $0.5$ represents neutral balance. In physics, velocity $u$ lives on the interval $(-1,1)$, where $0$ represents rest.

Slide and rescale the probability coordinate so that neutral balance sits at zero:

$$u=2p-1.$$

Under this transformation, absolute refutation sits at $u=-1$, neutral uncertainty sits at $u=0$, and absolute confirmation sits at $u=+1$.

Because $p=(u+1)/2$, differentiating gives $dp=du/2$. A standard change of variables in calculus requires multiplying the density by the Jacobian factor $|dp/du|=1/2$ to preserve the total area under the curve:

$$f_U(u)=f_P\left(\frac{u+1}{2}\right)\cdot \frac{1}{2}=\frac{1}{\pi \sqrt{\left(\frac{1+u}{2}\right)\left(\frac{1-u}{2}\right)}}\cdot \frac{1}{2}=\frac{1}{\pi \sqrt{1-u^2}}.$$

Look at that result! Set it alongside the Lorentz factor from relativity:

$$\boxed{\frac{E}{m{c}^2}=\gamma (u)=\pi f_U(u).}$$

The centered Jeffreys prior of probability theory and the Lorentz factor of special relativity are the exact same mathematical curve!

The physics curve $\gamma (u)$ is simply the probability density $f_U(u)$ scaled by $\pi$. Or viewed from the other direction, the Jeffreys prior is the Lorentz factor normalized so that the total area underneath it equals 1. In the uncentered coordinate $p$, the relation is $\gamma (2p-1)=\frac{\pi }{2}{f}_P(p)$.

A probability curve meets relativity

Make the two curves meet.

Centered uProbability pMultiply the prior by π

● Relativistic energy ratio● Jeffreys density

Centered coordinate u

0.600

$\gamma =1.250$$f_U=0.398$$\rho =0.693$

$$\frac{E}{m{c}^2}=\gamma (u)=\pi f_U(u)$$

Multiply the teal curve to reveal the shared shape. Switching coordinates changes density per unit of the horizontal axis.

$E$ is total relativistic energy; $m$ is positive rest mass; $c$ is the speed of light. Compare signed speed $u=v/c$ with centered probability $u=2p-1$. The curve height is $\gamma$ (gamma); $\rho =\mathrm{artanh}u$ (rho) is rapidity.

Use the interactive bridge above to test the connection. Slide along the curve, toggle between the centered coordinate $u$ and the standard probability $p$, and watch the multiplier transition from $\pi$ to $\pi /2$.

Why does a circle get a cameo?

Where did the factor of $\pi$ come from?

To normalize the Lorentz factor into a valid probability density, we must integrate it over its entire domain from $u=-1$ to $u=+1$:

$${\int }_{-1}^1\frac{du}{\sqrt{1-u^2}}=\pi .$$

Why does integrating a curve from physics and probability produce the geometric ratio of a circle’s circumference to its diameter?

The answer reveals itself through trigonometric substitution. Let $u=\sin \theta$, where the angle $\theta$ ranges from $-\pi /2$ to $+\pi /2$.

Differentiating gives $du=\cos \theta d\theta$.

Meanwhile, the square root simplifies via the Pythagorean identity:

$$\sqrt{1-u^2}=\sqrt{1-{\sin }^2\theta }=\sqrt{{\cos }^2\theta }=\cos \theta .$$

Substitute both into the integral:

$${\int }_{-1}^1\frac{du}{\sqrt{1-u^2}}={\int }_{-\pi /2}^{\pi /2}\frac{\cos \theta d\theta }{\cos \theta }={\int }_{-\pi /2}^{\pi /2}d\theta =\frac{\pi }{2}-(-\frac{\pi }{2})=\pi .$$

The speed factor $\sqrt{1-u^2}$ in the denominator perfectly cancels the geometric contraction $\cos \theta$ in the numerator! The entire integral collapses into the length of the angle interval: $\pi$.

You can visualize this substitution using a flashlight and a semicircle.

Imagine a half-circle of unit radius. Place evenly spaced tick marks along its curved rim (say, every 10 degrees). Now shine a light straight down from above, projecting the shadow of each tick mark onto the horizontal diameter.

Near the middle of the diameter, the shadows are spaced widely apart. But near the left and right edges, the curvature of the circle causes many tick marks to project onto tiny, crowded slivers of the line.

If you pick an angle uniformly at random along the semicircle ($\mathrm{\Theta }\sim \mathrm{Uniform}(-\pi /2,\pi /2)$) and drop its shadow $U=\sin \mathrm{\Theta }$ onto the diameter, the probability density of that shadow is our centered Jeffreys curve:

$$f_U(u)=\frac{1}{\pi \sqrt{1-u^2}}.$$

This is why mathematicians also call this the arcsine distribution: the inverse trigonometric function $\arcsin (u)$ unrolls the shadows back onto the uniform arc of a circle.

Even though the density diverges to infinity at the two edges ($u\to \pm 1$), the area remains strictly finite and equal to 1, because the width of the shadow intervals shrinks to zero just as fast as the curve rises.

A spring has been drawing this curve all along

We do not need to go to particle accelerators or abstract probability spaces to see this curve in action. A simple mechanical spring in a physics laboratory has been tracing it all along.

Consider a mass $m$ oscillating back and forth on an ideal spring with amplitude $A$ and angular frequency $\omega$:

$$x(t)=A\cos (\omega t).$$

Here $x(t)$ is the displacement of the mass from equilibrium at time $t$. Divide by the amplitude to obtain the normalized position $u(t)=x(t)/A=\cos (\omega t)$, which oscillates back and forth within $[-1,1]$.

Suppose you photograph this vibrating mass at a random moment chosen uniformly across a complete cycle. What is the probability density of finding the mass at position $u$?

Differentiate the position to find its velocity:

$$\frac{du}{dt}=-\omega \sin (\omega t),\left|\frac{du}{dt}\right|=\omega |\sin (\omega t)|=\omega \sqrt{1-u^2}.$$

The speed of the mass through position $u$ is $|du/dt|=\omega \sqrt{1-u^2}$.

Now pause and think about the physical motion:

As the mass passes through the center ($u=0$), the spring is relaxed, and the mass travels at its maximum speed $\omega$. It sweeps past the center in a flash.

As the mass approaches the turnaround points ($u=\pm 1$), the spring exerts maximum restoring force, slowing the mass down until it halts momentarily and reverses direction.

Because the time $dt$ spent in any narrow window $du$ is inversely proportional to speed ($dt=du/|du/dt|$), a random photograph is far more likely to capture the mass near the ends of its trajectory than near the center:

$$f_U(u)\propto \frac{1}{|du/dt|}=\frac{1}{\omega \sqrt{1-u^2}}.$$

Normalizing this time distribution over $[-1,1]$ yields our exact centered Jeffreys density:

$$f_U(u)=\frac{1}{\pi \sqrt{1-u^2}}.$$

A spring, a circle, a familiar curve

Where does a spring spend its time?

Play the spring

Fraction of the cycleJeffreys probability in each bin

Phase

60°

Position / amplitude

0.500

$$u=\cos \theta ,f_U(u)=\frac{1}{\pi \sqrt{1-u^2}}$$

Drag the circle's gold dot, play the motion, or scrub the phase. Arrow keys turn the focused dot by 5°. Its horizontal projection becomes the spring's position; both small axes run from −1 to 1. Here A is the spring's amplitude and θ is the circle angle.

The mass rushes through the center and lingers near the ends. Each teal bar is the exact fraction of a cycle spent in that bin. Each rose outline is the Jeffreys prior probability in the same bin, using u = 2p − 1. They agree in every bin, including the turning points.

Use the simulation above to watch the oscillator in real time. Notice how the histogram builds up tall spikes at the turning points and stays shallow in the middle.

Four completely different physical and mathematical phenomena have arrived at the exact same equation:

1. The relativistic energy factor of a particle accelerating toward light speed.
2. The invariant prior distribution for a Bayesian observer facing a coin toss.
3. The geometric shadow cast by a uniform circle onto its diameter.
4. The time budget of a mechanical mass bobbing on a spring.

Why do these four disparate worlds share the exact same mathematical signature?

The ruler inside a probability model

To find the unifying bridge, we look to information geometry, a field pioneered by Ronald Fisher and C. R. Rao.

Information geometry asks a geometric question: how distinguishable are two neighboring probability distributions?

If a coin has success probability $p$, and another coin has probability $p+dp$, how many experimental trials do you need to tell them apart?

For a single Bernoulli trial, this sensitivity is quantified by the Fisher information:

$$I(p)=\frac{1}{p(1-p)}.$$

Look at what this formula says:

If $p=0.50$, shifting the probability to $0.51$ produces an imperceptible change in outcomes. You would need thousands of tosses to detect the difference. The denominator $p(1-p)=0.25$ is at its maximum, making Fisher information small ($I=4$).

Now suppose $p=0.99$, and someone claims the true probability is $1.00$ (a coin that never fails). If you flip the coin and see even a single tail, the hypothesis $p=1.00$ is refuted instantly! Near the boundaries, tiny shifts in probability produce massive, easily detectable differences in outcomes. As $p\to 0$ or $p\to 1$, the denominator $p(1-p)$ vanishes, and Fisher information explodes toward infinity.

Fisher information defines a natural Riemannian metric (a statistical ruler) on the parameter space:

$$d{s}^2=I(p)d{p}^2=\frac{d{p}^2}{p(1-p)}.$$

Now transform this statistical ruler into our centered coordinate $u=2p-1$, where $dp=du/2$:

$$d{s}^2=\frac{(du/2{)}^2}{\left(\frac{1+u}{2}\right)\left(\frac{1-u}{2}\right)}=\frac{d{u}^2}{1-u^2}=\gamma (u{)}^{2}d{u}^2.$$

Taking the square root gives an equation of singular beauty:

$$ds=\gamma (u)du.$$

The relativistic Lorentz factor squared, $\gamma (u{)}^2$, is the metric coefficient of Bernoulli information geometry!

Its square root, $\gamma (u)$, is the conversion factor that translates an ordinary step in probability into an invariant step of statistical distinguishability.

When Sir Harold Jeffreys looked for an uncommitted prior that would treat all parameterizations equally, his mathematical criterion was to set the prior density proportional to the square root of the Fisher metric: $f(u)\propto \sqrt{g(u)}=\gamma (u)$.

The Jeffreys prior matches the Lorentz factor because both are governed by the same geometry of distinguishability.

We can make this geometry literal. Consider the square root probability vector:

$$\mathit{q}(p)=(\sqrt{p},\sqrt{1-p}).$$

Because $(\sqrt{p}{)}^2+(\sqrt{1-p}{)}^2=p+(1-p)=1$, the vector $\mathit{q}(p)$ always has unit Euclidean length. As $p$ varies from $0$ to $1$, $\mathit{q}(p)$ traces an arc along the unit circle in the positive quadrant!

Let $\varphi$ be the angle along this circle, so that $\sqrt{p}=\cos \varphi$ and $\sqrt{1-p}=\sin \varphi$.

Then $p={\cos }^2\varphi$, and differentiating gives $dp=-2\cos \varphi \sin \varphi d\varphi$.

Substitute this into the statistical distance formula:

$$d{s}^2=\frac{d{p}^2}{p(1-p)}=\frac{4{\cos }^2\varphi {\sin }^2\varphi d{\varphi }^2}{{\cos }^2\varphi {\sin }^2\varphi }=4d{\varphi }^2.$$

Taking the square root:

$$ds=2d\varphi .$$

Statistical distance along the space of Bernoulli probabilities is simply twice the angle along the circle!

Integrating across the entire positive quadrant from $\varphi =0$ to $\varphi =\pi /2$:

$$\text{Total statistical length}={\int }_0^{\pi /2}2d\varphi =2\cdot \frac{\pi }{2}=\pi .$$

The entire universe of binary probabilities has a total statistical length of exactly $\pi$.

The mysterious factor of $\pi$ was never a coincidence: it is the total geometric circumference of the space of Bernoulli beliefs.

Log-odds arrive wearing a spacesuit

Special relativity offers another coordinate that makes calculations remarkably simple: rapidity, denoted $\rho$ (rho).

Because no particle can exceed the speed of light, velocities cannot add linearly: if rocket A travels at $0.8c$ and fires probe B forward at $0.8c$, the probe does not travel at $1.6c$.

Einstein solved this by mapping velocity $u\in (-1,1)$ to an unbounded coordinate $\rho \in (-\mathrm{\infty },\mathrm{\infty })$ using the inverse hyperbolic tangent:

$$\rho =\mathrm{artanh}(u)=\frac{1}{2}\ln \left(\frac{1+u}{1-u}\right).$$

Under consecutive collinear boosts, rapidities add linearly:

$${\rho }_{\mathrm{combined}}={\rho }_1+{\rho }_2.$$

Now substitute our centered probability coordinate $u=2p-1$ into rapidity:

$$\rho =\frac{1}{2}\ln \left(\frac{1+(2p-1)}{1-(2p-1)}\right)=\frac{1}{2}\ln \left(\frac{2p}{2(1-p)}\right)=\frac{1}{2}\ln \left(\frac{p}{1-p}\right).$$

Look at that expression!

The ratio $p/(1-p)$ is the odds of success (for example, $p=3/4$ gives odds of $3:1$).

The logarithm $\ln (p/(1-p))$ is the log-odds, also known as the logit function in statistics and machine learning.

Epistemic rapidity is exactly half the log-odds:

$$\rho =\frac{1}{2}\mathrm{logit}(p).$$

Now consider Bayes’ theorem for updating beliefs with new independent evidence:

$$\text{Posterior Odds}=\text{Prior Odds}\times \text{Likelihood Ratio}.$$

Take the natural logarithm of both sides:

$$\ln (\text{Posterior Odds})=\ln (\text{Prior Odds})+\ln (\text{Likelihood Ratio}).$$

In Bayesian statistics, independent evidence adds linearly in log-odds.

In relativity, independent boosts add linearly in rapidity.

The two operations are mathematically identical!

What happens when you add two rapidities ${\rho }_1+{\rho }_2$ and convert back to centered velocity $u$?

Using the hyperbolic identity for $\tanh ({\rho }_1+{\rho }_2)$:

$$u_{\mathrm{combined}}=\tanh ({\rho }_1+{\rho }_2)=\frac{\tanh {\rho }_1+\tanh {\rho }_2}{1+\tanh {\rho }_1\tanh {\rho }_2}=\frac{u_1+u_2}{1+u_1{u}_2}.$$

Einstein’s relativistic velocity addition law is Bayes’ rule in disguise!

Let us test this with numbers.

Suppose your prior probability is $p_1=0.75$ (odds $3:1$, centered velocity $u_1=2(0.75)-1=0.50$).

New independent evidence arrives with a likelihood ratio of $3:1$ (which corresponds to an evidence velocity $u_2=0.50$).

Using standard Bayes’ rule, the new odds are $3\times 3=9:1$, giving posterior probability $p=9/10=0.90$.

Now calculate the combined velocity using Einstein’s formula:

$$u_{\mathrm{combined}}=\frac{0.50+0.50}{1+(0.50)(0.50)}=\frac{1.00}{1.25}=0.80.$$

Convert centered velocity back to probability:

$$p=\frac{u+1}{2}=\frac{0.80+1}{2}=0.90.$$

The relativistic velocity addition formula produces the exact Bayesian posterior probability!

Try the update yourself. Change the starting probability or the likelihood ratio, then apply fresh evidence with the same ratio again. The population panel makes the parallel with natural selection visible: relative reproductive fitness multiplies population odds just as a likelihood ratio multiplies hypothesis odds.

The odds route

Dock three worlds with the same boost

Restart routes

Try a route Free exploration

1 / 3 · Cross the midpoint. Turn one-to-three odds into three-to-one odds. Choose evidence tiles to reach 75.0%. You can undo every move.

Belief probability25.0%p

Variant A's share25.0%two-type selection

Ship velocity / c-0.5002 × probability − 1

← Divide odds by 3opposing evidence · inverse boost Multiply odds by 3 →supporting evidence · forward boost

Undo last tile Clear route Your route starts here.

Destination: belief and population 75.0%, ship velocity 0.50c. The gold dock marks their shared rapidity.

Why these three worlds move together

A tile multiplies Bayesian odds by a likelihood ratio L. In a population it changes the relative fitness of variant A; on the ship it supplies a collinear boost with velocity tanh(½ ln L). Reversing the tile uses 1/L.

$$O=\frac{p}{1-p},O^{\prime }=LO,\rho =\frac{1}{2}\ln O,{\rho }^{\prime }=\rho +\frac{1}{2}\ln L$$ $$u=2p-1=\tanh \rho ,u^{\prime }=\frac{u+w}{1+uw},w=\tanh (\frac{1}{2}\ln L)$$

The three models share this update algebra. Inference assumes new evidence with the specified likelihood ratio at each step; selection assumes two variants with the chosen relative fitness; physical boosts act along one axis. These tiles reweight hypothesis odds rather than adding Bernoulli counts to a Beta distribution.

Two maps of the same territory

We now have two natural coordinate systems for a belief:

1. The Fisher angle $\theta =\arcsin (u)$, which measures intrinsic statistical distance.
2. The rapidity $\rho =\mathrm{artanh}(u)$, which measures additive Bayesian evidence.

Both coordinates straighten the geometry of beliefs, but they straighten different features:

The Fisher angle straightens distinguishability. It treats both endpoints as reachable boundaries, placing them at a finite distance of $\pi /2$ from the neutral center.

Rapidity straightens evidence accumulation. It treats both endpoints as infinite horizons, requiring an infinite sum of evidence increments to reach absolute certainty ($u\to \pm 1$).

Just as a topographic map can show both horizontal trail distance and vertical elevation gain, the geometry of belief carries both metrics simultaneously.

This distinction sharpens the original surprise. We found an exact shared curve, and then discovered several operations that give it different jobs. That is a fertile place to start inventing:

• What if learning algorithms take steps in statistical angle $\theta$, moving along geodesics of the Fisher information metric to guarantee invariant progress regardless of parameterization?
• What if interactions between communicating agents compose in rapidity $\rho$, multiplying odds like relativistic boosts?
• What if total concentration $\tau$ supplies the scale that both coordinates leave out, providing the inertia of accumulated experience?

What would happen if a simulated world organized its histories this way? Could an autonomous agent represent particles as interacting evidence notebooks, where physical collisions exchange small units of evidence called tracelets? Such collisions would conserve total evidence and momentum while redistributing directional bias. We could test whether these exchange rules produce classical Newtonian inertia, relativistic kinematics, or entirely unexpected dynamics.

When probability is treated as a bare scalar number, these geometric symmetries remain hidden. But as soon as we preserve the evidence counts in a Beta notebook $(\alpha ,\beta )$, tracking the lineage of evidence from incoming tracelets and trials, the full apparatus of spacetime emerges:

Total evidence ${\tau }_{\mathrm{E}}=\alpha +\beta$ acts as forward-ticking epistemic time.

Net balance $x_{\mathrm{E}}=\alpha -\beta$ acts as epistemic spatial position.

Certainty acts as the speed of light ($c_{\mathrm{E}}=1$).

And the clash of opposing evidence generates epistemic rest mass ($m_{\mathrm{E}}=2\sqrt{\alpha \beta }$), providing the inertia that stabilizes our convictions against sudden whims.

The factor of $\pi$ between the Jeffreys prior and the Lorentz factor was not a mathematical coincidence. It was a doorway into a unified physical geometry of learning. Step through with a circle in one pocket and an evidence notebook in the other.