--- title: 'Entry-Exit Sequences: Theory & Applications' url: https://www.emergentmind.com/topics/entry-exit-sequences type: topic --- # Entry-Exit Sequences: Theory & Applications Searching arXiv for the supplied works and closely related "entry-exit" papers to ground the article in current arXiv records. Entry-exit sequences are formal descriptions of systems whose evolution is organized by an entrance event, an internal phase, and a corresponding exit event. Across the cited arXiv literature, the term appears in several technically distinct senses: subject matching between entry and exit observations in surveillance [1909.05145]; sequential optimal stopping rules for opening and liquidating positions or projects [2309.16008, 1503.01285]; Hoare-like assertions for program segments with multiple entry and exit points [1408.2955]; stationary flows of firms and metastable or recurrent stochastic systems [1910.14023, 1603.03483, 2403.00619]; and entry-exit maps or functions governing delayed loss of stability in slow-fast dynamics [1511.01815, 1910.06318, 2203.10357, 2510.02770]. This suggests a common abstract pattern: entry and exit are coupled not merely as two timestamps, but by a mathematically specified relation such as a score, a stopping boundary, an invariant measure, a hitting-time identity, or a return map. ## 1. Semantic entry-exit matching in surveillance In computer vision, the entry-exit sequence problem is posed as appearance invariant subject recognition for Entry-Exit surveillance applications. Kumar and Nagabhushan define a Semantic Entry-Exit matching model that uses ancillary information about subjects such as height, build, complexion and clothing color to endorse exit of every subject who had entered private area. The model is designed to be robust to variations in clothing, and it assigns equal weight to each describing attribute when computing the matching score [1909.05145]. Each subject \(i\) is represented by four soft-biometric descriptors. Height is normalized as \(h_i = H_i/H_{\max}\in[0,1]\). Body-build is computed from the height-width ratio \(r_i=H_i/W_i\), stabilized by a vertical-projection profile of the silhouette, and normalized as \(b_i=r_i/R_{\max}\in[0,1]\). Skin complexion is extracted from the head region in YCbCr space using standard skin thresholds, with normalized feature vector \[ c_i=\bigl[\mu_{Cb,i}/255,\;\mu_{Cr,i}/255\bigr]^T\in[0,1]^2. \] Clothing color is represented by a \(96\)-dimensional normalized histogram formed by concatenating torso-Cb, torso-Cr, leg-Cb, and leg-Cr histograms, with \(\sum_{j=1}^{96} f_i^{color}[j]=1\). For each attribute \(f\in\{\text{height, build, complexion, color}\}\), the method learns an LDA projector \(T_f\) by maximizing Fisher’s criterion \[ \operatorname{tr}(T_f^T D_f T_f)\;/\;\operatorname{tr}(T_f^T S_f T_f), \] with pooled within-class scatter \(S_f\) and between-class scatter \(D_f\). After projection \(u^f=T_f^T v^f\), entry-exit similarity is measured by Euclidean distance \(d_f(E,X)=\lVert u_E^f-u_X^f\rVert_2\), converted to an attribute-confidence score \[ Cf_f(E_j,X)=\frac{n-\operatorname{rank}_f(E_j,X)+1}{n}, \] and averaged equally: \[ \operatorname{Score}(E_j,X)=\frac14\sum_f Cf_f(E_j,X). \] The online pipeline detects and segments the person, normalizes the image to \(128\times 64\), computes the four descriptors, ranks the gallery for each attribute, and outputs the top-\(k\) matches. Representative results reported for the EnEx benchmark and the authors’ own two-camera \(360^\circ\) Entry-Exit database are as follows. | Setting | Method | Reported matching rates | |---|---|---| | EnEx, gallery size \(25\) | RS-KISS | rank-1 \(0.079\), rank-5 \(0.316\), rank-10 \(0.486\) | | EnEx, gallery size \(25\) | Ensemble-Learning | rank-1 \(0.094\), rank-5 \(0.367\), rank-10 \(0.491\) | | EnEx, gallery size \(25\) | Proposed model | rank-1 \(0.231\), rank-5 \(0.489\), rank-10 \(0.867\) | | EnEx, gallery size \(50\) | Proposed model | rank-1 \(0.217\), rank-5 \(0.466\), rank-10 \(0.812\) | | Two-camera database, \(n=25\) | Proposed model | rank-1 \(0.366\), rank-10 \(0.891\) | | Two-camera database, \(n=50\) | Proposed model | rank-1 \(0.342\), rank-10 \(0.838\) | The reported limitation is that rank-1 accuracy is modest, approximately \(20\%\)–\(36\%\), because height, build and complexion can be uniform among co-located subjects and because of occlusions such as full-body coverings. Skin-feature extraction can fail under heavy head coverings, height can be affected by head-accessories, and clothing histograms are vulnerable if garments are actually changed. The intended use is therefore a two-stage pipeline: Stage 1 performs soft-biometric narrowing and outputs top-\(k\) gallery candidates, and Stage 2 applies more expensive face and/or gait recognition only on those \(k\ll n\) candidates. ## 2. Sequential optimal stopping in finance and real options In mathematical finance, entry-exit sequences are formulated as nested stopping problems. For a one-dimensional Markov process \(X\) with infinitesimal generator \[ \mathcal L f(x)=b(x)f'(x)+\tfrac12\sigma^2(x)f''(x), \] optimal spread trading is posed as a pre-entry decision followed by a post-entry liquidation decision. The post-entry value is \[ G(x)=\sup_{\nu\ge0} E_x\!\bigl[e^{-\hat r\nu}(X_\nu-\hat c)\bigr], \] and the pre-entry value is \[ V(x)=\sup_{\tau\ge0} E_x\!\bigl[e^{-r\tau}(G(X_\tau)-c)\bigr]. \] The free-boundary characterization gives an exit barrier \(G^*\) with continuation region \(\{xa^*\}\) and stopping region \(\{x\le a^*\}\). Smooth-pasting conditions are imposed at both boundaries [2309.16008]. The same work solves the sequential problem by a signature optimal stopping method. A truncated path signature \[ S^N(X_{[0,t]}) \] generates linear signature policies \(\theta(X_{[0,t]})=\langle \theta,S^N(X_{[0,t]})\rangle\), and randomized stopping times are defined through the accumulated quadratic functional of this policy. In practice, the implementation replaces the random threshold by a constant \(k>0\), smooths the indicator by a sigmoid, optimizes a discrete-time loss by Adam or L-BFGS, and then uses the learned \(\theta^*\) in testing to generate sequential stopping times \(\tau_0^*,\nu_0^*,\tau_1^*,\nu_1^*,\dots\). Training paths can be obtained by fitting an OU model, by block-bootstrap, or by a time-series GAN. In simulated OU experiments with \(\mu=10,\theta=10,\sigma\in\{0.1,1.0,2.0\}\), the signature optimal stopping value is reported to cluster near the true optimum. In real-market tests on four US-equity pairs over 2021–22, the signature strategy achieved annualized returns \(>40\%\) versus \(10\%\) for baseline, Sharpe \(=0.18\) versus \(0.05\), and reduced drawdown. Transaction costs enter additively in both stopping problems; increasing \(c\) shifts the entry boundary downward, increasing \(\hat c\) raises the exit barrier, and doubling \(c\) from \(0.01\) to \(0.02\) widened the no-trade region by approximately \(15\%\) in the simulated OU model. A closely related real-options formulation studies entry-exit decisions with implementation delay under uncertainty. There, the underlying price process is geometric Brownian motion, \[ dX_t=\mu X_t\,dt+\sigma X_t\,dW_t, \] entry is decided at \(\tau_I\), exit at \(\tau_O\ge \tau_I\), but implementation occurs at \(\tau_I+\delta\) and \(\tau_O+\delta\). The expected NPV is transformed into an equivalent instant-implementation problem with modified coefficients \(k_1,k_0,l_1,l_0\), after which the problem again decomposes into a post-entry stopping problem \(G\) and a pre-entry stopping problem \(H\). On continuation regions, \(G\) and \(H\) satisfy linear ODEs, with value-matching and smooth-pasting at the optimal thresholds. If \(\lambda_1<0<1<\lambda_2\) are the roots of \[ \tfrac12\sigma^2\lambda(\lambda-1)+\mu\lambda-r=0, \] then \(G\) and \(H\) admit closed forms in terms of \(x^{\lambda_1}\) and \(x^{\lambda_2}\). The paper distinguishes the cases \(\Delta=I+J\ge0\) and \(\Delta<0\): when \(\Delta\ge0\), there is a single upward trigger \(p_I\); when \(\Delta<0\), there may be two entry triggers \(00\), and join only when expected value is high enough. Within each period, incumbents first decide whether to exit, then entrants decide whether to enter, then surviving firms produce and earn profits, and finally productivities update [1910.14023]. The incumbent value function is written as \[ V(\phi,p)=\pi(\phi,p)+\beta\cdot\max\Bigl\{0,\int V(\phi',p)\,\Gamma(\phi,d\phi')\Bigr\}, \] with outside option \(0\). The exit threshold is \[ \phi^*(p)=\min\Bigl\{\phi\ge0\;\Bigm|\;\int \bar V(\phi',p)\Gamma(\phi,d\phi')\ge0\Bigr\}, \] and free entry requires \[ \int \bar V(\phi,p)\,\gamma(d\phi)=c_e. \] A stationary recursive equilibrium is a triple \((p,M,\mu)\) satisfying goods-market clearing, the entry condition, the exit cutoff, balanced entry-exit \(M=\mu(\{\phi<\phi^*(p)\})\), and the invariance equation \[ \mu(B)=\int_{\phi\ge\phi^*(p)}\Gamma(\phi,B)\,\mu(d\phi)+M\gamma(B). \] Existence and uniqueness are characterized by finiteness of lifetime output \(\ell(p^*)\); when equilibrium exists, the invariant distribution is unique, aggregate supply is \(M^*\ell(p^*)\), and a Pitman-occupation formula represents \(\mu\) as expected occupation time up to the exit time. Under multiplicative productivity growth with \(E[A^\alpha]=1\), the invariant distribution has a Pareto right tail: \[ P(X>x)\sim \bar C x^{-\alpha}. \] A different stochastic meaning of entry-exit sequencing appears in reversible finite-state Markov chains with metastable states in series. The model assumes three distinguished states \(x_2\), \(x_1\), and \(x_0\), with \(H(x_2)>H(x_1)>H(x_0)\), common metastable barrier \(\Gamma_m\), and a series condition \[ \lim_{\beta\to\infty} P_{x_2}(\tau_{x_0}<\tau_{x_1})=0. \] Starting from the higher metastable state \(x_2\), the process must therefore visit \(x_1\) before reaching the stable state \(x_0\) in the low-temperature limit. The theory expresses mean hitting times in terms of capacities and proves sharp asymptotics \[ E_{x_2}[\tau_{\{x_1,x_0\}}]\sim (1/k_2)e^{\beta\Gamma}, \qquad E_{x_1}[\tau_{x_0}]\sim (1/k_1)e^{\beta\Gamma}, \] together with the addition rule \[ E_{x_2}[\tau_{x_0}] \sim \Bigl(\frac1{k_2}+\frac1{k_1}\Bigr)e^{\beta\Gamma}. \] This turns a serial entry-exit path through metastable states into an asymptotically additive decomposition of exit times. In the zero-chemical-potential Blume-Capel model, the states “all \(-1\),” “all \(0\),” and “all \(+1\)” play the roles of \(x_2\), \(x_1\), and \(x_0\), with critical droplet size \(\ell_c=\lfloor 2/h\rfloor+1\) and barrier \[ \Gamma_c = 4\ell_c - h[\ell_c(\ell_c-1)+1] \] [1603.03483]. ## 5. Entrance and exit chains for Markov processes For general Markov chains, entry-exit sequences can be isolated as induced chains observed only when a set \(A\) is crossed from its complement. Let \(Y=(Y_n)_{n\ge0}\) be a time-homogeneous Markov chain on a Polish space \(X\) with known \(\sigma\)-finite invariant measure \(\mu\), and fix a Borel set \(A\subset X\). The entrance times are \[ T_0^{\rangle A}:=0,\qquad T_n^{\rangle A}:=\inf\{k>T_{n-1}^{\rangle A}: Y_{k-1}\in A^c,\;Y_k\in A\}, \] and the entrance chain is \[ Y_n^{\rangle A}:=Y_{T_n^{\rangle A}}. \] Using the same stopping times, the exit chain records the last position in \(A^c\) immediately before entering \(A\): \[ Y_n^{A^c\rangle}:=Y_{T_n^{\rangle A}-1}. \] Under the minimal assumption that each \(T_n^{\rangle A}\) is almost surely finite, both induced processes are themselves time-homogeneous Markov chains with explicit kernels [2403.00619]. The paper studies these chains by inducing on the two-step cylinder \[ C=\{\omega=(\omega_0,\omega_1,\dots): (\omega_0,\omega_1)\in A^c\times A\}. \] If \(\mu\) is recurrent for \(Y\) and both \(A\) and \(A^c\) have positive \(\mu\)-mass, the invariant measures of the induced chains are explicit: \[ \mu_A^{\rm entr}(dx)=P_x(\widehat Y_1\in A^c)\,\mu(dx)\quad\text{on }A, \] where \(\widehat Y\) is a dual chain under \(\mu\), and \[ \mu_{A^c}^{\rm exit}(dx)=P_x(Y_1\in A)\,\mu(dx)\quad\text{on }A^c. \] A Kac-type lifting formula reconstructs \(\mu\) from the induced chain by integrating occupation times between successive entrances. Under the additional assumptions that \(Y\) is weak Feller, topologically irreducible, and topologically recurrent, there is a bijection between locally finite invariant measures of \(Y\) and those of the entrance and exit chains. Hence uniqueness and ergodicity of the original invariant measure are inherited by the induced chains. The main applications concern random walks on \(\mathbb R^d\), viewed as stationary chains started under Lebesgue–Haar measure on the closed subgroup generated by the increment support. For one-dimensional oscillating random walks and the choice \(A=[0,\infty)\cap Z\), the successive overshoots above zero form an entrance chain with invariant measure \[ \pi_+(dx)=(1-P(X_1\le x))\,\lambda(dx)\quad(x\ge0), \] while the corresponding undershoots have \[ \pi_-(dx)=(1-P(X_1>x))\,\lambda(dx)\quad(x<0). \] The overshoot chains are ergodic under these invariant laws. This ergodic structure is then used to prove a central limit theorem for the number \(L_n\) of zero-level crossings: \[ \lim_{n\to\infty} P_x\!\Bigl( \frac{\sigma L_n}{E|X_1|\sqrt n}\le y \Bigr) = 2\Phi(y)-1, \] assuming \(E[X_1]=0\) and \(E[X_1^2]=\sigma^2\in(0,\infty)\). ## 6. Entry-exit functions in slow-fast dynamics In singular perturbation theory, the entry-exit function describes how an orbit that is attracted to a critical manifold before a turning point remains close to it after the stability has changed and only later leaves. For the planar slow-fast system \[ \dot x=\epsilon,\qquad \dot z=h(x,z,\epsilon)z, \] with \[ h(x,0,0)<0\quad(x<0),\qquad h(x,0,0)>0\quad(x>0), \] the critical manifold \(z=0\) is normally attracting for \(x<0\) and normally repelling for \(x>0\). If the orbit enters at \((x_{\rm in},z_0)\) with \(x_{\rm in}<0\), its return to the line \(z=z_0\) occurs at \(x_{\rm out}(\epsilon)=p_\epsilon(x_{\rm in})\), and the limiting entry-exit relation is \[ \int_{x_{\rm in}}^{\,p_0(x_{\rm in})} h(x,0,0)\,dx=0. \] Geometric singular perturbation theory explains this by blow-up of the axis \(\{z=\epsilon=0\}\), analysis in entry, central, and exit charts, and matching along the blown-up cylinder. The same picture holds for the quadratic variant \[ \dot x=\epsilon,\qquad \dot z=H(x,z,\epsilon)z^2, \] and the linear case can be reduced to the quadratic one by the singular substitution \[ z=\kappa(w)= \begin{cases} \exp(-1/w),& w>0,\\ 0,& w\le 0. \end{cases} \] The corresponding return map is generically \(C^\infty\)-smooth in \((x,\epsilon,\epsilon\log\epsilon)\), and under a flatness property \(\mathcal F\) it is \(C^\infty\) in \((x,\epsilon)\) without \(\epsilon\log\epsilon\)-dependence [1511.01815]. The scalar theory extends to multi-dimensional fast variables in systems \[ \dot p=\epsilon f(p,z,\epsilon)+h(p,z,\epsilon),\qquad \dot z=g(p,z,\epsilon), \] with invariant critical manifolds \(M_i=\{z=z_i\}\), fast heteroclinics \((B_{i-1},z_{i-1})\to (A_i,z_i)\), and a turning-point condition under which exactly one fast component changes stability on each \(M_i\). Local charts yield entry maps \(\pi_i\) and slow exit maps \(\widehat Q_i\), and chaining them around an \(N\)-corner singular loop produces a return map \[ \widetilde P:A_1\times\Lambda_1\to A_1\times\Lambda_1. \] Its fixed points encode relaxation oscillations, and their orbital stability is determined by the spectral radius of \(D\widetilde P\): \(\rho<1\) implies orbital asymptotic stability, while \(\rho>1\) implies orbital instability. The paper applies this framework to predator-prey systems with rapid ecological evolutionary dynamics, obtaining both stable and unstable relaxation cycles [1910.06318]. A related semistable setting arises in the two-dimensional system \[ \dot x=1,\qquad \dot y=\frac1\varepsilon\,G(x,y,\varepsilon),\qquad G(x,y,\varepsilon)=\sqrt{m^2+y^2}\,(f(x)-y), \] with \(m(\varepsilon)=e^{\rho/\varepsilon}\), \(\rho<0\). The reduced problem has slow branches \(y=f(x)\) and \(y=0\), where \(y=0\) is semi-stable. The exit point \(\Theta(\theta;\rho)\) associated with entrance at \(x=\theta\) is determined by the area rule \[ \int_\theta^{\,\Theta(\theta;\rho)} f(s)\,ds=-2\rho. \] This defines an entry-exit map \(E_\varepsilon\), whose fixed points are canard trajectories; \(N\)-turn canards arise from fixed points of the \(N\)-fold composition \(E_\varepsilon^N\). In the related linear model of Katriel, the same mechanism yields a unique globally-attracting periodic solution whose Floquet multiplier tends to zero as \(\varepsilon\to0\) [2203.10357]. The most delicate case is a degenerate turning point on the invariant line \(y=0\), with reduced slow flow \[ x'=-x^{2n}(1+o(1)) \] at \(\epsilon=0\). For \(n=1\), the leading-order entry-exit relation is a principal-value equation \[ I_{\rm slow}(x_{\rm out}^b,x_{\rm in}^b)+I_{\rm cyl}(\lambda)=0, \] and the Dulac map \[ \Delta(x_{\rm in},\epsilon)=\Delta_0(x_{\rm in})+\phi\!\bigl(x_{\rm in},\epsilon,\epsilon\ln\tfrac1\epsilon\bigr) \] is \(C^k\) in \((x_{\rm in},\epsilon,\epsilon\ln(1/\epsilon))\). For \(n\ge2\), the analysis shows that a well-defined Dulac map generally does not exist without additional control parameters; in a codimension-one case, varying a single unfolding parameter can restore a well-defined entry-exit relation. The \(n=1\) theory is applied to the graphics \((I_2^1)\) and \((I_4^1)\) in the Dumortier–Roussarie–Rousseau program [2510.02770]. ## 7. Conceptual unification and recurrent misconceptions The literature does not support a single universal meaning of “entry-exit sequence.” In surveillance it is a ranked matching problem over ancillary descriptors; in optimal stopping it is a nested pair of stopping rules; in Hoare-like logic it is a correctness judgment for a program segment with designated entry and exit indices; in firm dynamics it is a stationary turnover mechanism; in induced Markov-chain theory it is a subsequence sampled on crossings of a set boundary; and in slow-fast dynamics it is a return or Dulac map determined by delayed loss of stability. Treating these as if they shared a single formalism would be misleading. A second recurring misconception is that entry-exit constructions are intrinsically two-point descriptions. In the cited works they generally include an internal mechanism: LDA projection and rank aggregation in surveillance [1909.05145]; variational inequalities, smooth pasting, or signature filtrations in stopping problems [2309.16008, 1503.01285]; concatenation and repetition rules in program logic [1408.2955]; invariant measures or capacities in stochastic models [1910.14023, 1603.03483, 2403.00619]; and blow-up charts, principal-value integrals, or return-map derivatives in singular perturbation theory [1511.01815, 1910.06318, 2203.10357, 2510.02770]. A third misconception is that exit should occur exactly at the loss of local stability. The slow-fast literature shows the opposite: orbits can remain near a repelling branch until an accumulated balance condition is met, yielding integral relations such as \[ \int h(x,0,0)\,dx=0 \quad\text{or}\quad \int f(s)\,ds=-2\rho. \] Likewise, in the economic and financial stopping problems, entry and exit are not arbitrary thresholds but free boundaries determined jointly by continuation values and transaction costs. Taken together, these results show that entry-exit sequencing is a cross-disciplinary pattern for linking boundary-crossing events to a mathematically constrained continuation mechanism. The specific objects differ sharply across fields, but the repeated appearance of thresholds, invariant laws, matching equations, and return maps indicates that the technical significance of an entry-exit sequence lies in the rule that mediates between entering and leaving, rather than in either event viewed in isolation. Source: https://www.emergentmind.com/topics/entry-exit-sequences