PaperlayerView exact revision on arXiv

Source-linked AI summary

Closed-Form Subspace Loss and Snapshot-Limited Dimension Selection for Beamspace MVDR Radar

Moris S. Mukwayi, Michael A. van Wyk, Andre M. McDonald

arXiv:2610.37089v1eess.SP

TL;DR

Large-array MVDR beamspace processing trades subspace loss against covariance-estimation loss as the beam window grows under limited training support. The paper derives a closed-form loss and an oracle window-selection rule for unloaded SMI. In the tested cases with multiple runnable windows, the rule selected the best window in all 30 cases.

  • Problem

    For a given number of training snapshots K, window size is commonly chosen by simulation, single-source captured energy, or data-driven rules, while a geometry-based subspace-loss form before data collection is not provided.

    Source: §I · p. 1

  • Method

    The paper derives the DFT-window subspace loss for point jamming and combines it with the fixed-transform scaled-Beta SMI law to obtain B⋆(K) for unloaded SMI.

    Source: §I · p. 1, §IV · p. 3

  • Results

    B⋆(K) selected the best window for unloaded SMI in all 30 tested cases where more than one window could run.

    Source: D. Choosing the window · p. 6, §VIII · p. 7

  • Takeaways & Limitations

    For the tested unloaded-SMI scenarios, window choice follows from subspace loss and training support, while noise-power loading favored windows no smaller than B⋆(K).

    Source: §VIII · p. 7

    B⋆(K) is an oracle rule requiring jammer directions and powers or the true covariance to obtain the subspace loss.

    Source: §IV · p. 3

Abstract

from arXiv · show

A beamspace minimum variance distortionless response (MVDR) processor adapts a window of B beams instead of all N elements of a large array, trading a subspace loss that falls with B against a covariance estimation loss that rises with B for a fixed number K of training snapshots. We give the subspace loss of a DFT beam window in point jamming in closed form: exactly for any jammer power, and in the high interference-to-noise ratio (INR) limit through the Dirichlet kernel of the array. Combined with the known scaled-Beta law of sample matrix inversion (SMI) with a fixed transform, it gives a window size B*(K) for unloaded SMI. On a simulated 32x32 array with carriers from 4 to 8 GHz, noise jammers and Billingsley land clutter, the exact form is within 0.04 dB of the clairvoyant subspace loss for 25- and 49-beam windows at jammer levels from 10 to 50 dB. For a 9-beam window whose three columns are taken by three jammers near the target's elevation, the clutter that the form leaves out costs up to 5 dB. With the nominal K, B*(K) picks the lowest-loss window in all 30 cases with more than one runnable window. The rule is an oracle: it needs the jammer parameters or the true covariance, and errors of 0.25 bin rms in the jammer directions add a mean loss of 0.04 dB. With diagonal loading the best window was larger still wherever the rule chose more than the smallest window.

I. INTRODUCTION

Large-array SMI requires substantial homogeneous training support, motivating reduced-dimension beamspace processing. This paper derives a geometry-based subspace-loss form and a snapshot-dependent window rule, and specifies the array and interference model used for analysis.

I. INTRODUCTION

  • For a 32 × 32 array, SMI needs about 2000 homogeneous range cells in one dwell to approach the optimum, while land clutter is seldom homogeneous over that many cells.The paper describes reduced-dimension processing as adapting after a fixed transform maps the array onto a few channels.Source: §I · p. 1
  • Windowed and tiled beamspace processing has also been used to reduce the cost of massive multiuser MIMO reception.Earlier work studied window-size trade-offs and tiled beamspace processing.Source: §I · p. 1
  • The paper derives a closed-form subspace loss for a DFT window under point jamming and combines it with the scaled-Beta SMI law to select a window size from K.The form is exact at any INR, with a high-INR limit and a one-jammer form through the array's Dirichlet kernel; the resulting B⋆(K) rule is an oracle.Source: §I · p. 1

A. Array and interference

  • The modeled 32 × 32 uniform rectangular array has 1024 elements, and its interference-plus-noise covariance includes noise, independent noise jammers, and land clutter.Training snapshots come from nearby cells and contain no target.Source: A. Array and interference · pp. 1–2

B. MVDR and sample matrix inversion

The paper frames MVDR performance through output SINR and the estimation conditions of sample matrix inversion. In beamspace, a DFT window restricts the processor to a lower-dimensional transform, whose loss with the true covariance is expressed as a projection onto the transformed subspace.

B. MVDR and sample matrix inversion

C. Beamspace processing

  • The beamspace transform keeps a W × W window of DFT beams around the look direction, with its position set by the look direction's fractional bin indices.The transform has B = W^2 orthonormal columns.Source: C. Beamspace processing · p. 2
  • With the true covariance, beamspace subspace loss is the SINR ratio between the beamspace processor and the element-space optimum, with equality when R−1a lies in range(T).The loss is expressed using projection onto the subspace range(R1/2T).Source: C. Beamspace processing · p. 2

III. CLOSED-FORM SUBSPACE LOSS

The section derives a closed-form subspace loss for a DFT beam window without clutter, valid for any jammer power, and gives its high-INR projection limit. For one jammer, that limit is expressed through the array’s Dirichlet response and the jammer energy captured by the window.

III. CLOSED-FORM SUBSPACE LOSS

  • Without clutter, Proposition 2 gives the DFT-window subspace loss for any jammer powers, with the finite-power expression requiring no rank condition.For one jammer, the expression is ρ = [1−γ²/(η+x)]/[1−γ²/(1+x)], where x = 1/(N INR1).Source: §III · p. 2
  • When c² < 1, the loss tends at high jammer powers to a form built from projections onto the jammer spans in element and beamspace.The limit uses pseudo-inverses if A or Ab lacks full column rank.Source: §III · p. 2
  • For one jammer, the high-INR expression depends on γ, the array’s quiescent response toward the jammer, and η, the share of jammer energy captured by the window.With offset (∆p, ∆q) in bins, the separable array factors are Dirichlet kernels.Source: §III · p. 2
  • The limiting loss is small when the window captures much more jammer energy outside its centre beam than the array receives outside the window; multiple-jammer losses must be evaluated jointly.The closed forms assume point sources, an ideal separable manifold, and one carrier; clutter requires computing the loss from the full covariance.Source: §III · p. 3

IV. LOSS DECOMPOSITION AND WINDOW SIZE

For unloaded SMI with a fixed beamspace transform, the loss separates into subspace loss and a finite-snapshot estimation factor, yielding a window-size rule. The rule is an oracle, and its assumptions do not extend exactly to data-dependent selection or diagonal loading.

Loss decomposition and window size

  • The SMI-to-optimum SINR ratio is the subspace factor ρ(T) times a Beta-distributed estimation factor for K ≥ B independent Gaussian snapshots.The factor follows from the fixed transform making beamspace snapshots independent Gaussian vectors in B dimensions.Source: §IV · p. 3
  • Maximizing the product of subspace and estimation factors gives B*(K), the optimum unloaded-SMI window under Proposition 3’s assumptions.The rule requires jammer directions and powers or the true covariance to obtain ρ(B).Source: §IV · p. 3
  • A selector that estimates ρ(B) from the same snapshots used to estimate the covariance makes the transform data-dependent, so the decomposition no longer follows.The passage says an in-sample estimate would be expected to favor larger windows; splitting snapshots restores the assumption but reduces K.Source: §IV · p. 3
  • Clutter training snapshots are not independent Gaussian samples, so the study represents their estimation effect with an equivalent Gaussian support Keff.Keff is inferred from observed loss, not counted as independent snapshots, and Monte Carlo scatter can put it slightly above K.Source: §IV · p. 3
  • Diagonal loading departs from Proposition 3; when loading lies between noise and interference eigenvalues, estimation loss depends on interference rank rather than B.Once that rank stops growing, a larger window costs little, and the unloaded rule does not give an exact optimum for loaded SMI.Source: §IV · p. 3

V. SIMULATION SETUP

The simulations use a checked Simulink testbench to compare beamspace processors across a jamming-only reference and ground-radar clutter scenarios. Scenario G includes Billingsley land clutter, while the evaluation varies carrier, dwell, and training-snapshot conditions.

V. SIMULATION SETUP

  • The testbench’s SMI matches (4) at K = 2N within 0.002 dB, and Simulink matches an independent MATLAB reference within 1.5 × 10^-8 relative.Scenario P is jamming-only; scenario G models a ground radar over low-relief rural land.Source: §V · p. 3
  • Scenario G uses 1° clutter patches over ±90° with Billingsley reflectivity and intrinsic motion, plus unit-mean texture that varies by clutter model.Texture can be held over blocks of 4 or 16 range cells; G2 includes jammers and a clutter ridge inside the window, while G1 uses clutter alone.Source: §V · pp. 3–4
  • Scenario P holds the carrier at 41 frequencies from 4 to 8 GHz for 500 dwells, while the closed-form test uses 21 frequencies and 20 dwells.Scenario G results for Figure 1 and Table 4 hop pseudo-randomly across the band, with 500 dwells per run.Source: §V · p. 4
  • The comparison includes unloaded SMI when K > B, noise-power-loaded SMI, the clairvoyant beamspace processor, and element-space SMI as reference.Losses are averaged over dwells in linear units and reported in decibels.Source: §V · p. 4

A. Closed-form subspace loss

The closed-form point-jammer loss is tested against clairvoyant subspace loss and against its high-INR limit. It is accurate for larger windows in the reported cluttered case, but omitted clutter and tightly packed jammers can materially alter loss.

A. Closed-form subspace loss

  • 0.04 dB: the exact form’s error for B = 25 and 49 stays within this bound across INR and clutter levels in G2.The limit is within 0.05 dB from 20 dB INR upward and within 0.26 dB at 10 dB; the comparison is against clairvoyant loss.Source: A. Closed-form subspace loss · p. 4, Table 2 · p. 4
  • With three jammers near the target elevation, the 3 × 3 window’s three target-row columns are consumed, and the limit gives −10.1 to −16.2 dB across the band.Removing any one jammer yields −1.0 to −4.3 dB; the two outer jammers leak only weakly into the window, so the limit is reached slowly.Source: A. Closed-form subspace loss · p. 4
  • At 50 dB INR, the cluttered 3 × 3 case has clairvoyant loss of −13.6 to −14.6 dB versus −11.9 dB from the exact form.The omitted clutter ridge adds loss after the jammers are nulled; at 10 dB INR, band means are −3.6 dB exact-form and −3.7 dB clairvoyant.Source: A. Closed-form subspace loss · p. 4
  • Across 104 random ideal-array draws, the high-INR limit’s median error against the exact form is 0.17, 0.02, and 0.002 dB for W = 3, 5, and 7.Where κ exceeded 100, error was at most 0.02 dB; in the G2 3 × 3 window, κ = 100 still leaves 0.5 to 1.1 dB error.Source: A. Closed-form subspace loss · p. 4
  • Under manifold perturbations, the limit’s median error stays within 0.14 dB in P and G2, but spreading jammers over five tones yields 1.68 dB error in G2.A look direction 0.2 beamwidth off target gives median errors of 0.28 dB in P and 0.46 dB in G2.Source: A. Closed-form subspace loss · p. 4

B. Loss decomposition

For scenario P, the loss decomposition closely matches simulated unloaded SMI at K = 2B across the tested windows and target azimuths.

B. Loss decomposition

  • Within 0.02 dB: band means from the decomposition and simulation agree for all three windows and all three target azimuths at K = 2B.The reported differences are within the Monte Carlo noise of the 500-dwell means.Source: B. Loss decomposition · p. 5

C. Estimation loss and equivalent support in clutter

In clutter, estimation support generally tracks the independent-snapshot prediction for 5×5 windows, while texture and window size affect support in specific cases. Pooling pulses and loading can change the estimation loss when training data are limited.

C. Estimation loss and equivalent support in clutter

  • For B = 25 with jammers, measured estimation loss closely matches prediction across K = 32–128, with texture changing the result by at most 0.05 dB.At K = 32, 48, 64, 96, and 128, measured losses are −5.62, −2.94, −2.03, −1.23, and −0.88 dB, respectively.Source: C. Estimation loss and equivalent… · p. 5
  • With clutter alone, Rayleigh and K-distributed clutter keep B = 25 equivalent support within 3% of K from K = 32 to 128.In Weibull clutter, B = 25 support falls to 0.87K at K = 128; its estimation loss is −1.04 dB, versus −0.88 dB in Rayleigh clutter.Source: C. Estimation loss and equivalent… · p. 5
  • In Weibull clutter, B = 9 equivalent support falls to 0.64K at K = 128, while B = 25 support falls to 0.87K.With jammers, the 3×3 window also loses support in Weibull and K-distributed clutter where jammer leakage is weak.Source: C. Estimation loss and equivalent… · p. 5
  • For B = 25, pooling 4 or 16 pulses gives 98–99% of nominal support with jammers, while clutter-only support ranges from 86–98%.Pulses from one cell see nearly the same clutter, so pooling adds independent noise samples faster than independent clutter samples.Source: C. Estimation loss and equivalent… · p. 5
  • With noise-power loading, B = 25 reaches −1.85 dB estimation loss at K = 16 with jammers, when unloaded SMI cannot run.In clutter alone, loaded B = 25 has total loss −2.52 dB at K = 16 and −1.09 dB at K = 128, versus −10.66 and −2.83 dB for loaded element-space SMI.Source: C. Estimation loss and equivalent… · pp. 5–6

D. Choosing the window

The best window depends on training support and loading: the oracle rule identifies the best unloaded-SMI window in the tested cases, while loading favors larger windows. The 49-beam window remains close to the best loaded choice.

D. Choosing the window

  • B⋆(K) selects the best unloaded-SMI window in all 30 cases with more than one runnable window.Across all 39 cases, many choices were settled in advance because only the 3×3 window could run or it began about 10 dB down.Source: D. Choosing the window · p. 6, Table 4 · p. 6
  • With clutter alone and K ≥ 32, the smaller estimation loss keeps the 3×3 window best; its margin over the next choice shrinks to a tie at K = 128.The tie is at −1.60 dB and is within Monte Carlo noise.Source: D. Choosing the window · p. 6
  • For loaded SMI, the preferred window is B = 49 with jammers and B = 25 or 49 from K = 32 with clutter alone.Where B⋆ exceeds the smallest window, the best loaded window is larger still; the 49-beam window is never more than 0.2 dB worse than the best.Source: D. Choosing the window · p. 6

E. Sensitivity of the rule to its inputs

The window rule is sensitive chiefly to jammer-direction inputs and remains an oracle requiring jammer parameters or the true covariance. Its tested conclusions are bounded by the simulation settings and assumptions.

E. Sensitivity of the rule to its inputs

  • With 0.25-bin rms direction errors, the rule selects the best window in 86% of cases, with 0.04 dB mean regret and 1.6 dB maximum regret.Errors of ±10 dB in each INR have little effect; omitting one jammer gives 83% best-window selection and 1.3 dB maximum regret.Source: E. Sensitivity of the rule to its inputs · p. 6, Table 5 · p. 7
  • Direction errors matter most; in scenario P, fixed B = 3×3 already has 0.14 dB mean regret versus 0.99 dB for fixed B = 5×5.The route using true covariance in clutter is not assessed, and the study does not estimate rule inputs from data.Source: E. Sensitivity of the rule to its inputs · p. 6

VII. DISCUSSION

  • The discussion recommends evaluating all jammers together, since a single-jammer form can miss windows whose rows or columns several jammers occupy.The high-INR limit is safe only when κ is large relative to c2/(1−c2).Source: §VII · p. 6
  • The propositions are exact only under their stated assumptions; other findings are numerical observations from one laydown, array size, and loading level.Clutter was 5.5 dB below the noise per element in G1, and stronger clutter could make the omitted clutter term matter more.Source: §VII · p. 7

VIII. CONCLUSION

  • In tested scenarios, the closed form gives 25- and 49-beam subspace loss within 0.04 dB, and B⋆ chooses the best window in all 30 cases with multiple runnable windows.For noise-power-loaded SMI, the best window was never smaller than B⋆; estimating rule inputs from training snapshots remains future work.Source: §VIII · p. 7
Loading 2610.37089v1…