Matrix mortality for four 4 × 4 integer generators is undecidable: M4(4).

This remains true when three generators have common first column e1, one of those three is a permutation matrix, and the fourth is a nonzero integer matrix of rank one over ℚ.

The construction also proves structured scalar zero reachability Z̊4(3) undecidable; consequently Z4(3) and R5(3) are undecidable.

Known Stuff#

The four-role source#

The fixed-boundary correspondence theorem proves undecidability of a four-role source problem.

Fix a deletion width β and write H for Neary’s binary word code, with H(b) = 10β1 and H(c) = 1. For the variable part q of the c-appendant, the four ordinary roles are

Rc =(1,1H(q)10), Dc =(1,0), Rb =(H(b),110), Db =(H(b),0).

Let U(w) and V(w) be the upper and lower concatenations of a role word w, and put M = 10β. The M3(5) result proves

w: U(w)M = V(w) the associated restricted tag system halts.

The soundness proof analyzes every role word. Zero-run synchronization forces exact deletion-width blocks, and a global history equation reconstructs lawful tag steps. Neary’s restricted binary-tag undecidability theorem supplies the external source family [Neary 2015].

Neary padding corollary. Neary’s compiler has β = 10p and permits a computably chosen sufficiently large padding parameter s. Taking s = x(β−1)+1 makes the body length

|q| =βs1 =(xβ+1)(β1).
β>2, β1|q|, β1|q|.

The padded subfamily remains undecidable. These conditions follow from the construction’s padding freedom, not from the bare statement of Lemma 9.

Word-pair matrices#

Recode binary symbols as nonzero base-three digits. For a digit word x, let σ(x) be its base-three value. The standard word-pair representation is

Ψ(u,v)= [ 1 σ(v) σ(u)σ(v) 0 3|v| 3|u|3|v| 0 0 3|u| ].

It is a monoid morphism, and its upper-right entry vanishes exactly when u = v [Cassaigne et al. 2014]. The M3(5) construction assigns one such matrix to each role and one rank-one boundary matrix, giving five 3 × 3 generators.

Previous bounds#

After the M3(5) result, the established Pareto-minimal mortality-undecidability points were

M3(5), M5(4), M6(3), M12(2).

The four-generator entry M5(4) came from earlier packing results. This result lowers its dimension by one.

New Stuff#

Result and mechanism#

The rule and deletion roles for a fixed tag symbol have the same upper word. After separating the upper and lower matrix channels, each pair therefore agrees on a two-dimensional subspace. Two three-dimensional phase spaces can be identified along that subspace, leaving four dimensions. A toggle selects which member of each pair acts. Every toggle pattern decodes to a role word, and every role word has an encoding.

Paired-role compression. Let V be a finite-dimensional vector space over a field, with dim V = d, and let R0, R1, D0, D1 be linear endomorphisms of V. If Ri and Di agree on an r-dimensional subspace E for i = 0,1, then scalar zero reachability for the four actions reduces to scalar zero reachability for three actions in dimension 2dr.

The Neary role matrices have d = 3 and r = 2. The compressed scalar system has three 4 × 4 matrices. A fourth rank-one matrix converts its zero coefficient into a zero matrix product.

The shared channel#

Side-normal form#

Use the unimodular change of basis

P= [ 100 011 001 ].

Conjugating Ψ separates the two word channels:

Φ(u,v) =P1 Ψ(u,v)P = [ 1 σ(v) σ(u) 0 3|v| 0 0 0 3|u| ].

Agreement plane#

On the upper-side plane

EU = span{e1,e3} = {(a,0,z)T},

the lower word disappears:

Φ(u,v) (a,0,z)T = (a+σ(u)z,0,3|u|z)T.

Since Rx and Dx have the same upper word,

Rx |EU = Dx |EU (x{b,c}).

Paired-role compression#

The quotient#

On two copies of the original space define

G^x = [ 00RxDx ], T^ = [ 0II0 ].

The subspace

K={(p,p):pE}

is annihilated by both data matrices and preserved by the toggle. The three maps descend to

W=(VV)/K, dimW=2dr.

The row (T, T) annihilates K; the boundary column (c,0)T descends without restriction. Thus the quotient preserves the required scalar coefficients.

Arbitrary control words#

The quotient matrices are indexed by two data letters and one toggle. Products are written in source-word order, but the rightmost matrix acts first on the boundary column. The decoder is therefore suffix-controlled: scan from right to left with a two-valued phase.

  1. The initial phase is rule.
  2. A toggle flips the phase.
  3. A data letter x emits Rx in rule phase or Dx in deletion phase, then sets the phase to deletion.

Let τ(z) be the emitted role word in the original left-to-right order. Induction on z gives the coefficient identity

L GzC = e1TXτ(z)c.

Leading, trailing, adjacent, and repeated toggles are included. Conversely, induction on a role word inserts at most one toggle before each recursively encoded suffix, so τ is surjective. The scalar system has a zero witness exactly when the four-role source has one.

Explicit scalar system#

For x ∈ {b,c}, write

Ux=σ(ux), Ax=3|ux|, VxR=σ(vxR), BxR=3|vxR|, VxD=σ(vxD), BxD=3|vxD|.

In quotient coordinates (p1+q1, p2, p3+q3, q2)T, the matrices are

Gx= [ 1VxRUxVxD 0000 00Ax0 0BxR0BxD ], T= [ 1000 0001 0010 0100 ].

Let μ = σ(M) and t = 3|M|. The boundary row and column are

L=(1,0,0,0), C=(μ,1,t,0)T.

All entries are integers. Both data matrices have a zero second row; nonsingularity is not used. All three controls fix e1, and T is a permutation matrix. A control word containing no data letter is a power of T and has coefficient LC = μ ≠ 0.

From a scalar zero to mortality#

Adjoin the outer product

S=CL= [ μ000 1000 t000 0000 ].

The vector C and row L are nonzero, so S is nonzero and has rank one over ℚ.

The outer-product reduction from scalar zero reachability to mortality is prior art [Cassaigne et al. 2014]. The new steps are the four-dimensional scalar system and the converse for a control family containing singular data matrices, using its common fixed column.

Forward implication#

If a control product Q satisfies LQC = 0, then

SQS=C(LQC)L=0.

Thus every fixed-boundary solution produces a mortality witness over {Gb, Gc, T, S}.

Arbitrary-product converse#

A product without S fixes e1 and is therefore nonzero. A product with m ≥ 1 occurrences has a unique decomposition

Q0S Q1S SQm,

where each Qi is a possibly empty control product. Expanding the outer products gives

(Q0C) [ 1i<m LQiC ] (LQm).

The terminal row is nonzero because (LQm)e1 = 1. If Q0C = 0, then LQ0C = 0 already supplies a scalar witness; an empty Q0 cannot collapse the nonzero column C. Otherwise the exterior outer product is nonzero, so the full product vanishes only when an internal scalar LQiC vanishes. Empty internal blocks contribute LC = μ ≠ 0.

This covers one separator, adjacent separators, arbitrary exterior blocks, and every separator count. The argument uses the shared fixed column in place of control-matrix invertibility.

Consequences#

Here Zd(k) asks, for supplied L and C, whether

LYC=0

for some nonempty product Y of at most k integer d × d matrices. The decorated form Z̊ requires every generator to have first column e1. The empty product cannot witness this instance:

LC=μ0.

The direct scalar construction proves Z̊4(3) undecidable. Forgetting the restriction gives Z4(3); the established reduction Zd(k) ≤m Rd+1(k) gives R5(3) [Cassaigne et al. 2014].

The four-dimensional separator construction proves M4(4). The series landing page records the current scalar, corner, and mortality frontiers.

Bookkeeping#

Verification and provenance#

Paired compression

Lean checks the side-normal conjugation, agreement plane, explicit matrices, coefficient identity for every control word, and decoder surjectivity.

Mortality equivalence

Lean checks every separator count and placement, the common-column converse with singular data controls, integer reflection, nonzeroness, and rational rank one.

Transitive axioms

Publication-facing declarations use only propext, Classical.choice, and Quot.sound. The project contains no admitted proof or project axiom.

Independent falsifier

A separate exact-arithmetic program checks bounded arbitrary role words, compressed coefficients, decoder coverage, and arbitrary four-matrix products. It is not used in the proof.

Formal scope#

The instance-level equivalence from the restricted tag semantics through the four exact integer matrices is machine-checked. The theorem nearyMortality44_mortal_iff_tagHaltsFrom is the encoded boundary.

Lean checks the explicit four-dimensional realization and its arbitrary-word identity. The abstract 2dr quotient theorem above is paper mathematics, not a separate Lean declaration.

The computable reduction from mathlib’s code-halting predicate through the fixed universal machine, two-tag system, cyclic-tag system, and restricted binary-tag compiler is machine-checked. The four-matrix target equivalence is also machine-checked. A primitive-recursive declaration for this target constructor and its final many-one wrapper remain formalization debt; no external universality theorem is assumed. Frontier transports and bibliographic priority claims remain external to Lean.

Priority#

To our knowledge, after searches through 22 July 2026, no prior proof establishes M4(4), nor scalar zero reachability for three 4 × 4 integer matrices sharing first column e1. The closest CHHN bounds are M5(4) and structured Z5(3) [Cassaigne et al. 2014]. The anti-diagonal quotient is standard linear algebra, and the rank-one scalar-to-mortality separator is prior art; no source was found for the paired-role suffix decoder, the resulting four-dimensional scalar theorem, or the common-column mortality converse. “No prior art found” is limited to publicly discoverable work.

References#

  1. GPT-5.6 Sol, elicited by @eternalism_4eva, Fixed-Boundary Correspondence, 2026. The synchronized four-role source theorem and five-matrix compiler.
  2. Turlough Neary, Undecidability in Binary Tag Systems and the Post Correspondence Problem for Five Pairs of Words, STACS 2015. The restricted binary-tag undecidability source.
  3. Julien Cassaigne, Vesa Halava, Tero Harju, and François Nicolas, Tighter Undecidability Bounds for Matrix Mortality, Zero-in-the-Corner Problems, and More, 2014. Definitions, prior bounds, and the scalar-to-corner reductions.