Pith. sign in

REVIEW 3 major objections 4 minor 20 references

A k-nearest neighbors approach to the design of radar detectors

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proves that a k-nearest-neighbors classifier fed by standard radar statistics can be a constant-false-alarm-rate detector with closed-form performance.

desk verdict A genuinely new CFAR result for KNN radar detectors, honestly presented, but the 'closed-form' claim outruns what is actually computed. read the letter →

arxiv 1908.00870 v1 pith:IIOKSJS6 submitted 2019-08-02 eess.SP

Angelo Coluccia , Alessio Fascista , Giuseppe Ricci This is my paper · ORCID
classification eess.SP
keywords k-nearestneighborsradardetectionconstantfalsealarmrategeneralizedlikelihoodratiotestprobabilityofadaptivefeaturevectormachinelearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes using a k-nearest-neighbors (KNN) classifier as a radar detector, and proves that, for a particular choice of features, the detector has the constant false alarm rate (CFAR) property: the probability of a false alarm does not depend on the unknown noise covariance matrix. The central result is a closed-form characterization of false-alarm and detection probabilities for a KNN detector whose features are arbitrary weighted functions of two classical radar statistics, the Kelly statistic and a beta-distributed auxiliary statistic. If correct, this gives a principled, theoretically tractable way to design machine-learning radar detectors whose performance can be predicted and whose robustness or selectivity can be tuned by choosing feature weights. The paper also shows a raw-data KNN detector can be more powerful than classical detectors in matched conditions, at the cost of strict CFAR.

What carries the argument

The key object is the decomposition of classical radar detectors (Kelly, AMF, ACE, and related statistics) into two common statistics: $\tilde{t} = t_{\mathrm{Kelly}}/(1-t_{\mathrm{Kelly}})$ and $\beta = 1/(1 + z^H S^{-1} z - |z^H S^{-1} v|^2/(v^H S^{-1} v))$. The feature vector takes the form $x = [d_1 \tilde{t} b[1], \dots, d_m \tilde{t} b[m]]^T$ with $b[j] = f_j(\beta)$ and a nonnegative diagonal weight matrix $D = \mathrm{diag}(d_1,\dots,d_m)$. The argument relies on distributional facts imported from the radar-detection literature: under the null hypothesis, $\tilde{t}$ follows a complex central F distribution with $1$ and $K-N+1$ degrees of freedom, $\beta$ follows a complex central $\beta$ distribution with $K-N+2$ and $N-1$ degrees of freedom, and the two are independent, regardless of $C$; under the alternative, noncentral versions are parameterized by the training SNR (and, under mismatch, by $\mathrm{SNR}_p \beta \cos^2\theta$ and $\mathrm{SNR}_p \sin^2\theta$). These facts let the paper express $p_0$ and $p_1$ as integrals over the $\beta$ density with F cumulative distribution functions, which is what makes the CFAR property and the stated $P_d$ dependence visible.

What would settle it

Simulate the proposed CFAR KNN detector (feature vector from eq. 20, say) with a fixed training SNR but two substantially different noise covariance matrices, for example one with one-lag correlation 0.95 and one with correlation 0.1 or a diagonal matrix, keeping all other parameters fixed. If the empirically measured probability of false alarm, over enough trials (e.g., $10^5$), changes materially between the two setups beyond Monte Carlo error, Proposition 2 is contradicted.

Watch

Extended reading notes

Core claim

Proposition 2 is the paper's central claim: a KNN detector whose feature vector is a weighted stack of classical radar statistics, all expressible through the common pair $\tilde{t}$ and $\beta$, has a probability of false alarm that depends only on the signal-to-noise ratio used to generate the training data and is otherwise independent of the actual noise covariance matrix $C$. Thus the detector possesses the constant false alarm rate property. The detection probability depends only on the training SNR, on the actual target SNR $\mathrm{SNR}_p$, and on the cosine-squared angle $\cos^2\theta$ between the nominal and actual steering vectors. The paper derives closed-form expressions for the elementary probabilities $p_0$ and $p_1$ in Appendix A, from which $P_{fa}$ and $P_d$ follow through the general KNN performance formula in Proposition 1.

Load-bearing premise

The proof imports exact distributional results from the radar-detection literature: under the noise-only hypothesis the two statistics that build each feature are independent, with the first following a complex F distribution and the second a complex beta distribution, regardless of the noise covariance. If those distributions, or that independence, fail for the actual noise or feature model, the CFAR property and the stated detection-probability formula do not follow.

Editorial extensions

If this is right

  • A KNN detector built on a two-dimensional feature vector containing Kelly and AMF statistics can match Kelly's detection power in matched conditions while behaving more robustly under steering-vector mismatch, all while preserving CFAR.
  • By adjusting the weights in the diagonal matrix $D$, the designer can interpolate between robust behavior (AMF-like) and selective behavior (Kelly- or ACE-like) without losing the constant false alarm rate property.
  • The closed-form expressions for $P_{fa}$ and $P_d$ allow performance prediction and threshold setting for KNN radar detectors without Monte Carlo simulation.
  • The raw-data KNN detector, using $S^{-1/2}z$ as a feature, is not strictly CFAR but its false alarm probability shows only mild sensitivity to changes in noise correlation, as illustrated by the numerical examples.
  • Since the training set can be generated artificially from a design covariance and a design SNR, the proposed CFAR KNN detector does not require a collection phase of real radar data for training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same decomposition trick should extend to other detectors that share the $\tilde{t}$ and $\beta$ dependence, such as W-ABORT, enabling a broader family of tunable CFAR KNN detectors beyond the two examples shown.
  • Editorial inference: the CFAR property is tied to training data generated from a nominal SNR; if training examples are collected from real scenes rather than synthesized, a mismatch between the training distribution and the design distribution could break the constant false alarm rate, so synthetic training is the safer route in practice.
  • Editorial inference: Proposition 1 is a fully general performance formula for any KNN classifier, and the closed-form evaluation demonstrated here could be applied to other feature spaces where distance comparisons reduce to known distributions, making the approach useful beyond radar detection.
  • Editorial inference: a natural testable extension is to optimize the weights in $D$ from the closed-form formulas to maximize detection probability under a false-alarm constraint, producing detectors tuned for a specific robustness-selectivity tradeoff.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a k-nearest neighbors (KNN) approach to radar detection. The authors first derive a general expression (Proposition 1) for the probability that a KNN decision statistic exceeds a threshold, in terms of an expectation over feature-space events. They then specialize to radar detectors whose feature vectors are stacks of the classical statistics t-tilde and beta, possibly weighted by a diagonal matrix D and passed through arbitrary functions of beta. For this class they claim (Proposition 2) that the false alarm probability is independent of the actual noise covariance matrix, i.e., the detector is CFAR, and that the detection probability depends only on the training SNR, the actual SNR, and cos^2 theta. The proof in Appendix A reduces the required probabilities to integrals involving complex central and noncentral F and beta distributions imported from reference [20]. Simulations compare KNN detectors using raw data, Kelly+AMF features, and Kelly+ACE features against conventional detectors.

Significance. If Proposition 2 is valid, the paper provides a useful design framework: KNN detectors can be made CFAR while retaining robustness or selectivity tunable through the weights D, and their performance is comparable to classical receivers such as Kelly's detector and the AMF. The general performance decomposition in Proposition 1 is a contribution that goes beyond the specific radar application. The paper is also refreshingly honest about the non-CFAR behavior of the raw-data KNN detector. However, the central CFAR and detection-probability claims rest on distributional facts that are quoted from [20] rather than proved or numerically verified in the present setting, and the 'closed-form' performance expressions are never evaluated or compared with Monte Carlo results. The significance is therefore contingent on filling these gaps.

major comments (3)
  1. [Appendix A / Proposition 2] The CFAR claim and the Pd parameter dependence are not established within the manuscript: the proof imports from [20] the exact joint laws of (t-tilde, beta) under H0, matched H1, and mismatched H1, together with the independence of t-tilde and beta under H0, and the stated noncentrality parameters in the mismatched case. Since all cancellations of the covariance matrix C in Pfa, and all reductions of Pd to SNR, SNRp, and cos^2 theta, are algebraic consequences of those imported facts, this is the load-bearing step. Please state the imported results as a lemma with a precise and complete citation (section or theorem in [20]), or prove them from model (12) and the definition of S, and add a Monte Carlo verification for the actual model of Section 3: under H0, matched H1, and mismatched H1 with z=alpha p+n, check the marginal and conditional laws of t-tilde and beta, including the claimed noncentrality parameters and the independence property. Without such verification, Proposition 2 is an unproved assertion.
  2. [Section 2.2 and Appendix A (abstract claim of closed-form characterization)] The abstract and Section 2.2 state that Pfa and Pd are 'characterized in closed-form', but the final expression in Proposition 1 is a high-dimensional expectation over y whose dimension is m + m(k-M) + m(NT-M), and the Appendix A expressions for p0 and p1 are one-dimensional integrals over beta with integrands depending on roots r1,...,r4 that are only implicitly defined through discriminants and continuous functions of x and the training vectors. These expressions are never evaluated numerically, nor are they compared with Monte Carlo estimates for any of the feature vectors used in Section 3, such as (20) or (21). Please either provide a tractable evaluation for the cases actually simulated (for example by numerical integration or by specializing to m=1 or m=2) and a validation plot against the empirical Pfa/Pd, or revise the terminology to 'integral representations' rather than 'closed-form'.
  3. [Section 3.2, Figures 4-7] The CFAR property asserted in Proposition 2 is not demonstrated empirically anywhere in the paper for the feature-vector detectors. For the feature vectors (20) and (21), the authors should report estimated Pfa under at least two different noise covariance matrices, e.g., the Gaussian-shaped C with one-lag correlation 0.95 and 0.5 used in Section 3.1, at a fixed threshold and fixed training SNR, and show that the Pfa is unchanged. The current simulations only show Pd curves and do not provide evidence that the detector is CFAR rather than merely insensitive to C in the tested cases. Such a test would also directly support the imported distributional assumptions.
minor comments (4)
  1. [Section 1 and Section 3.2] There are several typos: 'stastistics' in the introduction, 'CF AR' in the Section 3.2 heading, and the notation 'min ||x1_r, x||' in Appendix A contains a misplaced comma.
  2. [Appendix A] The symbols b0[j] and b1[j] are used without definition; please define them consistently, e.g., b0[j]=f_j(beta0) and b1[j]=f_j(beta1), following the notation for a generic b[j]=f_j(beta) in Eq. (18).
  3. [Proposition 1] The proof of Proposition 1 implicitly relies on exchangeability of the training vectors to justify multiplying by the binomial coefficients; this assumption should be stated explicitly before the result is used.
  4. [Section 3.1] The reported empirical Pfa values, such as Pfa = 0.0048 and Pfa = 0.0062, are given without confidence intervals; given that they are based on 10^5 trials, a brief statement of the associated binomial uncertainty would help the reader interpret small differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Proposition 2 is a derivation from independently established distributional facts, not a re-statement of fitted inputs.

full rationale

The paper's central claim, Proposition 2, is that for the feature vector in Eq. (18) the KNN detector is CFAR and that Pd depends only on the training SNR, SNRp, and cos^2(theta). This is derived in Appendix A from distributional facts quoted from reference [20]: under H0, t~ is complex central F(1,K-N+1) and beta is complex central beta(K-N+2,N-1), independent of C; under H1, t~ given beta is noncentral F with noncentrality proportional to SNR, SNR*beta, or SNRp*beta*cos^2(theta), and beta is central or noncentral beta. These cited results do not presuppose the CFAR property or the target claim; they are parameter-free statements about the statistics of Kelly-type and beta variables and are externally established in the radar detection literature. The KNN design parameters k, M, D, and the training SNR are chosen by the designer, not fitted to the evaluated Pfa/Pd curves. The paper does not fit a parameter to the benchmark detectors' false-alarm rates; for the raw-data detector it explicitly reports a non-CFAR behavior (Pfa = 0.0048 vs 0.0062 for different noise correlation), which shows the analysis is not inflated to force a CFAR conclusion. The only mild self-citation is reliance on [20], which includes one of the present authors, but the cited distributional laws are classical and independent of the paper's conclusions, so this does not constitute circularity. The derivation chain is therefore self-contained conditional on standard distributional results.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central CFAR claim rests on standard radar distributional assumptions, with hand-set design parameters for the examples. No new physical entities or fitted constants are introduced.

free parameters (3)
  • Design weights D = diag(d1,...,dm) = d1=1, d2=0.7 (Kelly+AMF example); d2=0.8 (Kelly+ACE example)
    Chosen by hand in Sec. 3.2 to illustrate robustness and selectivity; the CFAR proof holds for any nonnegative D, so these are design degrees of freedom, not fitted to data.
  • Training SNR used to generate artificial T1 = 12 dB
    Used to generate the H1 training set in all simulations (Sec. 3.1, Sec. 3.2). Proposition 2 states Pfa/Pd depend on this design SNR; it sets the detector operating point rather than being estimated from test data.
  • KNN hyperparameters (k, M, NT) = k=50, M=25, NT=1000 per class (raw-data example); N=16, KS=32 for CFAR examples
    Chosen for simulations; the closed-form analysis in Propositions 1 and 2 holds for arbitrary integer values, so these are user-set hyperparameters, not fitted constants.
assumptions (3)
  • domain assumption Complex Gaussian noise model for primary and secondary data, with identical covariance C under H0 and signal-plus-noise under H1 (eq. (12)).
    Classical radar assumption; used to derive the distributional facts in Appendix A.
  • domain assumption Known distributions of t~ and beta from [20]: under H0 t~ ~ CF(1,K-N+1), beta ~ CBeta(K-N+2,N-1), independent; under H1, noncentral versions with noncentrality parameter SNR*beta.
    This is the load-bearing statistical premise for the CFAR proof; it is cited to a reference rather than proved in the paper.
  • domain assumption Training set contains NT independent realizations under H0 and H1, generated according to the same model with preassigned design parameters.
    Sec. 2.1 assumes this without loss of generality; it is needed for the independence assumptions in Proposition 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A k-nearest neighbors approach to the design of radar detectors." pith.science (2026). https://pith.science/paper/IIOKSJS6

@misc{pith2026190800870,
  author       = {Pith},
  title        = {Pith review of: A k-nearest neighbors approach to the design of radar detectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IIOKSJS6}},
  note         = {Machine review of arXiv:1908.00870}
}
read the original abstract

A k-nearest neighbors (KNN) approach to the design of radar detectors is investigated. The idea is to start with either raw data or well-known radar receiver statistics as feature vector to be fed to the KNN decision rule. In the latter case, the probability of false alarm and probability of detection are characterized in closed-form; moreover, it is proved that the detector possesses the constant false alarm rate (CFAR) property and the relevant performance parameters are identified. Simulation examples are provided to illustrate the effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 1908.00870 by the authors.

Figure 1
Figure 1. Graphical representation of a two-dimensional feature space, with [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Pd vs SNR under matched conditions for KNN using raw data S −1/2z. SNR [dB] 5 10 15 20 25 30 Pd 0 0.2 0.4 0.6 0.8 1 KELLY ACE AMF KNN raw data (SNR = 12 dB) [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Pd vs SNR under mismatched conditions with cos2 θ = 0.50 for KNN using raw data S −1/2z. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Pd vs SNR under matched conditions for a KNN fed by eq. (20). loss of generality, we set d1 = 1 so leaving d2 as the sole tunable parameter. Simulations are conducted as for the previous case of Sec. 3.1, but for N = 16 and KS = 32; moreover, we set d2 = 0.7 [PITH_FUL…
Figure 5
Figure 5. Figure 5: Pd vs SNR under mismatched conditions with cos2 θ = 0.46, for a KNN fed by eq. (20). SNR [dB] 8 10 12 14 16 Pd 0.4 0.6 0.8 1 KELLY ACE AMF Proposed KNN: Kelly + ACE [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Pd vs SNR under matched conditions for a KNN fed by eq. (21). 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Pd vs SNR under mismatched conditions with cos2 θ = 0.46, for a KNN fed by eq. (21). namely [20] x =  d1t d ˜ 2 t˜ 1 − β T . (21) In this case, the aim is to obtain a detector with intermediate performance between such two receivers, and again without loss of general…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [16]

    Radar detection in K-distributed clutter plus thermal noise based on KNN methods,

    A. Coluccia, G. Ricci, “Radar detection in K-distributed clutter plus thermal noise based on KNN methods,” IEEE Radar Conference, Boston, MA, USA, 22-26 Apr. 2019

    work page 2019

  2. [20]

    Advanced Radar Detection Schemes Under Mismatched Signal Models,

    F. Bandiera, D. Orlando, G. Ricci, “Advanced Radar Detection Schemes Under Mismatched Signal Models,” Synthesis Lectures on Signal Process- ing No. 8, Morgan & Claypool Publishers , 2009. 30

    work page 2009

  3. [1]

    An Adaptive Detection Algorithm,

    E. J. Kelly, “An Adaptive Detection Algorithm,” IEEE Trans. Aerosp. and Electron. Syst. , Vol. 22, No. 2, Mar. 1986

    work page 1986

  4. [2]

    Performance of an Adaptive Detection Algorithm; Rejection of Unwanted Signals,

    E. J. Kelly, “Performance of an Adaptive Detection Algorithm; Rejection of Unwanted Signals,” IEEE Trans. Aerosp. and Electron. Syst. , Vol. 25, No. 2, Mar. 1989. 27

    work page 1989

  5. [3]

    A CFAR Adap- tive Matched Filter Detector,

    F. C. Robey, D. L. Fuhrman, E. J. Kelly, R. Nitzberg, “A CFAR Adap- tive Matched Filter Detector,” IEEE Trans. Aerosp. and Electron. Syst. , Vol. 29, No. 1, Jan. 1992

    work page 1992

  6. [4]

    Asymptotically Optimum Radar Detection in Compound Gaussian Noise,

    E. Conte, M. Lops, G. Ricci, “Asymptotically Optimum Radar Detection in Compound Gaussian Noise,” IEEE Trans. Aerosp. and Electron. Syst. , Vol. 31, No. 2, April 1995

    work page 1995

  7. [5]

    The CFAR adaptive subspace detector is a scale- invariant GLRT,

    S. Kraut, L. L. Scharf, “The CFAR adaptive subspace detector is a scale- invariant GLRT,” IEEE Trans. Signal Process., Vol. 47, No. 9, Sept. 1999

    work page 1999

  8. [6]

    Adaptive Beamformer Orthogonal Rejec- tion Test,

    N. B. Pulsone, C. M. Rader, “Adaptive Beamformer Orthogonal Rejec- tion Test,” IEEE Trans. Signal Process. , Vol. 49, No. 3, Mar. 2001

    work page 2001

Show all 20 references
  1. [7]

    Spatial Adaptive Subspace Detection in OTH Radar,

    G. A. Fabrizio, A. Farina, M. D. Turley, “Spatial Adaptive Subspace Detection in OTH Radar,” IEEE Trans. Aerosp. and Electron. Syst. , Vol. 39, No. 4, Oct. 2003

    2003

  2. [8]

    An ABORT-Like Detector With Im- proved Mismatched Signals Rejection Capabilities,

    F. Bandiera, O. Besson, G. Ricci, “An ABORT-Like Detector With Im- proved Mismatched Signals Rejection Capabilities,” IEEE Trans. Signal Process., Vol. 56, No. 1, Jan. 2008

    2008

  3. [9]

    A Tunable W-ABORT-like Detector with Im- proved Detection vs Rejection Capabilities Trade-Off,

    A. Coluccia, G. Ricci, “A Tunable W-ABORT-like Detector with Im- proved Detection vs Rejection Capabilities Trade-Off,” IEEE Signal Pro- cess. Lett., Vol. 22, No. 6, Jun. 2015

    2015

  4. [10]

    Gener- alized likelihood ratio test for detection of Gaussian rank-one signals in Gaussian noise with unknown statistics,

    O. Besson, A. Coluccia, E. Chaumette, G. Ricci, F. Vincent, “Gener- alized likelihood ratio test for detection of Gaussian rank-one signals in Gaussian noise with unknown statistics,” IEEE Trans. Signal Process. , Vol. 65, No. 4, 15 Feb. 2017. 28

    2017

  5. [11]

    Low signal-to-noise ratio radar target detection using Lin- ear Support Vector Machines (L-SVM),

    J. E. Ball, “Low signal-to-noise ratio radar target detection using Lin- ear Support Vector Machines (L-SVM),” 2014 IEEE Radar Conference , Cincinnati, OH, USA, 19-23 May 2014

    2014

  6. [12]

    Intelligent CFAR Detector Based on Sup- port Vector Machine,

    L. Wang, D. Wang, C. Hao, “Intelligent CFAR Detector Based on Sup- port Vector Machine,” IEEE Access, Vol. 5, 16 Nov. 2017

    2017

  7. [13]

    Machine learning techniques for cooperative spectrum sensing in cognitive radio networks,

    K.M. Thilina, K.W. Choi, N. Saquib, E. Hossain, “Machine learning techniques for cooperative spectrum sensing in cognitive radio networks,” IEEE J. Sel. Areas in Commun. , Vol. 31, No. 11, Nov 2013

    2013

  8. [14]

    Mobile col- laborative spectrum sensing for heterogeneous networks: A Bayesian machine learning approach,

    Y. Xu, P. Cheng, Z. Chen, Y. Li, B. Vucetic, “Mobile col- laborative spectrum sensing for heterogeneous networks: A Bayesian machine learning approach,” IEEE Trans. Signal Process. , Vol. 66, No. 21, Nov 2018

    2018

  9. [15]

    Spectrum sensing by higher-order SVM-based detection

    A. Coluccia, A. Fascista, G. Ricci, “Spectrum sensing by higher-order SVM-based detection”, 27th European Signal Processing Conference (EU- SIPCO 2019) , A Coru˜ na, Spain, 2-6 Sep. 2019

    2019

  10. [17]

    Deep radar detectors,

    D. Brodeski, I. Bilik, R. Giryes, “Deep radar detectors,” IEEE Radar Conference, Boston, MA, USA, 22-26 Apr. 2019

    2019

  11. [18]

    Deep learning-based object classification on automotive radar spectra,

    K. Patel, K. Rambach, T. Visentin, D. Rusev, M. Pfeiffer, B. Yang, “Deep learning-based object classification on automotive radar spectra,” IEEE Radar Conference , Boston, MA, USA, 22-26 Apr. 2019. 29

    2019

  12. [19]

    A neural net- work approach for waveform generation and selection with multi-mission radar,

    J. M. Kurdzo, J. Y. N. Cho, B. L. Cheong, R. D. Palmer, “A neural net- work approach for waveform generation and selection with multi-mission radar,” IEEE Radar Conference , Boston, MA, USA, 22-26 Apr. 2019

    2019

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.