Frobenius-Regularized Linear Assignment

• FRA is a relaxation framework that replaces traditional Euclidean projection with Frobenius regularization to mitigate geometric misalignment in the doubly stochastic set.
• It introduces a tunable parameter θ to explicitly control assignment sharpness, reconciling scale sensitivity with normalization in iterative QAP solvers.
• Integrated via the SDSN method, FRA offers unique convergence guarantees and competitive performance in both dense and sparse graph matching applications.

Frobenius-regularized Linear Assignment (FRA) is a relaxation framework for assignment and graph matching that replaces a scale-sensitive Euclidean projection onto the doubly stochastic set with a Frobenius-regularized optimization step. In the formulation introduced in “FRAM: Frobenius-Regularized Assignment Matching with Mixed-Precision Computing,” FRA is used as the projection subproblem inside an iterative Quadratic Assignment Problem (QAP) solver for graph matching. Its purpose is twofold: to mitigate the geometric misalignment introduced by relaxing permutation matrices to the doubly stochastic polytope, and to restore algorithmic scale invariance through explicit normalization and a tunable regularization parameter $\theta>0$ (Shen et al., 26 Jul 2025).

1. Problem setting and motivation

Let $\Pi_n$ denote the set of $n\times n$ permutation matrices. Its convex hull is the Birkhoff polytope, namely the set of doubly stochastic matrices

$$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$

A standard linear assignment problem (LAP) is

$\min_{P\in \Pi_n}\langle C,P\rangle,$

or equivalently $\min_{X\in DS}\langle C,X\rangle$ when the relaxed solution is later discretized, for example by the Hungarian algorithm (Shen et al., 26 Jul 2025).

In graph matching, the relevant objective is typically a QAP. For two attributed graphs $G=(V,E,A,F)$ and $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$ with $|V|=|\tilde V|=n$, a standard relaxation over $DS$ is

$\Pi_n$0

with

$\Pi_n$1

where $\Pi_n$2 are symmetric nonnegative edge-attribute matrices, $\Pi_n$3 are node-attribute matrices, and $\Pi_n$4 balances the two terms. Projection-based methods update via

$\Pi_n$5

followed by projection to $\Pi_n$6 and a convex combination:

$\Pi_n$7

The paper identifies two errors induced by the standard relaxation-and-projection paradigm. First, the doubly stochastic polytope strictly contains $\Pi_n$8, so iterates may remain in the interior rather than near permutation matrices; this is described as geometric misalignment. Second, Euclidean projection onto $\Pi_n$9 is not invariant to positive rescaling: for $n\times n$0, $n\times n$1. These two issues motivate FRA as a replacement for the “naked” projection step (Shen et al., 26 Jul 2025).

2. FRA objective and geometric role

FRA reformulates the projection step as a Frobenius-regularized optimization over $n\times n$2. Starting from the identity

$n\times n$3

the Euclidean projection of $n\times n$4 onto $n\times n$5 can be viewed as solving

$n\times n$6

FRAM makes this dependence explicit by normalizing $n\times n$7 by its maximum entry and introducing a tunable scale parameter $n\times n$8 (Shen et al., 26 Jul 2025).

The primal projection form of FRA is

$n\times n$9

and the equivalent concave form is

$$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$0

The first term, $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$1, is the linear assignment score inherited from the QAP gradient. The second term, $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$2, biases the optimizer toward larger Frobenius norms within $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$3. On the doubly stochastic set, $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$4 is maximized by permutation matrices, with value $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$5, and minimized by the uniform matrix $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$6, with value $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$7. Consequently, larger $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$8 induces sharper, near-permutation assignments, while smaller $$DS := \{X \in \mathbb{R}^{n\times n}: X \ge 0,\; X\mathbf{1}=\mathbf{1},\; \mathbf{1}^T X=\mathbf{1}^T\}.$$9 yields softer assignments. In the paper’s interpretation, this directly counteracts the feasible-region inflation caused by the doubly stochastic relaxation (Shen et al., 26 Jul 2025).

This construction separates numerical scale from assignment sharpness. Rather than allowing arbitrary scaling of the gradient-like matrix to implicitly alter the projection behavior, FRA uses the normalization $\min_{P\in \Pi_n}\langle C,P\rangle,$0 and lets $\min_{P\in \Pi_n}\langle C,P\rangle,$1 alone govern the trade-off between softness and discreteness. The intended effect is scale invariance at the algorithmic level together with explicit geometric control over the relaxed solution.

3. Theoretical properties

The central equivalence result is stated as Theorem 1:

$\min_{P\in \Pi_n}\langle C,P\rangle,$2

Because the maximization is strictly concave in $\min_{P\in \Pi_n}\langle C,P\rangle,$3 over the convex compact set $\min_{P\in \Pi_n}\langle C,P\rangle,$4, existence and uniqueness hold. This gives FRA a well-posed inner problem with a unique solution for each $\min_{P\in \Pi_n}\langle C,P\rangle,$5 and input $\min_{P\in \Pi_n}\langle C,P\rangle,$6 (Shen et al., 26 Jul 2025).

The paper also provides an explicit bound relative to the unregularized linear assignment score. Let

$\min_{P\in \Pi_n}\langle C,P\rangle,$7

and define

$\min_{P\in \Pi_n}\langle C,P\rangle,$8

For nonnegative $\min_{P\in \Pi_n}\langle C,P\rangle,$9, Theorem 2 states that

$\min_{X\in DS}\langle C,X\rangle$0

The interpretation given is that the loss in linear score induced by the Frobenius regularization is bounded by the regularizer weight; larger $\min_{X\in DS}\langle C,X\rangle$1 reduces this gap.

The limiting behavior is also characterized. Theorem 3 states that as $\min_{X\in DS}\langle C,X\rangle$2, $\min_{X\in DS}\langle C,X\rangle$3 converges to the unique minimizer of $\min_{X\in DS}\langle C,X\rangle$4 on the optimal face

$\min_{X\in DS}\langle C,X\rangle$5

If the optimal permutation is unique, then $\min_{X\in DS}\langle C,X\rangle$6 converges to that permutation. Theorem 4 states that as $\min_{X\in DS}\langle C,X\rangle$7, $\min_{X\in DS}\langle C,X\rangle$8, the uniform doubly stochastic matrix. FRA therefore interpolates between the barycenter of the relaxation and a near-discrete assignment; the paper describes moderate $\min_{X\in DS}\langle C,X\rangle$9 as yielding a “soft-yet-sharp” assignment that can mitigate premature overconfidence (Shen et al., 26 Jul 2025).

A KKT sketch is given for the concave form. With equality multipliers for row and column sums and complementary slackness for nonnegativity, stationarity yields

$G=(V,E,A,F)$0

with $G=(V,E,A,F)$1 chosen to enforce the doubly stochastic constraints. The paper notes that these conditions are consistent with the closed-form affine projection operator used in the solver.

FRA is also presented as a proximal step:

$G=(V,E,A,F)$2

where $G=(V,E,A,F)$3 is the indicator of $G=(V,E,A,F)$4. This links FRA to Euclidean proximal methods on polytopes and clarifies that the regularization is not an external heuristic but a reformulation of the projection step with explicit control over its geometry (Shen et al., 26 Jul 2025).

4. SDSN: the solver for FRA

The FRA subproblem is solved by Scaling Doubly Stochastic Normalization (SDSN), an alternating-projection method between two sets: the affine subspace of matrices with row and column sums equal to $G=(V,E,A,F)$5, and the nonnegative orthant. The method initializes with

$G=(V,E,A,F)$6

then alternates between an affine projection $G=(V,E,A,F)$7 and an elementwise nonnegativity projection $G=(V,E,A,F)$8 until a deviation metric $G=(V,E,A,F)$9 is below a threshold (Shen et al., 26 Jul 2025).

The nonnegativity projection is

$\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$0

elementwise. The affine projection onto $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$1 has the closed form

$\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$2

Theorem 5 states that this is the exact Euclidean projection for general asymmetric $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$3.

The stopping metric is

$\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$4

Using von Neumann’s lemma for alternating projections and the structure of $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$5, the paper states that $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$6 decreases monotonically to $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$7. In FRAM’s outer loop, convergence is monitored by the normalized Frobenius decrease

$\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$8

Each SDSN iteration has $\tilde G=(\tilde V,\tilde E,\tilde A,\tilde F)$9 time and $|V|=|\tilde V|=n$0 memory cost. Theorem 6 gives an iteration bound: for $|V|=|\tilde V|=n$1, SDSN needs

$|V|=|\tilde V|=n$2

iterations to obtain $|V|=|\tilde V|=n$3 with $|V|=|\tilde V|=n$4, where $|V|=|\tilde V|=n$5 is the linear convergence rate constant of DSN/SDSN (Shen et al., 26 Jul 2025).

The paper contrasts SDSN with Sinkhorn-style entropic normalization. Entropic regularization solves

$|V|=|\tilde V|=n$6

through multiplicative row and column scalings and depends on exponentials such as $|V|=|\tilde V|=n$7. By contrast, FRA/SDSN uses Euclidean regularization and additive projections with clipping. The stated consequences are that it tolerates sign and sparsity in $|V|=|\tilde V|=n$8, is robust to wide dynamic ranges under low precision, and avoids the overflow and underflow issues associated with exponentials (Shen et al., 26 Jul 2025).

5. Integration into FRAM and computational profile

Within FRAM, FRA is the projection stage of an outer QAP solver. The method initializes with $|V|=|\tilde V|=n$9 and, at iteration $DS$0, computes

$DS$1

when node features are present; edge-only tasks set $DS$2. It then solves

$DS$3

updates

$DS$4

stops when $DS$5, and finally discretizes $DS$6 to a permutation matrix $DS$7 using the Hungarian algorithm (Shen et al., 26 Jul 2025).

The dominant cost in dense settings is forming $DS$8, which is $DS$9. Each SDSN call is $\Pi_n$00 per inner iteration, and Hungarian discretization has worst-case $\Pi_n$01 complexity. The overall FRAM iteration complexity is therefore $\Pi_n$02 time and $\Pi_n$03 memory.

The paper recommends $\Pi_n$04 for stable, fast progress, as in DSPFP, and notes that moderate $\Pi_n$05 improves stepwise gains while avoiding overly confident early assignments. It also recommends normalizing $\Pi_n$06, $\Pi_n$07, and $\Pi_n$08 by a common scale before iteration. Specifically, with

$\Pi_n$09

the pre-scaling is

$\Pi_n$10

This is done in FP64 and is intended to keep $\Pi_n$11 in a numerically safe range for low-precision computation (Shen et al., 26 Jul 2025).

The mixed-precision architecture assigns FP64 to preconditioning, state update, convergence checks, and final Hungarian discretization; TF32 on GPU to forming $\Pi_n$12; and FP32 to SDSN iterations. Theorem 7 states that if $\Pi_n$13 at SDSN step $\Pi_n$14, with $\Pi_n$15 the low-precision truncation residual, then under alternating $\Pi_n$16 the residual is progressively corrected and vanishes as $\Pi_n$17 increases. This is the paper’s theoretical explanation for why TF32/FP32 inner loops can preserve final accuracy when coupled with FP64 accumulation and stopping (Shen et al., 26 Jul 2025).

6. Empirical behavior and practical use

The empirical evaluation covers real-world image matching, the CMU House sequence, and the Facebook-ego social network. For the real-image datasets, the paper uses eight sets—viewpoint, scale, blur, JPEG, and illumination—represented as dense graphs with node and edge attributes and with $\Pi_n$18 up to $\Pi_n$19. The metric is

$\Pi_n$20

For CMU House, which uses attribute-free edges, runtime and error are reported for $\Pi_n$21 up to approximately $\Pi_n$22. For Facebook-ego, with $\Pi_n$23 and $\Pi_n$24, node accuracy $\Pi_n$25 is reported under $\Pi_n$26, $\Pi_n$27, and $\Pi_n$28 noise (Shen et al., 26 Jul 2025).

On CPU-FP64 benchmarks, the paper reports that for real images at $\Pi_n$29, FRAM has average runtime approximately $\Pi_n$30s versus DSPFP at $\Pi_n$31s, approximately $\Pi_n$32 faster, while achieving best or tied-best error on most sets. On CMU House, FRAM is reported at $\Pi_n$33s versus DSPFP at $\Pi_n$34s and ASM at $\Pi_n$35s, with lowest error $\Pi_n$36 versus DSPFP $\Pi_n$37. On Facebook-ego, FRAM achieves accuracies of $\Pi_n$38, $\Pi_n$39, and $\Pi_n$40 at $\Pi_n$41, $\Pi_n$42, and $\Pi_n$43 noise, respectively; it is reported as faster than ASM by a factor of $\Pi_n$44 and as having higher accuracy than all baselines, including a gain of $\Pi_n$45 versus DSPFP at $\Pi_n$46 noise (Shen et al., 26 Jul 2025).

For mixed precision, the paper reports that on NVIDIA RTX 4080 SUPER, FRAM on ubc(2000) yields a $\Pi_n$47 speedup versus GPU-FP64 and a $\Pi_n$48 speedup versus CPU-FP64; for problems with at most $\Pi_n$49 nodes, speedups are below $\Pi_n$50, which the paper describes as consistent with Amdahl’s law and fixed overheads. The abstract summarizes the acceleration as “up to 370X speedup over its CPU-FP64 counterpart,” with negligible loss in solution accuracy (Shen et al., 26 Jul 2025).

The practical guidance given in the paper recommends $\Pi_n$51, with small values for dense or clean graphs $\Pi_n$52–$\Pi_n$53 and larger values for sparse or noisy graphs $\Pi_n$54–$\Pi_n$55. A separate summary recommends $\Pi_n$56 for dense tasks and $\Pi_n$57 for sparse tasks. The suggested defaults are $\Pi_n$58–$\Pi_n$59 with $\Pi_n$60 default, $\Pi_n$61–$\Pi_n$62 with $\Pi_n$63 default, $\Pi_n$64 to $\Pi_n$65 for SDSN stopping, and $\Pi_n$66 to $\Pi_n$67 for FRAM stopping. The paper also states that results are not sensitive to $\Pi_n$68 and that $\Pi_n$69 is used throughout (Shen et al., 26 Jul 2025).

7. Relation to adjacent methods, scope, and limitations

FRA is positioned against several classes of assignment and graph-matching methods. Relative to the Hungarian algorithm, which is an exact solver for the unregularized LAP and is often used only for final discretization in QAP relaxations, FRA introduces squared Frobenius regularization within the doubly stochastic relaxation in order to bias intermediate iterates toward extreme points. Relative to entropic regularization, including Sinkhorn, SoftAssign, and GA, FRA replaces KL or entropy penalties and multiplicative scaling with Euclidean regularization and additive projections. The paper emphasizes that the two approaches have different numerical profiles: entropic methods can be sensitive to the regularization parameter $\Pi_n$70 and to scaling, may overflow or underflow, and use exponentials that are poorly suited to low precision; FRA is designed to sharpen assignments through $\Pi_n$71 without exponentials (Shen et al., 26 Jul 2025).

Relative to Frank–Wolfe and projected gradient methods on $\Pi_n$72, FRA replaces standard Euclidean projection with a proximal-with-bias step having explicit sharpness control through $\Pi_n$73 and a unique inner solution due to strong concavity. Relative to DSPFP and IPFP, the paper’s claim is that unregularized doubly stochastic projection is scale-sensitive and may keep iterates in the interior, whereas FRA stabilizes projection behavior, sharpens assignments, and enables mixed precision. Relative to OT and Gromov–Wasserstein variants such as GWL and S-GWL, the paper describes FRAM as more scalable and faster in the reported large-graph settings, with higher matching accuracy in those experiments (Shen et al., 26 Jul 2025).

The method’s limitations are also explicit. The parameter $\Pi_n$74 is chosen empirically rather than adaptively. The dense multiplication $\Pi_n$75 imposes $\Pi_n$76 time per outer iteration, so sparsity exploitation, structure-aware kernels, or blocked and tiled GPU GEMMs are identified as potential improvements. The paper further notes that extending FRA to higher-order matching, tensor QAP, and partial or rectangular matching with rigorous guarantees remains open. It also points to low-level optimizations such as kernel fusion, custom mixed-precision accumulators, and asynchronous pipelines as possible future work (Shen et al., 26 Jul 2025).

Within that scope, FRA serves as the mathematical core of FRAM: a Frobenius-regularized projection on the doubly stochastic set that interpolates between the uniform matrix and near-permutation assignments, is equipped with score-gap and limiting guarantees, admits an $\Pi_n$77 alternating-projection solver, and is designed to preserve scale invariance while remaining compatible with mixed-precision computation.