Generalized chi-squared detector for LTI systems with non-Gaussian noise

Justin Ruths; Navid Hashemi

REVIEW 3 major objections 4 minor 12 references

Generalized chi-squared detector for LTI systems with non-Gaussian noise

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 A chi-squared detector can be tuned to any target false-alarm rate even when the noise is not Gaussian, by representing the noise as a Gaussian mixture and integrating each mode over the detector's acceptance region.

desk verdict Useful non-Gaussian extension of chi-squared detector tuning, but Theorem 1's printed formula has a missing -1/2 and validation is only at a 48% false-alarm rate. read the letter →

arxiv 1909.01469 v1 pith:T437WH7J submitted 2019-09-03 eess.SY cs.SY

Navid Hashemi , Justin Ruths This is my paper · ORCID

classification eess.SYcs.SY MSC 93E1093C05

keywords chi-squareddetectorGaussianmixturemodelfalsealarmrateanomalydetectionlineartime-invariantsystemsnon-Gaussiannoiseresidualdistributionthresholdtuning

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 extends the standard chi-squared anomaly detector, whose threshold is normally tuned under a Gaussian noise assumption, to systems driven by arbitrary noise distributions. The key move is to approximate each noise density as a Gaussian mixture model, propagate the mixture analytically through an LTI estimator to obtain the residual distribution, and then integrate the residual density over the ellipsoidal region defined by the detector's threshold. The result is a closed-form expression for the expected false alarm rate as a mixture-weighted sum of sphere integrals, and this expression can be inverted to select a threshold for any desired false alarm rate. If the claim holds, precise detector tuning no longer requires assuming Gaussian noise, and the previous Gaussian tuning rule reappears as a special case.

What carries the argument

The central object is the Gaussian mixture model representation of arbitrary noise densities, combined with the characteristic-function identity that turns convolutions of noise terms into products. This carries the argument because it lets the steady-state residual distribution be written analytically as a sum of Gaussian modes; the detector's false alarm probability then becomes a sum over those modes of Gaussian integrals over an ellipsoid, evaluated in spherical coordinates as the $M_j$ integrals in the paper's equation (⋆).

What would settle it

Take a known non-Gaussian noise density with a heavy tail, draw a very large Monte Carlo sample, fit a Gaussian mixture, compute the threshold from Remark 2 for a target false alarm rate $A^*$, then run the closed-loop detector on a fresh independent sample under no anomaly and compare the empirical alarm frequency to $A^*$; a statistically significant mismatch across several thresholds would indicate the sphere-integral formula or the GMM tail representation is wrong.

Watch

Extended reading notes

Core claim

For an LTI system driven by arbitrary iid noise whose densities are given as Gaussian mixtures, the residual is also a Gaussian mixture, and the false alarm rate of the quadratic detector $z_k > \alpha$ is $A = 1 - \sum_{j=1}^{m} \pi_j M_j$, where each $M_j$ is the probability mass of the $j$-th Gaussian mode inside the ellipsoid $(r-\mu)^T \Sigma^{-1}(r-\mu) \le \alpha$, expressed in the paper as a sphere integral (the starred equation). Because this relation maps thresholds to false alarm rates, a root-finding step such as bisection inverts it to produce the threshold that delivers any desired false alarm rate. In the Gaussian limit $m=1$, the formula specializes to the known chi-squared tuning rule of Lemma 1, demonstrating backwards compatibility.

Load-bearing premise

The false-alarm calculation is only as accurate as the fitted Gaussian mixture's representation of the true noise density in the region that determines alarms, and the example prunes mixture modes using hand-chosen thresholds that have no stated error bound.

Editorial extensions

If this is right

A designer can invert $A(\alpha) = 1 - \sum_j \pi_j M_j$ to choose a threshold that yields a specified false alarm rate without assuming Gaussian noise.
Setting $m=1$ recovers the standard chi-squared tuning rule, so existing Gaussian implementations are a limiting case and need no separate treatment.
The steady-state residual distribution can be approximated to any specified error by taking the settling time $k^*$ large enough, giving an approximation guarantee for the false alarm computation.
The formula makes explicit how the false alarm rate depends on the system and observer through the mixture modes' means, covariances, and weights, which supports ROC analysis and detector comparison.

Reading between the lines

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

The derivation does not depend on how the noise GMM was fitted, so any fitting procedure that faithfully represents the noise tail would plug into Theorem 1; the practical bottleneck is tail estimation rather than the detector formula itself.
The same GMM-propagation machinery could be applied to other residual-based detectors, such as CUSUM, by feeding the Gaussian-mixture residual distribution into the cumulative-sum recursion, though the paper only treats the quadratic form.
Because the false alarm rate is computed mode-by-mode, each $M_j$ could be approximated by quadrature or sampling in higher-dimensional measurement spaces; the paper's example is scalar, so computational scaling in large $p$ remains untested.

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.

Desk Editor's Note

read the letter

Short version: the paper gives a genuinely useful analytic extension of chi-squared detector tuning to non-Gaussian noise via Gaussian mixtures, but the displayed formula in Theorem 1 has a missing -1/2 in the Gaussian exponent, and the example validates only near the median false alarm rate. Fix the formula, add tail validation, and it's a solid contribution.

The real new thing is Lemma 2: propagating a GMM representation of iid process and measurement noise through an LTI residual and obtaining the residual GMM analytically. That's not in the cited Gaussian-only tuning papers. Theorem 1 then integrates inside the ellipsoid of the quadratic detector to get the false alarm rate. The interpretation in Section IV, viewing the mixture as a set of chi-squared detectors with thresholds per mode, is nice and backwards-compatible. The example is honest: they fit a 6-mode GMM to 5e6 samples, reduce 6^10 modes to 282 by merging with heuristics, and get 0.478 predicted vs 0.484 empirical false alarm rate at alpha=0.75. That agreement is real evidence the machinery works at least at that operating point.

Soft spots, in order of severity. First, the printed Theorem 1 integral is wrong: e^{(CTρ+γ)^T K^{-1}(CTρ+γ)} needs a negative sign and a factor -1/2. As printed, the exponent is positive, so M_j is not a probability integral; any inversion is meaningless until this is fixed. I'm treating it as a typo because the proof in the text substitutes correctly, but the displayed formula has to be corrected.

Second, the validation is only at alpha=0.75, where the false alarm rate is ~0.48. That's the middle of the distribution. Security and fault detection usually need false alarm rates of 10^-2 or smaller, and tail accuracy is exactly where a GMM fitted to the body, followed by ad hoc merging, can drift. The authors don't provide sensitivity of d_mu, d_K, k* or the EM fit to the computed false alarm rate. They also don't ship code or data, so the 0.478 result can't be reproduced independently.

These are fixable. The architecture is sound, the derivation is mostly clear, and the self-citations to [4,5] are appropriate. I'd send it to a serious referee but ask for the correction and for a second example at a low alarm rate. If they add that, it's publishable.

Referee Report

3 major / 4 minor

Summary. This paper addresses threshold tuning of the quadratic-form (generalized chi-squared) anomaly detector (4) for LTI systems whose iid system and measurement noises are non-Gaussian. The authors represent the noise densities as Gaussian mixtures (6)-(7), propagate the mixture through the Luenberger observer to obtain a residual GMM (Lemma 2), truncate at a settling time k* (Lemma 3), and in Theorem 1 express the false-alarm probability for a threshold alpha as 1 - sum_j pi_j M_j, where M_j is an integral over the ellipsoidal region { (r - mu)^T Sigma^{-1} (r - mu) <= alpha } under the j-th residual mode. The claimed contribution is that (31) can be inverted to set alpha for any desired false-alarm rate without assuming Gaussian noise. A scalar-output example fits a six-mode GMM to a multimodal noise distribution and reports predicted vs empirical false-alarm rates of 0.478 and 0.484 at alpha = 0.75.

Significance. If fully correct and appropriately validated, this would be a useful tool: detector thresholds under multimodal/non-Gaussian noise could be set analytically rather than by brute-force Monte Carlo, and the backward-compatibility interpretation with Gaussian noise is appealing. The paper's main structural result, Lemma 2, is sound and the example's agreement is encouraging. However, the central formula (star) is printed with a sign/coefficient error, the Section IV covariance construction is not correct as stated, and the validation does not exercise the regime (small false-alarm rates) that motivates threshold inversion. These are fixable within a revision, but they affect the paper's central claims as currently written.

major comments (3)

[Section III-B, Theorem 1 and Eq. (star)] The exponent in the integrand of M_j is printed as (C^T rho_k + gamma_j)^T K_j^{-1} (C^T rho_k + gamma_j), but the Gaussian density requires -1/2 (C^T rho_k + gamma_j)^T K_j^{-1} (C^T rho_k + gamma_j). The same sign and coefficient omission appears in the derivation immediately preceding (star). With the printed positive exponent, M_j is not a probability integral and (31) cannot be used to invert for alpha. This is a load-bearing error in the main theorem, even if it is a typesetting mistake, and it must be corrected before the numerical example and Remark 2 can be checked from the printed equations.
[Section V and Remark 2] The method's usefulness rests on threshold inversion for arbitrary desired false-alarm rates, but the validation is confined to alpha = 0.75 with A approximately 0.478 versus 0.484, i.e., near the center of the distribution rather than a tail event. No quantitative bound or estimate is given for how the three approximation steps perturb the computed false-alarm probability: EM fitting of the noise GMM from 5e6 samples, the mode-merging thresholds d_mu = 0.0747 and d_K = 0.0917, and the truncation k* = 10. Since small false-alarm rates depend precisely on the tail of the residual distribution, the unquantified tail error leaves the central claim unsupported for the regime in which such detectors are typically used. Please provide tail-focused validation over a range of alpha values (for example down to P(z <= alpha) near 1 - 1e-4 or 1 - 1e-6) and/or an a priori bound on |P_pred(z <= alpha) - P_true(z <= alpha)|.
[Section IV, Eq. (35)] The covariance formula C_j = E^{-1} K_j E^{-T} does not follow from the stated construction. If the residual is r = sum_kappa A_kappa eta_kappa with iid Gaussian measurement noise of covariance C_j, the residual covariance is sum_kappa A_kappa C_j A_kappa^T, not E C_j E^T with E = sum_kappa A_kappa, because cross terms appear in E C_j E^T. The proposed C_j therefore does not, in general, generate a Gaussian residual with covariance K_j. This affects the claimed backwards-compatibility interpretation and should be corrected or the interpretation qualified.

minor comments (4)

[Lemma 2 and Section V] The mode count should be typeset as m_1^k m_2^{k-1}; in the text the expression appears as an ambiguous superscript and the example should read 6^10 = 60,466,176 before the merging step reduces the count to about 282. Please disambiguate the exponent notation.
[Lemma 3] The statement says 'If the system is stable' but the proof uses rho(F - LC) < 1, which is a property of the estimator error dynamics, not of the plant (F, G) alone. Please state the hypothesis as 'if the estimator error dynamics are stable' or 'if F - LC is Schur.'
[Abstract and Introduction] The phrase 'arbitrary noise distributions' is too strong if the method requires a finite GMM approximation obtained by EM and then pruned with heuristic thresholds. Please either explicitly state that the noise densities are modeled as Gaussian mixtures or clearly qualify the approximation error as part of the method's assumptions.
[Figure 4 and Section V] The values 0.522 and 0.516 for the GMM and empirical P(z <= 0.75) appear only in the text; the figure would be easier to read if the threshold alpha = 0.75 and the two CDF values were annotated directly on the plot.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the false-alarm rate is a forward computation from an independently fitted GMM, validated against Monte Carlo; the only self-citation (Lemma 1) is a standard Gaussian special case used for interpretation, not for the main derivation.

full rationale

The derivation chain is self-contained. Theorem 1 computes A = 1 - sum_j pi_j M_j, where M_j is the integral of the j-th Gaussian residual mode over the ellipsoidal region D = {(r-mu)^T Sigma^{-1}(r-mu) <= alpha}. This is a direct application of the law of total probability to the GMM representation (29); the GMM itself is fitted to the noise distribution by EM (Section V), not fitted to the false-alarm rate. The target quantity A is therefore a forward functional of the fitted density, and the validation against independent Monte Carlo draws from the original noise is an out-of-sample check of the model, not a tautology. Lemma 1, cited from the authors' prior work [5], is the Gaussian special case used only in the interpretive Section IV to relate M_j to a hypothetical chi-squared threshold; the central tuning relation (31)-(star) does not rely on it. No uniqueness theorem or ansatz is imported from the authors' prior work. The displayed integral (star) appears to lack the minus sign and -1/2 in the exponent, and (39) may have swapped or inverted arguments; these are typographical/correctness issues, not circularity. The example validates only near the median of the distance measure (A ~ 0.48) and gives no tail-error bound, but that is a limitation of empirical validation, not circular reasoning. Overall circularity score: 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central false-alarm formula depends on fitted GMM parameters and on hand-picked merging thresholds; the paper provides no uncertainty propagation from the GMM fit, the EM convergence, or the truncation horizon.

free parameters (4)

Measurement noise GMM modes (m1=6) = Six modes as in Table I; all weights, means, variances listed there.
EM fit to the empirical noise density; the residual GMM and hence A=0.478 are computed from these values.
System noise GMM modes (m2) = None in example (no system noise)
General formulation allows m2 modes; if present they would enter Lemma 2 and Theorem 1.
Mode-merging thresholds d_mu and d_K = d_mu=0.0747, d_K=0.0917
Chosen by heuristic based on the spread to reduce 6^10 modes to 282; no error bound, and the false alarm estimate depends on the reduced model.
Truncation horizon k_star = 10
Treats the residual as converged to steady state; Lemma 3 says larger k_star improves accuracy, but k_star=10 is chosen in the example.

assumptions (5)

domain assumption Noises v_k and eta_k are iid across time and mutually independent
The central propagation in Eqs. (10)-(16) uses convolution of independent samples; correlated or time-dependent noise would break Lemma 2.
ad hoc to paper Noise densities admit a Gaussian mixture representation (Eqs. (6)-(7))
Remark 1 claims all densities can be expressed as GMM; only approximation is guaranteed by the cited literature, so detector tuning inherits GMM approximation error.
domain assumption Observer gain L is chosen so that F-LC is stable (system detectable)
Lemma 3 requires rho(F-LC) < 1 so A_kappa and B_kappa vanish and the residual reaches steady state.
domain assumption The residual covariance Sigma is nonsingular so its Cholesky decomposition exists
Theorem 1 uses Sigma^{-1} and C from Cholesky; singular Sigma would make the detector (4) and the sphere change undefined.
standard math Standard Gaussian characteristic function and affine transform identities hold
Used in the proof of Lemma 2, Eqs. (17)-(20).

invented entities (1)

Hypothetical Gaussian measurement noise for each residual mode (a_j, C_j in Eq. 35)
purpose: Interpret each residual GMM mode as generated by a Gaussian measurement noise in isolation, so each mode can be paired with a chi-squared detector threshold.
Constructed by Eq. (35) from residual mode parameters; no physical meaning, only an interpretation tool.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generalized chi-squared detector for LTI systems with non-Gaussian noise." pith.science (2026). https://pith.science/paper/T437WH7J

@misc{pith2026190901469,
  author       = {Pith},
  title        = {Pith review of: Generalized chi-squared detector for LTI systems with non-Gaussian noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T437WH7J}},
  note         = {Machine review of arXiv:1909.01469}
}

read the original abstract

Previously, we derived exact relationships between the properties of a linear time-invariant control system and properties of an anomaly detector that quantified the impact an attacker can have on the system if that attacker aims to remain stealthy to the detector. A necessary first step in this process is to be able to precisely tune the detector to a desired level of performance (false alarm rate) under normal operation, typically through the selection of a threshold parameter. To-date efforts have only considered Gaussian noises. Here we generalize the approach to tune a chi-squared anomaly detector for noises with non-Gaussian distributions. Our method leverages a Gaussian Mixture Model to represent the arbitrary noise distributions, which preserves analytic tractability and provides an informative interpretation in terms of a collection of chi-squared detectors and multiple Gaussian disturbances.

Figures

Figures reproduced from arXiv: 1909.01469 by the authors.

**Figure 2.** The 3D surface of the residual GMM probability density function (shaded to reveal height mapping) constructed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** The measurement noise distribution (top) is approxi [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The cumulative distribution function of the distance [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

IEEEtran_HOWTO

11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

work page 2015
[2]

" id="W5M0MpCehiHzreSzNTczkc9d

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

work page 2017
[3]

Z. Guo, D. Shi, K. H. Johansson, and L. Shi, `` Optimal Linear Cyber-Attack on Remote State Estimation ,'' IEEE Transactions on Control of Network Systems, vol. PP, no. 99, pp. 1--10, 2016

work page 2016
[4]

Mo and B

Y. Mo and B. Sinopoli, ``On the performance degradation of cyber-physical systems under stealthy integrity attacks,'' IEEE Transactions on Automatic Control, vol. 61, pp. 2618--2624, 2016

work page 2016
[5]

C. Z. Bai, F. Pasqualetti, and V. Gupta, ``Security in stochastic control systems: Fundamental limitations and performance bounds,'' in American Control Conference (ACC), 2015, pp. 195--200

work page 2015
[6]

Murguia and J

C. Murguia and J. Ruths, ``Characterization of a cusum model-based sensor attack detector,'' in proceedings of the 55th IEEE Conference on Decision and Control (CDC), 2016

work page 2016
[7]

------, ``Cusum and chi-squared attack detection of compromised sensors,'' in proceedings of the IEEE Multi-Conference on Systems and Control (MSC), 2016

work page 2016
[8]

Chen and R

J. Chen and R. J. Patton, Robust Model-based Fault Diagnosis for Dynamic Systems. 1em plus 0.5em minus 0.4em Norwell, MA, USA: Kluwer Academic Publishers, 1999

work page 1999

Show all 12 references

[9]

Astrom and B

K. Astrom and B. Wittenmark, Computer-controlled Systems (3rd Ed.). 1em plus 0.5em minus 0.4em Prentice-Hall, Inc., 1997

1997
[10]

A. G. Bacharoglou, ``Approximation of probability distributions by complex mixtures of gaussian measures,'' Proceedings of the American Mathematical Society, vol. 138, no. 7, pp. 2619--2628, jul 2010

2010
[11]

Xu and M

L. Xu and M. I. Jordan, ``On convergence properties of the em algorithm for gaussian mixtures,'' Neural Computation, vol. 8, no. 1, pp. 129--151, 1996

1996
[12]

1em plus 0.5em minus 0.4em Pearson Education,Inc., 2014

H.Stark and J.W.Woods, Probability, Statistics, and Random Processes for Engineers, 4th Edition. 1em plus 0.5em minus 0.4em Pearson Education,Inc., 2014

2014

Pith tools

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

[1] [1]

IEEEtran_HOWTO

11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

work page 2015

[2] [2]

" id="W5M0MpCehiHzreSzNTczkc9d

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

work page 2017

[3] [3]

Z. Guo, D. Shi, K. H. Johansson, and L. Shi, `` Optimal Linear Cyber-Attack on Remote State Estimation ,'' IEEE Transactions on Control of Network Systems, vol. PP, no. 99, pp. 1--10, 2016

work page 2016

[4] [4]

Mo and B

Y. Mo and B. Sinopoli, ``On the performance degradation of cyber-physical systems under stealthy integrity attacks,'' IEEE Transactions on Automatic Control, vol. 61, pp. 2618--2624, 2016

work page 2016

[5] [5]

C. Z. Bai, F. Pasqualetti, and V. Gupta, ``Security in stochastic control systems: Fundamental limitations and performance bounds,'' in American Control Conference (ACC), 2015, pp. 195--200

work page 2015

[6] [6]

Murguia and J

C. Murguia and J. Ruths, ``Characterization of a cusum model-based sensor attack detector,'' in proceedings of the 55th IEEE Conference on Decision and Control (CDC), 2016

work page 2016

[7] [7]

------, ``Cusum and chi-squared attack detection of compromised sensors,'' in proceedings of the IEEE Multi-Conference on Systems and Control (MSC), 2016

work page 2016

[8] [8]

Chen and R

J. Chen and R. J. Patton, Robust Model-based Fault Diagnosis for Dynamic Systems. 1em plus 0.5em minus 0.4em Norwell, MA, USA: Kluwer Academic Publishers, 1999

work page 1999