When Can We Trust the Matching Principle? Robust Deployment Geometry Under Finite-Sample and Model Uncertainty

1 Introduction

• On UCI HAR embeddings, always matching the estimated direction is worse than abstaining on all 6/6 cells because the estimate is not trustworthy.The passage distinguishes this estimation failure from a claim that Matching itself is false.Source: §1 · p. 1
• Under the LQ Matching response, oracle-relative drift for probes in the chosen top-r deployment subspace scales as O(τ^2) when τ < 1/2, with practical usefulness depending on constants.The bound is a theorem result; the separation-region condition does not establish that constants make every such τ numerically useful.Source: §1 · p. 2
• Confidence-Calibrated Matching (CCM) maps τ to directional matching, a soft blend, or even-spread abstention, using calibrated thresholds rather than a theorem-derived rule.The match threshold lies inside the τ < 1/2 separation region, while most of the soft band does not.Source: §1 · p. 2
• The paper studies oracle-relative regret for estimated projector Matching under the LQ response, not global optimality among regularizers or a nonlinear Matching extension.Deep and real-block results are experiments under a specified protocol.Source: §1 · p. 2

2 LQ objects (self-contained preliminaries)

• A label-preserving displacement Δ has second moment Σ = E[ΔΔ⊤], whose large eigenvalues mark directions in which deployment moves substantially.The LQ response is introduced for a training penalty Σ′, probe v, and weight λ > 0.Source: §2 · p. 2
• As λ → ∞, drift vanishes for every probe in range(Σ) if and only if range(Σ′) covers range(Σ); missing a deployment direction leaves a residual floor.This is the cover-or-floor condition.Source: §2 · p. 2
• Even-spread uses Σ′ ∝ I without estimating axes, while matching aligns Σ′ with Σ or its top subspace and is sharper when the axes are known.Wrong-axis penalties leave a permanent floor; the stated operational default is even-spread unless estimated geometry is trustworthy enough to match.Source: §2 · p. 2
• The paper studies how to choose Σ′ from a finite-sample estimate bΣ and organizes the choice by τ = ε/γ rather than proposing a new coverage theorem.The cover question is treated as given.Source: §2 · p. 2

3 Related work

• Rajput (2026b) and Rajput (2026a) provide the cover question and even-spread default; this paper instead asks when an estimate of that geometry is trustworthy enough to match..Source: §3 · p. 2
• Classical covariance bounds, spiked models, Davis–Kahan / Wedin (Davis and Kahan, 1970; Yu, Wang, and Samworth, 2015), and shrinkage are ingredients or baselines, not new matrix inequalities.The paper packages ε/γ as τ and proves a Matching drift-gap result under the LQ response.Source: §3 · p. 3
• CORAL (Sun and Saenko, 2016), DANN (Ganin et al., 2016), DRO and group-robust training, and IRM / DG address shift with different objectives, usually without a Matching coverage certificate.CCM is presented as a finite-sample gate on Matching, not a competitor to those programmes.Source: §3 · p. 3
• Landers (2026) studies pathwise volatility under a frozen predictor, whereas this paper studies oracle-relative Matching drift regret as a function of τ = ε/γ.The distinction passage describes Landers (2026) as regularizing along an estimated drift subspace and reporting tolerance to moderate misspecification.Source: §3 · p. 3
• Selective classification (Geifman and El-Yaniv, 2017) abstains on uncertain labels, while CCM abstains from directional matching and falls back to even-spread..Source: §3 · p. 3

4 Deployment geometry estimation

• The estimator forms bΣ from the observed displacements Δ1, . . . , Δn as their sample second moment.The supplied formula passage is partial; the text identifies bΣ as the estimate used in the subsequent bound.Source: §4 · p. 3
• Under standard sub-Gaussian assumptions, ∥bΣ − Σ∥op ≤ ε(n, δ) holds with probability at least 1 − δ.The paper treats ε as a cited ingredient and uses a proxy.Source: §4 · p. 3
• With δ = 0.05, Exp. 1 calibrates c = 0.5 on the synthetic spike grid, while ambient-dimensional ε over-abstains on high-dimensional real displacements.The passage says Exp. 6 reports a practical numerator alongside ambient-τ as a limitation.Source: §4 · p. 3

5 The trust ratio

• The trust ratio τ = ε/γ compares a high-probability covariance-error bound ε with the eigengap γ separating the top-r deployment subspace.A large gap can absorb moderate estimation error, while a tiny gap can make even small error consequential.Source: §5 · p. 3
• When τ ≪ 1, the direction is identifiable enough to match; when τ ≫ 1, spreading the penalty is the default, with τ = O(1) marking the soft band.The numerator measures estimate uncertainty and the denominator measures subspace identifiability.Source: §5 · p. 3
• The theorem uses population gap γ, while CCM substitutes the data-based plug-in gap bγ; an unstable bγ makes τ large and triggers abstention.The plug-in gap is bγ = bλr − bλr+1.Source: §5 · p. 4

6 What the theorem says

• The intellectual object is τ = ε/γ: Davis–Kahan converts it to projector error, and the LQ response converts that error to oracle-relative drift gap.Neither conversion is identified as the central idea.Source: §6 · p. 4

6.1 Perturbation lemma

• Davis–Kahan bounds projector error under 2ε < γ, equivalently τ < 1/2; at τ ≥ 1/2, the bound is uninformative and the direction is not identifiable at that precision.The probability guarantee is at least 1 − δ.Source: §6.1 · p. 4

6.2 Matching consequence

• For probes in the chosen top-r deployment subspace, estimated projector Matching has oracle-relative absolute drift gap O(τ^2) when τ < 1/2.Practical usefulness depends on constants.Source: §6.2 · p. 4
• Davis–Kahan gives projector error O(τ), and orthogonal-projector compression squares that error in the LQ response, yielding an O(τ^2) drift gap.Ambient probes outside the chosen subspace retain only the weaker O(τ) estimate.Source: §6.2 · p. 4
• The guarantee is oracle-relative under correct second-moment specification, not a claim of absolute optimality among regularizers or a numerical recipe for CCM thresholds.CCM’s match threshold lies inside the separation region, but most of its soft band does not.Source: §6.2 · p. 4
• The theorem analyzes projector matching; full estimated-covariance matching can add bulk eigenvalues that shrink task directions and is treated as the dangerous naive baseline.The projector is CCM’s match arm.Source: §6.2 · p. 5

7 How CCM uses it

• CCM maps τ to three penalty choices: directional match when geometry is trusted, a soft blend, or even-spread abstention when it is not.The thresholds are fixed by calibration rather than implied by Theorem 3.Source: §7 · p. 5

7.1 Inputs and three penalties

• CCM’s arms use the top-r estimated projector for matching, trace-scaled µI for even-spread abstention, and a projector-plus-isotropic blend for the soft choice.Full estimated covariance uses every eigenvalue and is treated as the dangerous naive baseline.Source: §7.1 · p. 5, §7.2 · p. 5, §7 · p. 5

7.3 Threshold-selection protocol

• Thresholds are selected once by a predetermined procedure and frozen, rather than adjusted per evaluation or derived from Theorem 3.The synthetic calibration requires all three arms to appear and CCM to beat always-full covariance and always-isotropic on mean dual cost in at least half of mid-n cells.Source: §7.3 · p. 5
• The synthetic thresholds (τlow, τhigh) = (0.35, 2.5) are frozen as a feasible, balanced-regime choice; the grid’s dual-cost minimizer is (0.5, 4.0).The minimizer does not change the reported qualitative Exp. 3–5 conclusions.Source: §7.3 · p. 6
• The match threshold τlow = 0.35 lies inside τ < 1/2, but most of the soft interval lies outside that separation region and is a calibrated heuristic.The theorem’s practical usefulness also depends on constants.Source: §7.3 · p. 6
• For real-data experiments, ambient-d ε forces permanent even-spread, so a practical numerator and separate Fashion development split set the frozen (1, 8) thresholds for Office-31 and UCI HAR.The Fashion development split uses half the loaded images, not the reported test split.Source: §7.3 · p. 6

8.1 Flagship: always-match hurts when τ forbids trust

• Practical τ correlates with subspace error at 0.87, yet always-match hurts on 6/6 UCI HAR cells; CCM beats always-match on 6/6 and B0 on 3/6.The authors attribute the entanglement to nuisance and task directions in the embedding; CCM stays soft or even-spread.Source: §8.1 · p. 6, §8 · p. 6
• On Fashion rotations, τ correlates with subspace error at 0.92, and always-match and CCM beat B0 on 6/6 sample-size cells.As n grows, CCM progresses from even-spread through soft to match.Source: §8.1 · p. 6
• Soft CCM is best among deployable arms, at 0.689 versus 0.674 for always-match.Ambient-d τ is 5–10× larger and would remain in even-spread; the practical numerator is used for these real-data decisions.Source: §8.1 · pp. 7–8

8.2 Supporting synthetic programme (Exp. 1–5)

• In synthetic Exp. 2, τ correlates with subspace error at approximately 0.83 and drift gap at approximately 0.91, tracking quantities in Theorem 3.Exp. 1 confirms that the proxy ε covers the operator-norm estimation error.Source: §8.2 · p. 8
• On the synthetic 3×3 gap-by-n grid, CCM beats isotropic on dual cost in 9/9 cells and both pure policies in 6/9.At n ≥400, CCM dual cost is 0.01–0.11 versus 0.30–0.44 for full bΣ; at n = 80, it pays a caution premium versus noisy bΣ.Source: §8.2 · p. 8, Table 1 · p. 9
• Across dimensions 16, 32, 64, and 128 and three fixed n/d ratios, CCM beats both pure policies in 9/12 cells and stays within 5% of hindsight on 84% of reps.This is the Exp. 5 stress test.Source: §8.2 · p. 8
• In controlled misspecification, same-Σ but different-P shifts are invisible to τ; wrong-family geometry breaks naive matching, while CCM’s soft or isotropic fallback is safer.The authors distinguish statistical confidence from model correctness and describe this as controlled misspecification, not the main product.Source: §8.2 · p. 8

• When the displacement covariance differs from deployment, small τ does not repair a misspecified geometry family, and abstention is not a repair of the wrong estimand.Source: §9 · pp. 9–10
• The same second moment can accompany different higher-order structure, so second-moment confidence is not model correctness.Source: §9 · p. 9
• With weak gaps or tiny n, τ is large and CCM abstains by design; its proxy can also over-abstain on high-dimensional real features.Sharper real-data bounds remain open.Source: §9 · p. 10
• Nonlinear Matching extensions, online or LLM Matching, new shrinkage estimators, and ambient-d certificates for Office, Fashion, and HAR are out of scope.Source: §9 · p. 10

• For probes in the chosen top-r deployment subspace, oracle-relative projector-matching drift scales as O(τ^2) when τ < 1/2, with practical usefulness depending on constants.Source: §10 · p. 11
• On UCI HAR embeddings, always-match is worse than abstain in 6/6 cells, while Fashion shows the identifiable complementary regime.Source: §10 · p. 11
• CCM maps τ to match, soft, or abstain using calibrated thresholds, which are not implied by the theorem; sharper real-data ε bounds, adaptive rank certificates, and trust gates beyond LQ remain open.Source: §10 · p. 11

A.1 Assumptions (restated)

• The assumptions require label-preserving displacements with correctly specified finite second moment, an LQ response, a positive top-r eigengap, and a finite-sample operator-norm covariance bound.The concentration radius is based on standard matrix concentration; the paper does not claim a new concentration theorem.Source: §A.1 · p. 11

A.2 Proof of Lemma 2

• When 2ε < γ, the top-r eigenspace remains separated and Davis–Kahan/Wedin bounds projector error with probability at least 1−δ; for τ ≥ 1/2, the bound is uninformative.CCM treats that regime as non-identifiable and abstains.Source: §A.2 · pp. 11–12

A.3 Projector response identity

• For projector penalties, the inverse response has the closed form I−αΠ, allowing the response difference to be bounded without the looser general resolvent factor that grows with λ.Here α = 2λ/(1+2λ).Source: §A.3 · p. 12

A.4 Proof of Theorem 3

• For probes in the chosen top-r subspace, the compressed projector gap is quadratic in subspace error, yielding O(τ^2) oracle-relative drift in the separation region τ < 1/2.The bound's practical usefulness depends on prefactor constants.Source: §A.4 · pp. 12–13
• For probes outside the chosen top-r subspace, the general ambient-probe estimate is only O(τ), and equality of the covariance range with the top-r subspace requires exact rank-r structure.If λr+1 > 0, the Matching cover game is larger and outside probes retain only the O(τ) estimate.Source: §A.4 · p. 13
• Matching the full estimated covariance adds bulk eigenvalues that can shrink task directions even when the top subspace is accurate, making full-covariance matching the dangerous naive baseline.Source: §A.4 · p. 13