By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

Universal Tensor Squares for Symmetric Groups

For every positive integer n∉{2,4,9}n \notin \{2, 4, 9\} there is a partition λ⊢n\lambda \vdash n such that g(λ,λ,ν)>0g(\lambda, \lambda, \nu) > 0 for every ν⊢n\nu \vdash n, resolving the tensor square conjecture for symmetric groups affirmatively.

For each positive integer n∉{2,4,9}n \notin \{2, 4, 9\}, choosing a partition λn\lambda_n supplied by Theorem 1.1, the unipotent irreducible complex character UqλnU_q^{\lambda_n} of GLn(Fq)\mathrm{GL}_n(\mathbb{F}_q) satisfies that Uqλn⊗UqλnU_q^{\lambda_n} \otimes U_q^{\lambda_n} contains UqνU_q^{\nu} for every ν⊢n\nu \vdash n and every prime power qq.

For M≥9M \geq 9, the partition L=(M+b+s−1,M−1+b,M−2,M−3,…,3,2b+1,1s)L = (M+b+s-1, M-1+b, M-2, M-3, \dots, 3, 2^{b+1}, 1^s) is self-conjugate of size nn, so the support of SL⊗SLS^L \otimes S^L is invariant under transposition of the target; if no parity-compatible M≥9M \geq 9 is available, then n≤64n \leq 64.

Let M≥9M \geq 9 and let λ⊢∣L∣\lambda \vdash |L| have column heights h1≥h2≥⋯h_1 \geq h_2 \geq \cdots and row lengths w1≥w2≥⋯w_1 \geq w_2 \geq \cdots. If integers d,qd, q satisfy 1≤d≤41 \leq d \leq 4, 0≤q≤80 \leq q \leq 8, ∑i=1dwi≥K\sum_{i=1}^d w_i \geq K, dq+7+d≤Kdq + 7 + d \leq K, and ∑i=1qmax⁡(0,⌊hi−d2⌋)≥r\sum_{i=1}^q \max(0, \lfloor \frac{h_i - d}{2} \rfloor) \geq r, then SλS^\lambda occurs in SL⊗SLS^L \otimes S^L.

Assuming M≥9M \geq 9, if the column heights h1≥h2≥⋯h_1 \geq h_2 \geq \cdots of ν⊢n\nu \vdash n (or of νt\nu^t) satisfy ∑i=14hi≤2M−2\sum_{i=1}^4 h_i \leq 2M - 2, then SνS^\nu occurs in SL⊗SLS^L \otimes S^L.

Grouping the length-three block, the b+1b+1 length-two blocks, and the ss singleton blocks in both layers and projecting to outputs in [1,3][1, 3] produces a module containing every partition in P3(c,s)\mathcal{P}_3(c, s), and this component can be induced independently with components on lengths 4,…,M−24, \dots, M-2 and with the big pair.

If M≥22M \geq 22, a target satisfying H4≥KH_4 \geq K and W4≥KW_4 \geq K passes the band test in one of its two orientations.

The first-four balance test and the band tests cover every target when 9≤M≤219 \leq M \leq 21, r>0r > 0, except possibly for the 19 pairs (M,r)∈{(10,8),…,(10,17)}∪{(12,14),…,(12,20)}∪{(14,22),(14,23)}(M,r) \in \{(10,8), \dots, (10,17)\} \cup \{(12,14), \dots, (12,20)\} \cup \{(14,22), (14,23)\}.

For every admissible M≥9M \geq 9 and r>0r > 0, every target partition of n=NM+2rn = N_M + 2r is supplied by the band and balance constructions for LL.

For every integer 1≤n≤641 \leq n \leq 64 other than 2,4,92, 4, 9, there is a self-conjugate partition λ⊢n\lambda \vdash n such that g(λ,λ,ν)>0g(\lambda, \lambda, \nu) > 0 for every ν⊢n\nu \vdash n.

Methods

Realize Specht modules by column-alternating tensors and detect irreducible constituents via nonzero contractions with dual block tensors (dual-polytabloid test), combined with ordinary induction on disjoint position sets and Littlewood–Richardson/Young support rules.

Source excerpts

“Lemma 2.1 (Dual-polytabloid test). Let z ∈ E⊗B and θ ⊢ |B| with at most D rows. Then Sθ occurs in C[SB]z if and only if z has a nonzero contraction with a dual block tensor whose blocks are the columns of θ, for some placement of the blocks and some common ordered basis of E∗.” · Show in paper

Separate a smaller cyclic staircase module from a width-two odd path and two attachments in WLW_L via a coordinate projection, and contract against a dual polytabloid whose endpoint placement prevents cancellation between path and attachment covectors.

Source excerpts

“Lemma 4.3 (Triangle quotient). There is a quotient of WL isomorphic to IndSB SD×SF (WM−2 ⊠ UF ), (4.2) where WM−2 is the specified cyclic module in Theorem 2.4.” · Show in paper

“Lemma 4.4 (Separated contraction). Retain from uF only terms using P on the path, X+ on E+, and X− on E−. The resulting tensor is, up to one nonzero overall sign, upath ⊗ z+ ⊗ z−. (4.3)” · Show in paper

Partition column blocks of LL into disjoint position sets and separate components in SL⊗SLS^L \otimes S^L by the sizes and sums of their pair-letter outputs z(a,a′)=a+a′−1z(a, a') = a + a' - 1, then apply Young's rule to the resulting trivial-factor size lists.

Source excerpts

“Lemma 5.1 (Separation by sizes and sums). Partition the column blocks of L into collections B1, . . . , Bv on disjoint position sets D1, . . . , Dv, using the same blocks in both factors. Suppose I1, . . . , Iv are nonempty sets of integers with every element of Ii smaller than every element of Ii+1. Let Wi be the image of YBi ⊗ YBi under the coordinate projection retaining words all of whose outputs lie in Ii.” · Show in paper

Use an exact Python capacity-verification program (`capacity_checks.py`) with diagonal-hook enumeration, safe prefix pruning, and columnwise big-pair allocation to exhaust all targets for intermediate and exceptional (M,r)(M,r) pairs.

Source excerpts

“The following standalone Python 3 file, verification/code/capacity_checks.py, contains both finite capacity checks. The mathematical justification of its tests and pruning rules is in Section 6.” · Show in paper

Compute exact residues of whole Kronecker-coefficient vectors g(λ,λ,ν)(modp)g(\lambda, \lambda, \nu) \pmod p for p∈{1000000007,1000000009}p \in \{1000000007, 1000000009\} using a recursive Murnaghan–Nakayama border-strip algorithm with 128-bit beta-set masks in C++ (`smallcode_submitted.cpp`) to certify all eligible degrees 1≤n≤641 \leq n \leq 64.

Source excerpts

“Calculations take place in Fp, where p ∈ {1000000007, 1000000009}. Both numbers are prime, as is checked by trial division through 31622. Since n ≤ 64 < p, every denominator in (7.1) is invertible modulo p.” · Show in paper

“This is the complete C++ verifier justified in Section 7. Compile with a compiler supporting unsigned __int128, for example g++ -O3 -std=c++17, and supply one integer degree 1 ≤ n ≤ 64, n /∈ {2, 4, 9}, as its argument.” · Show in paper