PaperlayerView exact revision on arXiv

Source-linked AI summary

Rician MIMO Channel- and Jamming-Aware Decision Fusion

D. Ciuonzo, A. Aubry, V. Carotenuto

arXiv:1702.07915v1cs.IT

TL;DR

The paper addresses channel-aware decision fusion when a multi-antenna DFC receives simultaneous sensor decisions over Rician channels with only statistical scattered-fading information. It develops low-complexity fusion rules and GLRT-like extensions for partially unknown jamming, finding regime-dependent rule advantages and significant gains from interference awareness.

  • Problem

    MIMO decision fusion with Rician channels lacks practical rules that handle exponential LRT complexity, incomplete scattered-channel knowledge, and unknown jamming interference.

  • Method

    The paper derives IS, NLOS, WL, and IGMM rules, then extends IS, NLOS, and IGMM through composite hypothesis testing and GLRT-like formulations for subspace jamming.

  • Results

    The rules exhibit regime-dependent equivalences and performance, while interference-aware rules significantly outperform interference-unaware counterparts at moderate-to-high SNR with non-negligible LOS interference.

  • Takeaways & Limitations

    IGMM is most appealing across the three interference-free channel scenarios, while interference-aware fusion provides substantial jamming suppression in the considered regimes.

Abstract

from arXiv · show

In this manuscript we study channel-aware decision fusion (DF) in a wireless sensor network (WSN) where: (i) the sensors transmit their decisions simultaneously for spectral efficiency purposes and the DF center (DFC) is equipped with multiple antennas; (ii) each sensor-DFC channel is described via a Rician model. As opposed to the existing literature, in order to account for stringent energy constraints in the WSN, only statistical channel information is assumed for the non-line-of sight (scattered) fading terms. For such a scenario, sub-optimal fusion rules are developed in order to deal with the exponential complexity of the likelihood ratio test (LRT) and impractical (complete) system knowledge. Furthermore, the considered model is extended to the case of (partially unknown) jamming-originated interference. Then the obtained fusion rules are modified with the use of composite hypothesis testing framework and generalized LRT. Coincidence and statistical equivalence among them are also investigated under some relevant simplified scenarios. Numerical results compare the proposed rules and highlight their jammingsuppression capability.

I. INTRODUCTION

The paper studies decision fusion over a Rician fading multiple-access channel with multiple DFC antennas, statistical knowledge of scattered fading, and energy-efficient anomaly-detection settings. It develops reduced-complexity fusion rules and extends them to partially unknown jamming interference.

  • I. INTRODUCTION: Decision fusion over a multiple-access channel uses simultaneous sensor transmissions and multiple DFC antennas to form a distributed or virtual MIMO channel.The setup is motivated by increased spectral efficiency and diversity against small-scale fading.
  • I. INTRODUCTION: The LLR is computationally impractical because its Gaussian-mixture complexity grows exponentially with the number of sensors and requires strong system knowledge.The mixture under each hypothesis contains 2^K components.
  • I. INTRODUCTION: The Rician model assumes that only the LOS component is known at the DFC, while scattered fading is not estimated.Unequal long-term received sensor powers are also represented through path loss and shadowing.
  • I. INTRODUCTION: The paper derives IS, NLOS, WL, and IGMM sub-optimal rules, followed by IS-GLRT, NLOS-GLRT, and IGMM-GLRT rules for jamming.The proposed rules address both complexity and incomplete model knowledge.
  • I. INTRODUCTION: The study compares rule performance and investigates asymptotic equivalences, Rician parameters, thermal noise, receive-antenna count, and complexity.Simulation studies are used for these comparisons and verifications.

III. FUSION RULES

The fusion-rule section starts from the optimal LLR and then introduces simpler rules based on ideal-sensor and NLOS design assumptions. These rules reduce implementation requirements while retaining channel-aware received-signal processing.

  • III. FUSION RULES: The optimal LLR compares the log likelihood ratio with a threshold to decide between H1 and H0.The threshold may be selected using Neyman–Pearson or Bayesian criteria.
  • III. FUSION RULES: Because each hypothesis induces a Gaussian-mixture received vector, direct LLR evaluation has O(2^K) complexity and differing component means and covariances.This motivates sub-optimal fusion rules.
  • III. FUSION RULES: The IS rule applies perfect local decisions only for design, producing a weighted combination of maximum-ratio combining and energy detection.It does not require sensor-performance probabilities for implementation.
  • III. FUSION RULES: The NLOS rule assumes κk = 0 at design time and uses the received energy ∥y∥2 as its decision statistic.This choice requires no sensor-performance information and is motivated by simplicity and earlier monotonicity results in a restricted case.

D. Widely-linear (WL) rules

The WL and IGMM rules approximate the mixture-valued detection problem through low-order received-signal statistics, with their relationships to IS and NLOS rules depending on channel and sensor regimes.

  • D. Widely-linear (WL) rules: WL processing uses an augmented received vector because the conditional received vector is improper complex-valued.The WL statistic is compared with a threshold and requires second-order knowledge of x|Hi.
  • D. Widely-linear (WL) rules: The optimized WL design is limited by non-closed-form optimization, non-trivial computation, false-alarm dependence, and the underlying Gaussian-mixture model.These drawbacks motivate the adopted simplified WL construction.
  • D. Widely-linear (WL) rules: WL weights are selected by maximizing normal or modified deflection measures, yielding a closed-form rule with lower complexity than globally optimized alternatives.The paper notes acceptable performance loss relative to the LLR in analogous settings.
  • F. Asymptotic equivalences: At low Rician factors, IS and IGMM become statistically equivalent to the NLOS rule, collapsing to received-energy detection.This corresponds to sensors approaching a purely scattered, NLOS condition.
  • F. Asymptotic equivalences: Under weak LOS, IGMM can exhibit the same linear-quadratic dependence as IS, while NLOS lacks the corresponding mean-dependent contribution.The weak-LOS regime is also likely at low SNR and with good-quality sensors.
  • F. Asymptotic equivalences: Under the IS assumption, IGMM is statistically equivalent to IS and therefore attains optimum performance, whereas WL retains only the widely-linear part.For ideal sensors, the paper consequently expects WL performance loss relative to IS and IGMM.

IV. JAMMER (SUBSPACE) INTERFERENCE ENVIRONMENT

The jammer extension models interference as a Rician-channel signal added to the useful received vector and treats unknown jammer parameters through composite hypothesis testing. The resulting framework covers distributed, co-located, and multiple-jammer configurations.

  • IV. JAMMER (SUBSPACE) INTERFERENCE ENVIRONMENT: The received signal under jamming is modeled as ys = y + sJ, where the jammer has an r-dimensional transmitted signal and a Rician channel.The jammer channel follows the same general Rician structure as the WSN channel.
  • IV. JAMMER (SUBSPACE) INTERFERENCE ENVIRONMENT: The jammer model includes steering, scattered-channel, Rician-factor, and large-scale-fading matrices, with distributed or co-located transmitting antennas.The distributed case includes multiple jammers.
  • IV. JAMMER (SUBSPACE) INTERFERENCE ENVIRONMENT: The jammer signal is classified as a constant jammer, although its continuously emitted radio signal may change with time and remains unknown at the DFC.The model is presented as a first step toward rules robust to smarter jammers.
  • IV. JAMMER (SUBSPACE) INTERFERENCE ENVIRONMENT: The DFC is assumed to know only the jammer steering matrix, not its Rician factors, large-scale fading, or transmitted signal.The resulting unknown deterministic interference parameters motivate sub-optimal fusion rules.

A. Clairvoyant LRT and GLRT

The clairvoyant LRT is optimal but impractical because jamming parameters are unknown. GLRT addresses the resulting composite hypotheses, yet its exact implementation is computationally infeasible, motivating simplified GLRT-like rules.

  • The clairvoyant LRT assumes the jamming parameters {ψ, DJ, RJ} are known and serves as the comparison benchmark.
  • The LRT is uniformly most powerful, so no other fusion rule is expected to outperform it.
  • Unknown jamming parameters prevent practical LRT implementation and create a composite hypothesis testing problem.
  • The GLRT maximizes the likelihood under both hypotheses with respect to the unknown jamming parameters.
  • Exact GLRT implementation is infeasible because each hypothesis likelihood is a 2^K-component Gaussian mixture and requires nonlinear optimization.
  • The paper therefore exploits GLRT philosophy together with simplifying assumptions to construct computationally efficient, jamming-robust rules, including IS-GLRT.

C. NLOS-GLRT rule

The NLOS-GLRT simplifies the jammer-aware test by assuming no sensor LOS paths and applying Gaussian moment matching. Under this model, the decision depends monotonically on projected received-signal energy, avoiding explicit jammer-power estimation.

  • The NLOS-GLRT begins by imposing κk = 0 for every sensor, eliminating LOS components from the conditional received-signal model.
  • Even under simplifying assumptions, the received-signal density remains a 2^K-component complex Gaussian mixture, making direct evaluation computationally intractable.
  • Gaussian moment matching fits the received-signal density under each hypothesis to a proper complex Gaussian model.
  • The derivation estimates the jammer waveform-related parameter ζ under each hypothesis and forms concentrated matched likelihoods.
  • The test statistic is an increasing function of ∥r0∥2 independently of the unknown jammer variance σ2_J, so estimating σ2_J can be avoided.
  • NLOS-GLRT is uniformly most powerful under the NLOS assumption after moment matching, despite not requiring σ2_J estimation.

E. Asymptotic equivalences in the presence of jammer

With jammer interference, the three GLRT-derived rules generally use different design criteria, but simplifying regimes reveal shared behavior and exact equivalences. Under the ideal-sensor assumption, IGMM-GLRT, IS-GLRT, and the exact GLRT are statistically equivalent.

  • In the jammer setting, NLOS-GLRT implicitly estimates ζ but avoids estimating σ2_J, whereas IS-GLRT and IGMM-GLRT estimate σ2_J.
  • Because their design criteria differ, identical performance among NLOS-GLRT, IS-GLRT, and IGMM-GLRT is not generally expected in NLOS conditions.
  • After projecting out the jammer LOS component, IS-GLR and IGMM-GLR increase with the received-signal energy ∥r0∥2 over the stated variance interval.
  • When the statistic is safely approximated as increasing in ∥r0∥2, IS-GLR and IGMM-GLR attain the same performance as NLOS-GLRT.
  • If jammer variance uncertainty yields overlapping overall-variance intervals under both hypotheses, discrimination cannot be achieved through simple variance estimation.
  • Under the IS assumption, IGMM-GLRT is statistically equivalent to IS-GLRT and therefore attains exact GLRT performance.
  • With ideal sensors, the equivalence follows because the covariance structure does not change between hypotheses; with good-quality sensors, NLOS-GLRT may lose performance by ignoring LOS components.

V. COMPLEXITY ANALYSIS

The optimum LLR is computationally unfeasible for large sensor networks, whereas the proposed rules have polynomial complexity. IGMM-GLRT is the most computationally demanding because it requires quadratic forms and polynomial-root solving.

  • The LLR is unfeasible, especially when the number of sensors K is very large, while all proposed rules have polynomial complexity.
  • IS and NLOS rules scale linearly with the number of receive antennas N, with NLOS-GLRT additionally requiring projection onto the jammer-orthogonal subspace.
  • IGMM requires O(N^2) complexity because its statistic is a quadratic form of the received vector.
  • IGMM-GLRT additionally computes vi and solves a polynomial of order pord = 2N + 4(N − r) − 1.
  • Using a Sturm-based approach, polynomial solving has complexity O(pord^4τ^2), where τ depends on coefficient bit resolution.

A. Setup description and measures of performance

The study evaluates Rician MIMO decision fusion under no-jamming and jamming conditions, using simulated sensor deployments, varying channel factors, noise, and DFC antenna counts. Performance is measured through false-alarm and detection probabilities, while comparisons examine rule behavior across LOS, Intermediate, and NLOS scenarios.

  • Network and channel setup: Sensors are randomly deployed around the DFC with bounded distances, log-normal large-scale fading, and unequal received powers modeled through path loss and shadowing.The simulated area has radius rmax = 1000 m, minimum sensor distance rmin = 100 m, path-loss exponent L = 2, and shadowing parameters (µP, σP) = (15, 2).
  • Network and channel setup: The DFC uses a half-wavelength-spaced uniform linear array, while jammer parameters are generated analogously with higher received-power parameters to represent non-negligible interference.The jammer model uses (µP, σP) = (25, 2), and jammer devices are distributed in angular space.
  • Scenario generation: Rician factors define LOS, Intermediate, and NLOS WSN scenarios, generated respectively over [10, 20] dB, [−10, 10] dB, and a lower-factor interval.The supplied passage explicitly identifies the first two intervals and introduces the NLOS interval, whose endpoint is truncated.
  • Measures of performance: Performance is evaluated using system probabilities of false alarm and detection for statistics Λ and thresholds γ, under conditionally independent sensor decisions.The experiments compare rules across noise, interference, Rician scenarios, and receive-antenna counts.
  • No-interference comparisons: In no-interference comparisons, the LLR performs best overall; WL rules are closest in LOS, while NLOS, IS, and IGMM behavior becomes nearly coincident in NLOS conditions.Increasing antenna count benefits LLR, WL, and IGMM generally, whereas IS and NLOS benefit mainly at low SNR or in NLOS settings.
  • Interference comparisons: Under interference, IS-GLRT, NLOS-GLRT, and IGMM-GLRT generally outperform their interference-unaware counterparts, while IGMM-GLRT is worse in the NLOS case because unknown-parameter supports overlap.For IS-GLRT and NLOS-GLRT, detection probability increases with antenna count and their advantage over unaware rules also grows; suppression can partially cancel sensor contributions.

VII. CONCLUSIONS

The paper develops channel-aware and interference-aware fusion rules for Rician MIMO decision fusion, then compares their behavior across LOS, intermediate, and NLOS WSN setups. Results show scenario-dependent rule preferences, benefits from more receive antennas, and asymptotic equivalences under simplified conditions.

  • Contributions: The study develops five sub-optimal fusion rules for Rician MIMO decision fusion when only LOS channel components are known at the DFC.The rules address the exponential complexity of the LRT and limited system knowledge.
  • No-jamming performance: In LOS setups, WL rules are the most convenient alternative to the LLR, whereas their performance degrades severely in NLOS conditions.The NLOS rule is mainly appealing in NLOS setups, while IS also performs satisfactorily in weak-LOS conditions.
  • Interference-aware performance: IGMM performs best in LOS setups, while IS-GLRT and NLOS-GLRT are the strongest alternatives in intermediate and NLOS setups, with IS-GLRT slightly outperforming NLOS-GLRT.The LOS result is associated with a significant pseudo-covariance structure change between hypotheses; in intermediate and NLOS cases, IS-GLRT and NLOS-GLRT are preferred.
  • Contributions: Interference-aware IS-GLRT, NLOS-GLRT, and IGMM-GLRT rules extend IS, NLOS, and IGMM fusion to partially unknown Rician jamming interference.The extension uses composite hypothesis testing and the GLRT framework.
  • Interference-aware performance: Interference-aware rules significantly outperform their interference-unaware counterparts at moderate-to-high SNR, while suppression can be unfavorable in noise-dominated LOS conditions.At high SNR in LOS conditions, interference suppression improves performance even if some sensor contributions are partially eliminated.
  • Interference-aware performance: Increasing the number of receive antennas generally improves interference-aware detection and its gain over interference-unaware counterparts.The reported exception is IGMM-GLRT in an NLOS case, attributed to lack of identifiability.
  • Equivalences: The paper establishes asymptotic equivalences among the proposed rules in selected interference-free and interference-prone scenarios.Under the IS assumption, IGMM-GLRT is statistically equivalent to IS-GLRT and attains exact GLRT performance.
  • Future work: Future work targets theoretical performance analysis and fusion schemes robust to smarter jammers.These directions define the stated boundary of the present study.

APPENDIX

The appendix derives the mean, covariance, and pseudo-covariance characterization of the received vector under each hypothesis. It also explains why augmented-form processing is needed when the pseudo-covariance is nonzero.

  • Second-order characterization: The appendix evaluates the conditional mean vector E{y|Hi} using the received-signal model and statistical independence assumptions.The derivation exploits zero-mean noise and independence between fading coefficients and sensor decisions.
  • Second-order characterization: The conditional covariance matrix is formed from the transformed decision covariance and the channel, fading, and noise terms.The displayed expression centers y by its hypothesis-dependent mean before taking the conditional covariance.
  • Second-order characterization: The conditional pseudo-covariance is derived separately and simplifies because the NLOS fading vector and receiver noise are circular.The appendix uses E{w w^T}=O_N and related null cross terms in the simplification.
  • Second-order characterization: A nonzero pseudo-covariance motivates augmented-form processing of the received vector.Equation (77) is explicitly identified as non-null.
Loading 1702.07915v1…