Theorem 3.4 — simultaneous separation of an aligned partial direct sum
OpenGraphTensorCP.Supplement.theorem_3_4direct-sumgraph-tensorsp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1tensor-degenerationtensor-rank
Let be any field, let and , and let be the aligned partial direct sum of order- tensors , sharing coordinates on the first legs and occupying separate sectors on the final two legs. Set
Then the full direct sum of copies of formally degenerates to
The direct sum on the right separates every tensor leg. Zero blocks are permitted and retain their specified ambient dimensions.
This theorem supplies the exact linear-cost simultaneous separation step used by the paper’s entropy inequality.
Formalization Note Legs are 0-based, so the final two legs are k-2 and k-1. MME.Degenerates X Y means that the target degenerates from the source .
Differences from the source:
- generalization: the source uses coordinate-aligned block inclusions → Lean uses the published invariant predicate
IsAlignedPartialSum, which also permits transported blocks and zero padding; reason: every source partial sum satisfies the predicate, so the Lean theorem specializes to the printed theorem. - encoding: the source writes , , and in 1-based coordinates → Lean uses
bigAdd,kron,Degenerates, andFin k; reason: these are the corresponding published tensor operations and a change of index base.
Preamble
import Mathlib import Definitions.Def_GraphTensorCP_IsAlignedPartialSum import Definitions.Def_GraphTensorCP_wt import Definitions.Def_mme_degeneration universe u
Formal statement
namespace GraphTensorCP
namespace Supplement
/-- **Theorem 3.4 (simultaneous separation)**, Zhou, *Asymptotic Rank of Weighted Graph
Tensors via Completely Positive Covers*, arXiv:2610.34273 (version of 2026-11-03), pp.31–32.
If `A` is an aligned partial direct sum of `M` order-`k` blocks sharing the first `k-2` legs,
then the full direct sum of `D_{k,M}=1+(k-1)(M-1)` copies of `A` degenerates to the full direct
sum of the blocks after each is tensored with an `M`-dimensional edge on the last two legs.
**Formalization Note.** Legs are 0-based, so the paper's `(k-1,k)` are `k-2,k-1` here.
`MME.Degenerates X Y` means the target `X` formally degenerates from the source `Y`. `bigAdd`
is the full direct sum and `kron` the Kronecker product. No nonzero hypothesis is imposed: zero
blocks retain the ambient spaces already stored in the tensor objects. -/
theorem theorem_3_4 {K : Type u} [Field K] {k M : ℕ} (hk : 2 ≤ k) (hM : 1 ≤ M)
(A : MME.TensorObj K k) (B : Fin M → MME.TensorObj K k)
(hAB : IsAlignedPartialSum A B) :
MME.Degenerates
(MME.TensorObj.bigAdd fun h : Fin M =>
MME.TensorObj.kron (B h)
(edgeTensor K ⟨k - 2, by omega⟩ ⟨k - 1, by omega⟩ M))
(MME.TensorObj.bigAdd fun _ : Fin (1 + (k - 1) * (M - 1)) => A) := by sorry
end Supplement
end GraphTensorCP
Source
Zhou, Asymptotic Rank of Weighted Graph Tensors via Completely Positive Covers, arXiv:2610.34273 (version of 2026-11-03), Theorem 3.4, pp.31–32