Tensor semantics · Learning · Languages

A grammar that
means something.

Give every wire a space, every box a map, and every rewrite a proof. Then ask what a learned system can compute.

01 · Birdtracks with components

A closed wire has a value.

A crossing is a specified permutation. An adjoint line contracts a normalized generator basis. The coefficients and dimensions remain part of the equation.

A₂ = (I − SWAP)/2 acts on ℂᴺ ⊗ ℂᴺ. The closed identity wire evaluates to N.
Alternating rank
Idempotence residual
Closed loop

The fundamental color contraction

Σₐ (Tᵃ)ⁱⱼ (Tᵃ)ᵏₗ = ½(δⁱₗδᵏⱼ − δⁱⱼδᵏₗ/N)

Tr(TᵃTᵇ) = δₐᵦ/2. Choose external indices; the left side sums actual complex matrix entries.

Generator contraction
Delta contraction
Residual

What the identity guarantees

The antisymmetrizer is an orthogonal projector in characteristic zero. Its trace equals its rank, N(N−1)/2. The Fierz identity follows from an orthonormal basis of all complex matrices with the scalar identity removed. These finite representation identities supply color factors; a Yang–Mills continuum measure and mass gap require the separate analytic arguments described in the paper.

Integral Schur modules require the earlier straightening construction when factorials cannot be inverted. Open the wedge and standard-monomial examples →

02 · A semantic certificate

Change the circuit.
Check the same map.

The local grammar cancels adjacent inverse gates on exactly the same ordered wires. The full complex matrix is evaluated before and after the rewrite, including its phase.

Gate count
Compared matrix size
Maximum entry difference
Cancelled pairs

Nonzero output probabilities from the input |0…0⟩
Basis stateProbability
Soundness, termination, and what is formalized

Every cancellation uses U†U = I. Composition in a typed context preserves equality, and each cancellation decreases the gate count by two. The simplifier terminates, although it does not find every possible equality. A learned policy may choose among certified rules without weakening their soundness.

The Lean 4 module checks a typed sequential linear-map syntax, its contexts, and finite rewrite derivations. The paper proves tensor-context and measurement consequences. This browser uses double-precision arithmetic; its residual is a finite numerical check. Qubit 0 is the most-significant basis bit. The explicit limit is six qubits and 80 gates.

03 · Expressivity depends on the definition

A simple path can spell
a non-context-free language.

A six-port nonterminal grows three paths together. The starting graph connects them with two fixed separator edges. Erasing those separators yields aⁿbⁿcⁿ.

Three-component derivation

Left-to-right path yield

Graph treewidth: 1.

The unbounded string language is not context free. “Context-free graph grammar” names a graph replacement formalism; it does not classify every string yield as Type 2.

Different questions require different classifications
ModelClassified objectResult
This bounded circuit labFinite serializationsFinite, hence regular
Fixed-interface graph replacementGenerated graphsBounded width under the paper’s hypotheses
Unrestricted effective graph rewritingExistence of an accepting derivationRecursively enumerable; no general termination algorithm
Uniform polynomial quantum circuitsBounded-error measurement decisionsBQP; a complexity class, not another Chomsky level

04 · An explicit neural mechanism

Recursion can be
an invariant of the state.

This fixed ReLU recurrence tracks bracket balance and records whether any prefix went negative. Its weights are constructed analytically. No training claim is attached to this example.

Use ( and ). The count c is updated by +1 or −1; the error flag is retained once a prefix is invalid.

Accepted
Final count
First invalid prefix
Why this is a counterexample to a categorical computational prohibition

Under exact integer arithmetic the same fixed recurrence recognizes the unbounded one-bracket Dyck language. The proof is by induction on the count and persistent error flag. A finite automaton cannot recognize that language. The growing count uses O(log L) bits on length L; fixed bounded hardware still has finite resources.

That proves neural representability. It does not show that gradient descent discovers this recurrence, nor that a particular language model generalizes to arbitrary depth. The manuscript separates those questions and reviews the primary linguistic experiments and their published critiques.

05 · Mechanism and information

The same observations.
Different counterfactuals.

Both models produce X = Y with equal probabilities for 0 and 1. Their responses to an intervention differ. The finite Boolean mechanisms have exact ReLU implementations.

Mechanism A

X = U; Y = X

Mechanism B

X = U; Y = U

With U unobserved and fair, do(X = 0) gives Y = 0 surely in A, but P(Y = 1) = ½ in B. No observational-only learner can distinguish these identical input distributions. Additional causal assumptions or interventions are required.

The argument in the paper

A direct critique with explicit obligations.

The chapter rejects Chomsky’s categorical claim that neural methods cannot yield understanding in principle. It proves concrete capabilities, distinguishes idealized expressivity from learned behavior, and includes Hunter’s critique of the “impossible languages” experiment. An operational semantic specification can be tested; a universal claim about consciousness is not inferred from it.