--- title: Covert Quickest Change Detection url: https://www.emergentmind.com/topics/covert-quickest-change-detection type: topic --- # Covert Quickest Change Detection Searching arXiv for the cited topic and related quickest change detection papers. arxiv_search query: "covert quickest change detection" Covert quickest change detection (CQCD) studies sequential detection of an unknown change point under an additional covertness constraint that alters either the detector’s own sensing behavior or the adversary’s post-change behavior. In the classical backbone, one seeks a stopping time that minimizes detection delay subject to false-alarm constraints, typically under Lorden, Pollak, or Bayesian criteria. The covert dimension enters through pre-change duty-cycle constraints on observation usage, relative-entropy budgets for active probing, sparse communication or observation control, or adversarially weakened post-change laws chosen as a function of the false-alarm parameter. Across these formulations, the main technical issue is how covertness changes the usual likelihood-ratio recursions, KL-driven asymptotics, and optimality guarantees that underlie standard QCD [2104.04186][2412.20207][2605.16140][2601.20022][2509.17778]. ## 1. Classical sequential detection backbone The standard QCD model observes a sequential stream with an unknown change point. In the i.i.d. setting, observations follow \(f_0\) before the change and \(f_1\) after the change; in non-i.i.d. models, the conditional law changes at the change point. The pointwise likelihood ratio is \(\Lambda_t = f_1(X_t)/f_0(X_t)\), the log-likelihood ratio is \(\ell(X_t)=\log \Lambda_t\), and the KL information \(I=D(f_1\Vert f_0)\) governs first-order asymptotics in the i.i.d. case [2104.04186]. The principal performance criteria are Lorden’s worst-case average detection delay, \[ \mathsf{WADD}(T)=\sup_{\nu\ge 1}\operatorname{ess\,sup}\mathbb{E}_\nu[(T-\nu)^+\mid \mathcal{F}_{\nu-1}], \] Pollak’s conditional average detection delay, \[ \mathsf{CADD}(T)=\sup_{\nu\ge 1}\mathbb{E}_\nu[T-\nu\mid T\ge \nu], \] and, in Bayesian formulations, the average detection delay \[ \mathsf{ADD}(T)=\mathbb{E}[(T-\Gamma)^+] \] together with \(\operatorname{PFA}(T)=\mathbb{P}(T<\Gamma)\). False alarms are also expressed through an Average Run Length constraint \(\mathbb{E}_\infty[T]\ge \gamma\) [2104.04186]. Classical procedures remain the structural templates for covert variants. Page’s CUSUM uses \[ S_0=0,\qquad S_t=\max\{0,S_{t-1}+\log \Lambda_t\}, \] with stopping rule \(T_{\mathrm{CUSUM}}=\inf\{t\ge 1:S_t\ge h\}\). Shiryaev–Roberts uses \[ R_0=r,\qquad R_t=(1+R_{t-1})\Lambda_t, \] with stopping rule \(T_{\mathrm{SR}}=\inf\{t\ge 1:R_t\ge A\}\). For unknown post-change models, GLR-CUSUM and mixture-CUSUM replace the fixed log-likelihood with a supremum or integral over the post-change family. In the i.i.d. case, CUSUM is exactly optimal for Lorden’s criterion, and first-order lower bounds give \[ \inf_{T:\operatorname{PFA}(T)\le \alpha}\mathsf{WADD}(T)\ge \frac{|\log \alpha|}{I}(1+o(1)), \] with CUSUM matching this asymptotically [2104.04186]. ## 2. What “covert” means in QCD Current work uses “covert” in more than one technically distinct sense. One line constrains the detector’s own monitoring footprint: sampling only intermittently, probing a channel probabilistically, or reporting sparsely so that sensing remains resource-efficient or hard to detect. Another line studies a covert adversary that tailors the post-change law to the detector’s false-alarm requirement, so that the change itself becomes progressively harder to distinguish as the detector becomes more conservative [2104.04186][2412.20207][2605.16140][2601.20022][2509.17778]. In detector-side formulations, covertness is quantified either by a sampling budget or by an information-leakage budget. In the robust sampling-control framework, the pre-change duty cycle is \[ \mathrm{PDC}(\Psi)=\limsup_k \frac{1}{k}\mathbb{E}_\infty\!\left[\sum_{i=1}^{k-1}M_i\mid \tau\ge k\right], \] where \(M_i\in\{0,1\}\) indicates whether \(X_i\) was used. The constraint \(\mathrm{PDC}(\Psi)\le \beta\) is a direct proxy for covertness: lower \(\beta\) means less sampling, lower exposure, and larger delay [2412.20207]. In the Bayesian active-sensing formulation, covertness is imposed through the expected covertness budget \[ \mathrm{ECB}(\Pi^A)=\chi_{2,0}\,\mathbb{E}\!\left[\sum_{i=1}^{\min(\tau^A,\Gamma)}\beta_i^2\right] +\chi_{2,1}\,\mathbb{E}\!\left[\sum_{i=\Gamma+1}^{\tau^A}\beta_i^2\right], \] which upper bounds the relative entropy between Eve’s observations under active and innocent sensing. Imposing \(\mathrm{ECB}(\Pi^A)\le \delta\) ensures \(D(P_Z^{\Pi^A}\Vert P_Z^{\Pi^0})\le \delta\) [2605.16140]. In adversary-side formulations, covertness is encoded in the vanishing distinguishability of the post-change law. In discrete time, the adversary chooses \(q_\gamma\to q\) and covert scaling is characterized by \(\gamma D(q\Vert q_\gamma)=\Theta(1)\), which forces the optimal ADD to scale as \(\Theta(\gamma)\) rather than \(\Theta(\log \gamma)\). In continuous time, a Brownian drift adversary selects \(\mu(\gamma)\), and covertness holds if and only if \(\mu(\gamma)=O(\gamma^{-1/2})\), equivalently \(\limsup_{\gamma\to\infty}\gamma\mu(\gamma)^2<\infty\) [2601.20022][2509.17778]. A common clarification follows from these formulations: covert QCD is not a single metricized problem. It is a family of sequential detection problems in which the covertness constraint may act on sensing, communication, or the post-change signal itself. ## 3. Sampling-controlled robust CQCD and least favorable laws A non-Bayesian robust formulation studies i.i.d. observations \(\{X_n\}\) with pre-change density \(f\), unknown change point \(\nu\), and post-change density \(g\in G\), where only the family \(G\) is known. The policy \(\Psi=\{\tau,\phi_1,\dots,\phi_\tau\}\) chooses both when to sample and when to stop, minimizing \[ \inf_\Psi \sup_{G\in G}\mathrm{WADD}^G(\Psi) \] subject to \(\mathrm{FAR}(\Psi)\le \alpha\) and \(\mathrm{PDC}(\Psi)\le \beta\), where \[ \mathrm{FAR}(\Psi)=\frac{1}{\mathbb{E}_\infty[\tau]}. \] This realizes covert QCD through the duty-cycle constraint \(\beta\) [2412.20207]. The robust design hinges on a least favorable law \(\bar G\) with density \(\bar g\). Writing \[ \ell(X)=\log \frac{\bar g(X)}{f(X)}, \] the family \(G\) has a least favorable law if for all \(G\in G\), \[ \mathcal{L}(\ell(X),G)\succ \mathcal{L}(\ell(X),\bar G), \] where \(\succ\) denotes first-order stochastic dominance. Under this condition, the detector can use a single recursive statistic based on \(\bar g/f\), avoiding GLR or mixture computations while retaining minimax robustness [2412.20207]. The resulting Robust Data-Efficient CUSUM (RDE-CUSUM) uses parameters \(A>0\), \(\mu>0\), and \(h\ge 0\), initializes \(\bar D_0=0\), and updates \[ \bar D_{n+1}=(\bar D_n+\ell_{n+1})^{h+}\quad \text{if } \bar M_{n+1}=1, \] \[ \bar D_{n+1}=\min\{\bar D_n+\mu,0\}\quad \text{if } \bar M_{n+1}=0, \] with \((x)^{h+}=\max\{x,-h\}\), sampling rule \[ \bar M_{n+1}=1 \ \text{if } \bar D_n\ge 0,\qquad \bar M_{n+1}=0 \ \text{otherwise}, \] and stopping rule \[ \bar\tau_{\mathrm{rdc}}=\inf\{n\ge 1:\bar D_n\ge A\}. \] When the statistic is negative, the procedure skips observations and recovers deterministically toward zero with slope \(\mu\). The parameter \(h\) truncates the undershoot and bounds skip lengths [2412.20207]. The false-alarm and duty-cycle controls decouple. Setting \(A=|\log \alpha|\) guarantees \(\mathrm{FAR}\le \alpha\) for any fixed \(\mu,h\). For duty cycle, the paper gives a general sufficient condition \[ \mu \le \frac{\beta}{1-\beta}\frac{C_2(h)}{C_1}\ \Rightarrow\ \mathrm{PDC}\le \beta, \] and the asymptotic simplification \[ \mu \le \frac{\beta}{1-\beta}D_{KL}(f\Vert \bar g), \qquad \mathrm{PDC}\approx \frac{\mu}{\mu + D_{KL}(f\Vert \bar g)}. \] The robust minimax lower bound is \[ \sup_{G\in G}\mathrm{WADD}^G(\Psi)\ge \frac{|\log \alpha|}{D_{KL}(\bar g\Vert f)}(1+o(1)), \] and RDE-CUSUM achieves \[ \sup_{G\in G}\mathrm{WADD}^G(\bar\Psi_{\mathrm{rdc}}) = \frac{|\log \alpha|}{D_{KL}(\bar g\Vert f)}(1+o(1)), \qquad \alpha\to 0, \] for fixed \(\beta\), establishing asymptotic robust optimality [2412.20207]. The least favorable construction is explicit in several families. For \(f=\mathcal{N}(0,1)\) and \(g=\mathcal{N}(\mu,1)\) with \(\mu\ge \bar\mu\), the LFL is \(\bar g=\mathcal{N}(\bar\mu,1)\). For \(f=\mathrm{Pois}(\lambda_0)\) and \(g=\mathrm{Pois}(\lambda_1)\) with \(\lambda_1\ge \bar\lambda_1\), the LFL is \(\bar g=\mathrm{Pois}(\bar\lambda_1)\). Simulations show that at \(\beta=0.5\), RDE-CUSUM’s WADD closely matches robust CUSUM with full sampling, while fractional sampling is inferior. In county-level COVID-19 onset detection with \(f=\mathrm{Pois}(1)\), \(\bar g=\mathrm{Pois}(2)\), and \(A=\log(1000)\), robust CUSUM and RDE-CUSUM both detect outbreaks within about a week of significant case rises, while RDE-CUSUM reduces sampling cost [2412.20207]. ## 4. Bayesian covert sensing under a warden A distinct Bayesian infinite-horizon formulation models active probing of a discrete memoryless channel \(W_{Y,Z|X,\theta}\) by a legitimate detector observed by an adversary. The state \(\theta_t\in\{0,1\}\) changes at an unknown time \(\Gamma\) with geometric prior \(\pi_k=\rho(1-\rho)^{k-1}\), and the detector chooses \(X_t\in\{0,1\}\), where \(X_t=0\) is innocent and \(X_t=1\) is active probing. Alice observes \(Y_{t-1}\) with unit delay before choosing \(X_t\), while Eve observes \(Z_t\) immediately [2605.16140]. The model assumes “no free passive sensing,” \[ P_1^0(\cdot)=P_0^0(\cdot), \] so innocent inputs reveal no change to Alice. Active sensing provides finite information gain \[ D=D(P_1^1\Vert P_0^1)\in (0,\infty), \qquad V=\mathbb{E}_{Y\sim P_1^1}\!\left[\left(\ln\frac{P_1^1(Y)}{P_0^1(Y)}\right)^2\right]<\infty, \] and Eve’s channel satisfies \(Q_\theta^1\ll Q_\theta^0\) with \(\chi_{2,\theta}(Q_\theta^1\Vert Q_\theta^0)>0\), especially \(\chi_{2,1}>0\) post-change [2605.16140]. The performance objective is \[ \mathrm{ADD}^\star(\alpha,\delta)=\inf_{\Pi^A\in\mathcal P}\mathrm{ADD}(\Pi^A) \] subject to \(\mathrm{PFA}(\Pi^A)\le \alpha\) and \(\mathrm{ECB}(\Pi^A)\le \delta\). The baseline innocent policy \(\beta_t\equiv 0\) and deterministic stop at \(N_\alpha=\lceil |\ln \alpha|/d\rceil\), where \(d=|\ln(1-\rho)|\), gives \(\mathrm{PFA}\le \alpha\), \(\mathrm{ECB}=0\), and \[ \mathrm{ADD}(\Pi^0)=\frac{|\ln \alpha|}{d}+O(1). \] This fixes the passive first-order benchmark [2605.16140]. Two information inequalities govern the covert tradeoff. With unmodulated LLR \(\ell_t=\ln(P_1^1(Y_t)/P_0^1(Y_t))\), modulated LLR \(L_t=X_t\ell_t\), and post-change sensing effort \[ B(\Pi^A)=\mathbb{E}\!\left[\sum_{t=\Gamma+1}^{\tau^A}\beta_t\right], \] the paper proves \[ d\cdot \mathrm{ADD}(\Pi^A)+D\cdot B(\Pi^A)\ge |\ln \alpha|-\frac{d}{\rho}, \] and \[ B(\Pi^A)\le \left(\frac{\delta}{\chi_{2,1}}\right)^{1/2}\bigl(\mathrm{ADD}(\Pi^A)\bigr)^{1/2}. \] Combining them yields the second-order asymptotic converse \[ \mathrm{ADD}(\Pi^A)\ge \frac{|\ln \alpha|}{d} -\frac{D}{d^{3/2}\sqrt{\chi_{2,1}}}\sqrt{\delta |\ln \alpha|}+O(1), \qquad \alpha\to 0. \] The leading term is unchanged by covert sensing; the gain appears only at square-root order [2605.16140]. Achievability is obtained by a constant-sensing-probability Shiryaev-type policy. While running, \(\beta_t\equiv \beta\in(0,1]\), and the stopping time is \[ \tau^A=\inf\{t\ge 1:p_t\ge 1-\alpha\} =\inf\left\{t\ge 1:\Lambda_t\ge \frac{1-\alpha}{\alpha}\right\}, \] where \(p_t=P(\Gamma0\). The CuSum recursion is \[ Z_n=\log \frac{q_\gamma(X_n)}{q(X_n)},\qquad W_0=0,\qquad W_n=\max\{0,W_{n-1}+Z_n\}, \] and \(T_\gamma=\tau_{h(\gamma)}=\inf\{n\ge 1:W_n\ge h(\gamma)\}\). Under vanishing overshoot conditions, the exact asymptotics are \[ E_{1,\gamma}[\tau_h]\sim \frac{e^{-h}+h-1}{D(q_\gamma\Vert q)}, \qquad E_0[\tau_h]\sim \frac{e^{h}-h-1}{D(q\Vert q_\gamma)}. \] If \(h^\star(\gamma)\) satisfies \(E_0[\tau_{h^\star}]=\gamma\), then the optimal ADD obeys the regime split \[ n(\gamma)\sim \begin{cases} \dfrac{\log(\gamma D(q\Vert q_\gamma))}{D(q_\gamma\Vert q)}, & \gamma D(q\Vert q_\gamma)\to\infty,\\[1.2ex] \dfrac{\gamma D(q\Vert q_\gamma)}{D(q_\gamma\Vert q)}\cdot \dfrac{G(y)}{y}, & \gamma D(q\Vert q_\gamma)\to y\in(0,\infty),\\[1.2ex] \dfrac{\gamma D(q\Vert q_\gamma)}{D(q_\gamma\Vert q)}, & \gamma D(q\Vert q_\gamma)\to 0, \end{cases} \] where \[ G(y)=e^{1+y+W_{-1}(-e^{-1-y})}-W_{-1}(-e^{-1-y})-y-2. \] Covertness is defined by \(n(\gamma)=\Theta(\gamma)\), and this occurs precisely when \[ \gamma D(q\Vert q_\gamma)=\Theta(1). \] The threshold itself changes regime: \[ h^\star(\gamma)\sim \begin{cases} \log(\gamma D(q\Vert q_\gamma)), & \gamma D(q\Vert q_\gamma)\to\infty,\\[0.7ex] -1-y-W_{-1}(-e^{-1-y}), & \gamma D(q\Vert q_\gamma)\to y\in(0,\infty),\\[0.7ex] \sqrt{2\gamma D(q\Vert q_\gamma)}, & \gamma D(q\Vert q_\gamma)\to 0. \end{cases} \] Thus the classical \(O(\log \gamma)\) delay law fails when the adversary can make the post-change law converge to the pre-change law [2601.20022]. The paper works out Gaussian and exponential perturbations explicitly. For \(q=\mathcal N(0,1)\) and \(q_\gamma=\mathcal N(\mu(\gamma),1+\sigma^2(\gamma))\), the divergences satisfy \[ D(q_\gamma\Vert q)\sim \frac{1}{2}\mu^2+\frac{1}{4}\sigma^4, \qquad D(q\Vert q_\gamma)\sim \frac{1}{2}\mu^2+\frac{1}{4}\sigma^4, \] and covertness arises when \(\mu(\gamma)=\Theta(\gamma^{-1/2})\) and/or \(\sigma^2(\gamma)=\Theta(\gamma^{-1/2})\). For exponential models \(q=\mathrm{Exp}(\lambda)\), \(q_\gamma=\mathrm{Exp}(\lambda_\gamma)\), covertness occurs when \(|1-\lambda_\gamma/\lambda|=\Theta(\gamma^{-1/2})\) [2601.20022]. In continuous time, the observation process satisfies \[ dX_t=\mathbf{1}_{\{t\ge \nu\}}\mu(\gamma)\,dt+dW_t, \] where \(W_t\) is standard Brownian motion and \(\mu(\gamma)\) is chosen by the covert adversary as a function of the false-alarm constraint \(\gamma\). The continuous-time CuSum statistic is \[ Y_t=\sup_{0\le s\le t} Z_{s,t}=U_t-M_t, \] with \[ Z_{s,t}=\mu(\gamma)(X_t-X_s)-\tfrac{1}{2}\mu(\gamma)^2(t-s), \] \[ U_t=\mu(\gamma)X_t-\tfrac{1}{2}\mu(\gamma)^2 t, \qquad M_t=\inf_{0\le r\le t}U_r, \] and stopping time \(T_h=\inf\{t\ge 0:Y_t\ge h\}\). Exact formulas hold: \[ E_\infty[T_h]=\frac{2}{\mu(\gamma)^2}(e^h-h-1), \qquad E_0[T_h]=\frac{2}{\mu(\gamma)^2}(e^{-h}+h-1). \] The threshold solving \(E_\infty[T_h]=\gamma\) has the Lambert \(W\)-form \[ h(\gamma)=-1-\tfrac{1}{2}\gamma\mu(\gamma)^2 -W_{-1}\!\left(-e^{-1-\frac{1}{2}\gamma\mu(\gamma)^2}\right), \] and the worst-case delay is \[ n(\gamma)=\frac{2}{\mu(\gamma)^2}G\!\left(\tfrac{1}{2}\gamma\mu(\gamma)^2\right). \] The covert regime is characterized by \[ \limsup_{\gamma\to\infty}\gamma\mu(\gamma)^2<\infty, \] equivalently \(\mu(\gamma)=O(\gamma^{-1/2})\), under which \(n(\gamma)=\Theta(\gamma)\). If \(\mu(\gamma)=o(\gamma^{-1/2})\), then \(n(\gamma)\sim \gamma\) and \(h(\gamma)\to 1\). If \(\mu(\gamma)=c/\sqrt{\gamma}\), then \[ n(\gamma)\sim \frac{2}{\theta}G\!\left(\tfrac{1}{2}\theta\right)\gamma,\qquad \theta=c^2, \] and for \(\theta=1\), \[ \frac{n(\gamma)}{\gamma}\to 2G\!\left(\tfrac{1}{2}\right)\approx 0.563. \] The adversary’s total damage is \(D(\gamma)=\mu(\gamma)n(\gamma)\), and among power-law scalings \(\mu(\gamma)=\gamma^{-\delta}\), it is asymptotically maximized at \(\delta=1/2\), i.e., at the covertness boundary \(\mu(\gamma)=\Theta(\gamma^{-1/2})\) [2509.17778]. ## 6. Scope, assumptions, and research directions Several assumptions recur across the literature. Sampling-controlled robust CQCD assumes known pre-change density \(f\), a post-change family \(G\) admitting a least favorable law, i.i.d. observations conditioned on the change point, and LLRs with finite positive moments up to second order [2412.20207]. Bayesian covert sensing assumes a known DMC, a geometric prior, causality with unit-delay feedback, finite \(D\) and \(V\), and Eve-channel absolute continuity with finite \(\chi^2\)-divergences [2605.16140]. Discrete-time covert-adversary asymptotics assume a stationary post-change law \(q_\gamma\) and vanishing expected overshoots; the paper provides sufficient conditions and verifies them for Gaussian and exponential families [2601.20022]. Continuous-time results are specific to the Brownian drift-change model, where exact CuSum expressions are available [2509.17778]. These assumptions delimit where exact theory is currently available. Robustness may degrade if the least favorable law structure fails; the Bayesian ECB formulation does not address unknown post-change or unknown channels; and the covert-adversary analyses rely on stationary \(\gamma\)-dependent post-change laws rather than adaptive nonstationary ones [2412.20207][2605.16140][2601.20022]. This suggests that a unified theory of covert QCD remains incomplete: detector-side covertness is presently best developed through sampling control and ECB constraints, while adversary-side covertness is best understood through asymptotics under vanishing post-change information. The main open directions already identified in the source material include composite pre-change models, nonparametric settings, non-i.i.d. or Markov models, communication-constrained environments, non-Bayesian covert criteria of Lorden or Pollak type for active sensing, adaptive sensing probabilities, constrained POMDP formulations, multi-channel sensing, learning-based policies, and generalized structural conditions guaranteeing vanishing overshoots beyond the Gaussian and exponential examples [2412.20207][2605.16140][2601.20022][2104.04186]. Taken together, the literature shows that covertness does not merely add an engineering side constraint. It can fundamentally change the asymptotic regime. Under detector-side covert sensing, first-order delay laws may remain classical while the attainable gain appears only at second order; under adversary-side covert scaling, the classical logarithmic delay law can collapse to linear scaling in the false-alarm horizon. Source: https://www.emergentmind.com/topics/covert-quickest-change-detection