# Dimensional Relational Closure Theory (DRCT) v1

## Status

This document separates three layers:

1. **Proved mathematical core** — finite deterministic systems, plus explicit
   counterexamples outside that regime.
2. **Established neighboring mathematics** — partition refinement/bisimulation,
   automata future equivalence, lumpability, observability, Takens delay
   embeddings, computational mechanics.
3. **DRCT research program** — a proposed synthesis centered on residual-guided
   refinement of an insufficient observation until predictive closure.

It does **not** prove that physical reality is deterministic, nor that the
universe implements DRCT. It proves the internal mathematics of the proposed
representation and states a falsifiable physical hypothesis.

---

# 1. Primitive objects

Let

\[
\Omega
\]

be a state space and

\[
F:\Omega\to\Omega
\]

a deterministic discrete-time evolution.

A context-dependent observation is

\[
P_C:\Omega\to Y_C.
\]

For the formal finite core we write simply \(P\).

The observation identifies states:

\[
x\sim_0 y
\iff
P(x)=P(y).
\]

Thus an observed value is not a complete state. It is an equivalence class
(fiber) of complete states.

---

# 2. Canonical predictive equivalence

Define

\[
x\sim_\* y
\iff
P(F^t x)=P(F^t y)\quad\forall t\ge0.
\]

Call the equivalence classes of \(\sim_\*\) the **canonical predictive closure**
of the observation \(P\), denoted \(\mathcal C_\*\).

This is the deterministic analogue of grouping states by identical observable
future behavior.

## Theorem 1 — Canonical Closure

\(\mathcal C_\*\) is:

1. an equivalence relation;
2. a refinement of the observation partition \(\mathcal C_0\);
3. forward invariant:
   \[
   x\sim_\*y\Rightarrow F(x)\sim_\*F(y);
   \]
4. the unique coarsest forward-invariant refinement of \(\mathcal C_0\).

### Proof

Items 1–3 follow directly from equality of future observation sequences.

For 4, let \(\mathcal R\) be any forward-invariant refinement of
\(\mathcal C_0\). If \(x\sim_{\mathcal R}y\), refinement gives
\(P(x)=P(y)\). Forward invariance gives
\(F^t x\sim_{\mathcal R}F^t y\) for every \(t\ge0\), hence
\(P(F^t x)=P(F^t y)\) for every \(t\). Therefore
\(x\sim_\*y\). Thus every closed refinement \(\mathcal R\) refines
\(\mathcal C_\*\), so \(\mathcal C_\*\) is the coarsest one. QED.

---

# 3. Exact refinement algorithm

Let \(\mathcal C_0\) be the partition induced by \(P\). Recursively define

\[
\boxed{
\mathcal C_{n+1}
=
\mathcal C_n\vee F^{-1}(\mathcal C_n)
}
\]

or equivalently

\[
x\sim_{n+1}y
\iff
x\sim_n y
\ \text{and}\
F(x)\sim_n F(y).
\]

## Theorem 2 — Finite-Horizon Meaning

\[
\boxed{
x\sim_n y
\iff
P(F^j x)=P(F^j y)
\quad
j=0,\dots,n.
}
\]

### Proof

Induction.

For \(n=0\) this is the definition of observation equivalence.

Assume true for \(n\). Then \(x\sim_{n+1}y\) iff:

- their length-\((n+1)\) observable words starting at \(x,y\) agree; and
- the length-\((n+1)\) words starting at \(F(x),F(y)\) agree.

Together these are exactly equality at times \(0,\dots,n+1\). QED.

## Theorem 3 — Finite Convergence

If \(|\Omega|=N<\infty\), the refinement sequence stabilizes after at most

\[
N-|\mathcal C_0|
\]

strict refinements, and its fixed point is \(\mathcal C_\*\).

### Proof

Each strict refinement increases the number of blocks by at least one. No
partition has more than \(N\) blocks. Thus there can be at most
\(N-|\mathcal C_0|\) strict refinements.

At a fixed point \(\mathcal C\),

\[
\mathcal C=\mathcal C\vee F^{-1}\mathcal C,
\]

so \(\mathcal C\) is forward invariant. By Theorem 1 it must refine
\(\mathcal C_\*\). But every finite-stage partition groups states with equal
finite future words, so \(\mathcal C_\*\) refines every \(\mathcal C_n\).
Hence the fixed point equals \(\mathcal C_\*\). QED.

---

# 4. Quotient dynamics

Let

\[
\pi_\*:\Omega\to\Omega/\mathcal C_\*
\]

map a state to its predictive class.

## Theorem 4 — Exact Reduced Law

There exists a unique deterministic quotient map

\[
\bar F:\Omega/\mathcal C_\*\to\Omega/\mathcal C_\*
\]

satisfying

\[
\boxed{
\pi_\*\circ F=\bar F\circ\pi_\*.
}
\]

It is given by

\[
\bar F([x]_\*)=[F(x)]_\*.
\]

### Proof

Forward invariance of \(\mathcal C_\*\) makes the definition independent of the
representative \(x\). Uniqueness follows because \(\pi_\*\) is surjective. QED.

This is the exact mathematical meaning of “find a representation in which the
observed dynamics close.”

---

# 5. Projection defect and emergent uncertainty

Let a random complete state \(X\sim\mu\). Define

\[
Y=P(X),\qquad Y^+=P(F(X)).
\]

At the complete-state level,

\[
H(Y^+\mid X)=0.
\]

At the projected level the one-step uncertainty is

\[
\boxed{
H(Y^+\mid Y).
}
\]

This can be positive even though \(F\) is completely deterministic.

Therefore apparent stochasticity in a reduced description can be generated by
unresolved distinctions inside observation fibers.

This does **not** imply that all physical stochasticity is epistemic. It says
only that projection is one rigorous source of apparent noise.

---

# 6. Predictive deficit — rigorous reinterpretation of D-distance

Let \(Z=Z(X)\) be any descriptor available to an observer. Define

\[
\boxed{
D_\mu(Z)
=
H(C_\*\mid Z),
}
\]

where \(C_\*\) is the random canonical predictive class.

Interpretation:

- \(D_\mu(Z)>0\): descriptor \(Z\) merges predictively distinct states.
- \(D_\mu(Z)=0\): \(Z\) identifies the predictive class almost surely.

This is **predictive truth distance**, not metaphysical distance to complete
reality.

## Theorem 5 — Monotonicity Under Genuine Added Information

For an added observable \(Q\),

\[
\boxed{
D_\mu(Z,Q)\le D_\mu(Z).
}
\]

Moreover,

\[
D_\mu(Z)-D_\mu(Z,Q)
=
I(C_\*;Q\mid Z).
\]

### Proof

This is the standard identity

\[
H(C_\*\mid Z)-H(C_\*\mid Z,Q)
=
I(C_\*;Q\mid Z)\ge0.
\]

QED.

Thus a proposed new “dimension” has a measurable information gain:

\[
\boxed{
G(Q\mid Z)=I(C_\*;Q\mid Z).
}
\]

---

# 7. No-Information-From-Nothing Theorem

Suppose

\[
Q=g(Z)
\]

is only a deterministic re-expression of information already present in \(Z\).

## Theorem 6

\[
\boxed{
D_\mu(Z,Q)=D_\mu(Z)
}
\]

and

\[
\boxed{
I(C_\*;Q\mid Z)=0.
}
\]

### Proof

Since \(Q\) is a deterministic function of \(Z\),

\[
H(Q\mid Z)=0.
\]

Hence

\[
I(C_\*;Q\mid Z)\le H(Q\mid Z)=0.
\]

By Theorem 5 the deficit reduction is exactly this conditional mutual
information. QED.

### Consequence

No algebraic transform of the same present observation can recreate distinctions
already erased by that observation.

A successful refinement must obtain genuinely additional information from at
least one of:

- history/path;
- another simultaneous observable;
- finer measurement resolution;
- external context;
- intervention/control;
- another sensor/channel.

This is the rigorous version of “something cannot come from nothing.”

---

# 8. Minimum auxiliary information

Inside an observed fiber \(P^{-1}(y)\), let

\[
m_y
=
\#\{
\text{distinct }\mathcal C_\*\text{ classes contained in }P^{-1}(y)
\}.
\]

Let \(Q\) be an auxiliary finite-valued coordinate such that \((P,Q)\) exactly
identifies \(\mathcal C_\*\).

## Theorem 7 — Exact Fixed-Width Bound

The minimum possible alphabet size of \(Q\) is

\[
\boxed{
|Q|_{\min}
=
\max_y m_y.
}
\]

Therefore the minimum fixed-width binary budget is

\[
\boxed{
B_{\max}
=
\left\lceil
\log_2\max_y m_y
\right\rceil.
}
\]

### Proof — lower bound

In a fiber containing \(m_y\) distinct predictive classes, \(Q\) must assign
different values to all \(m_y\) classes; otherwise two predictively distinct
states would remain merged. Thus \(|Q|\ge m_y\) for every \(y\).

### Proof — achievability

Within each observation fiber, number its predictive subclasses
\(0,\dots,m_y-1\). Reuse the same labels in different observation fibers.
Then \((P,Q)\) uniquely determines the predictive class and uses exactly
\(\max_y m_y\) symbols. QED.

---

# 9. Average missing predictive information

Under distribution \(\mu\),

\[
\boxed{
B_\mu
=
H(C_\*\mid P)
}
\]

is the Shannon conditional uncertainty in the predictive class after observing
\(P\).

If an auxiliary message \(Q\) permits exact recovery of \(C_\*\) from \((P,Q)\),

\[
H(C_\*\mid P,Q)=0.
\]

Then

\[
H(C_\*\mid P)
=
I(C_\*;Q\mid P)
\le
H(Q\mid P).
\]

Hence:

## Theorem 8

\[
\boxed{
H(Q\mid P)\ge H(C_\*\mid P).
}
\]

So \(H(C_\*\mid P)\) is an information-theoretic lower bound on average
additional information required for exact predictive closure.

---

# 10. Closure spectrum

Define

\[
\boxed{
\delta_n
=
H(\mathcal C_{n+1}\mid\mathcal C_n).
}
\]

Because \(\mathcal C_{n+1}\) refines \(\mathcal C_n\),

\[
\delta_n\ge0.
\]

## Theorem 9 — Telescoping Defect Decomposition

For a finite deterministic system,

\[
\boxed{
\sum_{n\ge0}\delta_n
=
H(C_\*\mid P).
}
\]

### Proof

Since each \(\mathcal C_n\) is a deterministic function of
\(\mathcal C_{n+1}\),

\[
H(\mathcal C_{n+1}\mid\mathcal C_n)
=
H(\mathcal C_{n+1})-H(\mathcal C_n).
\]

Summing telescopes from \(\mathcal C_0=P\) to \(\mathcal C_\*\):

\[
\sum_n\delta_n
=
H(C_\*)-H(P)
=
H(C_\*\mid P),
\]

where the final equality holds because \(P\) is a deterministic function of
\(C_\*\). QED.

This spectrum says **at which future depth the observation first reveals
distinctions it had collapsed**.

The entropy machinery is standard; “closure spectrum” is proposed DRCT
terminology.

---

# 11. Path repair

For invertible \(F\), define the observed past coordinate

\[
\Pi_k(x)
=
(P(x),P(F^{-1}x),\dots,P(F^{-k}x)).
\]

Let \(\mathcal H_k\) be its partition.

Define

\[
\boxed{
\kappa
=
\min\{k:\mathcal H_k\succeq\mathcal C_\*\}.
}
\]

We call \(\kappa\) the **past closure depth**.

## Theorem 10 — Exact Path-Repair Criterion

A depth-\(k\) observed history is sufficient for exact predictive closure iff

\[
\boxed{
\mathcal H_k\succeq\mathcal C_\*.
}
\]

This is immediate from the definition of predictive sufficiency: equal histories
must never merge distinct predictive classes.

## Theorem 11 — Finite Permutations Are Finitely Repairable

If \(\Omega\) is finite and \(F\) is bijective, then

\[
\boxed{\kappa<\infty.}
\]

A coarse universal bound is

\[
\boxed{\kappa\le |\Omega|^2-1.}
\]

### Proof sketch

Consider any pair \((x,y)\) of states. Under the bijection
\((F^{-1},F^{-1})\), the pair moves on a finite permutation of
\(\Omega\times\Omega\), which has at most \(|\Omega|^2\) pair states.

If two states agree in observed past for \(|\Omega|^2\) consecutive pair
positions, the pair orbit has completed a repetition without encountering an
observation mismatch. Because the pair dynamics is a permutation, the entire
pair orbit—including its forward segment—has equal observations. Thus the two
states are predictively equivalent.

Therefore any predictively distinct pair is separated by some past observation
within \(|\Omega|^2-1\) lags. QED.

The bound is intentionally loose; tighter system-specific bounds are possible.

---

# 12. History is not universally sufficient

Let

\[
\Omega=\{0,1\}^{\mathbb Z}
\]

be bi-infinite bit sequences and let \(F\) be the left shift:

\[
(Fx)_j=x_{j+1}.
\]

Observe only

\[
P(x)=x_0.
\]

Choose \(x,y\) satisfying

\[
x_j=y_j\quad\forall j\le0
\]

but

\[
x_1\ne y_1.
\]

Their entire observed pasts are identical:

\[
P(F^{-k}x)=P(F^{-k}y)
\quad\forall k\ge0,
\]

yet their next observations differ:

\[
P(Fx)=x_1\ne y_1=P(Fy).
\]

## Theorem 12 — New-Observable Necessity Exists

There are deterministic systems for which no amount of observed history repairs
the projection.

Thus the decoder has a mathematically forced branch:

\[
\boxed{
\text{projection defect}
\rightarrow
\begin{cases}
\text{increase path/history depth},\\
\text{acquire genuinely new observable information}.
\end{cases}
}
\]

---

# 13. Conditional decoder convergence

Let a current descriptor \(Z\) induce a partition that is coarser than
\(\mathcal C_\*\).

Suppose a candidate library \(\mathcal L\) has the **separation property**:

> whenever \(D_\mu(Z)>0\), at least one available \(Q\in\mathcal L\) strictly
> separates some predictive classes currently merged by \(Z\).

A decoder repeatedly accepts only a \(Q\) with

\[
I(C_\*;Q\mid Z)>0.
\]

## Theorem 13 — Finite Exact Convergence Under a Separating Library

For finite \(\mathcal C_\*\), if every accepted coordinate strictly refines the
current partition toward \(\mathcal C_\*\), the decoder reaches

\[
D_\mu(Z)=0
\]

after finitely many accepted refinements.

### Proof

Every strict accepted refinement increases the number of descriptor classes by
at least one without splitting a true \(\mathcal C_\*\) class. The number of
classes is bounded by \(|\mathcal C_\*|\). Thus only finitely many strict
refinements are possible before the descriptor partition equals
\(\mathcal C_\*\). QED.

This theorem is conditional. It does not guarantee that a chosen sensor/operator
library contains the missing information.

---

# 14. The DRCT master descriptor

The earlier dimensional descriptor can now be sharpened to

\[
\boxed{
\mathfrak D
=
(Z,R,\Gamma,\Pi,C,\epsilon)
}
\]

with:

- \(Z\): current observables;
- \(R\): closure/residual diagnostics;
- \(\Gamma\): algebraic or structural grade;
- \(\Pi\): path/history;
- \(C\): observational/interventional context;
- \(\epsilon\): resolution.

A proposed added coordinate is valuable only insofar as it causes

\[
\boxed{
D_\mu(\mathfrak D,Q)
<
D_\mu(\mathfrak D).
}
\]

The exact information gain is

\[
\boxed{
I(C_\*;Q\mid\mathfrak D).
}
\]

A complexity-penalized search principle is

\[
\boxed{
Q^\star
=
\arg\max_Q
\left[
I(C_\*;Q\mid\mathfrak D)-\lambda K(Q)
\right].
}
\]

---

# 15. D0 redefined

The original intuition used \(D_0\) as “truth.”

The rigorous predictive version is:

\[
\boxed{
D_0
\iff
H(C_\*\mid\mathfrak D)=0.
}
\]

This does **not** mean the observer knows the complete microscopic universe.

It means:

> relative to the specified observation target and dynamics, the descriptor
> contains zero unresolved predictive information.

This distinction is essential.

---

# 16. Physical hypothesis

The mathematics above is proved independently of physics.

The proposed physical hypothesis is:

> **Dimensional Relational Closure Hypothesis.**  
> Observed physical states are context-dependent projections of richer states.
> Apparent randomness, memory and non-Markovian behavior can arise when the
> observation merges states belonging to different predictive closure classes.
> Adding genuinely informative relational, historical, contextual or
> higher-resolution coordinates should monotonically reduce the predictive
> deficit when such information is physically available.

This hypothesis is falsifiable.

It would be weakened by systems where:

1. the proposed new coordinates repeatedly fail to reduce forward-tested
   predictive deficit; and
2. irreducible stochastic models remain necessary even after experimentally
   sufficient state reconstruction.

The theory does **not** assume this cannot happen.

---

# 17. Relation to established mathematics

The following components have strong prior art and must not be claimed as new:

- coarsest stable partition / partition refinement / bisimulation;
- automata future equivalence and minimization;
- Markov-chain lumpability;
- observability and system identification;
- delay-coordinate reconstruction and Takens-type embedding;
- computational mechanics causal states and minimal predictive sufficient
  statistics;
- Shannon conditional entropy and mutual information.

The candidate research contribution is the **combined reverse-engineering
calculus**:

\[
\boxed{
\text{projection}
\to
\text{canonical closure}
\to
\text{deficit}
\to
\text{closure spectrum}
\to
\text{path-repair test}
\to
\text{minimum missing-bit bound}
\to
\text{new-observable search}
\to
D_0.
}
\]

Novelty requires a careful literature review before publication.

---

# 18. What is proved vs not proved

## Proved here

For the stated mathematical models:

- canonical predictive closure exists;
- finite deterministic refinement converges exactly;
- quotient dynamics are well-defined;
- deterministic post-processing cannot recover erased information;
- fixed-width missing-bit bound is exact;
- conditional entropy lower bound holds;
- closure-spectrum entropy telescopes;
- finite bijective systems are repairable by finite observed history;
- deterministic systems exist where even infinite past history is insufficient;
- a separating refinement library yields finite convergence.

## Not proved

- that physical reality is deterministic;
- that all physical noise is hidden signal;
- that complete microscopic reality is finitely reconstructible;
- that DRCT is a new law of physics;
- that the proposed terminology is novel;
- that any real-world decoder can reach 100% prediction with available sensors.

Those are empirical and literature questions, not consequences of the finite
theorems.