WorksPhysics and Astronomy
Delay-induced anomalous density-polarization relation in chiral active matter
Abstract
Aligning active matter can spontaneously exhibit diverse forms of ordered collective motion, and higher local density is generally expected to enhance local polarization. Meanwhile, delays in agents' responses are widespread in real active systems, raising the question of how delayed alignment affects collective motion and, in particular, whether higher local density still strengthens polar order. Here, using a chiral active particle (CAP) model with delayed alignment, we uncover a counterintuitive anomalous density-polarization relation (DPR) within droplets: the local polarization first increases and then decreases with increasing local density. Analysis of the deviation-angle dynamics reveals a delay-induced supercritical discrete Hopf bifurcation. Noise-sustained oscillations before the bifurcation and self-sustained oscillations afterward broaden the deviation-angle distribution and reduce polarization. Increasing the delay shifts the bifurcation to lower densities, enhancing the anomalous DPR in agreement with simulations. Using the nonequilibrium landscape-flux theory, we further identify the nonequilibrium flux as the dynamical origin of the anomalous DPR. Our findings not only establish time delay as a control knob for collective behavior in chiral active matter, but also offer a useful analytical perspective on anomalous density-order relations in other nonequilibrium systems exhibiting collective order.
APSmanus.pdf
Active matter systems consist of agents that consume energy to generate directed motion [1]. They range from bacterial suspensions and bird flocks [2, 3] to artificial systems such as Janus particles and Quincke rollers [4, 5]. These systems exhibit diverse collective phenomena, including flocking, clustering, and active turbulence [6, 7, 8]. In aligning active matter, local interactions tend to align the orientations of neighboring agents. By bringing more neighbors within alignment range, higher density can favor polar order (e.g., in Quincke roller bands and mobbing jackdaw flocks [3, 9]). The Vicsek model [10] provides a minimal description of polar alignment. It has been extensively studied [11–15] and has spawned many variants [16–18]. Among these, Liebchen et al. Liebchen et al. [19] proposed a chiral variant of the Vicsek model, called the chiral active particle (CAP) model, which describes active matter systems with kinetic chirality commonly found in nature [20–23]. The introduction of chirality in the CAP model not only generates unique dynamic and thermodynamic behaviors [24–26] but also gives rise to rich patterns such as micro-flocks and macro-droplets [19, 27]. As we show below, local polarization increases monotonically with local density within droplets of the CAP model. We refer to this monotonic relation as the normal density–polarization relation (DPR).
In most Vicsek-like models (including the CAP model), an agent receives directional signals from its neighbors and changes its own direction simultaneously. In real-world systems, however, agents require time to respond due to the finite speeds of information transfer, processing, and actuation, which naturally introduce delays [28]. Remarkably, time delays have significant effects on active matter systems, such as inducing chirality [29] and forming collective patterns [30] in active Brownian particles with delayed attractions. Moreover, in flocking models [31] such as the Vicsek model [32, 33], time delays provide an effective means to switch between dynamical states, with short delays promoting order and long delays promoting disorder. These effects of delay on collective order motivate us to ask how delayed alignment reshapes the DPR within ordered droplets and what mechanisms underlie this change.
In this work, by incorporating delayed alignment into the CAP model, we develop a delayed CAP model and uncover a counterintuitive behavior within the formed droplets: local polarization first rises and then falls as local density increases, leaving the densest core less polarized than the periphery, contradicting the normal DPR in the delay-free CAP model. We term this counterintuitive behavior the anomalous density–polarization relation. Focusing on this behavior, we derive the dynamical equation for the deviation angle of particles inside droplets and analyze it via bifurcation theory [34]. We identify a supercritical discrete Hopf bifurcation in the deviation-angle dynamics, producing a periodic orbit with self-sustained oscillations above a critical density. Even below this density, noise-sustained oscillations occur around the stable fixed point. Both oscillations broaden the deviation-angle distribution and reduce polarization, accounting for the anomalous DPR. As the delay increases, the critical density decreases, shifting the bifurcation to lower densities, thereby making the anomalous DPR more pronounced, in agreement with the simulation results. Moreover, using the nonequilibrium landscape-flux theory [25, 35–39], we identify the nonequilibrium flux and its density derivative as the dynamical origin of the anomalous DPR and the signature of the bifurcation, respectively. In addition, we also explore the influence of the alignment strength and chirality strength on the anomalous DPR.
Model and Methods.– In the delayed CAP model, each chiral active particle experiences an alignment force after a delay following the reception of neighbor signals and the computation of the alignment force. To be more precise, the dynamics of the -th particle is governed by
where , , , and are respectively the position vector, the self-propulsion speed, the angle, and the moving direction of the -th particle. The angular equation comprises chirality, alignment, and noise as its three terms. is the chirality strength, representing the intrinsic rotation frequency of the particle. In the alignment term, and are the alignment strength and radius, respectively. denotes the set of all particles in the neighborhood of particle (the region within ), including particle itself. is the noise intensity, and is the standard Gaussian white noise satisfying . denotes the delay time. For , the delay effect disappears, and the model reduces to the original CAP model.
In our simulation, we use and as basic units for length and time, respectively. We discretize the equations (1) by using the Euler integration method and perform numerical simulations with particles placed in a two-dimensional box of size with periodic boundary conditions, corresponding to the global density . We set , , , and , where is the time step. We set and , if not otherwise stated. The initial construction of the delayed CAP system is random. All sampled data are collected after the system reaches steady state ( time steps).
Delay-induced anomalous DPR in ordered droplets.
To quantitatively describe the anomalous DPR, we introduce two parameters: the local density () and the local polarization (), where is the area of a local grid cell, and is the number of particles in . Accordingly, regions with and correspond to the ordered droplet and the disordered gas phase, respectively.
For the original and delayed CAP systems under our parameter setting, the ordered droplet phase can spontaneously emerge, as shown in Figure 1. In both systems, we observe that is higher in regions closer to the droplet core [Fig. 1(c,d)]. However, the spatial distribution of differs between the two systems. In the undelayed system [Fig. 1(f)], regions with higher near the core also exhibit higher . In contrast, in the delayed system [Fig. 1(e)], we observe a novel behavior: the region with the highest at the core actually shows a lower than the peripheral regions.

Figure 1. Anomalous density–polarization relation (DPR) induced by time delay. (a–f) Spatial distributions of the ordered droplet phase. (a,c,e) Snapshot, local density , and local polarization , respectively, for the delayed CAP system (). (b,d,f) Snapshot, , and for the delay-free CAP system. (g) Mean local polarization vs at various . (h) (orange), (black), and (red) as functions of .
For clarity in describing the anomalous DPR, we plot the mean local polarization against for various in Figure 1(g). In the absence of delay (), increases monotonically with . This is consistent with the intuitive picture that a higher local density provides more neighbors, which strengthens the alignment force and thus enhances the local polarization. However, for nonzero , the anomalous DPR is induced: first increases and then decreases with increasing , exhibiting a non-monotonic dependence. As increases, decreases globally, and its peak shifts to lower , implying a more pronounced anomalous DPR.
To quantify the degree of the DPR anomaly, we define , which measures the polarization deficit of the densest region relative to the most polarized region within the droplet. Here, is the maximum mean local polarization and is the maximal local density. Thus, corresponds to a normal
monotonic DPR, whereas a larger indicates a more pronounced anomalous DPR. Figure 1(h) summarizes the dependence of (orange), (black), and (red) on . The quantity decreases linearly with increasing , indicating that the delay reduces the overall polarization inside the droplet. Meanwhile, decreases more rapidly than does, but its decline slows down at larger . Consequently, increases with first, and tends to saturate at large delays, confirming that the delay enhances the anomalous DPR in general.
Dynamical bifurcation.— With the neighborhood quantities defined, the alignment term in the angular dynamics of Eq. (1), inspired by the Kuramoto model [40], takes the form: , where , , and are the density, polarization, and average angle of CAPs in the -th particle’s neighborhood, respectively (see Supplemental Material (SM) for details). Inside the steady-state droplet, we define the deviation angle , and derive its evolution from Eq. (1) under the steady-state approximation (see SM for details):
where is the effective alignment strength of the neighborhood ( and are time-independent under our approximation explained in SM). Thus, the original multi-particle coupled Eq. (1) is reduced to an equivalent single-particle equation in the neighborhood, Eq. (2). Regardless of a particle’s position within the droplet, as long as it experiences the same and , the evolution of its deviation angle follows Eq. (2).
To reveal the dynamical mechanism of anomalous DPR, we consider the deterministic part of Eq. (2). Discretizing it in time yields a single-variable delay-difference equation:
Through a standard bifurcation analysis [34], we characterize the dynamics of Eq. (3) (see SM for details). For , at the critical point , the system undergoes a supercritical discrete Hopf bifurcation. In the limit , . For , the fixed point is locally asymptotically stable. For , the fixed point loses stability and a periodic orbit emerges, corresponding to self-sustained oscillations. Linearizing the noisy version of Eq. (3) around predicts a polarization maximum at (see SM for details). As , approaches a constant (). For , noise continuously excites the damped oscillatory mode that loses stability at , sustaining quasicycles [41, 42]. As increases, weaker damping amplifies the response to noise, enhancing fluctuations in and reducing polarization (see Figure S2 in SM). In short, the dynamics of is determined by . Over the simulated
Figure S2. Deviation-angle power spectra obtained from simulation data for . Parameters: , , , and .
density range, increases monotonically with (see Figure S3 in SM), giving a one-to-one correspondence. We thus define the optimal and critical densities and as the densities corresponding to and , respectively.

Figure S3. The equivalent alignment strength versus the local density for .
To intuitively show the bifurcation, we plot the deterministic solution of Eq. (3) and the steady-state probability distribution obtained from simulations, both in the phase plane for different . Take as an example. For [Figure 2(a,Figure 2b)], remains a point and remains unimodal. As density increases, broadens, consistent with stronger quasicycles and reduced polarization. For [Figure 2(c)], becomes a periodic orbit and develops a ring-shaped structure that expands with density, further reducing polarization. By contrast, for sufficiently large in the delay-free case, remains a point and remains unimodal [Fig. 2(d)], which yields a narrow unimodal distribution and a large , consistent with the normal DPR. Furthermore, the time autocorrelation of in the delayed CAP system displays three distinct behaviors: overdamped relaxation in the gas, and heavily and weakly damped oscillations in the droplet periphery and core, respectively (see SM for details).

Figure 2. Bifurcation, theoretical polarization curves, and scaled master curves for the anomalous DPR. (a–d) Probability distribution (color map) and deterministic solution (red) in the phase plane, for neighborhood densities (a) , (b) , and (c) with delay , and (d) without delay. (e) Theoretical polarization and fitted simulated polarization versus for . (f) Scaled polarization curves for different in the scaled coordinates and . Inset of (f): the maximum polarization (black dots) and the optimal density (orange dots) as functions of ; solid lines are fits. In (e, f), the gray and red vertical lines mark the optimal density and the critical density , respectively, and the red dashed line denotes the theoretical result.
Scaled curves.— For quantitative comparison with simulations, we compute the theoretical polarization from the stationary distribution obtained numerically from Eq. (2). Combining with the corresponding density for each yields the theoretical DPR –. The theoretical curve agrees well with the simulation data for [Fig. 2(e)]. Moreover, coincides with the density at which reaches its maximum, marking the onset of the polarization decline. Meanwhile, exactly coincides with the density at the most negative value of (see Fig. S5 in SM), indicating the most rapid decrease of with near . In addition, the – curves for different (see Fig. S6 in SM) are qualitatively consistent with the – curves in Figure 1(g).
When plotted in scaled coordinates (, ), the – curves for different merge onto a single master curve, as shown in Figure 2(f). The red dashed line represents the scaled theoretical curve – (see Figure S7 in SM), which also merges onto the same master curve. In scaled coordinates, the values of for different nearly coincide, implying that is independent of . The scaling parameters reflect the degree of the DPR anomaly (). As increases, the vertical scale decreases linearly, whereas the horizontal scale decays as shifted , as shown in the inset of Figure 2(f). For small , decreases much more rapidly than , resulting in a much stronger horizontal compression of the master curve than vertical. Consequently, in Figure 1(h) drops faster than with increasing , causing a sharp increase in . For large , the decay of slows down, resulting in a comparable vertical compression to horizontal compression. Hence, and decrease at similar rates, thereby saturating the growth of .

Figure S7. Theoretical scaled curves of versus for nonzero delays. The gray and red lines mark and , respectively, using and for .
Dynamical mechanism of anomalous DPR.— To uncover the dynamical origin of the anomalous DPR, we apply the nonequilibrium landscape-flux theory [25, 35–39] in the -dimensional delay-embedded space spanned by at different , where is the number of delay steps in the discretized system (see SM for details). In steady state, the probability density defines the nonequilibrium potential landscape , which generates the gradient force , tending to drive the system toward potential minima.
Figure 3. Landscape-flux analysis for . (a,b) 2D potential and flux at and 30, respectively. (c) The averaged flux magnitude, ; inset: . (d) max and max for the curl (green) and gradient (orange) forces at different , respectively. Red and gray lines mark and , respectively.
potential minima. The nonequilibrium flux determines the curl force , which drives circulation in this delay-embedded space. Here, is the probability velocity field. For visualization, we perform dimensionality reduction and construct the two-dimensional potential and flux in the plane (see SM for details). Below [xref]Fig. 3(a), circulates around the stable fixed point in the single basin of , tending to broaden the basin and counteracting the attraction toward the fixed point. Above [Fig. 3(b)], develops a ring-shaped valley that attracts the system toward the periodic orbit. circulates along this valley, sustaining the oscillations.
The averaged flux magnitude, , increases with [Fig. 3(c)], indicating stronger circulation that contributes to the polarization decline above . Moreover, the slope of the average flux, , reveals three distinct regimes [inset of Fig. 3(c)]. Below , the attractive basin dominates, while both and remain small, corresponding to the normal DPR regime. For , both and rise sharply with , so that the flux dominates, leading to a decrease in and thus to the anomalous DPR regime. peaks near and then decreases slowly, indicating that grows most rapidly near the bifurcation and making a dynamical signature of it. Furthermore, we evaluate , the largest real part of the Jacobian eigenvalues for the gradient and curl forces (see SM for details). Near , and [Fig. 3(d)], revealing a competition between the stabilizing tendency of the potential landscape and the destabilizing tendency of the flux. Together, these results support the flux as the dynamical origin of the anomalous DPR.
Influence of and on anomalous polarization.– In addition to , we also examine the influence of and on the anomalous DPR. Figure 4(a) shows for different values of and . For fixed , increases with , making the DPR more anomalous. The reason is that a larger raises , which facilitates the transition into the bifurcation regime. For fixed , first increases and then decreases with , reaching a maximum at [Figure 4(a)]. This non-monotonic dependence cannot be explained by the equivalent single-particle equation Eq. (2) alone, because it neglects spatial coupling. In fact, it originates from the non-monotonic dependence of on , since also peaks at [Figure 4(b)]. When is small, alignment interactions dominate the particle dynamics, allowing neighboring particles to rapidly phase-lock into a synchronously rotating macro-droplet. Due to the rotation-enhanced polarization effect [19], faster rotation enables the droplet to absorb randomly directed particles more effectively, thereby increasing its density. Consequently, for , increases with [Figure 4(b)]. However, as increases further, the rotation timescale becomes comparable to or even shorter than the alignment timescale [19]. This generates significant phase differences between neighboring particles, frustrating phase-locking and driving the system into the micro-droplet phase [19]. These phase differences weaken polar alignment, reduce the droplet size, and consequently decrease the droplet density. Hence, for , declines with [Figure 4(b)]. Overall, peaks at for fixed , causing to reach its minimum, which in turn gives rise to a non-monotonic dependence of on with a maximum at the same point.

Figure 4. Influence of and on the anomalous DPR. (a) in the – parameter space. (b) in the – parameter space. Each data point is averaged over 20 ensembles. Parameters: , , .
Conclusion.– In summary, by incorporating time delay into the chiral active particle (CAP) model, we have uncovered a counterintuitive anomalous density-polarization relation (DPR) in ordered droplets: the densest core is less polarized than the lower-density periphery. As density increases beyond the optimal density , noise-sustained quasicycles around the stable fixed point broaden the deviation-angle distribution and reduce polarization. At the critical density , a supercritical discrete Hopf bifurcation gives rise to a periodic orbit, whose self-sustained oscillations further reduce polarization. Quantitatively, the density–polarization curve obtained from theory agrees well with those obtained from simulations. Furthermore, and decrease with increasing delay, bringing the bifurcation to lower densities, which makes the anomalous DPR more pronounced, in line with the simulation observation. More importantly, the nonequilibrium landscape-flux analysis identifies the nonequilibrium flux as the dynamical origin of the anomalous DPR and the density slope of the flux as a dynamical signature of the bifurcation. In addition, we explore the effects of alignment strength and chirality strength on the anomalous DPR. Our findings establish time delay as a mechanism for anomalous DPR and advance the theoretical understanding of collective behavior in chiral active matter. The resulting predictions are experimentally testable using programmable light-controlled robots [43, 44] and may guide the design of collective patterns. More broadly, our theoretical framework offers a route toward a unified understanding of density-modulated collective dynamics across diverse nonequilibrium systems, including polar active liquids [45], chemotactic bacteria [46], pedestrian crowds [47, 48], and confluent epithelia [49].
Dongrun Jian, Jie Su, and Jun Wang gratefully acknowledge funding from National Natural Science Foundation of China (Nos. 12404237 and 12234019), the Zhejiang Key Laboratory of Soft Matter Biomedical Materials (Nos. 2025ZY01036 and 2025E10072) and Wenzhou Institute of Biomaterials and Engineering (No. WIUCASQD2022012).
Supplemental Material: Delay-induced anomalous density-polarization relation in chiral active matter
Dongrun Jian; Jie Su; Jun Wang; Jin Wang
Derivation of the dynamical equation for the deviation angle
The parameter transformation
For a given particle , let denote the set of all particles within the alignment radius of particle , including particle itself. We define the average orientation vector of the particles in as:
where denotes the orientation angle of particle , is the neighborhood polarization, is the number of particles in set , and is the average angle over . Multiplying both sides by , taking the imaginary part, and replacing by the neighborhood density , we obtain the parameter transformation:
where corresponds to the deviation angle used in the main text.
The steady-state approximation
Inside the steady-state droplet (), the neighborhood density and polarization vary little over timescales of order . We quantify their relative variations by
where , , and the averages are taken over time and particles. As shown in Figure S1, both and are much smaller than 1 for , indicating that and remain approximately constant over time intervals of order . We therefore approximate them as time-independent, , .

Figure S1. (a) Relative variations of the local polarization and the local density as functions of . (b) Ratio of to , , as a function of .
Defining the deviation angle , the angular equation Eq. (1) of the main text becomes
where is the effective alignment strength of the neighborhood. Inside the droplet, fluctuates weakly around . To quantify the magnitude of this fluctuation relative to , we define
where the averages are taken over time and particles. As shown in Fig. S1(b), for , indicating that . We may therefore neglect the term , yielding
Bifurcation analysis of the deterministic delayed difference equation
To reveal the dynamical mechanism of anomalous DPR, we consider the deterministic part of (9). We define and take , where the integer is the number of delay steps. Discretizing in time yields a single-variable delay-difference equation:
The fixed points of (10) satisfy . We restrict the deviation angle to the range , hence the fixed points are or .
Introducing a small perturbation , let . The fixed point is unstable for any . Linearizing (10) around yields the linear difference equation:
Substituting the ansatz , where is the growth rate per unit time, yields the characteristic equation:
We define the characteristic function by
(12) can then be written as .
Taking the critical oscillatory eigenvalue , where denotes the oscillation frequency, and substituting into Eq. (12), the phase and amplitude conditions give:
where the sign is matched with the sign of (plus for , minus otherwise), so that is a negative real number. Taking the lowest-frequency branch (with the plus sign), we obtain the bifurcation critical point:
The following discrete Hopf analysis concerns . Let . Differentiating the characteristic equation (S9) with respect to yields
Evaluating at the critical eigenvalue , we obtain , indicating that at the critical point the eigenvalue crosses the imaginary axis, confirming that this bifurcation is a discrete Hopf bifurcation.
Since the solution near the bifurcation point has its principal frequency near , we project the solution onto the critical eigenmode (the complex amplitude representation):
Reducing the nonlinear delay-difference equation (S7) to the two-dimensional center manifold and retaining terms through cubic order, we obtain the normal form equation:
where is the complex amplitude at step , and and denote the linear growth coefficient and the cubic saturation coefficient, respectively. Their real parts are explicitly given by and , with . The first Lyapunov coefficient is , whose negativity indicates that the system undergoes a supercritical discrete Hopf bifurcation at the critical point. Since the principal modal frequency is a rational fraction of , resonance occurs, forming a periodic orbit in delay-embedded space with a period of (which can be regarded as a discrete version of a limit cycle).
In summary, the dynamical properties of the deterministic delayed difference equation are:
For , the fixed point is locally asymptotically stable.
At , the system undergoes a supercritical discrete Hopf bifurcation.
For , the fixed point loses stability and a period- orbit emerges, which is the “limit cycle” of the discrete system.
Steady-state variance and the maximum point of polarization
As in the main text, we define the theoretical polarization as , where the average is taken over the stationary deviation-angle distribution at a given neighborhood density. Throughout this section, we consider fluctuations around the stable deterministic fixed point for . Within the linear approximation, this distribution is a zero-mean Gaussian [1] with variance , yielding
Within this approximation, the polarization maximum corresponds to the variance minimum, and increasing variance implies decreasing polarization. We derive the stationary variance below and interpret its increase as approaches from below.
Steady-state variance of the continuous system
Taking the continuous limit at fixed in the discrete critical value Eq. (15), the bifurcation point becomes . Linearizing the continuous dynamical equation Eq. (9) around yields:
Its stationary variance can be derived from the Fokker–Planck equation [2, 1]:
Minimizing with respect to , i.e., setting and , gives , whose solution is the Dottie number . Thus we obtain:
The ratio of the minimum point to the bifurcation point is a constant.
Steady-state variance of the discrete system
We now consider the discrete-time case. Writing , the Euler–Maruyama discretization of Eq. (20) in the stable fixed-point regime reads:
Here are independent standard Gaussian variables, with and .
We now construct the Yule–Walker equations. For convenience, define , , and . Multiplying Eq. (23) by and taking the expectation yields:
Since is independent of for , multiplying Eq. (23) by and taking the expectation gives:
Combined with the symmetry , we obtain a closed linear system for . Solving for yields the steady-state variance . is a rational function of and possesses a minimum point in the interval . The steady-state variance minimum points for different are listed in Supplementary Table S1.
As increases, approaches the constant value of the continuous limit . Moreover, in the continuous limit with fixed, the Yule–Walker solution reproduces the continuous variance of the previous subsection, , and correspondingly . The steady-state variance functions for discrete system are as follows:
For :
| 1 | 46.557 | 100 | 0.4656 |
| 2 | 28.954 | 61.803 | 0.4685 |
| 3 | 20.892 | 44.504 | 0.4694 |
| 4 | 16.318 | 34.730 | 0.4699 |
| 5 | 13.380 | 28.463 | 0.4701 |
Table S1. Steady-state variance minimum point and its comparison with the critical value for different delay steps . All values are computed with .
For :
For :
For :
For :
For , the continuous-time variance decreases monotonically with . The finite minimum arises from Euler–Maruyama discretization and diverges as . To avoid the resulting variance upturn, the delay-free comparison should satisfy .
Noise-sustained oscillations below the bifurcation point
For , the deterministic fixed point remains stable, but noise continuously excites the damped oscillatory mode that becomes unstable at the bifurcation. These noise-sustained oscillations are termed quasicycles [3, 4]. To understand how these fluctuations reduce polarization, we examine their power spectrum . Fourier transforming Eq. (23) gives
where is the characteristic function defined in Eq. (13), and gives the squared response to a unit harmonic perturbation at angular frequency . The variance is the integrated spectral power: .
The oscillatory mode that loses stability at has a characteristic root , satisfying . Here, is its decay rate, and is its oscillation frequency, distinct from the spectral frequency variable . As approaches , and . In the linear approximation, at fixed noise intensity, the resonant peaks have heights proportional to and widths proportional to , so their integrated power grows as . Thus, weaker damping amplifies the response to noise, increasing the variance of and reducing polarization. The predicted divergence at the threshold signals the breakdown of the linear approximation.
Figure S2 shows the power spectra obtained from simulation data at three effective alignment strengths below . As increases, the peak of becomes more pronounced and the integrated spectral power increases. These observations support that noise sustains increasingly strong oscillatory fluctuations before the deterministic fixed point loses stability, broadening the deviation-angle distribution and lowering polarization.
Monotonicity of
Taking as an example (Figure S3), the simulation data all fall within the monotonically increasing regime of , ensuring a one-to-one correspondence between and .
Correlation in the delayed chiral active particle system
To capture the relaxation dynamics of , we compute its time autocorrelation function:
where the average is taken over time and over all particles whose neighborhood density lies within a prescribed window . We denote , , and for the gas phase (), the droplet periphery (), and the droplet core (), respectively. In the stationary regime, approaches a long-time plateau:
which corresponds to the theoretical average polarization of the particles in the window. To clearly compare different correlation behaviors, we define the normalized autocorrelation :
Take as an example. The autocorrelation functions for three different density windows exhibit different relaxation behaviors in Figure S4 (a). shows a weakly damped oscillation that does not vanish within the observation time. This oscillation with period is the temporal fingerprint of the period-22 orbit created by the bifurcation in the droplet core. exhibits a heavily damped oscillation: its period is the same as that of , but the envelope shrinks quickly toward , corresponding to the complex eigenvalues with of the linearized dynamical equation Eq. (11) (the stable fixed-point regime) in the droplet periphery. For the gas phase, exhibits an overdamped relaxation, indicating that decorrelates from its history on a short timescale. Although Eq. (11) is only valid for the droplet phase (), it can still qualitatively explain that the dominant eigenvalue becomes real and negative at low density, leading to the overdamped relaxation toward . Moreover, as the density increases, the plateau first increases and then decreases, signaling the anomalous density–polarization relation reported in the main text.

Figure S4. Time autocorrelations of the deviation angle at . (a) Autocorrelation (solid) with the plateau value (dashed). The superscripts , , and denote the gas, droplet-periphery, and droplet-core regimes, respectively. (b) Normalized autocorrelation for different delays .
In Figure S4 (b), we present the normalized autocorrelations in the droplet core for different delays . The oscillation period can be written as , precisely consistent with the period- orbit. Moreover, the oscillation envelope decays faster for larger , indicating that the sustained oscillation is damped more strongly as the delay increases.
Derivative versus the density $\rho_N`
In Figure S5, the theoretical polarization-maximum density lies close to the zero crossing of , which identifies the polarization maximum in the simulation data. The bifurcation density agrees well with the minimum of on the simulation curve, where , indicating the most rapid decrease of with near .
Figure S5. Derivative of the average polarization, , as a function of the density for .
versus for different delays
In Figure S6, unlike the delay-free case, the average neighborhood polarization for first increases and then decreases with increasing neighborhood density . As the delay increases, decreases globally, and its peak shifts to lower , implying a more pronounced anomalous density–polarization relation (DPR). The – relation is qualitatively consistent with the – curves in the main text.

Figure S6. The average neighborhood polarization versus the neighborhood density for different delays
Scaling collapse of theoretical polarization curves
For the nonzero delays considered, the theoretical curves approximately collapse onto a single master curve when expressed in the scaled coordinates , as shown in Figure S7.
Markov property of the delay-embedded dynamics
Markov process in the -dimensional delay-embedded space
Let and . The Euler–Maruyama discretization of Eq. (9) reads
where are independent standard Gaussian variables, as in Eq. (23). We augment the state with the preceding values of to form the -dimensional state vector
The dynamics of can be written as
where and is the deterministic driving force.
The last components of are deterministic shifts of the first components of . The first component depends on and the new noise , which is independent of the current and all previous states. Thus, the discretized reduced equation Eq. (35) defines a Markov process in the -dimensional delay-embedded space.
Numerical test of the Markov property in the delay-embedded space
We set and and test the Markov property in delay-embedded spaces of different dimensions, focusing on the six-dimensional case. We quantify history dependence in the embedded trajectories using conditional mutual information (CMI):
where and is the embedding dimension. For , the preceding analysis predicts that the embedded dynamics are Markovian. For a Markov process, the past and future are conditionally independent given the present, so . A positive indicates history dependence. To assess whether the embedded trajectories are consistent with the Markov property, we generate Markov trajectories by resampling and compare their CMI with that of the embedded trajectories. During resampling, we randomly reassign the values preceding similar embedded states while retaining each embedded state and its subsequent value. The reference range spans the minimum and maximum CMI estimates from the resampled Markov trajectories. Because of finite-sample effects, the CMI estimates for the Markov trajectories fluctuate around zero.
As varies, the six-dimensional estimated CMI for the embedded trajectories remains close to zero and lies within the reference ranges of the Markov trajectories [Figure S8]. Near , the estimated CMI decreases with increasing embedding dimension and becomes comparable to that of the Markov trajectories at six dimensions [Figure S8]. These results support the Markov property of the six-dimensional delay-embedded dynamics.
Figure S8. CMI tests at and . (a) Six-dimensional CMI versus density. (b) CMI versus embedding dimension near . Circles and squares denote CMI estimates from embedded trajectories and locally resampled Markov trajectories, respectively. Gray error bars show the reference ranges.
Landscape–flux theory and Jacobian spectra
Landscape, flux, and force decomposition
The nonequilibrium landscape-flux theory has been widely used to study nonequilibrium systems, with applications in many areas [5, 6, 7, 8, 9, 10].
We consider a stochastic system whose state vector evolves according to the following Langevin equation:
where is the deterministic driving force and is the stochastic force with , where is the diffusion tensor. For the delay-embedded dynamics, the state vector is , with components for . Owing to the equivalence between the Langevin and Fokker–Planck descriptions, Eq. (39) can be rewritten as the Fokker–Planck equation for the probability distribution :
where is the nonequilibrium probabilistic flux. At the nonequilibrium steady state, the distribution becomes time independent, so that , where the subscript ss denotes the steady state. In contrast to the vanishing flux at equilibrium, the nonequilibrium steady state only requires to be divergence-free, and hence can persist as a rotational field. From the Fokker–Planck equation, reads:
In analogy with the equilibrium potential, the nonequilibrium potential landscape in the steady state is defined as
Different from the equilibrium case, in which the driving force derives solely from the potential gradient, can be decomposed into three parts:
where is the curl component of the driving force (a rotational flux force), is the gradient force of the potential landscape, and is the diffusion force arising from the spatially dependent noise. Note that the curl force equals the probability velocity field (computed via central differences), i.e., , so that . For notational simplicity, we henceforth drop the subscripts on , , , and (the subscript ss for the first three and neq for ) in the following and in the main text.
In the main text, we take and as an example to show the anomalous behavior of the delayed CAP system. The corresponding delay-embedded state is , where for . To visualize the six-dimensional landscape and flux, we project onto the two-dimensional plane and integrate over the remaining coordinates . Hence, the projected landscape and flux are
Here, and denote the probability density and flux in the six-dimensional delay-embedded space, respectively.
To quantify the global circulation of the flux, we define the mean flux as the spatial average of over the delay-embedded domain :
The mean flux measures the average strength of the flux field. In the main text, is computed in the 6-dimensional delay-embedded space for each neighborhood density , and its density slope provides a dynamical signature of the bifurcation.
Jacobian spectra and stability
For , the linearized delay-difference equation (S8) for perturbations about can be written in the delay-embedded space as
where and
Here, denotes the identity matrix.
Let denote the Jacobian eigenvalues of the driving force averaged over , and the eigenvalues of the one-step map . In the corresponding continuous-time linear dynamics, the transient driving force has Jacobian . Its eigenvalues satisfy
At the bifurcation, crosses zero from below, indicating that the fixed point loses stability.
Applying the same construction to each component force in Eq. (43), we obtain the Jacobian eigenvalues of the corresponding transient force, where . Near the bifurcation, remains positive for the curl force, negative for the gradient force, and slightly positive for the diffusion force, as shown in Figure S9. Thus, the potential landscape tends to stabilize the fixed point, whereas the flux tends to destabilize it. Together, these results support the flux as the dynamical origin of the anomalous DPR.
Figure S9. Largest real parts of the Jacobian eigenvalues of the transient curl (green), gradient (orange), total (blue), and diffusion (purple) forces at the six-dimensional origin, for and . The red line marks .
References
- [1]T. D. Frank, P. J. Beek, and R. Friedrich. Fokker–Planck perspective on stochastic delay systems: Exact solutions and data analysis of biological systems. Physical Review E, 68(2):021912, 2003.pubmed.ncbi.nlm.nih.gov/14525011
- [2]Uwe Küchler and Beatrice Mensch. Langevins stochastic differential equation extended by a time-delayed term. Stochastics and Stochastic Reports, 40(1–2):23–42, 1992.DOI
- [3]Carlos A. Lugo and Alan J. McKane. Quasicycles in a spatial predator-prey model. Physical Review E, 78(5):051911, 2008.pubmed.ncbi.nlm.nih.gov/19113159
- [4]A. J. McKane and T. J. Newman. Predator-prey cycles from resonant amplification of demographic stochasticity. Physical Review Letters, 94(21):218102, 2005.DOI
- [5]Jin Wang, Li Xu, and Erkang Wang. Potential landscape and flux framework of nonequilibrium networks: Robustness, dissipation, and coherence of biochemical oscillations. Proceedings of the National Academy of Sciences, 105(34):12271–12276, 2008.DOI
- [6]Jin Wang. Landscape and flux theory of non-equilibrium dynamical systems with application to biology. Advances in Physics, 64(1):1–137, 2015.DOI
- [7]Xiaona Fang, Karsten Kruse, Ting Lu, and Jin Wang. Nonequilibrium physics in biology. Reviews of Modern Physics, 91(4):045004, 2019.DOI
- [8]Jie Su, Zhiyu Cao, Jin Wang, Huijun Jiang, and Zhonghuai Hou. Dynamical and thermodynamical origins of motility-induced phase separation. Cell Reports Physical Science, 5(2):101817, 2024.DOI
- [9]Ruizhe Yan, Jie Su, and Jin Wang. Mechanism of the nonequilibrium phase transition in self-propelled particles with alignment. Physical Review Research, 7(3):033082, 2025.arxiv.org/abs/2411.06818
- [10]Dongrun Jian, Jie Su, Jun Wang, and Jin Wang. Uncovering the underlying mechanisms of phase transitions in chiral active particles. Communications Physics, 8(1):386, 2025.arxiv.org/abs/2503.18388




