Formation Matching Metric Overview

• Formation matching metric is a collection of methods that quantify similarity between configurations using metrics, pseudo-metrics, and learned descriptors.
• It encompasses distinct formulations including ordinal formation assignment with distortion guarantees, descriptor-based manifold matching, and learned pseudo-metrics in causal inference.
• Geometric approaches such as tree duality and calibration modulo 2 offer certifications and practical parallels to classical set distances like the Earth Mover’s Distance.

Formation matching metric denotes several closely related constructions for comparing, assigning, or certifying correspondences between configurations in metric or learned metric spaces. In the cited literature, the object being matched may be a bijection between agents and target positions, two finite point sets represented through permutation-invariant descriptors, observational units embedded in a learned latent space, or an even-cardinality metric space reduced to a tree-like geometry. The common theme is that matching quality is encoded by a metric, pseudo-metric, or descriptor-based objective, but the formal role of that objective differs sharply across ordinal assignment, deep metric learning, causal matching, and geometric duality [(Anari et al., 2023); (Dai et al., 2021); (Johnson et al., 2019); (Petrache et al., 2014)].

1. Scope and formal viewpoints

A first formulation arises in metric matching for formation assignment. Here agents $A=\{a_1,\dots,a_n\}$ and items $I=\{x_1,\dots,x_n\}$ lie in a shared metric space $(X,d)$, and a perfect assignment is a bijection $M:A\to I$ with cost

$C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$

When only ordinal rankings of items are available, the central performance measure is distortion, namely the worst-case ratio between the cost of a mechanism and the optimal metric cost over all metrics consistent with the rankings.

A second formulation treats a formation as a finite set and compares formations through learned descriptors in an embedding space. In that setting, a metric generator $g_w:\mathbb R^D\to\mathbb R^n$ induces the pullback metric

$d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$

and matching is performed through the centroid and $p$-diameter of sets under $d_w$, rather than through explicit pointwise correspondences.

A third formulation appears in high-dimensional matching estimation for observational causal inference. There one learns a low-dimensional representation $f_\theta:\mathbb R^p\to\mathbb R^q$ and matches units using the induced pseudo-metric

$I=\{x_1,\dots,x_n\}$0

It is a pseudo-metric because distinct covariate vectors may map to the same embedding.

A fourth formulation is geometric and unlabeled. For a finite metric space $I=\{x_1,\dots,x_n\}$1 with $I=\{x_1,\dots,x_n\}$2, the matching number is

$I=\{x_1,\dots,x_n\}$3

where $I=\{x_1,\dots,x_n\}$4 is the set of perfect matchings. This object admits a dual description through 1-Lipschitz maps to metric trees and, under additional assumptions, through level-set calibrations modulo $I=\{x_1,\dots,x_n\}$5.

Formulation	Core matched object	Metric or objective
Ordinal formation assignment	Agents to target positions	$I=\{x_1,\dots,x_n\}$6, distortion
Manifold matching	Two point sets or sample sets	$I=\{x_1,\dots,x_n\}$7, centroid, $I=\{x_1,\dots,x_n\}$8, $I=\{x_1,\dots,x_n\}$9
High-dimensional causal matching	Treated and control units	$(X,d)$0
Geometric matching theory	Even-cardinality metric space	$(X,d)$1

2. Ordinal metric matching for formation assignment

In the ordinal model, each agent reports only a ranking of items by increasing distance, consistent with an unknown shared metric on $(X,d)$2. The optimal benchmark is

$(X,d)$3

and for a mechanism $(X,d)$4,

$(X,d)$5

For randomized mechanisms, the numerator becomes the expected cost (Anari et al., 2023).

A classical baseline is Serial Dictatorship. Caragiannis et al. proved that its worst-case distortion is at most $(X,d)$6, and the same value is attained on a lower-bound instance. The newer deterministic mechanism RepMatch improves the upper bound to $(X,d)$7. RepMatch begins from singleton sets $(X,d)$8, each with representative $(X,d)$9 and parameter $M:A\to I$0. While two sets $M:A\to I$1 have representatives whose favorite $M:A\to I$2 and $M:A\to I$3 items intersect, the sets are merged; the new representative is the one with larger $M:A\to I$4, and the new $M:A\to I$5 is either the maximum of the two previous values or that maximum plus $M:A\to I$6 when they are equal. At termination, each set $M:A\to I$7 is assigned arbitrarily to the representative’s top $M:A\to I$8 items, which are disjoint across sets by construction.

The analysis is organized around the invariant

$M:A\to I$9

where $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$0, together with the fact that the representative’s top $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$1 items lie within radius

$C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$2

for at least $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$3 distinct items $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$4, and the size-parameter relation $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$5. These estimates yield the final bound

$C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$6

The corresponding lower-bound theory shows that every ordinal matching mechanism, deterministic or randomized, has worst-case distortion $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$7. The construction uses a tree metric with $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$8 agents at the leaves and items at internal nodes, with edge weights doubling toward the root. For an adversarially chosen leaf, the optimal matching cost is $C(M) \coloneqq \sum_{i=1}^n d\big(a_i,M(a_i)\big).$9, while any mechanism incurs cost at least $g_w:\mathbb R^D\to\mathbb R^n$0, giving distortion $g_w:\mathbb R^D\to\mathbb R^n$1.

The truthfulness landscape is less favorable. A mechanism is serializable if there exist permutations $g_w:\mathbb R^D\to\mathbb R^n$2 of $g_w:\mathbb R^D\to\mathbb R^n$3 such that, on structured instances with the preference pattern specified in the paper, the mechanism matches $g_w:\mathbb R^D\to\mathbb R^n$4 to $g_w:\mathbb R^D\to\mathbb R^n$5 for all $g_w:\mathbb R^D\to\mathbb R^n$6. Every serializable mechanism has worst-case distortion at least $g_w:\mathbb R^D\to\mathbb R^n$7. Any deterministic mechanism implemented by Deferred Acceptance with fixed item-side preferences is serializable, so DA-derived truthful mechanisms inherit the same exponential lower bound.

The paper also isolates a deterministic-versus-randomized gap through rounding fractional matchings. If $g_w:\mathbb R^D\to\mathbb R^n$8 is the best metric-oblivious rounding factor from a fractional matching to an integral perfect matching, and $g_w:\mathbb R^D\to\mathbb R^n$9 is the best thinness factor of a perfect matching with respect to all cuts, then

$d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$0

The lower inequality comes from cut metrics, and the upper inequality from Bourgain’s embedding. A plausible implication is that progress on thin matchings would directly sharpen deterministic guarantees for formation assignment from ordinal data.

3. Learned manifold metrics and descriptor-based formation comparison

In the deep metric learning formulation, real data are modeled as a manifold $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$1. Two networks are trained jointly: a distribution generator $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$2, which pushes a prior forward to a generated distribution, and a metric generator $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$3, which induces the pullback metric

$d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$4

The stated goal is to make the generated measure condensed around $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$5 while making $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$6 as straight as possible under the learned metric, so that embedding-space Euclidean distances approximate intrinsic geodesic distances (Dai et al., 2021).

Formation comparison is based on permutation-invariant shape descriptors. For a finite set $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$7, the Fréchet mean under the learned metric is represented in practice by the mean embedding

$d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$8

and the empirical $d_w(x,y)\coloneqq \|g_w(x)-g_w(y)\|_2,$9-diameter is instantiated at $p$0 as

$p$1

Given formations $p$2 and $p$3, the descriptor-matching loss is

$p$4

This objective explicitly enforces translation alignment through centroid matching and scale alignment through diameter matching.

The metric generator is trained by deep metric learning. The basic triplet loss is

$p$5

and the direction-regularized APN variant is

$p$6

Anchors and positives are real samples, while negatives are generated samples. The paper states that explicit graph geodesics are not used; a $p$7-NN graph and shortest-path approximation are described only as an optional adaptation.

At inference time, a formation distance can be written as

$p$8

where $p$9 are embedding centroids and $d_w$0 are embedding diameters. Optional centering and scaling provide scale invariance, and optional orthogonal Procrustes alignment can enforce rotation invariance. Standard set distances such as Chamfer distance and Earth Mover’s Distance can also be computed under the learned metric $d_w$1.

The reported empirical evidence is tied to generative modeling and super-resolution rather than classical robotic formations. On CelebA $d_w$2, the method achieved $d_w$3, and on LSUN Bedroom $d_w$4, $d_w$5. Descriptor distances, including centroid, diameter, and Hausdorff distances, decreased during training. Descriptor ablations found that centroid-only matching learned shallow patterns, diameter-only matching could suffer misalignment, and combining both was the most stable. In single-image super-resolution, replacing GAN loss with the manifold-matching objective improved PSNR and SSIM consistently and improved LPIPS and NIQE in most cases.

4. Learned pseudo-metrics for high-dimensional matching estimation

In observational causal inference, matching is used to approximate each unit with similar peers under the alternative treatment status. The setup consists of covariates $d_w$6, treatment $d_w$7, observed outcome $d_w$8, potential outcomes $d_w$9, and the assumptions of unconfoundedness, overlap, continuity, IID sampling, and SUTVA. For ATT, a nearest-neighbor matching estimator under a generic distance $f_\theta:\mathbb R^p\to\mathbb R^q$0 is

$f_\theta:\mathbb R^p\to\mathbb R^q$1

with $f_\theta:\mathbb R^p\to\mathbb R^q$2 placing mass on nearest controls (Johnson et al., 2019).

The motivating difficulty is the curse of dimensionality. Abadie and Imbens are cited for the fact that the nonparametric bias of matching on $f_\theta:\mathbb R^p\to\mathbb R^q$3 continuous covariates is of order $f_\theta:\mathbb R^p\to\mathbb R^q$4, which can invalidate asymptotic inference. The proposed response is to learn a low-dimensional pseudo-metric. One learns $f_\theta:\mathbb R^p\to\mathbb R^q$5, with $f_\theta:\mathbb R^p\to\mathbb R^q$6, and matches using

$f_\theta:\mathbb R^p\to\mathbb R^q$7

The terminology “pseudo-metric” is exact: identity of indiscernibles can fail when different inputs share the same embedding.

Two learned spaces are constructed. A prognostic embedding $f_\theta:\mathbb R^p\to\mathbb R^q$8 is trained to encode outcome-relevant similarity, and a propensity embedding $f_\theta:\mathbb R^p\to\mathbb R^q$9 is trained to encode treatment-assignment similarity. In the MLP-based method, the final hidden layers are scaled by the output-layer weights: $I=\{x_1,\dots,x_n\}$00 In the siamese method, embeddings are concatenated directly: $I=\{x_1,\dots,x_n\}$01 The outcome-target siamese loss is

$I=\{x_1,\dots,x_n\}$02

and the propensity-target contrastive loss is

$I=\{x_1,\dots,x_n\}$03

The central theoretical statement is Theorem 2.1. Under compact support, positivity, unconfoundedness, continuity, and IID sampling, nearest-neighbor ATT matching is asymptotically unbiased if there exists $I=\{x_1,\dots,x_n\}$04 such that for all $I=\{x_1,\dots,x_n\}$05, either

$I=\{x_1,\dots,x_n\}$06

or

$I=\{x_1,\dots,x_n\}$07

For ATUT, the corresponding prognostic condition is

$I=\{x_1,\dots,x_n\}$08

The paper’s interpretation is that small learned distances must imply closeness in a balancing score or prognostic score.

A key negative result concerns heterogeneous treatment effects. If one learns an outcome-based metric on pooled treated and control outcomes, inconsistency can result. The paper gives an explicit example in which the pooled-outcome metric converges to

$I=\{x_1,\dots,x_n\}$09

so that a treated unit at $I=\{x_1,\dots,x_n\}$10 is equally close to controls at $I=\{x_1,\dots,x_n\}$11, despite materially different treatment effects. This is the reason the outcome embedding is trained on controls only for ATT and on treated units only for ATUT.

The empirical results are reported for simulations, the NSW/LaLonde benchmark, and IHDP. In the sparse linear and sparse-linear-with-squared-terms simulations, the NN and SNN methods achieve low RMSE relative to propensity-only baselines and variable-selection methods. On NSW, the experimental benchmark is approximately $I=\{x_1,\dots,x_n\}$12; the SNN estimate is $I=\{x_1,\dots,x_n\}$13 with standard error $I=\{x_1,\dots,x_n\}$14, and the NN estimate is $I=\{x_1,\dots,x_n\}$15 with standard error $I=\{x_1,\dots,x_n\}$16. On IHDP with ATT $I=\{x_1,\dots,x_n\}$17, the NN and SNN methods report average estimates $I=\{x_1,\dots,x_n\}$18 and $I=\{x_1,\dots,x_n\}$19, respectively.

5. Tree duality, calibrations modulo 2, and unlabeled formation metrics

For an even-cardinality metric or pseudometric space $I=\{x_1,\dots,x_n\}$20, the matching number

$I=\{x_1,\dots,x_n\}$21

is the canonical unlabeled matching cost. The principal structural theorem states that for every finite pseudometric $I=\{x_1,\dots,x_n\}$22 on $I=\{x_1,\dots,x_n\}$23, there exists a tree-like pseudometric $I=\{x_1,\dots,x_n\}$24 on the same set such that

$I=\{x_1,\dots,x_n\}$25

Equivalently, there is a metric tree $I=\{x_1,\dots,x_n\}$26 and a 1-Lipschitz map $I=\{x_1,\dots,x_n\}$27 with

$I=\{x_1,\dots,x_n\}$28

This gives an unoriented Kantorovich duality in which the dual objects are 1-Lipschitz maps to trees rather than only real-valued 1-Lipschitz potentials (Petrache et al., 2014).

Tree-like spaces are characterized by the four-point condition

$I=\{x_1,\dots,x_n\}$29

The dual formulation can be written as

$I=\{x_1,\dots,x_n\}$30

Under homological conditions such as $I=\{x_1,\dots,x_n\}$31 or $I=\{x_1,\dots,x_n\}$32, the paper also proves

$I=\{x_1,\dots,x_n\}$33

where $I=\{x_1,\dots,x_n\}$34 is defined through parity cut-counting of level sets of 1-Lipschitz functions.

This duality is coupled to a calibration theory for rectifiable 1-chains with coefficients in $I=\{x_1,\dots,x_n\}$35. If $I=\{x_1,\dots,x_n\}$36 satisfies $I=\{x_1,\dots,x_n\}$37, and if $I=\{x_1,\dots,x_n\}$38 and $I=\{x_1,\dots,x_n\}$39 are as in the theorem, then

$I=\{x_1,\dots,x_n\}$40

with equality exactly for unions of geodesic segments that realize a minimal matching. The quantity $I=\{x_1,\dots,x_n\}$41 is therefore a global calibration modulo $I=\{x_1,\dots,x_n\}$42.

The paper explicitly connects this geometry to a formation matching pseudo-metric for unlabeled formations $I=\{x_1,\dots,x_n\}$43 and $I=\{x_1,\dots,x_n\}$44 of equal size: $I=\{x_1,\dots,x_n\}$45 where $I=\{x_1,\dots,x_n\}$46 is an invariance group such as rigid motions or similarities. This construction is unoriented: after alignment, the minimal pairing is not required to respect the partition into $I=\{x_1,\dots,x_n\}$47 and $I=\{x_1,\dots,x_n\}$48. If cross-form pairing is required, the paper states that one should use the oriented variant, which becomes the classical Earth Mover’s Distance or Wasserstein-1 distance.

Several technical limitations are explicit. The main results assume even cardinality; odd-cardinality settings require adding or removing a point or adopting another convention. The maximizing tree embedding is generally not unique. The paper provides existence and certification results rather than a full algorithm for constructing the optimal $I=\{x_1,\dots,x_n\}$49 and $I=\{x_1,\dots,x_n\}$50. It also notes that the approach does not extend in the same way to coefficients in $I=\{x_1,\dots,x_n\}$51 for $I=\{x_1,\dots,x_n\}$52.

6. Comparative themes, misconceptions, and open directions

A recurring source of confusion is that “formation matching metric” does not designate a single invariant formula across the literature. In ordinal assignment, the relevant object is a worst-case approximation guarantee relative to an unknown shared metric, measured by distortion (Anari et al., 2023). In manifold matching, the metric is the pullback distance $I=\{x_1,\dots,x_n\}$53 together with descriptor losses based on centroid and diameter (Dai et al., 2021). In causal matching, the learned object is deliberately only a pseudo-metric on the original covariate space (Johnson et al., 2019). In geometric matching theory, the central quantity is the matching number and its certification by tree duality and calibrations modulo $I=\{x_1,\dots,x_n\}$54 (Petrache et al., 2014).

A second misconception is that pointwise correspondence is always the primitive object. The descriptor-based manifold formulation explicitly avoids one-to-one assignment in favor of permutation-invariant set statistics. Conversely, the ordinal assignment model is precisely about bijective matching, and its guarantee is framed against the optimal perfect assignment $I=\{x_1,\dots,x_n\}$55. The tree-duality framework further distinguishes oriented from unoriented pairing, and the latter may pair points without preserving source-target labels.

A third issue concerns truthfulness and mechanism design. In the ordinal setting, truthfulness is not a free improvement: serializable mechanisms, including deterministic DA mechanisms with fixed item-side priorities, retain worst-case distortion at least $I=\{x_1,\dots,x_n\}$56. The data block explicitly notes that, in robot formation assignment, strategic reporting is typically not a concern; this suggests that truthful mechanisms may be unnecessarily restrictive in robotic applications even when they are indispensable in human-centric matching environments.

Several open problems are stated. For ordinal metric matching, the current gap is between the lower bound $I=\{x_1,\dots,x_n\}$57, the randomized upper bound $I=\{x_1,\dots,x_n\}$58, and the deterministic upper bound $I=\{x_1,\dots,x_n\}$59. The thin-matchings program asks whether $I=\{x_1,\dots,x_n\}$60, or at least $I=\{x_1,\dots,x_n\}$61, which would imply substantially tighter deterministic rounding guarantees. There is also an open truthfulness question: whether non-serializable truthful mechanisms with polynomial distortion exist.

The data block also identifies a relaxation through near-perfect matching. Truncated Random Serial Dictatorship matches only $I=\{x_1,\dots,x_n\}$62 agents and has expected distortion

$I=\{x_1,\dots,x_n\}$63

Setting $I=\{x_1,\dots,x_n\}$64 yields expected distortion $I=\{x_1,\dots,x_n\}$65. A plausible implication is that applications permitting a small fraction of dropped or reallocated positions can circumvent the worst behavior of full perfect matching.

Across the learned-metric formulations, limitations are equally explicit. Centroid and $I=\{x_1,\dots,x_n\}$66 may fail to capture anisotropy or higher-order structure, and the manifold-matching paper states that augmenting them with covariance-based descriptors or additional moments could improve performance. In causal matching, representation learning requires sufficient sample size, and learned features trade interpretability for downstream balance and bias reduction. These differences make the term “formation matching metric” best understood as a family of technically distinct constructions unified by a common objective: to encode similarity so that matching, assignment, or comparison is meaningful in the geometry induced by the problem.