Narrative Self Café v22: Revenge of The Drunken Golfers (Stochastic Geometric Mechanics vs. Newton)

Posted: May 4, 2026 | Author: Dr. Ernie | Filed under: AI-Powered Essays | Tags: competition, models, philosophy, science, systems |Leave a comment

Sequel to: Narrative Self Café v21b Interlude: The Drunken Golfer (Demystifying The Three Entropies)

Write a skit where Isaac Newton is trying to play a traditional round of golf, but his ball keeps “drifting” toward the wide end of the “shaking tray” triangle (+x) despite him not hitting it. He suspects Sakthivadivel/Ramstead/Friston of “cheating” with magnets or invisible strings.
The trio must explain the “Big Deal”: that they have derived Free Energy without Energy. They must explain that the ball moves because the Information Divergence between the golfer’s start position and the board’s geometry creates a Probability Flux that Newton mistakes for a “Force.”

Prelude: The Shaking Tray

A triangular golf course hums softly—not with energy, but with possibility.

Each position x is a well. Each well branches:

– 1 path left
– b paths right

The tray stutters—not violently, but rhythmically. A sampling. A reconsideration.

A golf ball sits at x = 0.

Without warning, it drifts to x = +1.

---

Scene 1: Newton Loses His Patience

Isaac Newton (slamming his cane):

“This is intolerable. A body at rest does not wander off without cause. Where is the force? Who struck it?”

(He crouches, inspecting the ground.)

“Magnets? Strings? A hidden gradient? Show yourselves!”

(The ball drifts again. x = +2.)

(Newton stands abruptly.)

“Acceleration without force is heresy.”

---

Scene 2: The Architect Speaks

Dalton Sakthivadivel (calmly sketching in the sand):

“You’re looking for a push. There isn’t one.”

(He draws a node:)

– Left: 1 path
– Right: b paths

“From any position, the ratio of forward to backward transitions is b:1.”

(He circles the diagram.)

“The action of a path is not force times distance. It is the log-likelihood of choosing it.”

(He writes:)

Action ∼ log(b^Δx)

“Your ball is not being pushed. It is selecting among paths—and most paths go right.”

---

Scene 3: Newton Doubles Down

Newton (furious):

“Selection? The ball has no mind!”

(He grabs it, places it back at x = 0.)

“Observe: I impose initial conditions.”

(The tray stutters.)

(The ball moves. x = +1.)

(Newton freezes.)

“…again.”

---

Scene 4: The Evangelist Steps In

Maxwell Ramstead (grinning):

“You’re assuming laws are fundamental. They’re not. They’re summaries.”

He gestures to the branching wells.

“This is a Markov process. No memory. Just transitions.”

He taps the ground.

“The drift you see is a probability flux J.”

He draws arrows:

J ∝ (b − 1)

“When b > 1, there are more ways forward than back. So probability flows forward.”

He looks directly at Newton.

“You call that force. We call it imbalance.”

---

Scene 5: Newton Searches for the Missing Quantity

Newton (muttering):

“Flux… imbalance… where is the magnitude? Where is the conserved quantity?”

He looks up sharply.

“Where is the energy?”

Silence.

---

Scene 6: The Grandfather Arrives

Karl Friston (softly):

“You’re still thinking in terms of substance.”

He kneels beside the ball.

“There is no energy here. Only surprisal.”

He traces two positions: x=0 and x=+3.

“The divergence between where the ball is and where it is likely to be…”

He writes:

D_KL(P_current ∥ P_geometry)

“…must resolve.”

The tray stutters.

“The system samples its own possibilities at rate P. That is your ‘shaking.’”

He looks at Newton.

“Each stutter is a question: ‘Where next?’”

---

Scene 7: The Reveal: Force Without Force

Sakthivadivel (finishing the thought):

“The answer is biased.”

Ramstead:

“Because the geometry prefers the right.”

Friston:

“So the divergence decreases by moving right.”

The ball rolls to x = +4.

Newton watches, silent.

---

Scene 8: Newton’s Last Stand

Newton (quietly, dangerously):

“You are telling me… that what I call force…”

He gestures at the drifting ball.

“…is merely the consequence of counting paths?”

Ramstead:

“Yes.”

Sakthivadivel:

“More precisely: the log ratio of paths.”

Friston:

“And the imperative to reduce surprisal.”

---

Scene 9: The Stutter Explained

Newton (suspicious):

“And this… shaking?”

Friston:

“Not heat.”

Sakthivadivel:

“Not energy.”

Ramstead:

“It is the sampling rate P.”

Sakthivadivel:

“The faster the sampling, the faster the system explores its branching structure.”

Friston:

“And thus, the faster it resolves divergence.”

Newton:

“So time itself is… sampling?”

They all nod.

---

Scene 10: The Entropic Force

Newton stares at the widening triangle ahead.

More doors. More paths.

He whispers:

“So the ball moves… because there are more ways to be there than here.”

Friston:

“Yes.”

Ramstead:

“That is the entropic force.”

Sakthivadivel:

“A force with no force.”

---

Scene 11: Collapse of the Old World

Newton grips his cane.

His voice is softer now.

“My laws…”

He looks at the drifting ball.

“…require force.”

Ramstead:

“They emerge when b → 1 and P → 0.”

Sakthivadivel:

“When the landscape becomes symmetric.”

Friston:

“When uncertainty vanishes.”

---

Scene 12: Final Shot

Newton places the ball one last time.

He does not swing.

The tray stutters.

The ball rolls forward.

He does not stop it.

---

Scene 13: Epilogue

Newton (grudgingly):

“Perhaps…”

A pause.

“…gravitation is not a pull…”

He looks out over the triangular course.

“…but a preference.”

Friston smiles.

Ramstead whispers:

“Welcome to the shaking tray.”

Sakthivadivel closes his notebook.

---

Final Line

Newton (walking off):

“Very well.”

A long pause.

“…but I refuse to call it golf.”

---

---

Appendices

---

Appendix I: Beyond F=ma

I.1 The Newtonian Complaint

Newton wants motion to mean:

F = ma

Something pushes.
Something accelerates.
Something resists.

But on the Shaking Tray, there is no:

• mass
• gravity
• club
• physical energy
• hidden field

Only:

• position x
• branching factor b
• sampling probability P
• probability distribution over wells
• information divergence
• probability flux

So the question becomes: What looks like force when nothing is pushing?

I.2 The Entropic Replacement

The “force” is not a cause. It is a statistical asymmetry.

At each well:

• there is 1 way to move left
• there are b ways to move right

So the rightward state-space volume is larger.

The ball drifts right not because it is compelled by force, but because repeated sampling makes rightward futures more numerous.

This is the core of an entropic force:

apparent motion caused by unequal numbers of accessible configurations.

I.3 The Stutter P

The shaking probability P is not temperature.

It is the sampling rate of geometry.

Each stutter asks:

Given the local branching structure, which neighboring well is sampled next?

So:

• low P: slow exploration
• high P: rapid exploration
• P = 0: Newton gets his frozen world back

No shaking, no updating.
No updating, no flux.

I.4 Probability Flux J

The probability current J is the net directional flow of likelihood across wells.

In the simplest cartoon:

J ∼ P (b/(b+1) − 1/(b+1))

So:

J ∼ P (b−1)/(b+1)

This says:

• when b = 1, no drift
• when b > 1, drift toward +x
• when P = 0, no motion
• when P increases, drift appears faster

This is the “acceleration” Newton mistakes for force.

I.5 Information Divergence

The board encodes a preferred distribution:

Q(x) ∝ b^x

There are exponentially more rightward micro-paths than leftward ones.

If the ball starts sharply localized at x=0, its actual distribution P_ball is far from the board’s combinatorial geometry.

That mismatch is:

D_KL(P_ball ∥ Q)

The tray’s stutter reduces this mismatch by letting probability flow toward states with more accessible continuations.

Motion becomes: the relaxation of informational mismatch.

I.6 Where Newton Reappears

Newton is not wrong. He is a limiting case.

His world reappears when:

• b ≈ 1
• P is very small or effectively averaged out
• fluctuations are ignored
• path-counting is replaced by smooth trajectories
• probability flux is re-described as deterministic acceleration

Then the stochastic geometry hardens into classical mechanics.

What was once:

J = probability flux

gets renamed:

a = acceleration

And what was once:

∇ D_KL

gets renamed:

F

I.7 The Big Deal

This is free energy without energy.

Not free energy in joules.
Not thermodynamic heat.
Not temperature.

Rather: a purely informational tendency for distributions to move toward lower divergence.

The “force” is the slope of possibility-space.

The “mass” is whatever slows updating.

The “trajectory” is the most probable path through a branching geometry.

---

Appendix II: From Classical to Geometric Action

II.1 The Old Action: “Which Path Costs Least?”

In classical physics, the universe does not just ask:

What force is acting right now?

It asks a deeper question:

Which whole path is best?

This is the idea of least action.

Instead of calculating motion step by step using:

F = ma

we assign every possible path a score called action:

S = ∫ L dt

where L is the Lagrangian.

Usually:

L = T − V

meaning:

• T: kinetic energy
• V: potential energy

The real path is the one where S is stationary: a tiny change in the path does not improve it.

II.2 Why This Sounds Like Magic

Imagine throwing a ball.

Newton says:

At each moment, forces determine acceleration.

Least action says:

The ball takes the path that makes the whole journey fit together most efficiently.

These are equivalent in ordinary mechanics, but least action is more powerful because it talks about paths, not just pushes.

That matters because the Shaking Tray is all about possible paths.

II.3 The New Setting: No Energy Allowed

On the Shaking Tray, we removed:

• mass
• gravity
• kinetic energy
• potential energy
• joules
• temperature

So we cannot use:

L = T − V

There is no T.
There is no V.

But we still have paths.

The ball can go:

0 → 1 → 2 → 3

or:

0 → −1 → 0 → 1

or many other sequences.

So the question becomes: If there is no energy, what scores a path?

II.4 Path Counting Replaces Energy

At each well x:

• there is 1 door left
• there are b doors right

So moving right is not “easier” energetically. It is simply more numerous.

The score of a path comes from how many ways that path can happen.

If a path moves right R times and left L times, its multiplicity is roughly:

b^R

because each rightward move has b choices.

Taking logs gives:

log(b^R) = R log b

That is the geometric action.

Not energy accumulated over time.
Information accumulated along a path.

II.5 The Geometric Action

Classical action:

S = ∫ L dt

Geometric action:

S_geo ∼ −log(path probability)

A highly likely path has low action.
A very unlikely path has high action.

So instead of:

nature chooses the path of least energy-cost

we say:

the system samples paths with lowest surprise-cost

This connects directly to information theory.

II.6 The Role of the Stutter P

The stutter P does not push the ball.

It determines how often the tray samples the next step.

Think of P as asking:

Does the system update now?

If no, the ball stays in the same well.
If yes, it samples the local geometry.

Then:

Pr(right) = b/(b+1)

Pr(left) = 1/(b+1)

So P controls the rate of updates, while b controls the bias of those updates.

II.7 From Random Steps to Smooth Motion

At first, the ball jitters.

Right.
Right.
Left.
Right.
Pause.
Right.

It looks random.

But over many steps, the average motion becomes smooth.

The expected step is:

⟨Δx⟩ = P(b/(b+1))(+1) + P(1/(b+1))(−1)

So:

⟨Δx⟩ = P(b−1)/(b+1)

That average drift is what Newton mistakes for acceleration.

II.8 Why Calculus Enters

Calculus lets us pass from tiny jumps to smooth curves.

Instead of discrete wells:

x = 0, 1, 2, 3, …

we approximate position as continuous:

x(t)

Instead of summing step-scores:

S_geo = ∑ step cost

we write:

S_geo = ∫ ℒ_geo(x, ẋ) dt

This looks like ordinary action, but the meaning has changed.

The “Lagrangian” is no longer energy difference.
It is an information-cost rate.

II.9 The Big Translation

Classical mechanics says:

path → energy action

Geometric mechanics says:

path → information action

Classical least action:

δS = 0

Geometric least action:

δS_geo = 0

Same grammar.
Different alphabet.

II.10 Newton’s World as a Special Case

Newton gets his familiar world when the random sampling becomes so regular that it can be approximated by a smooth deterministic path.

Then:

• probability flux looks like velocity
• changes in flux look like acceleration
• information gradients look like force
• geometric action looks like classical action

So Newton is not destroyed.

He is compressed.

His laws are what the Shaking Tray looks like after the randomness is averaged, smoothed, and renamed.

---

Appendix III: Relative Entropy & The Arrow of Time

III.1 The Missing Fourth Entropy

The previous “three entropies” were roughly:

• Shannon entropy: uncertainty in a message
• Boltzmann entropy: number of hidden microstates
• Thermodynamic entropy: macroscopic irreversibility

But the Shaking Tray needs a fourth:

• Relative entropy: mismatch between two probability distributions

That is:

D_KL(P ∥ Q)

It does not ask:

How random is the ball?

It asks:

How wrong is the ball’s current distribution compared to the geometry?

III.2 The Two Distributions

On the Shaking Tray, there are always two stories.

III.2.1 The Ball’s Story

The ball says:

P(x)

“I am here.”

Perhaps sharply localized at x=0.

III.2.2 The Board’s Story

The board says:

Q(x) ∝ b^x

“There are more ways to be over there.”

The farther right you go, the more branching futures exist.

III.3 Relative Entropy Is Tension

So D_KL(P ∥ Q) measures the tension between:

• where the ball currently is
• where the geometry says probability naturally accumulates

If P is concentrated at the narrow end, but Q favors the wide end, the mismatch is high.

The system does not need energy to move.

It only needs unresolved mismatch.

III.4 The Arrow of Time

The arrow of time is not merely:

entropy increases

On the Shaking Tray, it is sharper:

relative entropy decreases

The system moves from:

P_now ≠ Q_geometry

toward:

P_later ≈ Q_geometry

That relaxation defines “later.”

Time points in the direction of decreasing informational mismatch.

III.5 Why Newton Cannot Reverse It

Newton wants reversible laws.

If the ball moves right, he thinks:

Reverse the velocity and it should move left.

But the Shaking Tray has no velocity to reverse.

It has a branching asymmetry.

At every stutter:

• right has b doors
• left has 1 door

So the reverse path is not impossible.
It is just exponentially outnumbered.

That is irreversibility without force.

III.6 The Stutter Creates History

The stutter P samples the geometry.

Each sample converts possibility into path:

many possible futures → one actual next step

After many stutters, the ball has a history.

Not because a clock pushed it forward.

Because repeated sampling made one direction statistically dominant.

This is the Markov process version of becoming.

III.7 Contrast With the Three Entropies

Shannon entropy says:

How uncertain is the message?

Boltzmann entropy says:

How many hidden arrangements fit this macrostate?

Thermodynamic entropy says:

Why does heat-like disorder not naturally undo itself?

Relative entropy says:

How far is this system from the distribution implied by its world?

For the Drunken Golfer, relative entropy is the key one.

Because the ball is not merely becoming more random.

It is becoming better matched to the geometry.

III.8 The Crucial Distinction

High entropy is not the same as low relative entropy.

A distribution can be:

• very spread out but badly matched
• very concentrated but well matched
• random but still wrong
• ordered but still unstable

The Shaking Tray does not worship disorder.

It minimizes mismatch.

That is why “entropy increases” is too blunt.

The better phrase is:

the system relaxes toward its natural measure.

III.9 The Entropic Force Revisited

Newton sees:

motion

and infers:

force

But the New Guard sees:

∇ D_KL

a slope in mismatch-space.

The apparent force is just the direction in which relative entropy falls fastest.

The ball is not pulled by the future.

It is released from an improbable present.

---

Appendix IV: Who and Why SGM

IV.1 What Is “SGM” Actually?

Stochastic Geometric Mechanics (SGM) is not a single theory.
It is a convergence:

• stochastic processes (random sampling)
• geometric mechanics (structure of state space)
• information theory (probability & divergence)
• variational principles (action over paths)

In short:

SGM asks: what if motion is probability flow on a structured space?

Not particles in force fields.
Not energy being exchanged.

But distributions evolving on geometry.

IV.2 The Cast of Characters

IV.2.1 The Geometers

Darryl D. Holm

• Brings classical mechanics into the stochastic world
• Shows how randomness still obeys variational principles

His work answers:

Can noise still be lawful?

Answer: yes—if geometry constrains it.

IV.2.2 The Information Theorists

Dalton Sakthivadivel

• Replaces energy with probability
• Rewrites action as log-likelihood of paths

Inspired by:

• Information geometry
• Statistical manifolds

His question:

What if mechanics is just inference over trajectories?

IV.2.3 The Philosophers of Inference

Maxwell Ramstead and Karl Friston

They push the radical claim:

Systems behave as if they are minimizing surprise.

Using:

• Free Energy Principle
• Markov blankets
• Bayesian updating

Their move:

Turn physics into epistemology.

IV.2.4 The Emergent Gravity Rebels

Erik Verlinde

He asks:

What if gravity is not fundamental?

Answer:

• It emerges from entropy gradients
• Motion arises from counting microstates

This connects directly to the Shaking Tray:

more ways → more pull (without pulling)

IV.2.5 The Intuition Builders

Peter G. Wolynes

He gives us:

• energy landscapes
• funnel dynamics
• biased stochastic exploration

His insight:

randomness + geometry = reliable direction

IV.3 Why This Is Happening Now

This convergence did not happen by accident.

It solves real problems.

IV.3.1 Classical Mechanics Breaks Under Uncertainty

Newton works when:

• systems are isolated
• noise is negligible
• trajectories are smooth

But real systems:

• fluctuate
• interact
• adapt
• learn

SGM handles: motion under uncertainty

IV.3.2 Thermodynamics Lacked Geometry

Thermodynamics had:

• entropy
• irreversibility

But weak connection to:

• geometry of state space
• path structure

SGM adds: structure to probability

IV.3.3 Machine Learning Changed the Language

Modern AI relies on:

• optimization of probability distributions
• gradient descent
• likelihoods and divergences

SGM says:

physics may be doing the same thing

Not metaphorically—literally.

IV.3.4 Biology Forced the Issue

Living systems:

• maintain structure
• resist equilibrium
• adapt via feedback

Friston’s framework reframes life as: inference in action

SGM generalizes that idea beyond biology.

IV.4 What SGM Replaces

Classical Concept → SGM Replacement

• Force → Probability gradient
• Energy → Log-probability / divergence
• Mass → Resistance to updating
• Trajectory → Most probable path
• Time → Sampling process

Same roles.
Different ontology.

IV.5 What SGM Keeps

SGM does not throw everything away.

It preserves:

• variational principles
• conservation-like structure
• geometry of motion
• predictive power

Newton survives as a limit case.

IV.6 The Deep Motivation

All these researchers are circling the same question:

Why do systems move in structured ways despite randomness?

SGM answers:

because geometry biases probability, and probability flows

IV.7 The Risk

This framework is powerful—but not settled.

• Not yet a unified theory
• Competing interpretations
• Hard mathematics
• Easy to overextend philosophically

It can explain much—but must be used carefully.

IV.8 The Big Picture

SGM is an attempt to unify:

• physics
• information
• probability
• geometry

into one statement:

The world evolves by exploring its own possibilities under constraint.

---

Appendix V: Shaking Tray as SGM’s Atom

V.1 Why an “Atom” Is Needed

Every successful theory has a simplest unit:

• Newton → point mass
• Thermodynamics → gas particle
• Quantum mechanics → wavefunction

SGM needs its own primitive.

Not a particle.
Not a force carrier.

But a minimal system where probability, geometry, and dynamics are inseparable.

That primitive is the Shaking Tray.

V.2 What Makes It Fundamental

The Shaking Tray is irreducible because it contains, in the smallest possible form:

• State space: discrete wells x
• Geometry: asymmetric branching b:1
• Dynamics: stochastic updates via P
• Directionality: probability flux J
• Mismatch: relative entropy D_KL

Remove any one of these, and motion disappears.

V.3 The Minimal Ingredients

V.3.1 Geometry (b)

Defines the shape of possibility.

• b = 1: symmetric world → no drift
• b > 1: expanding world → drift toward +x

This is not a force.
It is a counting asymmetry.

V.3.2 Sampling (P)

Defines when updates happen.

• P = 0: frozen world
• P > 0: exploration begins

This replaces time as a background parameter.

Time = repeated sampling of structure

V.3.3 State (x)

Defines where the system is.

But crucially:

the system is not a point—it is a distribution over x

Even a single ball represents a probability evolving across wells.

V.3.4 Divergence (D_KL)

Defines why motion happens.

Mismatch between:

• current distribution P(x)
• geometric distribution Q(x)

creates pressure for change.

V.3.5 Flux (J)

Defines how motion appears.

J ∼ P (b−1)/(b+1)

Flux is not caused by force.

It is the visible trace of repeated biased sampling.

V.4 The Atomic Process

One full “tick” of the Shaking Tray:

1. The system samples (with probability P)
2. It evaluates local branching structure
3. It transitions according to ratios b:1
4. The distribution shifts
5. Relative entropy decreases (on average)

That is the entire engine.

No hidden variables.
No external drivers.

V.5 Why This Is Enough

From this tiny structure, you can build:

• drift (biased random walk)
• diffusion (variance over time)
• irreversibility (path asymmetry)
• effective forces (from gradients in b)
• smooth trajectories (via averaging)

In other words:

everything Newton needed—but without forces

V.6 Scaling Up

Real systems are just many coupled trays:

• multiple dimensions → lattices in x, y, z
• varying b(x) → curved geometry
• interacting distributions → fields

This leads toward:

• stochastic differential equations
• information geometry
• continuous manifolds

The simple tray becomes a full SGM universe.

V.7 Why It’s Not Just a Toy

The Shaking Tray is not a metaphor.

It is a minimal working model of:

• Markov processes
• biased diffusion
• entropic forces
• variational inference

It captures the same logic used in:

• statistical physics
• machine learning
• biological self-organization

V.8 The Deep Shift

In Newton’s atom:

motion is caused by forces acting on objects

In the Shaking Tray:

motion is caused by structure acting on probabilities

The object is secondary.
The distribution is primary.

V.9 The Compression

All of SGM can be compressed into the tray:

• geometry → b
• time → P
• dynamics → sampling
• causation → divergence reduction
• motion → flux

Nothing extra is needed.

---

Appendix VI: The Big Inversion

VI.1 The Analogy Breaks Here

In the macro world:

shaking adds energy, energy creates motion.

That is ordinary physics.

So if we picture a real tray, we accidentally smuggle in:

• work
• heat
• vibration
• kinetic energy
• hidden forcing

That is not the claim.

VI.2 The SGM Claim

In SGM, the order is reversed:

stochasticity is primitive.

Not energy first.
Not force first.
Not motion first.

Instead:

sampling + geometry → probability flux → apparent motion → effective force/energy language

VI.3 The Classical Story

Classical mechanics says:

energy gradients cause motion.

So we write:

F = −∇V

and then:

F = ma

Motion is downstream of force.

VI.4 The Inverted Story

The Shaking Tray says:

asymmetric possibilities create biased updating.

That biased updating appears as:

J

Then, after coarse-graining, Newton calls it:

F

So force is not fundamental.

It is the shadow of probability flux.

VI.5 Why “Shaking” Misleads

The tray is not physically shaking.

It is stuttering.

Each stutter means:

the system samples its next possible state.

So P does not add energy.

It sets the cadence of becoming.

VI.6 Energy as a Later Language

Energy becomes useful only after we compress many stochastic updates into smooth macroscopic behavior.

Then we can say:

• this path has “cost”
• this gradient has “potential”
• this resistance looks like “mass”
• this drift looks like “force”

But those are emergent bookkeeping concepts.

Not the original furniture of the model.

VI.7 The Big Inversion

Classical physics:

energy explains motion.

SGM:

structured stochasticity explains why energy language works.

---

Appendix VII: From Action to Energy

VII.1 Start With Action

On the Stuttering Tray, the primitive quantity is not energy.

It is geometric action:

S_geo(γ) = −log Pr(γ)

where γ is a path through the wells.

Action asks:

How surprising was this path?

VII.2 Step Action

For a single transition:

a_t = −log Pr(x_t+1 ∣ x_t)

So:

S_geo(γ) = ∑_t a_t

Right steps and left steps have different action-costs because the geometry has different path-counts.

VII.3 Energy as Action-Rate

Energy appears only after we ask:

How much action accumulates per stutter?

So define discrete effective energy as:

E_disc(t) = ΔS_geo = S_geo(t+1) − S_geo(t)

which is just:

E_disc(t) = a_t

In words:

energy is action-per-stutter.

VII.4 Expected Energy

Individual steps fluctuate.

A right step may have low action-cost.
A left step may have high action-cost.

So the conserved object is usually the expected action-rate:

𝔼[E_disc] = ∑_x’ Pr(x’ ∣ x)[−log Pr(x’ ∣ x)]

This is the local transition entropy.

VII.5 Stable Geometry

If the sampling geometry is stable:

• fixed b(x)
• fixed update rule
• fixed transition kernel
• fixed boundaries

then the action-cost structure does not change.

The distribution P_t(x) may evolve, but the rules assigning action-costs remain invariant.

VII.6 Conservation in the Stable Limit

So:

E_disc

is conserved only in the stable-rule sense:

the expected action-rate remains invariant under stationary stochastic geometry.

More carefully:

• along one sampled path: energy fluctuates
• across an ensemble: expected energy is stable
• under changing geometry: energy is not conserved

VII.7 Relation to Relative Entropy

Meanwhile:

D_KL(P_t ∥ Q)

can decrease over time.

That does not violate conservation, because D_KL is not the same thing as energy.

• Energy: action-rate under rules
• Relative entropy: mismatch between current distribution and geometry

A system can conserve expected action-rate while still reducing mismatch.

VII.8 The Translation to Classical Physics

Classical mechanics begins with:

E = T + V

SGM begins with:

S_geo = −log Pr(γ)

Then energy emerges as the time-translation invariant rate of action accumulation.

This mirrors Noether’s theorem:

stable rules imply conserved quantities.

VII.9 The Slogan

Action is primitive. Energy is derived. Conservation is symmetry.

Or sharper:

Energy is conserved action-rate under stationary stochastic geometry.

---