A belief is more than a number

Imagine two experimenters standing before you, each making a probability statement about a coin.

The first experimenter has just flipped a coin twice, recording one head and one tail. The second experimenter has spent five years logging ten thousand flips, recording five thousand heads and five thousand tails.

Ask each of them for their expected probability of heads on the next flip. Both give you the exact same answer: $50\mathrm{\%}$.

Now flip the coin once more, and watch it land heads. How should their beliefs respond?

For the first experimenter, this single new observation is an event of momentous proportion: their record shifts from an even split to a two-to-one majority, nudging their expectation upward to $62.5\mathrm{\%}$ with the half-count-per-channel prior we will introduce below.

For the second experimenter, the same new head is a drop in an ocean: their record shifts from $\mathrm{5,000}$ to $\mathrm{5,001}$ heads out of $\mathrm{10,001}$ tosses, nudging their expectation from $50.000\mathrm{\%}$ to $50.005\mathrm{\%}$.

Both observers held the exact same probability number. Yet their beliefs lived in completely different epistemic universes. One belief was a feather that spun at the slightest breeze; the other was a granite boulder that barely shuddered under an identical impact.

When I first began exploring these connections, this simple contrast was the point of departure. Classical probability theory routinely collapses an agent’s state of belief into a single scalar $p\in [0,1]$. In doing so, it erases the depth of history that holds that belief in place. A single number tells you where a belief points, but it cannot tell you how much evidence pins it there.

To give a belief its full geometric identity, we must look inside its evidence ledger.

Two notebooks, one experiment

Give both beliefs the same little push.

Reset comparison

Balanced trials in each column of the thicker notebook

40

A little evidence

$\mu =0.500$${\tau }_{\mathrm{E}}=5$

A lot of evidence

$\mu =0.500$${\tau }_{\mathrm{E}}=81$

Hollow dots mark the means before the last trial; solid dots mark them after it.

Add support to bothAdd opposition to both

At the same starting mean, the fuller notebook moves less. Keep adding the same outcome: the smaller notebook approaches the endpoint sooner, leaving less room for its next step. Reset to compare equal starting means again.

Use the interactive comparison above to experience this difference firsthand. Adjust the thickness of the second notebook, deliver a supporting trial to both, and watch how the same observation feels like a gentle nudge to a seasoned history and a violent shove to a fragile one.

Keep the shape of the uncertainty

In Epistemic Physics, our fundamental mathematical model for an uncertain Bernoulli probability $P\in (0,1)$ is the Beta distribution.

A Beta distribution does not summarize belief as a point. It describes a continuous curve of plausibility over all possible probabilities between 0 and 1. This curve is parameterized by two positive counts: $\alpha$ (alpha), representing accumulated support, and $\beta$ (beta), representing accumulated opposition.

When an observer begins with an uncommitted neutral prior of total concentration $n_0$, these parameters decompose into raw experimental counts plus prior scaffolding:

$$\alpha =Y+\frac{n_0}{2},\beta =N+\frac{n_0}{2}.$$

Here $Y$ is the total amount of positive supporting evidence, $N$ is the total amount of negative opposing evidence, and $n_0$ is the prior concentration. Throughout our work, we adopt the standard neutral prior $n_0=1$, which places a half-unit of prior evidence into each column: $({\alpha }_0,{\beta }_0)=(1/2,1/2)$. As we saw in our study of the Jeffreys prior, this choice has profound invariance properties under coordinate changes.

The probability density function of the Beta distribution is given by:

where $\mathrm{B}(\alpha ,\beta )={\int }_0^1{t}^{\alpha -1}(1-t{)}^{\beta -1}dt$ is Euler’s beta function, which normalizes the total area under the curve to 1.

From this distribution, we can extract two essential statistical summaries.

The first is the expectation (the mean probability), which represents the center of mass of the curve:

$$\mu =\mathbb{E}[P]=\frac{\alpha }{\alpha +\beta }.$$

The second is the variance, which measures the dispersion or spread of the curve around its mean:

$$\mathrm{Var}(P)=\frac{\alpha \beta }{(\alpha +\beta {)}^2(\alpha +\beta +1)}=\frac{\mu (1-\mu )}{\alpha +\beta +1}.$$

Now look closely at what happens when we compare different notebooks with identical means:

Consider a thin notebook with $(\alpha ,\beta )=(2,2)$. Its mean is $\mu =2/4=1/2$, and its variance is:

$$\mathrm{Var}(P)=\frac{(1/2)(1/2)}{4+1}=\frac{1/4}{5}=\frac{1}{20}=0.05.$$

Next, consider a moderately seasoned notebook with $(\alpha ,\beta )=(20,20)$. Its mean is still $\mu =1/2$, but its variance shrinks to:

$$\mathrm{Var}(P)=\frac{(1/2)(1/2)}{40+1}=\frac{1/4}{41}=\frac{1}{164}\approx 0.0061.$$

Finally, consider a thick notebook with $(\alpha ,\beta )=(200,200)$. Its mean remains $\mu =1/2$, while its variance collapses to:

$$\mathrm{Var}(P)=\frac{(1/2)(1/2)}{400+1}=\frac{1/4}{401}=\frac{1}{1604}\approx 0.00062.$$

This contrast resolves a classic philosophical ambiguity between two distinct kinds of uncertainty:

1. Aleatoric uncertainty (inherent randomness): Even if you know with absolute certainty that a fair coin has true bias $p=1/2$, any individual future toss remains completely unpredictable: heads and tails are equally likely. This is uncertainty about the outcome of the next trial.
2. Epistemic uncertainty (lack of knowledge): If you have only seen two flips of a coin, you do not yet know whether the coin is fair, double-headed, or loaded. Your belief distribution over $p$ is broad and easily perturbed. This is uncertainty about the underlying parameter itself.

A single probability number $\mu =1/2$ captures only the expected outcome. The full Beta notebook $(\alpha ,\beta )$ preserves both the direction of expectation and its resistance to changing that expectation.

Mode and trials: two questions about a belief

When a belief belongs to an agent, component, or neuron, one philosophical reading separates mode from trial amount. The mode, the location of a peak in the belief distribution, expresses a degree of true-ness or false-ness. A mode at 99% says “very sure it is true”; 1% says “very sure it is false”; 50% says “unsure.” Trial amount tells us how resistant that state is to changing its mind. We can call its ease of being swayed malleability: few trials leave a belief more malleable, while many make it resistant to the same new trial.

Mode and trial amount describe two aspects of the same state, which can also be expressed through trials for and trials against: support/opposition, true/false, or belief/disbelief. The mean $\mu$ used in the calculations here averages the distribution; the mode locates its peak.

These are among several useful interpretations. Trials can also represent accumulated evidence or observed outcomes. With a Dirichlet distribution, the two channels become weighted votes for mutually exclusive choices, such as A, B, or C. The geometry gives us a shared language in which to explore these readings.

Turn the notebook sideways: The spacetime light cone

Now let us perform a simple coordinate rotation that unlocks the physics of belief.

Instead of tracking support $\alpha$ and opposition $\beta$ along separate axes, take their sum and their difference:

$${\tau }_{\mathrm{E}}=\alpha +\beta ,x_{\mathrm{E}}=\alpha -\beta .$$

Let us define what these quantities mean:

• Epistemic time (${\tau }_{\mathrm{E}}$): The sum $\alpha +\beta$ measures the total accumulated concentration of evidence, including the prior. Every incoming observation adds weight to the notebook, pushing ${\tau }_{\mathrm{E}}$ forward like a one-way clock.
• Epistemic position ($x_{\mathrm{E}}$): The difference $\alpha -\beta$ measures the net signed balance between confirmation and refutation. If support exceeds opposition, ; if opposition dominates, ; if both sides balance evenly, $x_{\mathrm{E}}=0$.

Because and (guaranteed by any non-zero prior ), the triangle inequality ensures that:

In a spacetime diagram where $x_{\mathrm{E}}$ is plotted horizontally and ${\tau }_{\mathrm{E}}$ is plotted vertically, every valid belief notebook sits strictly inside the open cone:

This is the forward light cone of special relativity, with an epistemic speed of light equal to unity ($c_{\mathrm{E}}=1$)!

The ratio of horizontal position to vertical time defines the epistemic velocity:

$$v_{\mathrm{E}}=\frac{x_{\mathrm{E}}}{{\tau }_{\mathrm{E}}}=\frac{\alpha -\beta }{\alpha +\beta }=\frac{2\alpha -(\alpha +\beta )}{\alpha +\beta }=2\mu -1.$$

Epistemic velocity $v_{\mathrm{E}}$ is simply the mean probability rescaled to the symmetric physics interval $(-1,1)$. When an observer is completely agnostic ($p=1/2$), the velocity is zero: the belief is at rest. When evidence leans toward absolute confirmation ($p\to 1$), $v_{\mathrm{E}}\to +1$. When evidence leans toward absolute refutation ($p\to 0$), $v_{\mathrm{E}}\to -1$.

Now calculate the relativistic Minkowski spacetime interval:

$$s^2={\tau }_{\mathrm{E}}^2-x_{\mathrm{E}}^2.$$

Substitute the definitions of ${\tau }_{\mathrm{E}}$ and $x_{\mathrm{E}}$:

$$s^2=(\alpha +\beta {)}^2-(\alpha -\beta {)}^2=({\alpha }^2+2\alpha \beta +{\beta }^2)-({\alpha }^2-2\alpha \beta +{\beta }^2)=4\alpha \beta .$$

Taking the square root reveals the invariant epistemic rest mass of the belief:

$$\boxed{m_{\mathrm{E}}=\sqrt{{\tau }_{\mathrm{E}}^2-x_{\mathrm{E}}^2}=2\sqrt{\alpha \beta }.}$$

Look at how stunning this result is!

In Einstein’s special relativity, a particle’s invariant rest mass is given by $m=\sqrt{E^2/c^4-p^2/c^2}$. In Epistemic Physics, the invariant mass of a belief is twice the geometric mean of its supporting and opposing evidence!

Consider what this formula reveals about the nature of conviction:

If the opposing parameter itself reaches zero ($\beta =0$), the cone’s boundary has zero rest mass: $m_{\mathrm{E}}=0$. Like a photon of light, this boundary state travels at the maximum speed ($v_{\mathrm{E}}=1$). Our ordinary notebook retains its prior, however: even with no opposing trials, $\beta =1/2$ and . One-sided observations approach the boundary in velocity while the prior keeps every finite notebook inside!

Whenever both channels hold positive amounts ( and ), their combination generates positive epistemic rest mass. This mass places the belief in the interior of the light cone, endowing it with physical inertia that resists sudden changes in velocity.

Opposing evidence is not merely an inconvenience to be averaged away: it is the literal source of epistemic mass that stabilizes our convictions.

The law of persuasion: Why surprises pack a punch

How does a belief accelerate when it encounters a new trial?

Suppose an agent has accumulated a history $(\alpha ,\beta )$ with total evidence ${\tau }_{\mathrm{E}}=\alpha +\beta$ and epistemic velocity $v_{\mathrm{E}}=(\alpha -\beta )/{\tau }_{\mathrm{E}}$.

Now deliver one single unit of supporting evidence ($\alpha \to \alpha +1,\beta \to \beta$).

The new velocity is:

$$v_{\mathrm{E}}^{\prime }=\frac{(\alpha +1)-\beta }{(\alpha +1)+\beta }=\frac{x_{\mathrm{E}}+1}{{\tau }_{\mathrm{E}}+1}.$$

The shift in epistemic velocity $\mathrm{\Delta }{v}_{\mathrm{E}}=v_{\mathrm{E}}^{\prime }-v_{\mathrm{E}}$ can be simplified with elementary algebra:

$$\mathrm{\Delta }{v}_{\mathrm{E}}=\frac{x_{\mathrm{E}}+1}{{\tau }_{\mathrm{E}}+1}-\frac{x_{\mathrm{E}}}{{\tau }_{\mathrm{E}}}=\frac{{\tau }_{\mathrm{E}}(x_{\mathrm{E}}+1)-x_{\mathrm{E}}({\tau }_{\mathrm{E}}+1)}{{\tau }_{\mathrm{E}}({\tau }_{\mathrm{E}}+1)}=\frac{{\tau }_{\mathrm{E}}-x_{\mathrm{E}}}{{\tau }_{\mathrm{E}}({\tau }_{\mathrm{E}}+1)}.$$

Because $x_{\mathrm{E}}/{\tau }_{\mathrm{E}}=v_{\mathrm{E}}$, we can factor out ${\tau }_{\mathrm{E}}$:

$$\mathrm{\Delta }{v}_{\mathrm{E}}=\frac{{\tau }_{\mathrm{E}}(1-v_{\mathrm{E}})}{{\tau }_{\mathrm{E}}({\tau }_{\mathrm{E}}+1)}=\frac{1-v_{\mathrm{E}}}{{\tau }_{\mathrm{E}}+1}.$$

If instead the agent observes one single unit of opposing evidence ($\alpha \to \alpha ,\beta \to \beta +1$), the same algebra yields:

$$\mathrm{\Delta }{v}_{\mathrm{E}}=-\frac{1+v_{\mathrm{E}}}{{\tau }_{\mathrm{E}}+1}.$$

Notice the structure of these two equations:

First, both equations contain ${\tau }_{\mathrm{E}}+1$ in the denominator. As an agent’s total evidence ${\tau }_{\mathrm{E}}$ grows, the denominator increases, and the impact of any single new trial shrinks toward zero. This is the law of epistemic inertia: heavily backed beliefs are difficult to accelerate.

Second, examine the numerators. The distance $(1-v_{\mathrm{E}})$ is the remaining headroom between the current velocity and absolute confirmation ($+1$). The distance $(1+v_{\mathrm{E}})$ is the headroom to absolute refutation ($-1$).

Now compare the magnitude of the shift caused by an opposing trial to that caused by a supporting trial:

$$\frac{|\mathrm{\Delta }{v}_{\mathrm{opposing}}|}{|\mathrm{\Delta }{v}_{\mathrm{supporting}}|}=\frac{1+v_{\mathrm{E}}}{1-v_{\mathrm{E}}}=\frac{1+\frac{\alpha -\beta }{\alpha +\beta }}{1-\frac{\alpha -\beta }{\alpha +\beta }}=\frac{\frac{(\alpha +\beta )+(\alpha -\beta )}{\alpha +\beta }}{\frac{(\alpha +\beta )-(\alpha -\beta )}{\alpha +\beta }}=\frac{2\alpha }{2\beta }=\frac{\alpha }{\beta }.$$

Look at that equation:

$$\boxed{\frac{|\mathrm{\Delta }{v}_{\mathrm{opposing}}|}{|\mathrm{\Delta }{v}_{\mathrm{supporting}}|}=\frac{\alpha }{\beta }.}$$

The relative mechanical impact of an opposing observation compared to a supporting observation is exactly the odds of the belief!

Let us see this in concrete numbers:

Suppose an agent has accumulated odds of $9:1$ in favor of a hypothesis (for instance, $(\alpha ,\beta )=(9,1)$, giving mean $\mu =0.90$ and velocity $v_{\mathrm{E}}=+0.80$).

If one more supporting trial arrives, the velocity changes by:

$$\mathrm{\Delta }{v}_{\mathrm{supporting}}=\frac{1-0.80}{10+1}=\frac{0.20}{11}\approx +0.0182.$$

If instead a single opposing trial arrives (a surprise anomaly), the velocity changes by:

$$\mathrm{\Delta }{v}_{\mathrm{opposing}}=-\frac{1+0.80}{10+1}=-\frac{1.80}{11}\approx -0.1636.$$

The surprise kick is exactly $9$ times larger than the confirmatory nudge!

This is a fundamental mechanism of cognitive and artificial intelligence: an expected confirmation merely confirms what was already anticipated, producing negligible learning. But an event that contradicts a deeply held model delivers a massive, attention-grabbing jolt. The mathematics of Beta updates embeds this asymmetry naturally.

A confirmation or a surprise

Which observation packs the bigger punch?

Reset to 9:1 odds

Give the same starting belief either one supporting trial or one opposing trial. Predict which moves it farther, then compare their effects.

Starting mean probability

90.00%

Starting raw trial amount

9

Try support +1 Try opposition +1

Before the trial

Mean probability

90.00%

Epistemic velocity

0.8000

After one opposition trial

Mean probability

81.82%

Velocity change

−0.1636

The opposing trial moves the velocity 9.00× as far as the supporting trial. The less expected opposing outcome gives the larger push.

$$\mathrm{\Delta }{v}_{+}=\frac{1-v}{{\tau }_{\mathrm{E}}+1},\mathrm{\Delta }{v}_{-}=-\frac{1+v}{{\tau }_{\mathrm{E}}+1}$$

Here v is the starting epistemic velocity. The prior contributes half a trial to each channel, so prior-inclusive time is the raw trial amount plus 1. Increase that amount at a fixed mean: both pushes shrink, while their ratio stays the same. These are two alternatives from the same starting state.

Sometimes learning makes the curve wider

Gathering more data can make a belief harder to move while temporarily broadening its distribution.

Let us return to the variance formula:

$$\mathrm{Var}(P)=\frac{\mu (1-\mu )}{{\tau }_{\mathrm{E}}+1}.$$

When the mean $\mu$ remains constant, increasing the total evidence ${\tau }_{\mathrm{E}}$ will always decrease the variance, sharpening the Beta curve into a tall, narrow needle.

However, an unexpected trial does not hold the mean constant: it pulls the mean toward the center. And as $\mu$ moves toward $1/2$, the numerator $\mu (1-\mu )$ grows!

If the numerator grows faster than the denominator ${\tau }_{\mathrm{E}}+1$, the variance will actually increase.

Let us calculate an explicit example:

Start with an agent who has observed 9 consecutive successes and no failures, with the standard neutral prior of half a count in each column:

$$\alpha =9+0.5=9.5,\beta =0+0.5=0.5.$$

The total evidence is ${\tau }_{\mathrm{E}}=9.5+0.5=10.0$.

The mean probability is:

$$\mu =\frac{9.5}{10.0}=0.95.$$

The variance is:

$$\mathrm{Var}(P)=\frac{(0.95)(1-0.95)}{10.0+1}=\frac{(0.95)(0.05)}{11}=\frac{0.0475}{11}\approx 0.004318.$$

The agent’s belief is tightly concentrated near $0.95$.

Now the tenth trial arrives, and it is a failure! The agent updates the ledger:

$${\alpha }^{\prime }=9.5,{\beta }^{\prime }=0.5+1=1.5.$$

The new total evidence is ${\tau }_{\mathrm{E}}^{\prime }=11.0$.

The new mean is:

$${\mu }^{\prime }=\frac{9.5}{11.0}\approx 0.8636.$$

Now calculate the new variance:

$$\mathrm{Var}(P{)}^{\prime }=\frac{{\alpha }^{\prime }{\beta }^{\prime }}{({\tau }_{\mathrm{E}}^{\prime }{)}^2({\tau }_{\mathrm{E}}^{\prime }+1)}=\frac{(9.5)(1.5)}{{11}^2\cdot 12}\approx 0.009814.$$

Compare the two variances:

$$\frac{\mathrm{Var}(P{)}^{\prime }}{\mathrm{Var}(P)}=\frac{0.009814}{0.004318}\approx 2.27.$$

The variance has more than doubled!

More trials can make a wider curve

Break a perfect streak.

Reset to nine successes

A notebook has seen only successes. Add a failure, then keep experimenting. Watch the trial amount grow while the distribution first widens or narrows depending on what arrives.

Successes in the starting streak

9

Add a failure Add a success +1 Undo last trial

Added so far: 0 successes and 1 failures, out of 20 available steps. A failure adds one trial to opposition.

The original streak

The density diverges at p = 1. The curve is sampled only inside (0, 1).

Mean probability

95.00%

Variance

0.004318

After your new trials

Mean probability

86.36%

Variance

0.009814

The variance is now 2.27× its starting value. The last trial widened the distribution.

Raw trials now

10

Next trial's response weight

8.33%

The response weight is the fraction of the distance to the next outcome that the mean moves. It falls as the notebook accumulates trials, even when variance grows. Both plots use the same axes; the prior remains half a trial per channel.

$$\mathrm{Var}(P)=\frac{\mu (1-\mu )}{{\tau }_{\mathrm{E}}+1},\mathrm{\Delta }\mu =\frac{z-\mu }{{\tau }_{\mathrm{E}}+1}$$

Here z is 1 for a success and 0 for a failure. The mean is μ, and prior-inclusive epistemic time is τ_E.

Even though the agent absorbed an additional trial and holds more total evidence than before, their uncertainty about the true rate of success has dramatically widened.

This is a profound epistemic truth: encountering a rare exception shatters premature dogmatism. The observation forces the observer to realize that the world is more complex than a flawless streak implied, reopening possibilities that had seemed closed.

Experience as inertia: Tracelets and physical particles

We can take this physical correspondence further and imagine an epistemic physics where particles carry internal ledgers of experience.

In such a picture, the state of a particle is not just a spatial position and a momentum vector, but a lineage of historical interactions. When particles collide or exchange signals, they exchange discrete packets of evidence: tracelets.

Each tracelet transferred during an interaction updates the receiving particle’s internal $(\alpha ,\beta )$ ledger:

• A positive tracelet adds to $\alpha$, nudging its epistemic velocity toward $+1$.
• A negative tracelet adds to $\beta$, nudging its epistemic velocity toward $-1$.

As opposing tracelets accumulate along a particle’s worldline, its epistemic rest mass $m_{\mathrm{E}}=2\sqrt{\alpha \beta }$ grows. A particle with extensive, contradictory history becomes massive, developing high resistance to further acceleration. Conversely, a particle with pristine, uncontradicted history remains light and agile, reacting instantly to incoming signals.

Let us take the analogy seriously enough to build with it. Imagine a particle whose state consists of histories, with interactions transferring tracelets. Its motion reflects a balance, while accumulated opposing history gives it mass. A collision redistributes those histories and changes both participants’ tendencies. We could then test which exchange rules produce familiar inertial motion, and which create strange new dynamics.

At the scale of autonomous artificial systems or scientific institutions, this structure suggests a robust architecture for decision-making.

Systems should never report a naked probability $\mu$. They should report both the directional tendency $\mu$ (or velocity $v_{\mathrm{E}}$) and the backing mass $m_{\mathrm{E}}$ (or concentration ${\tau }_{\mathrm{E}}$). A medical diagnosis of “70% risk, derived from 5 patient trials” carries a radically different operational meaning from “70% risk, derived from 50,000 clinical trials.” Monitoring the ratio of mass to acceleration allows change-detection algorithms to distinguish between minor statistical fluctuations and genuine phase shifts in the environment.

Give the notebook a memory window

Up to this point, our notebooks have suffered from an occupational hazard: eternal memory.

If every historical trial is retained forever with equal weight, the total evidence ${\tau }_{\mathrm{E}}$ grows without bound. Eventually, the denominator ${\tau }_{\mathrm{E}}+1$ becomes so enormous that the notebook becomes completely deaf to the world. If the climate shifts, an eternal notebook requires thousands of trials just to admit that winter has arrived.

To survive in a non-stationary world, an agent must introduce a memory window.

Let us model a continuous evidence stream where raw evidence fades exponentially at a constant rate $\kappa$ (kappa), while the foundational neutral prior $n_0$ remains an immutable baseline.

Let supporting evidence arrive at continuous rate ${\lambda }_{+}$, and opposing evidence arrive at rate ${\lambda }_{-}$. We denote the time derivative with a dot:

$$\dot{Y}={\lambda }_{+}-\kappa Y,\dot{N}={\lambda }_{-}-\kappa N.$$

Under this linear decay law, an observation recorded at time $t_0$ retains a fractional weight of $e^{-\kappa (t-t_0)}$ at time $t$. The characteristic memory timescale is given by:

$$T_{\mathrm{mem}}=\frac{1}{\kappa }.$$

If $\kappa$ is very small, memory persists for a long time; if $\kappa$ is large, old observations fade quickly.

Let the total arrival rate of evidence be $\lambda ={\lambda }_{+}+{\lambda }_{-}$, and let the environment’s instantaneous driving velocity be:

$$u=\frac{{\lambda }_{+}-{\lambda }_{-}}{\lambda }\in (-1,1).$$

With prior concentration $n_0$, the total effective evidence at any time is ${\tau }_{\mathrm{E}}=n_0+Y+N$.

Differentiating the epistemic velocity $v_{\mathrm{E}}=(\alpha -\beta )/{\tau }_{\mathrm{E}}=(Y-N)/{\tau }_{\mathrm{E}}$ with respect to continuous time produces the dynamical equation of motion:

$${\dot{v}}_{\mathrm{E}}=\frac{\lambda (u-v_{\mathrm{E}})-\kappa n_0{v}_{\mathrm{E}}}{{\tau }_{\mathrm{E}}}.$$

Look at the forces competing in this equation:

1. The first term, $\lambda (u-v_{\mathrm{E}})$, is an informational drag force: the incoming environmental stream pulls the agent’s velocity toward the current environmental bias $u$.
2. The second term, $-\kappa n_0{v}_{\mathrm{E}}$, is an anchor force: as ancient observations fade away, the fixed neutral prior $n_0$ steadily pulls the belief back toward the center of indifference ($v_{\mathrm{E}}=0$).

Follow the seasons: The resonance of memory

What is the optimal memory window for an agent living in a cyclical environment?

Consider an environment that oscillates between favorable and hostile conditions like the passing of seasons:

$$u(t)=a\cos (\omega t),$$

where $a\in (0,1)$ is the amplitude of the seasonal swing and $\omega$ is its angular frequency.

A notebook for changing seasons

How much of the past should it keep?

Reset

The environment oscillates. Old observations fade while the neutral prior stays at one. Change the memory and see how much of the signal the belief follows. A matched resistor-capacitor circuit follows the same signal, with its voltage on the same scale.

Memory timescale

1.60 time units

One environmental cycle

8 time units

New evidence per time unit

12.0

Environment / input voltageBeliefCapacitor voltage

Signal amplitude retained59.2%

Delay behind environment1.14 time units

Settled concentration20.20

Very short memoryLargest responseLong memory

The sky dashes sit on the teal curve: the two systems have the same periodic response after their initial adjustment, with the evidence total settled. “Largest response” maximizes the output amplitude for these settings; its nearest slider value is shown. Long memory smooths the signal. Very short memory leaves the prior with a larger share.

A notebook with fading trials

$${\dot{v}}_{\mathrm{E}}=\kappa (gu-v_{\mathrm{E}})$$

Incoming balance u drives the centered mean v_E.

Memory time 1.60 Gain for slow signals 0.950

A capacitor with a ground leak

$$\dot{V}=\kappa (g{V}_{\mathrm{in}}-V)$$

Input voltage V_in drives capacitor voltage V.

Memory time 1.60 Gain for slow signals 0.950

A dot means change per unit time. The forgetting rate is κ = 1 / memory time, and g is the fraction of a very slowly changing input that gets through. The circuit uses matched units with capacitance C = 1, input conductance κg = 0.594, and ground conductance κ(1 − g) = 0.031. Conductance is the reciprocal of resistance; the ground path supplies the pull toward neutral.

$$\kappa =0.625,n_0=1,{\tau }_{\ast }=1+\lambda /\kappa =20.20$$

In the interactive lab above, adjust the forgetting rate $\kappa$ and watch how the belief tracks the sinusoidal wave. Notice the fundamental trade-off:

• If memory is too long (small $\kappa$), the notebook accumulates a huge history: it filters out noise, but its massive inertia causes it to lag far behind the seasons.
• If memory is too short (large $\kappa$), the notebook responds rapidly, but it forgets so quickly that almost no evidence survives above the prior, collapsing the amplitude of belief.

We can analyze this trade-off analytically.

Under a constant total arrival rate $\lambda$, the total evidence rapidly settles to its steady-state baseline:

$${\dot{\tau }}_{\mathrm{E}}=\lambda -\kappa (Y+N)=0⟹Y_{\ast }+N_{\ast }=\frac{\lambda }{\kappa }⟹{\tau }_{\ast }=n_0+\frac{\lambda }{\kappa }.$$

Substituting ${\tau }_{\ast }$ into the equation of motion for $v_{\mathrm{E}}$ gives a driven linear differential equation:

$${\dot{v}}_{\mathrm{E}}=\frac{\lambda u(t)-(\lambda +\kappa n_0)v_{\mathrm{E}}}{n_0+\frac{\lambda }{\kappa }}=\frac{\kappa }{\lambda +\kappa n_0}[\lambda u(t)-(\lambda +\kappa n_0)v_{\mathrm{E}}].$$

Simplifying the coefficients reveals a standard first-order relaxation equation:

$${\dot{v}}_{\mathrm{E}}+\kappa v_{\mathrm{E}}=\left(\frac{\kappa \lambda }{\lambda +\kappa n_0}\right)u(t).$$

When driven by the harmonic input $u(t)=a\cos (\omega t)$, the steady-state response is an oscillation $v_{\mathrm{E}}(t)=A_{\mathrm{out}}\cos (\omega t-\delta )$.

Using standard complex impedance methods from physics, the frequency response gain (the ratio of output amplitude to input amplitude, $A_{\mathrm{out}}/a$) evaluates to:

$$\boxed{\mathrm{Gain}(\kappa ,\omega )=\left(\frac{\lambda }{\lambda +\kappa n_0}\right)\left(\frac{\kappa }{\sqrt{{\kappa }^2+{\omega }^2}}\right).}$$

Look at the elegant structure of this gain formula:

The second factor, $\frac{\kappa }{\sqrt{{\kappa }^2+{\omega }^2}}$, is a classical low-pass response in signal frequency $\omega$: faster fluctuations are harder to follow. Increasing the forgetting rate $\kappa$ opens that response to faster changes; when $\kappa \gg \omega$, this term approaches 1 and the phase lag becomes small.

However, the first factor, $\frac{\lambda }{\lambda +\kappa n_0}=\frac{1}{1+\kappa n_0/\lambda }$, is an evidence retention efficiency factor: if the forgetting rate $\kappa$ becomes too high, evidence evaporates faster than it arrives. In the limit $\kappa \to \mathrm{\infty }$, this factor plunges to zero because the agent retains no observational evidence above the neutral prior $n_0$.

Between these two opposing effects, there exists a unique, optimal forgetting rate ${\kappa }^{\ast }$ that maximizes response amplitude for a fixed nonzero signal frequency!

To find it, we differentiate the square of the gain with respect to $\kappa$ and set it to zero:

$$\frac{d}{d\kappa }\left[\frac{{\kappa }^2}{(\lambda +\kappa n_0{)}^2({\kappa }^2+{\omega }^2)}\right]=0.$$

Working through the derivative yields the optimal forgetting rate:

$${\kappa }^{\ast }={\left(\frac{\lambda {\omega }^2}{n_0}\right)}^{1/3}.$$

What a striking connection between physics and epistemology!

The optimal memory timescale $T^{\ast }=1/{\kappa }^{\ast }$ scales with the two-thirds power of the seasonal period ($T_{\mathrm{season}}=2\pi /\omega$), balanced against the ratio of the arrival rate $\lambda$ to the prior weight $n_0$.

From migratory birds timing their departure with the changing length of days, to financial algorithms adjusting their moving averages during market volatility, to biological receptors adapting to shifting chemical gradients, this optimization gives us a concrete question to explore: when does maximizing response amplitude capture what a learning system needs?

A migrating animal, a receptor adapting to changing concentrations, a robot deciding which old map readings to trust: each gives this little model a different world to inhabit. We can vary the arrival rate, the timescale of change, or the cost of responding late, then watch which memory works best for that task.

For a trip into the engineering workshop, follow A workshop for beliefs: build this memory as an RC circuit, meet its Kalman-filter relatives, and watch a robot map settle like a network of springs.

The geometry of belief

A belief is never a mere point on a line.

When we preserve the full evidence ledger $(\alpha ,\beta )$:

• A belief becomes a point inside a Minkowski light cone, bounded by the speed of light.
• Clashing evidence generates epistemic rest mass $m_{\mathrm{E}}=2\sqrt{\alpha \beta }$, giving convictions physical inertia.
• Unexpected observations exert mechanical force proportional to odds ($\alpha /\beta$), creating a natural law of epistemic attention.
• Contradictory evidence can widen the belief curve, illustrating how genuine learning shatters premature certainty.
• Continuous memory creates an informational filter with an optimal forgetting rate for the amplitude of a given environmental season.

All of these rich physical phenomena emerge from a single foundational shift: refusing to throw away the information that “one half” left behind.

In our next excursion, we will follow these principles into the biochemistry of living cells in A cell that measures in rapidity, discovering how sensory receptors compute log-odds directly through thermodynamic equilibrium.