PaperlayerView exact revision on arXiv

Source-linked AI summary

Fundamental Limits of Adaptive Beamforming Under Finite Training

Zhiyong Cheng, Shengyao Chen, Di Song

arXiv:2609.06136v1cs.ITeess.SPmath.IT

TL;DR

Finite training lowers adaptive beamforming’s output SINR, raising the question of how much loss is unavoidable under a given training model. The paper derives a local minimax bound from an exact loss identity and gives matching constructions in two complex-Gaussian models. The finite-source ULA coefficient is at most K, while SMI attains N−1 in the unrestricted model.

  • Problem

    The known SMI benchmark gives its own first-order loss, but does not determine the minimum output-SINR loss permitted by the training model or which rule attains it.

    Source: §I · p. 1

  • Method

    An exact identity relates bounded output-SINR loss to clairvoyant MVDR-weight error, enabling a local asymptotic minimax converse and matching constructions in two complex-Gaussian models.

    Source: §I · pp. 1–2

  • Results

    For a ULA with K distinct point interferers and 2K+1 ≤ N, a split one-step beamformer attains Cθ ≤ K on fixed compact regular sets, while SMI attains (N−1)/n in the unrestricted Gaussian model.

    Source: §I · p. 2

  • Takeaways & Limitations

    The difference N−1−Cθ quantifies the first-order value of finite-source structure relative to fitting an unrestricted covariance.

    Source: §VIII · p. 9, §IX · p. 9

    Close sources can make compact-set constants large and delay the finite-sample onset of the local regime.

    Source: §VIII · p. 9

Abstract

from arXiv · show

Finite training reduces the output signal-to-interference-plus-noise ratio (SINR) of an adaptive beamformer, and a natural question is how much of this loss is unavoidable. This paper determines this question, providing a beamforming counterpart of the Cram\'er--Rao bound in spectral estimation. An exact identity expresses the SINR loss as a bounded function of the error in the clairvoyant minimum-variance distortionless-response (MVDR) weight. It yields a local asymptotic minimax lower bound over all measurable data-dependent beamforming rules, including biased and irregular rules. The first-order coefficient is $\tr(\Mb\Jb_{\rm eff}^{-1})$, where $\Jb_{\rm eff}$ describes the information in the training data and $\Mb$ measures the sensitivity of the output SINR. Matching constructions determine this coefficient in two complex-Gaussian models. For an $N$-sensor uniform linear array with $K$ distinct point interferers and $2K+1\le N$, a data-driven split one-step beamformer attains the coefficient $C_\theta\le K$ at every interior scene of a fixed compact regular parameter set. For unrestricted covariance matrices, sample matrix inversion (SMI) attains the coefficient $N-1$ through the classical Reed--Mallett--Brennan law. The difference quantifies the first-order value of finite-source structure. Geometric formulas and numerical results describe the dependence on interference power and array geometry, and the finite-sample departure near a weak-source boundary.

I. INTRODUCTION

Finite training makes the adaptive beamformer’s output SINR fall below the known-covariance MVDR value, motivating a model-wide lower bound on this loss. The paper links the exact loss to weight error, derives a local minimax coefficient, and gives matching constructions for two complex-Gaussian models.

I. INTRODUCTION

  • The finite-training benchmark question goes beyond SMI’s Reed–Mallett–Brennan result: what minimum output-SINR loss does the training model require, and which rule attains it?Under independent complex-Gaussian training, SMI has first-order mean-loss coefficient N−1 for N sensors.Source: §I · p. 1
  • An exact identity makes normalized output-SINR loss a bounded function of error in the clairvoyant MVDR weight, with a quadratic leading term near the optimum.This permits studying finite-training beamforming as MVDR-weight estimation without fixing a covariance estimator or beamforming algorithm.Source: §I · p. 1
  • The first-order coefficient combines per-snapshot Fisher information, which measures state distinguishability, with loss curvature, which measures how state error changes output SINR.In the regular fixed-dimensional regime, the loss rate is 1/n and the coefficient determines differing training requirements.Source: §I · p. 1
  • The local minimax lower bound applies to all measurable data-dependent beamforming rules, and matching constructions attain the first-order coefficient in finite-source ULA and unrestricted Gaussian models.For the ULA, the split one-step beamformer attains Cθ/n under the stated array and regularity conditions; SMI attains (N−1)/n in the unrestricted model.Source: §I · p. 2
  • For a ULA with K distinct point interferers and 2K+1 ≤ N, a fully data-driven split one-step beamformer attains the bound at every interior scene of a fixed compact regular parameter set.It fits a rough scene on a pilot block, then applies one Fisher-scoring correction using independent snapshots; source separation is not needed for achievability at fixed N and K.Source: §I · p. 2
  • Numerical results compare structured reconstructions with the finite-source coefficient, verify the sharp SMI coefficient, and examine finite-sample onset near a weak-source boundary.The boundary concerns an interferer whose power is barely sufficient for reliable localization.Source: §I · p. 2

A. Relations to Prior Art

Prior work studies finite-training performance for particular sample-covariance processors, structured beamformers, and MVDR-weight estimators. This paper distinguishes its task-level, local minimax bound from those processor-specific analyses and from bounds or robust designs addressing different losses or uncertainties.

A. Relations to Prior Art

  • Analyses of SMI and related processors, as well as loading, shrinkage, and random-matrix corrections, characterize specified methods but do not give the best first-order coefficient over the full rule class.These approaches provide detailed finite-sample performance or reduce covariance-inversion instability.Source: A. Relations to Prior Art · p. 2
  • Dominant-mode, reduced-rank, and parametric reconstruction methods exploit interference structure, while Section IV identifies the model-wide coefficient behind the resulting array-size reduction.The cited prior approaches restrict adaptation to an estimated interference subspace or estimate source directions and powers before forming the MVDR weight.Source: A. Relations to Prior Art · p. 2
  • Earlier work estimates the MVDR weight using CRBs, influence functions, or modified sample filters; this paper instead uses exact bounded output-SINR loss and a local minimax criterion over the full rule class.The paper also gives attaining rules for the two models it studies.Source: A. Relations to Prior Art · p. 2
  • Ziv–Zakai-family bounds address nonlocal or prior-dependent parameter mean-square error, whereas this paper bounds exact output-SINR degradation for a beamforming-weight action.Lemma 2 connects that task loss to the local asymptotic minimax bound.Source: A. Relations to Prior Art · p. 2
  • Robust adaptive beamforming addresses steering, covariance, or distributional uncertainty, while this paper fixes the training model and studies sampling uncertainty alone.The robust approaches use a prescribed uncertainty set or ambiguity model.Source: A. Relations to Prior Art · p. 2

B. Notation

The notation section sets conventions for vectors, matrices, norms, projections, derivatives, and probability distributions used throughout the paper.

B. Notation

  • Bold lowercase and uppercase letters denote vectors and matrices; ∗ and ⊤ denote Hermitian transpose and transpose, respectively.For Hermitian matrices, A ⪰ B means A−B is positive semidefinite, and H_N denotes the positive-definite Hermitian matrices.Source: B. Notation · p. 2
  • The paper uses Euclidean or spectral norms, Frobenius norm, covariance-weighted inner products, and positive-definite matrix square roots.It also defines orthogonal projections for nonzero vectors and full-column-rank matrices.Source: B. Notation · p. 2
  • Distributions CN(μ,R) and N(μ,Σ) denote circular complex and real Gaussians, while E_ϑ denotes expectation and ⇒ denotes convergence in distribution.Derivatives use real coordinates; constants in fixed-dimensional compact-set arguments may depend on the chosen set, N, and K.Source: B. Notation · p. 3

C. Organization

The paper proceeds from model and loss definitions to a general converse, then matching constructions, coefficient interpretation, numerical results, and closing discussion.

C. Organization

  • Section II defines the training model, output-SINR loss, and local risk; Section III gives the general local asymptotic minimax converse.Proofs are deferred to Appendices A–H.Source: C. Organization · p. 3
  • Sections IV and V establish matching achievability for the finite-source ULA and unrestricted Gaussian models, and Section VI interprets the finite-source coefficient.Source: C. Organization · p. 3
  • Section VII presents numerical results, followed by discussion and conclusion in Sections VIII and IX.Source: C. Organization · p. 3

II. STATISTICAL MODEL AND SINR LOSS

The section models independent training snapshots and defines beamforming loss relative to the clairvoyant MVDR solution. It connects local state uncertainty and the task’s SINR sensitivity to a minimax risk over nearby states.

Training model

  • The training model uses n independent, zero-mean snapshots with positive-definite covariance, and beamforming rules are measurable mappings from snapshots to weights.The look-direction steering vector is known.Source: §II · p. 3

SINR loss

  • The exact loss is bounded and depends on excess output power over the clairvoyant MVDR value; near the optimum, it is a quadratic weight error under the task metric.The leading metric is R(ϑ)/P⋆(ϑ).Source: §II · p. 3

Information and task metric

  • Per-snapshot Fisher information describes how well nearby states are distinguished, while loss curvature measures how state errors affect output SINR.Their trace pairing determines the first-order coefficient.Source: §II · p. 3

Nuisance coordinates

  • A Schur complement removes nuisance coordinates that change the data distribution without changing the optimum weight before evaluating the coefficient.The paper uses this adjustment for the finite-source model.Source: §II · p. 3

Local risk

  • The local risk tests one rule over states perturbed by h/√n, avoiding the trivial zero risk available when optimizing at a single fixed state.The first-order analysis keeps model dimension fixed as snapshots increase.Source: §II · p. 3

III. A GENERAL LOCAL ASYMPTOTIC MINIMAX CONVERSE

Under regularity conditions, the section establishes a local asymptotic minimax converse for the exact bounded output-SINR loss. Its coefficient is Cϑ and the bound applies even to biased or irregular measurable rules.

General converse

  • Theorem 1 gives a general local asymptotic minimax lower bound on output-SINR loss with coefficient Cϑ.The one-snapshot model must be differentiable in quadratic mean, have nonsingular Fisher information, and have a nearby continuously differentiable positive-definite covariance map.Source: §III · p. 4

Scope of the bound

  • The converse covers every measurable beamforming rule, including biased and irregular rules, without assuming unbiasedness, consistency, or asymptotic normality.It transfers a truncated quadratic-loss bound to the exact bounded SINR loss, including rules with zero look gain.Source: §III · p. 4

What determines the limit

  • The training model contributes Fisher information and the beamforming task contributes loss curvature, so equal covariance maps can yield different limits when observations carry different information.Source: §III · p. 4

IV. THE FINITE-SOURCE ULA LIMIT

For a finite-source ULA, the split one-step beamformer attains the local minimax coefficient at every interior scene in the compact regular construction. The coefficient is bounded by the number of interferers, and the section also characterizes the limiting loss distribution.

Model and regularity

  • The model has K distinct source directions, positive source powers and noise variance, with fixed N and K; the regular scene set is compact and ordered.The construction needs no quantitative lower bound on NΔ.Source: §IV · p. 4

Loss geometry

  • The eigenvalues λs quantify how strongly locally estimable parameter directions affect output-SINR loss.Source: §IV · pp. 4–5

Split one-step construction

  • The attaining rule fits a structured scene on a pilot block, then applies one Fisher-scoring correction using independent remaining snapshots before forming the MVDR weight.The split exponent lies between one-half and one, separating pilot estimation from the update.Source: §IV · p. 5

Attainment

  • At every interior point of the fixed compact parameter set, the split one-step beamformer attains the sharp first-order local minimax coefficient on each fixed local ball.The converse supplies the lower bound and the attaining rule supplies the upper bound.Source: §IV · p. 5

Dimension bound

  • With 2K+1 ≤ N, the finite-source coefficient satisfies Cθ ≤ K ≤ (N−1)/2 < N−1, and its upper bound depends on interferer count rather than array size.Source: §IV · p. 5

Loss distribution

  • The normalized loss has a limiting distribution expressed through 2K independent standard normal variables, with convergence of all moments.Its eigenvalues determine the limiting mean, first-order quantiles, and outage probabilities.Source: §IV · p. 5

Risk conventions

  • The bounded loss converse also bounds expected excess output power and ratio-of-expectations loss, and the one-step rule gives all three conventions the same first-order coefficient.Source: §IV · p. 5

V. THE UNRESTRICTED GAUSSIAN LIMIT

In the unrestricted complex-Gaussian covariance model, the SMI coefficient is both the classical RMB mean-loss coefficient and the best first-order guarantee over all measurable rules. The local minimax criterion compares a single rule across nearby covariance alternatives.

V. THE UNRESTRICTED GAUSSIAN LIMIT

  • The SMI sample covariance is almost surely positive definite, and its Gaussian experiment is quadratic-mean differentiable in covariance coordinates.The section defines the SMI weight and its risk as functions of the covariance chart.Source: §V · p. 5
  • N−1 is the first-order coefficient for unrestricted SMI risk and the best guarantee over all measurable rules in this model.The coefficient is independent of the covariance R.Source: §V · p. 6
  • The unrestricted-model converse evaluates the general coefficient on the full Hermitian covariance tangent space, while the exact SMI distribution is the classical RMB law.The RMB distribution and mean are independent of R.Source: §V · p. 6
  • A rule tuned to lower risk at one covariance may incur higher risk nearby because local minimax control covers all 1/√n alternatives at once.The result is local minimax, not a pointwise comparison of rules at a single covariance.Source: §V · p. 6

VI. GEOMETRY OF THE FINITE-SOURCE LIMIT

The finite-source coefficient is organized by source-wise angle and relative-power contributions, with coupling bounds and geometric limits describing its dependence on the scene. At high INR, power-estimation errors have vanishing first-order effect on output-SINR loss, leaving projected tangent geometry to set the limit.

VI. GEOMETRY OF THE FINITE-SOURCE LIMIT

  • Under quantitative separation, powers and noise stay within fixed ranges, while distinct source frequencies have gaps at least Δ with α0=NΔ≥C0K.For K≥2, packing with these gaps requires N ≳ K^2; the condition is only for dimension-uniform estimates.Source: §VI · p. 6
  • Csrc assigns each source an angle term and a relative-power term, and its coupling bound has vanishing relative error when εθ→0.This conclusion follows in particular when K/N→0 and K/α0→0 with the power and noise ranges fixed.Source: §VI · p. 7
  • When one interferer is present, its angle and relative-power contributions are determined by the corresponding whitened cosines.The angle contribution and relative-power contribution are distinguished in the per-source expression.Source: §VI · p. 7
  • At high INR, power-estimation errors' first-order effect on output-SINR loss vanishes, and the limit is set by projected tangent geometry.For a ULA with distinct source directions and 2K+1≤N, the stated independence conditions hold when the look direction differs from every source direction.Source: §VI · p. 7
  • The look-separation bound requires the look direction to be at least Δℓ from every source and bounds the sidelobe-scale contribution of separated interferers.The bound's constant depends on the stated parameter ranges and η.Source: §VI · p. 7

VII. NUMERICAL RESULTS

The simulations compare practical finite-source reconstructions with their converse coefficient and SMI with its sharp unrestricted coefficient. They examine coefficient accuracy, dependence on INR and array geometry, and the finite-sample onset near weak sources.

VII. NUMERICAL RESULTS

  • At n=2N, root-MUSIC+LS and SML are about 1.08 and 1.03 relative to Cθ, while SMI approaches N−1=19.Across the simulated range, both structured curves remain between about 0.96 and 1.08 relative to Cθ≈0.0133; error bars do not resolve deviations below a few percent. The scene uses N=20, K=3, and 3000 Monte Carlo trials.Source: §VII · p. 7, Figure 1 · p. 8
  • The relative coupling approximation error is 8.49×10^-5, with Csrc=0.01334408.The associated bound is conservative because it controls the full off-diagonal Fisher block in spectral norm.Source: §VII · p. 8
  • Above 10 dB, the finite-INR excess over the high-INR limit falls by about an order of magnitude per additional 10 dB; at 20 dB the difference is 1.18×10^-4.Figure 2 compares Cθ with the high-INR limit on a logarithmic scale.Source: §VII · p. 8, Figure 2 · p. 8
  • For one interferer, Cθ varies by several orders of magnitude across array-response sidelobes and nulls, while the 20 and 40 dB curves nearly coincide away from the look direction.The oscillations follow the centered-ULA identity.Source: §VII · p. 8, Figure 3 · p. 8
  • At n=60, root-MUSIC+LS loss is close to Cθ/n at moderate and high INR but rises rapidly when source powers become too small for reliable localization.The empirical onset—the largest INR where loss exceeds 2Cθ/n—moves lower as n increases; this finite-sample experiment concerns root-MUSIC+LS, whereas the converse is for each fixed regular scene.Source: §VII · p. 9, Figure 4 · p. 9

VIII. DISCUSSION

The coefficient Cθ separates training-model information from algorithmic efficiency and quantifies the first-order value of finite-source structure. The discussion extends the converse beyond ULAs under regularity and identifiability conditions, while delimiting the finite-source analysis to known, fixed model order.

VIII. DISCUSSION

  • N−1−Cθ measures the additional first-order cost of fitting an unrestricted covariance, while Cθ≤K shows the structural gain does not require widely separated interferers.Close sources can nevertheless enlarge compact-set constants and delay the finite-sample onset of the local regime.Source: §VIII · p. 9
  • The general converse applies to calibrated arrays when the training model is regular and locally identifiable; one-step achievability additionally requires compact-set regularity and pilot-localization conditions.For ULAs, Vandermonde uniqueness supplies identifiability, while Dirichlet-kernel estimates provide additional dimension-uniform bounds.Source: §VIII · p. 9
  • The finite-source limit assumes known, fixed K; unknown model order and source collisions require different analyses.Collisions produce a singular local model.Source: §VIII · p. 9

IX. CONCLUSION

The paper establishes matching first-order limits for finite-source ULA and unrestricted Gaussian beamforming, and relates their coefficient gap to finite-source structure. Its appendix proof also expresses exact SINR loss as a bounded function of an error quantity, with a local linear approximation for small errors.

IX. CONCLUSION

  • The local asymptotic minimax converse covers all measurable beamforming rules, including biased and irregular rules; matching constructions attain the finite-source and unrestricted coefficients.For regular ULA scenes, 0<Cθ≤K; in the unrestricted Gaussian model, SMI attains N−1, so N−1−Cθ quantifies the first-order value of finite-source structure.Source: §IX · p. 9

APPENDIX A PROOF OF LEMMA 1

  • With x=E/P⋆, the exact loss is x/(1+x)≤x and is x(1+o(1)) as x→0.This proof step gives a bounded loss expression and its small-error approximation.Source: Appendix A · p. 9

APPENDIX B PROOF OF THEOREM 1

The appendices provide proofs and technical foundations for the finite-source and unrestricted-covariance results, establishing ULA coherence and Fisher-information properties, regularity and loss-coefficient bounds, split one-step efficiency, and SMI attainment of the coefficient $N-1$.

APPENDIX B PROOF OF THEOREM 1

  • The proof normalizes arbitrary measurable beamformers, applies a local asymptotic minimax bound to the optimum-weight functional, and transfers it to the exact bounded SINR loss.The argument uses local asymptotic normality and the functional’s derivative, then extends bounds from finite parameter subsets to expanding local balls.Source: Appendix B · p. 10

APPENDIX C FISHER ORTHOGONALITY AND DIAGONAL DOMINANCE OF

  • For separated ULA sources with α0 ≳ K, distinct-node coherences remain O(ρoff) after whitening, and the Gram matrix differs entrywise from identity by O(ρoff).The corresponding inverse matrix has diagonal entries ≍1 and off-diagonal entries O(ρoff); the same coherence orders extend to S2 normalizers and, with look-direction separation, to source–look terms at scale ρℓ.Source: Appendix C · p. 11
  • For one source, angle–power and angle–noise Fisher entries are exactly zero; for K ≥ 2, normalized cross-block entries are O(ρoff), yielding angle–nuisance block decoupling up to that order.The same-source whitened coherence is exactly zero at K = 1 and bounded by Cρoff for multiple sources.Source: Appendix C · pp. 12–13
  • Under Assumption 1 with α0 ≳ K, Lemma 6 establishes diagonal dominance of the profiled efficient Fisher information and bounds its trace comparison when scene coupling is below one.The proof uses a normalized score Gram, a Gershgorin bound, and an exact rank-one profiling update.Source: Appendix C · p. 13
  • Under (A3), diagonal-dominance margins ensure JL ≻ 0 and support the claimed trace comparison.The argument controls the profiling update and cross-source terms before comparing the trace.Source: Appendix C · p. 14

FINITE-SOURCE BOUNDS

  • The finite-source results also include a look-separation bound alongside the high-INR asymptotic limit.The high-INR expansion is uniform over the stated bounded normalized-power range with fixed array geometry, N, K, and noise variance.Source: FINITE-SOURCE BOUNDS · pp. 14–16
  • At every interior regular scene, the finite-source coefficient satisfies 0 < Cθ ≤ K.Whitened-basis identities also establish per-source angle and power accounting, a high-INR limit, and a look-separation bound.Source: FINITE-SOURCE BOUNDS · pp. 14–16

APPENDIX E POINTWISE AND UNIFORM REGULARITY

  • At every interior regular scene, the covariance and MVDR functional are real-analytic, the Fisher information is positive definite, and the experiment is locally asymptotically normal at rate √n.The efficient information JL is positive definite, and the MVDR functional has ∂γψ = 0.Source: Appendix E · pp. 17–18
  • With 2K + 1 ≤ N, distinct interferer angles give nonsingular full and efficient Fisher matrices even without a quantitative separation condition.The proof uses the rank of the confluent Vandermonde system at distinct steering nodes.Source: Appendix E · p. 17
  • For fixed N and K on compact Θ, J and J_L are uniformly positive definite, with bounded Lipschitz inverses and bounded derivatives of R and ψ.The regularity holds throughout Θ, and the constants require no condition of the form NΔ ≥ C0K.Source: Appendix E · p. 18
  • Under Assumption 1, covariance and rescaled loss-relevant efficient information have dimension-uniform bounds.The rescaled information is bounded above and below by positive multiples of the identity.Source: Appendix E · pp. 18–19
  • The trace identity gives C_theta ≤ K, even without quantitative separation.This bound is established in Appendix E.Source: Appendix E · p. 19

APPENDIX F SPLIT ONE-STEP EFFICIENCY

  • With an o(n^-1/4) pilot localization rate, the split one-step update has a uniform asymptotic expansion under bounded local alternatives, with an all-order remainder.The pilot block localizes the parameter; the update block supplies score fluctuation, and Fisher-scoring cancels pilot error to first order.Source: Appendix F · p. 19
  • The lexicographically selected least-squares pilot is measurable and localizes at rate m_n^-1/2 with all moments.Covariance identification follows from unique Vandermonde decomposition of the rank-K signal term for K < N.Source: Appendix F · pp. 19–20
  • A least-squares pilot with m_n=min{ceil(n^ω), n-1}, 1/2<ω<1, uses a vanishing snapshot fraction while preserving full first-order Fisher information.Its rate is r_n=m_n^-1/2.Source: Appendix F · p. 21

APPENDIX G PROOFS OF THEOREM 2 AND COROLLARY 1

  • For regular ULA scenes, Lemmas 7 and 9 establish the conditions for Theorem 1 and the attaining construction, without quantitative separation.Theorem 2’s converse follows from Theorem 1.Source: Appendix G · p. 22

A. Achievability

  • The fixed-chart split one-step estimator achieves Theorem 2 by converting its quadratic loss expansion to SINR loss, with uniform moments controlling the remainders.The plug-in MVDR weight is distortionless, so the loss identity applies without an exceptional event.Source: A. Achievability · p. 22

B. Proof of Corollary 1

  • The split one-step loss has a limiting distribution and convergence of every moment, supported by Lyapunov convergence and uniform integrability.Uniform moment bounds on the one-step estimator ensure the scaled loss is uniformly integrable.Source: B. Proof of Corollary 1 · pp. 22–23

A. The exact SMI risk

  • The Reed–Mallett–Brennan law has covariance-independent output-SINR risk, since whitening reduces the risk to the same white-noise experiment, including under local alternatives.The whitening and unitary transformations preserve SINR loss.Source: A. The exact SMI risk · p. 23

C. Application of the general converse

D. Evaluation of the full trace

  • The 2(N-1) loss-relevant real directions each contribute half a unit, yielding trace coefficient N-1.Each direction has unit Fisher information and task curvature 1/2; the other Hermitian basis directions have zero task curvature.Source: D. Evaluation of the full trace · p. 24

E. Achievability by SMI

  • SMI matches the unrestricted converse at coefficient N-1 in the iterated expanding-neighborhood limit.The converse and achievability meet at the same coefficient.Source: E. Achievability by SMI · p. 24
Loading 2609.06136v1…