{
 "artifacts": null,
 "category": "optimization",
 "description": "Use the constraint matrix rank and nullity to set an explicit upper bound on the number of facial-reduction phases in an SDP layer representing structured binary decisions. Apply those phases before the main primal-dual solve, stopping after the rank–nullity budget and using the reduced face for all subsequent forward and backward computations.",
 "formulas_latex": [
  "$$P(A,b):=\\{x\\in\\{0,1\\}^{n}\\mid Ax=b\\}$$",
  "$$\\mathcal{R}(A,b):=\\left\\{Y=\\begin{pmatrix}1\u0026x^{T}\\\\x\u0026X\\end{pmatrix}\\succeq 0\\ \\middle|\\ \\operatorname{diag}(X)=x,\\ Ax=b,\\ AX=bx^{T}\\right\\}$$",
  "$$\\operatorname{sd}\\bigl(\\mathcal{R}(A,b)\\bigr)\\leq \\min\\{\\operatorname{rank}(A),\\ n-\\operatorname{rank}(A)\\}$$",
  "$$B(A):=\\min\\{r,n-r\\},\\qquad r=\\operatorname{rank}(A)$$"
 ],
 "id": 2857,
 "implementation": "1. Integration point: implement a differentiable structured-discrete layer whose input is neural logits l and whose output is a soft or rounded binary vector x satisfying Ax=b. Form the Shor/RLT SDP variable Y of size n+1 and solve an SDP projection or entropy-regularized SDP during the forward pass; use implicit differentiation or unrolled differentiation through the solver. This is most relevant to exact-cardinality MoE routing, graph selection, matching relaxations, or binary feature selection. 2. Precompute the numerical rank r of A using an SVD, counting singular values larger than tau_rank=max(m,n)*eps*||A||_2. Set B=min(r,n-r). Pseudocode: initialize the SDP cone and equality constraints; for k=1,...,B, solve the current exposing-vector subproblem for a PSD dual certificate Z_k; if \u003cZ_k,Y\u003e is below tau_expose for all primal iterates, parameterize the current face as Y=V_k S V_k^T with V_k spanning ker(Z_k), otherwise stop; solve the reduced SDP in S and recover Y. Add a small diagonal regularizer mu I inside the reduced cone only if the solver requires strict positivity. 3. The paper supplies B; estimate rank numerically, exposing certificates, and residual tolerances empirically. Log primal residual ||Ax-b||, RLT residual ||AX-bx^T||_F, minimum eigenvalue, number of Newton iterations, and backward-gradient norm. 4. First experiment: use a 32-variable exact-k routing layer in a 2-layer MLP on synthetic constrained assignment data, comparing a standard unreduced SDP solver against rank-budgeted facial reduction at equal accuracy. Test ranks r=1,4,8,16,24. Success means fewer solver iterations or fewer failed backward passes, lower condition numbers, and unchanged assignment feasibility; a useful target is at least 1.5x faster forward/backward solves or a clear reduction in NaN/indefinite-matrix failures. Also compare against naive truncation after 1 and n rounds to verify that the rank-based cap is informative.",
 "math_summary": "For a nonempty binary feasible set P(A,b)={x in {0,1}^n : Ax=b}, the paper defines the equality-generated SDP–RLT relaxation R(A,b)={Y=[[1,x^T],[x,X]] succeq 0 : diag(X)=x, Ax=b, AX=bx^T}. Let r=rank(A), with 0\u003cr\u003cn. The main result is sd(R(A,b)) \u003c= min{r,n-r}, where sd denotes singularity degree, the number of facial-reduction steps required to expose the minimal face containing the feasible SDP set; the paper states that this bound is attained for every rank. The transferable quantity is B=min{rank(A), n-rank(A)}. In an SDP layer, B is a certified maximum number of facial-reduction rounds for this equality-generated structure. A facial-reduction round finds a positive-semidefinite exposing matrix Z in the current dual cone with \u003cZ,Y\u003e=0 for every feasible Y, then replaces the PSD cone by the face F={Y succeq 0 : \u003cZ,Y\u003e=0}. The rank bound is exact in the worst case, so B should be used as a hard cap rather than an arbitrary iteration limit.",
 "math_tags": [
  "convex-analysis",
  "linear-algebra",
  "optimization"
 ],
 "ml_areas": [
  "moe-routing",
  "optimizer",
  "graph-nn"
 ],
 "paper": {
  "arxiv_id": "2608.29945",
  "arxiv_url": "https://arxiv.org/abs/2608.29945",
  "summary_what_math_gives_to_ml": "The paper gives an explicit rank–nullity bound on the singularity degree of an SDP relaxation built from binary variables and equality-generated RLT constraints: for A of rank r, at most min(r,n-r) facial-reduction rounds are needed, and this is sharp. This is transferable to neural systems that contain SDP-based differentiable layers for structured binary decisions, such as exact-cardinality selection, matching, routing, or discrete latent-variable inference. The practical asset is not the binary modeling itself but an a priori degeneracy budget that can control facial-reduction depth, solver tolerances, and numerical safeguards before expensive training begins. The idea is moderately narrow because it applies when a neural module genuinely solves an SDP relaxation rather than ordinary unconstrained network training.",
  "title": "Sharp Singularity-Degree Bounds for Equality-Generated SDP-RLT Relaxations of Binary Programs",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 7,
  "novelty": 8,
  "usefulness": 5
 },
 "solves": [
  "stability",
  "speedup",
  "memory"
 ],
 "title": "Rank-Budgeted Facial Reduction for Binary SDP Layers",
 "url": "https://synthcore.org/idea/2857/rank-budgeted-facial-reduction-for-binary-sdp-layers",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}