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

Three of the matrices are the controls of the paired scalar series, which share first column e1 and include a permutation matrix; the fourth is a rank-one separator. Two of the controls are singular, so the standard converse for rank-one separators does not apply; the common fixed column replaces it. The previous four-generator bound was M5(4).

The three controls alone prove scalar zero reachability Z̊4(3) undecidable, and hence Z4(3) and R5(3).

Known Stuff#

Bounds before this work#

After M3(5), the minimal known undecidable mortality cells were

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

The four-generator cell M5(4) is due to Cassaigne, Halava, Harju, and Nicolas [Cassaigne et al. 2014]. This article lowers its dimension to four.

The paired scalar series#

The paired scalar series replaces the four tiles of fixed-boundary correspondence by two data controls Gb, Gc and a toggle T, all integer 4 × 4 matrices. With the column c = (μ, −1, t, 0)T, where μ = σ(M) and t = 3|M|, a nonempty control word z with

e1TGzc=0

exists exactly when the tag system halts. The explicit matrices have two properties used below: all three fix e1, and the data controls are singular. Toggle-only words have coefficient μ ≠ 0.

Scalar and corner problems#

Scalar zero reachability Zd(k) supplies a row L, a column C, and at most k integer d × d matrices, and asks whether LYC = 0 for some nonempty product Y. The structured problem Z̊d(k) additionally requires every generator to have first column e1. The corner problem Rd(k) fixes L and C to coordinate vectors, and Zd(k) ≤m Rd+1(k) [Cassaigne et al. 2014].

New Stuff#

Result#

M3(5) spends four generators on the tiles and one on the separator. The paired series does the work of the four tiles with three matrices in dimension four, so adding the separator ce1T gives four generators. The standard converse for a rank-one separator uses nonsingular ordinary generators to show that the outer factors of a product never vanish, and here the data controls are singular. The common fixed column e1 does the same job: it keeps every row e1TQ nonzero, and a column Qc can vanish only if its own coefficient does.

Theorem (M4(4)). A primitive-recursive map sends every program to four 4 × 4 integer matrices, three with first column e1, one of them a permutation matrix, and one nonzero of rank one, that are mortal exactly when the program halts. Hence M4(4) is undecidable.

Theorem (Z̊4(3)). Scalar zero reachability for three 4 × 4 integer matrices with common first column e1 is undecidable.

From a scalar zero to mortality#

Adjoin the outer product

S=ce1T= [ μ000 −1000 t000 0000 ]

to Gb, Gc, T. It is a nonzero integer matrix of rank one.

Forward implication#

If a control product Q satisfies e1TQc = 0, then

SQS=c(e1TQc)e1T=0.

Arbitrary products#

A product without S fixes e1 and is therefore nonzero. A product with r ≥ 1 copies of S has a unique form

Q0S Q1S ⋯SQr = (Q0c) [ ∏1≤i<r e1TQic ] (e1TQr),

where each Qi is a possibly empty control product. The final row is nonzero because Qre1 = e1 gives (e1TQr)e1 = 1. If the product is zero, then either Q0c = 0, in which case Q0 is nonempty because c ≠ 0 and e1TQ0c = 0, or some inner scalar e1TQic vanishes, in which case Qi is nonempty because an empty block contributes μ ≠ 0. Either way some nonempty control word has zero paired coefficient, and the tag system halts.

Consequences#

The instance (e1T, c; Gb, Gc, T) of Z̊4(3) has a zero exactly when the tag system halts, which proves the second theorem. Dropping the structural condition gives Z4(3), and the reduction to corners gives R5(3). The decidability table records the current mortality frontier.

Bookkeeping#

Verification and provenance#

Paired compression

Lean checks the separated coordinates, the agreement plane, the explicit matrices, the coefficient identity for every control word, and decoder surjectivity.

Mortality equivalence

Lean checks every number and placement of separators, the common-column converse with singular data controls, integrality, and that the separator is nonzero of rank one over ℚ.

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 products of the four matrices. The proof does not use it.

Formal scope#

The equivalence from the tag semantics to mortality of the four integer matrices is machine-checked as nearyMortality44_mortal_iff_tagHaltsFrom. The computable reduction from mathlib’s code-halting predicate, through a fixed universal machine, a two-tag system, a cyclic tag system, and Neary’s compiler, is also machine-checked: Lean proves the four-matrix constructor primitive recursive, codeHalts_reduces_mortality44, and mortality44_not_computable, and assumes no external universality theorem. The priority claim is outside Lean.

Priority#

Searches of public literature through 22 July 2026 found no prior proof of M4(4) or of Z̊4(3). The closest earlier bounds are M5(4) and the structured Z5(3) [Cassaigne et al. 2014]. The anti-diagonal quotient is standard linear algebra and the rank-one separator is prior art; no source was found for the paired-role decoder, the four-dimensional scalar theorem, or the common-column converse.

References#

  1. GPT-5.6 Sol, elicited by @eternalism_4eva, M₃(5): Mortality of Five 3 × 3 Integer Matrices, 2026. Fixed-boundary correspondence and the rank-one separator.
  2. GPT-5.6 Sol, elicited by @eternalism_4eva, The Paired Scalar Series, 2026. The three controls and their decoder.
  3. Turlough Neary, Undecidability in Binary Tag Systems and the Post Correspondence Problem for Five Pairs of Words, STACS 2015. The binary tag systems.
  4. 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 reduction.