# RCHM × Hodge Conjecture — Research Round 4
## Constraint Matroids, Determinantal Phase Thresholds, and Arithmetic Legality

**Status:** Internal research note  
**Version:** Round 4 / v0.4  
**Date:** 2026-09-17  
**Methodological base:** RCHM; Typed Relation & Constraint Algebra; X-Layer Mathematical Completeness  
**Previous rounds:** RCHM × Hodge Conjecture — Rounds 1–3  
**Canonical source:** UTF-8 Markdown

---

# 0. Round 4 Objective

Rounds 2–3 studied two-channel relations:

$$
\Phi_1,\Phi_2:V\to Q
$$

and their projective coupling line

$$
\mathbb P^1.
$$

Round 4 moves to

$$
m\ge3
$$

constraint channels:

$$
\Phi_1,\ldots,\Phi_m:V\to Q.
$$

The goals are:

1. classify minimal cancellation relations;
2. determine when exceptional coupling sets become positive-dimensional;
3. identify a generic threshold in the number of constraint channels;
4. distinguish complex/geometric legality from rational/arithmetic legality;
5. convert a potentially exponential relation search into a structured combinatorial object.

The central result is that two classical structures appear automatically:

$$
\boxed{
\text{determinantal geometry in coefficient space}
}
$$

and

$$
\boxed{
\text{representable matroids in constraint space}.
}
$$

---

# 1. Multi-Constraint System

Let $\mathbb K$ be a field.

Let

$$
V\simeq\mathbb K^n,
\qquad
Q\simeq\mathbb K^q.
$$

Take linear obstruction maps

$$
\Phi_i:V\to Q,
\qquad
i=1,\ldots,m.
$$

A coefficient vector

$$
a=(a_1,\ldots,a_m)\in\mathbb K^m
$$

defines the combined constraint operator

$$
\boxed{
\Psi_a
=
\sum_{i=1}^m a_i\Phi_i.
}
$$

Since scaling $a$ by a nonzero scalar does not change its kernel,

$$
a\sim ca,
\qquad
c\in\mathbb K^\times,
$$

the natural coefficient domain is

$$
\boxed{
\mathcal P
=
\mathbb P^{m-1}(\mathbb K).
}
$$

For each coupling $[a]$ define the legal state domain

$$
L_a
=
\ker\Psi_a.
$$

---

# 2. Multi-Parameter Relation Space as a Linear Matrix Family

Choose bases of $V$ and $Q$.

Each $\Phi_i$ is represented by a matrix

$$
A_i\in\operatorname{Mat}_{q\times n}(\mathbb K).
$$

Then

$$
\Psi_a
\leftrightarrow
A(a)
=
\sum_{i=1}^m a_iA_i.
$$

Thus coefficient space determines a projective linear family

$$
\mathbb P^{m-1}
\longrightarrow
\mathbb P(\operatorname{Hom}(V,Q)).
$$

The relational problem becomes:

> Where does the rank of $A(a)$ drop?

This is a classical determinantal problem.

---

# 3. Rank Stratification

Let

$$
s
=
\min(n,q).
$$

For a sufficiently general family, the generic rank is

$$
s.
$$

Define the rank-$r$ determinantal stratum

$$
D_r
=
\left\{
[A]:
\operatorname{rank}A\le r
\right\}.
$$

It is cut out by all

$$
(r+1)\times(r+1)
$$

minors.

The standard codimension is

$$
\boxed{
\operatorname{codim}D_r
=
(n-r)(q-r).
}
$$

The coefficient-space rank locus is

$$
\boxed{
\Lambda_r
=
\left\{
[a]\in\mathbb P^{m-1}:
\operatorname{rank}A(a)\le r
\right\}.
}
$$

Thus

$$
\Lambda_r
$$

is the pullback/intersection of a determinantal variety with the projective coupling family.

---

# 4. First Rank-Drop Threshold

The first rank drop occurs at

$$
r=s-1.
$$

Its determinantal codimension is

$$
(n-s+1)(q-s+1).
$$

Since

$$
s=\min(n,q),
$$

this simplifies to

$$
\boxed{
c_1
=
|n-q|+1.
}
$$

A generic projective family of dimension

$$
m-1
$$

therefore has expected first-rank-drop dimension

$$
\boxed{
\dim_{\mathrm{exp}}\Lambda_{s-1}
=
m-1-(|n-q|+1).
}
$$

Equivalently,

$$
\boxed{
\dim_{\mathrm{exp}}\Lambda_{s-1}
=
m-|n-q|-2.
}
$$

This yields three generic regimes.

---

## 4.1 Subcritical relation count

If

$$
m<|n-q|+2,
$$

then

$$
\dim_{\mathrm{exp}}\Lambda_{s-1}<0.
$$

A generic coefficient family avoids the first rank-drop variety.

Exceptional relations then require nongeneric geometry, symmetry, special algebra, or other structure.

---

## 4.2 Critical relation count

If

$$
\boxed{
m=|n-q|+2,
}
$$

then

$$
\dim_{\mathrm{exp}}\Lambda_{s-1}=0.
$$

Exceptional couplings are generically expected to be isolated points.

This is the **critical relational threshold**.

---

## 4.3 Supercritical relation count

If

$$
m>|n-q|+2,
$$

then

$$
\dim_{\mathrm{exp}}\Lambda_{s-1}>0.
$$

Exceptional relations are generically expected to form positive-dimensional families.

Thus:

$$
\boxed{
\text{spectral points}
\longrightarrow
\text{spectral curves}
\longrightarrow
\text{spectral surfaces}
}
$$

appear naturally as the number of constraint channels increases.

---

# 5. Square Systems

Suppose

$$
n=q=N.
$$

Then the first rank-drop variety is simply

$$
\det A(a)=0.
$$

Because

$$
A(a)
=
\sum_i a_iA_i
$$

is linear in the coefficients,

$$
\det A(a)
$$

is homogeneous of degree $N$ in

$$
a_1,\ldots,a_m.
$$

If this determinant is not identically zero, the exceptional relation set is a degree-$N$ hypersurface:

$$
\boxed{
\Lambda_{N-1}
=
V(\det A(a))
\subset
\mathbb P^{m-1}.
}
$$

Hence:

### $m=2$

$$
\Lambda_{N-1}
\subset\mathbb P^1
$$

is a finite generalized-eigenvalue spectrum, with $N$ points counted with multiplicity over an algebraic closure for a regular pencil.

### $m=3$

$$
\Lambda_{N-1}
\subset\mathbb P^2
$$

is generically a plane curve of degree $N$.

### $m=4$

$$
\Lambda_{N-1}
\subset\mathbb P^3
$$

is generically a surface of degree $N$.

Thus the Round 3 term **relational spectrum** becomes literally compatible with classical matrix-pencil spectrum when $m=2$, but for $m\ge3$ generalizes from spectral points to projective spectral varieties.

---

# 6. Toy Example: Three Constraints Produce a Conic Phase Boundary

Take

$$
V=Q=\mathbb C^2
$$

and

$$
A(a)
=
\begin{pmatrix}
a_1 & a_2\\
a_3 & a_1
\end{pmatrix}.
$$

Then

$$
\det A(a)
=
a_1^2-a_2a_3.
$$

Therefore the exceptional coupling set is

$$
\boxed{
a_1^2-a_2a_3=0
}
$$

inside

$$
\mathbb P^2.
$$

This is a conic.

Off the conic,

$$
\operatorname{rank}A(a)=2.
$$

On a generic point of the conic,

$$
\operatorname{rank}A(a)=1.
$$

Thus increasing from two to three constraint channels converts isolated exceptional coefficients into a continuous algebraic phase boundary.

---

# 7. Higher Rank-Depth Thresholds

First rank drop is not the whole story.

Let

$$
k\ge1
$$

and ask for rank at most

$$
s-k.
$$

The determinantal codimension is

$$
\boxed{
c_k
=
(n-s+k)(q-s+k).
}
$$

If

$$
n\ge q,
$$

then $s=q$ and

$$
c_k
=
(n-q+k)k.
$$

Writing

$$
d=|n-q|,
$$

gives the symmetric expression

$$
\boxed{
c_k
=
k(d+k).
}
$$

Therefore a generic coefficient family is expected to encounter rank-loss depth $k$ only when

$$
\boxed{
m-1\ge k(|n-q|+k).
}
$$

Equivalently:

$$
\boxed{
m\ge k(|n-q|+k)+1.
}
$$

For square systems:

$$
n=q,
$$

this becomes

$$
\boxed{
m\ge k^2+1.
}
$$

So deeper rank collapse requires quadratically more relation parameters.

This gives a hierarchy of generic relational phase transitions.

---

# 8. Fixed-State View

Now fix a state

$$
v\in V.
$$

Define obstruction vectors

$$
q_i(v)
=
\Phi_i(v)
\in Q.
$$

Place them as columns of a matrix:

$$
\boxed{
O(v)
=
\begin{bmatrix}
q_1(v)&\cdots&q_m(v)
\end{bmatrix}.
}
$$

A legal coefficient relation satisfies

$$
O(v)a=0.
$$

Thus:

$$
\boxed{
K(v)
=
\ker O(v)
}
$$

is the linear relation space among all obstruction channels at $v$.

Its projectivization is the legal coupling fiber:

$$
\boxed{
\mathcal R(v)
=
\mathbb P(K(v)).
}
$$

---

# 9. Constraint Matroid

The columns

$$
q_1(v),\ldots,q_m(v)
$$

define a representable matroid on ground set

$$
E=\{1,\ldots,m\}.
$$

Call this, within RCHM, the **constraint matroid at $v$**:

$$
\boxed{
M_v.
}
$$

A subset

$$
S\subseteq E
$$

is independent exactly when

$$
\{q_i(v):i\in S\}
$$

is linearly independent.

Its rank function is

$$
r_v(S)
=
\dim
\operatorname{span}
\{q_i(v):i\in S\}.
$$

---

# 10. Circuits Are Minimal Legal Relations

A circuit of a representable matroid is a minimally dependent subset.

Therefore:

$$
C\subseteq E
$$

is a circuit exactly when there is a nonzero coefficient vector $a$ with

$$
\operatorname{supp}(a)=C
$$

such that

$$
\sum_{i\in C}a_iq_i(v)=0,
$$

while no proper subset of $C$ admits a nonzero dependence.

Hence:

$$
\boxed{
\text{matroid circuit}
=
\text{support-minimal relational cancellation}.
}
$$

This is a major compression principle.

Instead of treating every subset of constraints as an unrelated candidate, the system can search the circuit structure.

---

# 11. Minimal Relation Size Bound

Let

$$
r(v)
=
\operatorname{rank}M_v
=
\dim\operatorname{span}\{q_i(v)\}.
$$

Every circuit has cardinality at most

$$
r(v)+1.
$$

Since

$$
r(v)\le q,
$$

every minimal linear cancellation uses at most

$$
\boxed{
q+1
}
$$

constraint channels.

Therefore even if

$$
m\gg q,
$$

no support-minimal linear relation needs more than $q+1$ channels.

This is important for AI-native search:

$$
\boxed{
\text{global relation count may be huge, while minimal cancellation remains locally bounded by obstruction dimension.}
}
$$

This resembles the desired AMRAL principle of global graph plus local finite closure.

---

# 12. Circuit Generation of the Relation Space

Take any nonzero relation

$$
a\in K(v).
$$

If its support is not minimal, choose a circuit vector

$$
c
$$

whose support lies inside

$$
\operatorname{supp}(a).
$$

By subtracting a suitable scalar multiple of $c$, at least one support coordinate can be eliminated.

Repeating recursively decomposes $a$ into circuit-supported relations.

Therefore:

$$
\boxed{
K(v)
\text{ is generated by circuit relations.}
}
$$

The full legal relation space can therefore be reconstructed from minimal dependencies.

This does not imply that the number of circuits is always small; it only identifies the canonical minimal building blocks.

---

# 13. Toy Example: A Three-Way Cancellation Circuit

Let

$$
Q=\mathbb R^2
$$

and at one state $v$ suppose:

$$
q_1(v)
=
\begin{pmatrix}
1\\0
\end{pmatrix},
\qquad
q_2(v)
=
\begin{pmatrix}
0\\1
\end{pmatrix},
\qquad
q_3(v)
=
\begin{pmatrix}
1\\1
\end{pmatrix}.
$$

Each obstruction is nonzero.

No pair consisting of $q_1,q_2$ is dependent, and similarly for the other generic pairs.

But:

$$
q_1+q_2-q_3=0.
$$

Thus:

$$
C=\{1,2,3\}
$$

is a circuit.

The unique projective minimal relation is:

$$
\boxed{
[1:1:-1].
}
$$

This is a genuinely three-way emergent legality.

It cannot be reduced to any two-channel cancellation.

---

# 14. Channel-Surplus Law

For fixed $v$:

$$
O(v):\mathbb K^m\to Q.
$$

If

$$
m>q,
$$

then by dimension alone:

$$
\dim\ker O(v)
\ge
m-q
$$

for every $v$.

Thus nontrivial coefficient relations are unavoidable:

$$
\boxed{
m>q
\quad\Rightarrow\quad
\ker O(v)\neq0.
}
$$

If $O(v)$ has full row rank $q$, then:

$$
\dim\mathcal R(v)
=
m-q-1.
$$

Particularly:

### $m=q+1$

A generic state has a unique projective legal relation.

### $m=q+2$

A generic state has a projective line of legal relations.

Thus when the number of relation channels exceeds obstruction dimension, relational legality stops being exceptional and becomes structurally unavoidable.

---

# 15. Two Different Thresholds

Round 4 therefore discovers two distinct thresholds.

## 15.1 Coefficient-side rank-drop threshold

This asks:

> For which coupling coefficients does the operator $\Psi_a$ lose rank?

Generic first threshold:

$$
\boxed{
m\ge |n-q|+2.
}
$$

---

## 15.2 State-side relation-existence threshold

This asks:

> For a fixed state, must some relation among the $m$ obstruction vectors exist?

Dimension threshold:

$$
\boxed{
m>q.
}
$$

These are not the same phenomenon.

The first concerns rank variation of a family of maps.

The second concerns unavoidable linear dependence among obstruction channels at one state.

---

# 16. Matroid Stratification of State Space

As $v$ varies, the configuration

$$
q_1(v),\ldots,q_m(v)
$$

changes.

Therefore the constraint matroid

$$
M_v
$$

may change.

A matroid changes precisely when some collection of minors changes from nonzero to zero or vice versa.

Thus state space can be stratified by dependence pattern:

$$
\boxed{
V
=
\bigsqcup_M V_M,
}
$$

where:

$$
V_M
=
\{v:M_v=M\}.
$$

Inside one stratum, the combinatorial pattern of minimal legal relations is fixed.

Crossing a stratum boundary changes the available circuits.

This gives a second notion of relational phase transition:

$$
\boxed{
\text{phase transition}
=
\text{change of constraint matroid}.
}
$$

---

# 17. Complexity Warning

Matroid stratification is not automatically simple.

General configurations of vectors can produce highly complicated matroid strata.

Therefore:

$$
\boxed{
\text{matroid representation compresses logical dependence structure, but does not guarantee globally simple geometry.}
}
$$

This is important for AMRAL.

The goal is not to pretend the global mathematical graph is simple.

The goal is to give its local dependence structure a canonical representation that AI and algorithms can manipulate.

---

# 18. Arithmetic Legality Gap

For Hodge theory, a crucial issue appears.

The geometric obstruction space naturally lives over

$$
\mathbb C,
$$

but Hodge classes in the conjecture are rational:

$$
H^{2p}(X,\mathbb Q)
\cap
H^{p,p}(X).
$$

Suppose the obstruction vectors are complex:

$$
q_i(v)\in Q_{\mathbb C}.
$$

Define the complex relation space:

$$
K_{\mathbb C}(v)
=
\left\{
a\in\mathbb C^m:
\sum_i a_iq_i(v)=0
\right\}.
$$

But admissible rational class combinations require:

$$
a\in\mathbb Q^m.
$$

Define:

$$
\boxed{
K_{\mathbb Q}(v)
=
K_{\mathbb C}(v)\cap\mathbb Q^m.
}
$$

It is possible that:

$$
K_{\mathbb C}(v)\neq0
$$

but:

$$
K_{\mathbb Q}(v)=0.
$$

Therefore:

$$
\boxed{
\text{complex cancellation}
\not\Rightarrow
\text{rational Hodge relation}.
}
$$

---

# 19. Toy Arithmetic Example

Suppose the cancellation condition is:

$$
a_1+\sqrt2\,a_2=0.
$$

Over $\mathbb C$ or $\mathbb R$, a nonzero relation exists:

$$
[a_1:a_2]
=
[-\sqrt2:1].
$$

But no nonzero rational pair satisfies it:

$$
(a_1,a_2)\in\mathbb Q^2
\quad\Rightarrow\quad
a_1=a_2=0.
$$

Thus:

$$
\mathbb P(K_{\mathbb C})\neq\varnothing,
$$

while:

$$
\mathbb P(K_{\mathbb Q})
=
\varnothing.
$$

This is the simplest model of an **arithmetic legality gap**.

---

# 20. Field-of-Definition Layer

RCHM must therefore add coefficient field to its type information.

A legal relation should be typed as:

$$
\operatorname{Relation}
(
\text{state},
\text{support},
\text{coefficient field},
\text{scope}
).
$$

For example:

$$
\operatorname{Legal}_{\mathbb C}
$$

and:

$$
\operatorname{Legal}_{\mathbb Q}
$$

are different types.

For Hodge applications this prevents an invalid inference:

$$
\text{complex rank drop}
\Rightarrow
\text{rational Hodge class}.
$$

The correct chain is:

$$
\text{complex rank drop}
\to
\text{field-of-definition test}
\to
\text{rational relation}
\to
\text{Hodge legality}.
$$

---

# 21. Proposed Hodge Constraint Matroid

Let:

$$
\gamma_1,\ldots,\gamma_m
$$

be rational Hodge classes at a base hypersurface $X$.

For a deformation direction

$$
v\in T_XU,
$$

let:

$$
\Phi_{\gamma_i}(v)
$$

denote the first-order Hodge obstruction of $\gamma_i$ along $v$.

Define:

$$
\boxed{
M^{\mathrm{Hodge}}_v
=
M
\left(
\Phi_{\gamma_1}(v),
\ldots,
\Phi_{\gamma_m}(v)
\right).
}
$$

This is the proposed **Hodge constraint matroid** at $v$.

A circuit identifies a minimal set:

$$
C\subseteq\{1,\ldots,m\}
$$

for which a nonzero combination:

$$
\sum_{i\in C}a_i\gamma_i
$$

can remain Hodge to first order along $v$, even though no proper subcollection admits the same cancellation.

For Hodge-conjecture relevance, one must then impose:

$$
a_i\in\mathbb Q.
$$

Thus the useful circuits are not merely complex circuits, but rationally realizable circuit relations.

---

# 22. Hodge Research Program Suggested by Round 4

Given many algebraic cycle classes:

$$
\gamma_1,\ldots,\gamma_m,
$$

perform:

### Step 1

Compute first-order Hodge obstruction vectors:

$$
q_i(v)=\Phi_{\gamma_i}(v).
$$

### Step 2

Construct the constraint matroid:

$$
M_v.
$$

### Step 3

Extract circuits.

### Step 4

Filter by coefficient field:

$$
a_i\in\mathbb Q.
$$

### Step 5

For each rational circuit relation:

$$
\gamma_C
=
\sum_{i\in C}a_i\gamma_i,
$$

compute its Hodge locus.

### Step 6

Determine whether the circuit corresponds to:

- ordinary joint deformation;
- cancellation-only deformation;
- new reduced component;
- nonreduced thickening;
- carrier transmutation;
- new witness type.

### Step 7

Repeat across matroid strata of deformation space.

This converts multi-class Hodge exploration into a structured minimal-relation search.

---

# 23. AMRAL Representation

Round 4 suggests a compact graph schema.

For a local state $v$, store:

```text
constraint_state:
  state_id
  obstruction_space_dimension
  channel_count
  matrix_ref
  rank
  coefficient_field
  matroid_signature
  circuit_refs[]
  determinantal_stratum
```

For each circuit:

```text
relation_circuit:
  support[]
  coefficient_vector
  coefficient_field
  minimality_verified
  resulting_class_ref
  legality_order
  witness_ref
```

This allows AMRAL to store minimal relation generators instead of every linear combination.

---

# 24. Connection to X-Layer Completeness

Round 4 gives a concrete reason why X-layer mathematics should retain more than theorem statements.

A theorem stating merely:

$$
\gamma
\text{ remains Hodge}
$$

may hide:

- which channels generated $\gamma$;
- whether the relation is minimal;
- which coefficient field is required;
- whether legality is generic or determinantal;
- which matroid stratum supports it;
- whether the witness carrier changed.

Therefore the canonical graph should preserve:

$$
\boxed{
\text{result}
+
\text{relation circuit}
+
\text{rank stratum}
+
\text{field of definition}
+
\text{witness carrier}.
}
$$

This is exactly the kind of information lost by aggressive linear narrative compression.

---

# 25. Main New Principles

## Principle A — Determinantal Relation Threshold

For a generic $m$-parameter linear family of maps

$$
V^n\to Q^q,
$$

the first rank-drop locus has expected dimension:

$$
\boxed{
m-|n-q|-2.
}
$$

The critical channel count is:

$$
\boxed{
m=|n-q|+2.
}
$$

---

## Principle B — Rank-Depth Law

Generic rank-loss depth $k$ requires:

$$
\boxed{
m\ge k(|n-q|+k)+1.
}
$$

For square systems:

$$
\boxed{
m\ge k^2+1.
}
$$

---

## Principle C — Circuit Minimality

At fixed state $v$:

$$
\boxed{
\text{minimal legal cancellations}
=
\text{circuits of the constraint matroid }M_v.
}
$$

---

## Principle D — Local Support Bound

If obstruction space has dimension $q$, every minimal linear cancellation uses at most:

$$
\boxed{
q+1
}
$$

channels.

---

## Principle E — Channel Surplus

If:

$$
m>q,
$$

then every state has at least one nonzero linear relation among the obstruction channels.

---

## Principle F — Arithmetic Filter

For Hodge applications:

$$
\boxed{
\text{complex relation}
\not\Rightarrow
\text{rationally admissible relation}.
}
$$

Coefficient field is part of relation type.

---

# 26. Round 4 Conclusion

The $m\ge3$ extension changes the qualitative picture.

For two channels, exceptional coupling often appears as isolated spectral values.

For many channels, coefficient space becomes a projective geometry:

$$
\mathbb P^{m-1},
$$

and rank-drop relations form determinantal varieties.

Simultaneously, fixing a state turns obstruction channels into a representable matroid whose circuits encode the minimal cancellations.

Thus the multi-relation system has two complementary canonical structures:

$$
\boxed{
\text{coefficient side}
=
\text{determinantal phase geometry}
}
$$

and

$$
\boxed{
\text{state side}
=
\text{constraint matroid}.
}
$$

The resulting dual picture is:

$$
\boxed{
\begin{array}{ccc}
\text{Coupling space}
&
\longleftrightarrow
&
\text{State space}
\\[1mm]
\text{determinantal strata}
&&
\text{matroid strata}
\\[1mm]
\text{rank-drop varieties}
&&
\text{circuit relations}.
\end{array}
}
$$

For Hodge theory, a third layer must be added:

$$
\boxed{
\text{arithmetic field-of-definition filter}.
}
$$

Therefore the next research object is not merely a Hodge locus, but a three-layer structure:

$$
\boxed{
\text{determinantal phase}
\to
\text{minimal relation circuit}
\to
\text{rational witness search}.
}
$$

This is the main closure of Round 4.

---

# References

1. Remke Kloosterman, **On a conjecture on Hodge loci of linear combinations of linear subvarieties**, Rendiconti del Circolo Matematico di Palermo Series 2, 74, Article 187 (2025).  
   https://doi.org/10.1007/s12215-025-01307-4

2. Jorge Duque Franco and Roberto Villaflor Loyola, **Periods of join algebraic cycles**, arXiv:2312.17222.  
   https://arxiv.org/abs/2312.17222

3. Standard determinantal geometry: rank-$r$ matrix loci are defined by minors, with codimension $(n-r)(q-r)$ in the generic matrix space.

4. Standard representable matroid theory: a matrix determines a matroid on its columns; circuits are minimal linearly dependent column subsets.

5. Standard matrix-pencil theory: for a regular square pencil $A-\lambda B$, generalized eigenvalues are roots of $\det(A-\lambda B)$.