Graph-Represented Methods

Feasible GNN Bounds for Exact Max-Cut Branch-and-Bound

A note on a Max-Cut solver that replaces repeated SDP relaxation solves with a feasibility-preserving GNN surrogate, keeping branch-and-bound correctness by constructing dual-feasible upper bounds.

What Max-Cut Asks

For a weighted graph G=(V,E), Max-Cut assigns each vertex to one of two partitions and maximizes the total weight of edges crossing the partition. With xi∈{-1,1}, one common Laplacian form is:

maxx∈{-1,1}n 14 x⊤ L x.

The plain-language question is simple: which binary assignment makes as many important edges as possible disagree? This appears whenever the graph edge means “these two items prefer to be separated” or “separating this pair creates value.” Typical examples include two-way clustering with dissimilarity rewards, graph partitioning subroutines, anti-ferromagnetic Ising or spin-glass energy minimization, certain VLSI and circuit-layout abstractions, and approximation subproblems inside combinatorial optimization.

The important detail is that Max-Cut is not asking for a good-looking partition in a vague sense. It is optimizing a precise discrete objective. Every vertex must choose one of two sides, and each edge contributes only when its endpoints land on opposite sides.

Why It Is Hard

The difficulty is combinatorial. With n vertices, there are 2n binary assignments, up to a global sign flip. Local changes are misleading: moving one vertex may improve the cut for some incident edges while worsening it for others. Dense graphs, frustrated cycles, and mixed edge weights make the search landscape especially awkward.

Max-Cut is NP-hard, so exact solvers usually rely on branch-and-bound or branch-and-cut. At each node, the solver needs a lower bound from a feasible cut and an upper bound proving that a subtree cannot contain a better cut. SDP relaxations are attractive because they give strong upper bounds. They are also expensive when thousands or millions of branch-and-bound nodes must be evaluated.

This creates the bottleneck behind the paper: exact branch-and-bound may spend most of its time repeatedly solving similar SDP relaxations just to decide whether a node can be pruned.

How GNNs Have Usually Been Used

Graphs are the native input object, so GNNs are a natural candidate for Max-Cut. Prior learning-based approaches usually use GNNs in one of three ways.

First, a GNN can predict a primal cut directly. This is useful as a heuristic: it may produce a good feasible solution quickly, and that feasible solution gives a lower bound for the maximization problem. But a good cut alone does not certify that no better cut exists.

Second, a GNN can guide a solver. It can suggest branching variables, rank candidate cuts, or choose which local search move to try. This can improve runtime, but the classical relaxation solver still supplies the certifying bounds.

Third, a GNN can approximate scores used inside a heuristic search. This can be fast, but if the score is not tied to feasibility or dual validity, it is hard to use safely inside an exact pruning rule.

So the usual GNN role is advisory or heuristic. That is useful, but it is not the same as replacing the relaxation evaluation itself.

What Is Different Here

The interesting point in this paper is narrower: replace the repeated SDP relaxation solve inside branch-and-bound with a neural surrogate that still returns a valid bound.

That distinction matters. In exact branch-and-bound, a learned value predictor is dangerous if it can underestimate an upper bound. A fast but invalid bound can prune the optimal branch and destroy correctness. This work avoids that failure mode by making the GNN output dual-feasible SDP solutions after a projection step. The global optimality claim should therefore be read carefully. The neural network does not prove that it has found the optimal cut. It supplies a safe upper bound, and complete branch-and-bound keeps the exactness.

The computational bet is that evaluating more nodes with a cheap valid bound can be faster than evaluating fewer nodes with an expensive SDP solve. That bet only works if the learned bound remains valid. This paper is best understood as a feasibility-preserving neural bounding oracle for Max-Cut SDP relaxations.

SDP Relaxation and the Role of the Dual

The Max-Cut SDP relaxation replaces binary variables with a positive semidefinite matrix X whose diagonal entries are one:

maxX ⟨L,X⟩ subject to diag(X)=e, X⪰0.

The corresponding dual can be written as:

miny,S e⊤y subject to L-Diag(y)+S=0, S⪰0.

Equivalently, dual feasibility is:

Diag(y)-L⪰0.

This is the key constraint. If y satisfies it, then e⊤y is a valid upper bound on the SDP optimum, hence also on the integer Max-Cut optimum inside the current branch-and-bound node. The GNN does not need to output the optimal dual solution. It needs to output a feasible one that is tight enough to be useful.

What the GNN Is Asked to Predict

A standard node-level GNN is not an obvious fit for the SDP variable X, because X is not a vector of node labels. It is a matrix whose entries correspond to vertex pairs. The architecture therefore uses pairwise embeddings hij.

The initial pairwise token contains information such as the objective matrix entry and whether the pair is diagonal:

hij0 = INIT ( Cij , Ii=j ).

This is more natural than a purely node-level representation. The SDP relaxation is built from pairwise correlations. If the network only stores hi, it must reconstruct matrix-level structure indirectly. Pairwise tokens put the decision object closer to the representation.

The MC-MPNN update then lets an entry (i,j) aggregate information through intermediate indices u. The intuition is close to matrix multiplication: entry (i,j) is refined using information from entries (i,u) and (u,j). The sparse variant, δ-MC-MPNN, restricts part of this aggregation to nonzero graph edges.

The architecture is therefore graph-represented in a precise sense: it does not only pass messages over vertices; it passes messages over pairwise objects that match the SDP matrix.

Feasibility-Preserving Heads

The most important design choice is not the message passing layer by itself. It is the output parameterization.

For the primal SDP, the network predicts vectors oi and normalizes them:

oi ← o~i ||o~i||2 .

Then it constructs:

X^ij = oi⊤ oj.

This gives X^=OO⊤⪰0 and X^ii=1. Primal feasibility follows from the construction, not from a penalty term in the loss.

For the dual, the network first predicts an unconstrained vector y^. The raw slack matrix is:

S^ = Diag(y^) -L.

This raw matrix may not be positive semidefinite. The paper fixes that with a uniform eigenvalue shift:

δ = max { 0, -λmin (S^) }, y^feas = y^ + δe.

The corrected slack is:

S^feas = Diag(y^feas) -L = S^ +δI ⪰0.

The shift raises every eigenvalue by δ. If the smallest eigenvalue was negative, it is moved to zero. If the raw matrix was already PSD, nothing changes.

This is the mechanism that makes the bound safe.

Why This Does Not Change the Dual Problem

A tempting objection is that shifting the dual variable might be solving a different problem. It does not. The original dual feasible set is still:

{ y : Diag(y)-L⪰0 }.

The raw GNN output may lie outside this set. The projection constructs a new candidate inside the same set. Once y^feas is feasible, its objective cannot be smaller than the dual optimum, because the dual is a minimization problem. By weak duality:

zMaxCut* ≤ pSDP* ≤ dSDP* ≤ e⊤ y^feas.

So the corrected dual value is a valid upper bound for the original Max-Cut node. It may be loose, but it is safe. That is exactly the trade-off branch-and-bound can tolerate.

How Exactness Is Preserved in Branch-and-Bound

For a maximization problem, branch-and-bound tracks an incumbent lower bound from the best feasible cut and an upper bound for each unresolved node. A node can be pruned only when its upper bound proves that it cannot beat the incumbent.

The classical solver obtains this upper bound by solving the SDP relaxation at the node. The neural solver instead evaluates:

UB(node) = e⊤ y^feas.

Because the value is dual feasible, a wrong low upper bound is not introduced. The worst case is different: if the bound is too conservative, fewer nodes are pruned, and the search may become slower. Correctness is protected; performance is empirical.

This is the cleanest lesson of the paper. A neural component can sit inside an exact optimization algorithm if the interface exposes the right certificate. Here the certificate is dual feasibility.

Empirical Trade-Off

The reported results fit the expected pattern. The neural branch-and-bound evaluates more nodes than the vanilla SDP-based solver, because the learned upper bounds are typically looser than exact SDP optima. But each node is much cheaper to evaluate. On the reported Max-Cut instances, this gives speed-ups such as 5.6x on g05_60, 9.1x on g05_100, and 10.6x on w01_100 against the vanilla Mosek branch-and-bound baseline.

The GNN variant comparison is also informative. The sparse δ-MC-MPNN has a better asymptotic story, but the implementation still uses dense GPU operations. In the reported experiments, MC-MPNN and δ-MC-MPNN have similar objective gaps, and the dense MC-MPNN can be more stable in practice. The theoretical sparsity advantage does not automatically become wall-clock dominance.

This should temper the claim. The method is not presented as a full replacement for mature Max-Cut branch-and-cut solvers with cutting planes. Against solvers such as BiqCrunch with strong cuts, the neural method is not necessarily competitive. Its value is more specific: it shows that the SDP bounding step itself can be replaced by a learned surrogate without giving up validity.

Limitations

The first limitation is bound tightness. Dual feasibility guarantees:

pSDP* ≤ fGNN,

but it does not guarantee:

fGNN - pSDP* ≤ ε.

If the GNN is poorly calibrated or out of distribution, the branch-and-bound tree may grow dramatically.

The second limitation is the projection itself. A uniform shift is cheap and safe, but it can be conservative. It increases the dual objective by nδ. A more selective correction could be tighter, but computing it may reintroduce the optimization cost that the surrogate was meant to avoid.

The third limitation is distribution dependence. The training data are generated from graph families and branch-and-bound trajectories similar to those used in testing. For a practical solver, one would need to understand how the bound behaves across graph sizes, weight distributions, and structures not seen during training.

The fourth limitation is numerical. The certificate relies on the sign convention and on computing the smallest eigenvalue accurately enough. In implementation, a small positive tolerance is usually needed. A nearly feasible matrix with a slightly negative eigenvalue is harmless if corrected; an incorrectly accepted infeasible matrix is not.

Why This Matters Beyond Max-Cut

The broader message is useful for optimization with learned components. If a neural network is inserted into a solver only as a value predictor, it can be fast and still unsafe. If it is inserted as a certificate-producing module, it can accelerate part of the algorithm while preserving the logic of the solver.

That idea transfers beyond Max-Cut. In decomposition, Benders-type methods, SDDP, robust optimization, stochastic programming, and mixed-integer nonlinear workflows, repeated relaxation or recourse evaluation is often the bottleneck. A black-box surrogate for those values is risky. A surrogate that preserves feasibility, dual validity, or certified bounds is much more interesting.

For graph-represented optimization, this paper gives a concrete pattern: match the representation to the structured decision variable, then force the output to satisfy the certificate required by the solver.

The GNN is not trusted because it is neural. It is trusted only after its output has been turned into a valid SDP dual point. That is the right level of skepticism.