---
title: Behavioral-Feedback SIR Model Dynamics
url: https://www.emergentmind.com/topics/behavioral-feedback-sir-model
type: topic
---

# Behavioral-Feedback SIR Model Dynamics

Behavioral-feedback SIR model denotes a class of epidemic models in which the transmission process, the contact structure, or an auxiliary behavioral state evolves endogenously with the epidemic state rather than remaining fixed. In the cited literature, this class includes demographically open SIRS-vaccination systems with prevalence-dependent infectiousness \(\beta(I)\), SIR models with endogenous meeting rates \(q(t)\), network models with state-dependent interaction matrices \(A(x,y)\), and compartmental systems in which awareness, fear, compliance, or vaccination behavior has its own dynamics [2308.01038][2004.14533][2507.03852][1702.04999]. The unifying feature is a closed loop of the form epidemic state \(\to\) behavior, information, or control \(\to\) modified transmission or recovery \(\to\) new epidemic state, although the mathematical realization of that loop varies substantially across models.

## 1. Conceptual scope and relation to classical SIR

In the classical SIR model, the infection term is typically \(\beta S I\) or \(\beta S I/N\) with constant \(\beta\). Behavioral-feedback formulations replace that fixed coefficient by an endogenous object. One common route is prevalence-dependent transmission. In "Modeling the effects of adherence to vaccination and health protocols in epidemic dynamics by means of an SIR model" [2308.01038], the infectiousness parameter is made state dependent through a sigmoidal \(\beta(I)\) that either decreases with \(I\) under protective adaptation or increases with \(I\) under perverse relaxation. In "A feedback SIR (fSIR) model highlights advantages and limitations of infection-dependent mitigation strategies" [2004.13216], the effective reproduction factor becomes \(\mathcal{R}(i)=\mathcal{R}_0/(1+\kappa(i))\), where \(\kappa(i)\) is nonnegative, nondecreasing, and zero at \(i=0\). In "Behavioral-feedback SIR epidemic model: analysis and control" [2509.12257], the transmission rate is \(\beta(x,y)=b(x)/(1+a y)\), so it depends simultaneously on susceptible and infected fractions.

A second route is to endogenize contact intensity directly. In "Virus Dynamics with Behavioral Responses" [2004.14533], non-recovered agents choose an activity level \(q(t)\), and infection becomes proportional to \(q(t)^2 S(t)I(t)\) rather than to \(S(t)I(t)\) alone. A third route is structural: in the network behavioral-feedback SIR model, the infection matrix itself becomes a state-dependent map \(A:[0,1]^{2n}\to\mathbb{R}_+^{n\times n}\), so the incidence term is \([x]A(x,y)y\) rather than \([x]Ay\) with constant \(A\) [2507.03852].

Not every SIR model with behavioral heterogeneity is a full behavioral-feedback model in this stronger sense. "Analysis of SIR epidemic models with sociological phenomenon" [2207.08999] uses fixed group-specific transmission coefficients \(\beta_1,\beta_2\) and, in one version, constant switching rates \(\alpha_1,\alpha_2\); the epidemic does not itself alter those parameters. That literature is therefore more accurately described as static behavioral heterogeneity or exogenous behavioral switching than as a closed epidemic-behavior feedback system.

## 2. Canonical mathematical formulations

A representative prevalence-responsive compartmental formulation is the demographically open SIRS-vaccination system
\[
\frac{dS}{dt} =  \mu N - \frac{\beta S I}{N} - V - \mu S + \xi R,\qquad
\frac{dI}{dt} =  \frac{\beta S I}{N} - \mu I - \gamma I,\qquad
\frac{dR}{dt} =  \gamma I + V - \mu R - \xi R,
\]
with \(N=S+I+R\) constant because births and natural deaths balance. In its behavioral-feedback extension, \(\beta\) is replaced by a sigmoidal \(\beta(I)\), and the paper studies both \(V(t)=\nu N\) and \(V(t)=\nu S\) vaccination schemes [2308.01038]. This model is formally SIRS rather than permanent-immunity SIR, but it is explicitly presented as a behavioral-feedback extension of the SIR framework.

The activity-choice formulation modifies the infection technology rather than the compartmental structure. In the myopic benchmark of [2004.14533], the behavioral first-order condition is
\[
q(t)=\frac{1}{2a+C\beta\cdot \frac{I(t)S(t)}{S(t)+I(t)}},
\]
and the epidemic dynamics are
\[
\dot S(t)=-q(t)^2\beta S(t)I(t),\qquad
\dot I(t)=q(t)^2\beta S(t)I(t)-\kappa I(t),\qquad
\dot R(t)=\kappa I(t).
\]
Here behavior is not an auxiliary compartment but an equilibrium choice variable that feeds back multiplicatively into effective contact intensity.

The network behavioral-feedback SIR model generalizes both scalar SIR and fixed-matrix metapopulation SIR by writing
\[
\dot{x} = -[x]A(x,y)y,\qquad \dot{y} = [x]A(x,y)y-\gamma y,
\]
with \(A(x,y)\) nonnegative and \(C^1\) [2507.03852]. A central special case is rank-1 local feedback,
\[
A_{ij}(x,y)=g_i(x_i)f_j(y_j),
\]
which separates susceptibility/activity on the receiver side from infectivity/activity on the source side.

A further scalar BF-SIR specialization is
\[
\dot x = -\frac{b(x)}{1+a y}\,x y,\qquad
\dot y = \frac{b(x)}{1+a y}\,x y-\gamma y,
\]
with state-dependent reproduction number
\[
R(x,y)=\frac{x\,b(x)}{\gamma(1+a y)}.
\]
In that model, \(b(x)\) is non-decreasing and the denominator \(1+a y\) encodes prevalence-dependent caution [2509.12257].

## 3. Behavioral state variables, awareness, and social contagion

Many behavioral-feedback SIR models do not only deform \(\beta\); they enlarge the state space. In the Poletti-type model analyzed in "Geometric Singular Perturbation Theory Analysis of an Epidemic Model with Spontaneous Human Behavioral Change" [2006.09237], a new variable \(x\) denotes the fraction of susceptibles using normal behavior, while \(1-x\) use altered behavior. Transmission becomes
\[
\beta_{\mathrm{eff}}(x)=\beta_n x+\beta_a(1-x),
\]
and \(x\) evolves by imitation dynamics driven by the payoff difference
\[
p_n-p_a = k-(m_n-m_a)I.
\]
The resulting loop is explicit: \(I\) changes payoffs, payoffs change \(x\), and \(x\) changes effective transmission.

A local-information variant is the \(SS^FIR\) model, where \(S^F\) denotes a behaviorally protected susceptible class. Its reactions are
\[
S+I\xrightarrow{\beta} 2I,\qquad
S+I\xrightarrow{\beta_F} S^F+I,\qquad
S^F+I\xrightarrow{\gamma\beta} 2I,\qquad
I\xrightarrow{\mu} R.
\]
In that construction, infected neighbors transmit both pathogen and risk information: an \(S\) node can become infected directly or adopt self-protective behavior first [1509.01321].

Awareness-explicit SIRS models go further by duplicating every epidemiological class into aware and unaware compartments. In [1702.04999], the state is \((S_n,I_n,R_n,S_a,I_a,R_a)\). Awareness reduces susceptibility by \(\sigma_s\), reduces infectivity by \(\sigma_i\), accelerates recovery by \(\varepsilon r\), and changes immunity waning from \(\delta\) to \(\phi\delta\). Awareness itself spreads through private contact terms proportional to \(\alpha_j(S_a+I_a+R_a)X_n/N\) and public campaign terms \(\omega_j X_n\). The behavioral subsystem is therefore contagion-like rather than purely parametric.

A different explicit behavioral state appears in "Optimal Control of an SIR Model with Noncompliance as a Social Contagion" [2509.09075]. The population is partitioned into compliant \((S,I,R)\) and noncompliant \((S^*,I^*,R^*)\) compartments, and noncompliance spreads socially at rate \((\overline\mu-\mu)N^*\), where \(N^*=S^*+I^*+R^*\). Disease control and behavioral control are separated: \(\alpha\) reduces infectivity for compliant susceptibles, \(\eta\) increases recovery among compliant infecteds, while \(\mu\) and \(\nu\) act on noncompliance contagion and re-compliance.

History dependence can also be embedded directly into the behavioral law. In [2007.04425], vaccination is produced by a Preisach hysteresis operator,
\[
v(t)=v_{nat}+\iint_{\Pi} q(\alpha)\,\bigl({\mathcal R}_{\alpha}[\nu_\alpha^0]I\bigr)(t)\,d\alpha_1d\alpha_2,
\]
so the same prevalence \(I(t)\) can correspond to different vaccination levels depending on whether the epidemic is rising or falling. This yields a continuum of endemic equilibria and, under some conditions, periodic orbits.

## 4. Information structure, delay, spatiality, and networks

Behavioral feedback depends not only on what changes but also on what information drives the change. A clear taxonomy appears in [1107.0997]. Model I uses local prevalence-based information, with fear adoption
\[
S+I \xrightarrow{\beta_F} S^F+I.
\]
Model II uses global prevalence-based information,
\[
\lambda_{S\rightarrow S^F}^{II}=\beta_F\left(1-e^{-\delta I(t)}\right),
\]
and Model III adds belief-based contagion,
\[
S+S^F \xrightarrow{\alpha\beta_F} 2S^F.
\]
These alternatives separate local observation of infecteds, mass-media-like signals, and socially reinforced transmission of concern.

The timing of feedback can be as important as its sign. In "Oscillating behavior of a compartmental model with retarded noisy dynamic infection rate" [2301.12437], mitigation reduces transmission through \(f(j_d)=1+\alpha j_d\), where the delayed prevalence signal is
\[
j_d(t)=\int_{-\infty}^{t}K(t-\tau)j(\tau)\,d\tau.
\]
With fixed delay \(K(t)=\delta(t-\tau_0)\), sufficiently large \(\tau_0\) destabilizes the endemic equilibrium through a Hopf instability and generates persistent regular oscillations. With noise, the same mechanism yields irregular waves.

Spatial microfoundations alter the meaning of behavioral feedback by breaking the mass-action equivalence between prevalence and exposure. In [2102.10145], agents move in a two-dimensional unit square, meet only within contagion radius \(p\), and reduce contacts according to
\[
c=\alpha(I_t)\, d \Psi,\qquad
\alpha(I_t)=
\begin{cases}
1,& I_t\le \underline I,\\
\left(\dfrac{\underline I}{I_t}\right)^{1-\phi},& I_t>\underline I.
\end{cases}
\]
The paper shows that local interactions generate matching frictions and local herd immunity effects, so effective infection pressure is \(\beta_t\lambda(H_t;g)\) rather than a function of \(I_t/N\) alone.

Network formulations generalize the same point. In [2507.03852], behavioral response can be local or non-local, symmetric or asymmetric, rank-1 or full-rank, because all of it is absorbed into the state-dependent interaction matrix \(A(x,y)\). This permits reduced interactions with highly infected groups, voluntary distancing, self-protection, and even fatigue-like feedback through the shape of \(A(x,y)\).

## 5. Dynamical consequences, thresholds, and common misconceptions

The sign of the feedback is a primary determinant of qualitative behavior. In [2308.01038], decreasing \(\beta(I)\) yields a relatively simple threshold structure: typically one disease-free equilibrium and one endemic equilibrium exchanging stability in a transcritical bifurcation. Increasing \(\beta(I)\), by contrast, can generate multiple endemic equilibria, saddle-node bifurcations, hysteresis, Hopf bifurcations, and periodic epidemic waves. The same paper reports that protective behavior and vaccination are partially substitutable, whereas perverse feedback amplifies endemic burden and recurrent waves.

Risk compensation can reverse comparative statics even when the behavioral response is individually protective. In [2004.14533], the high infection risk condition
\[
C\beta \cdot \frac{S(t)I(t)}{S(t)+I(t)} > 2a
\]
implies that more current prevalence can reduce new infections because the induced drop in activity more than offsets the direct mechanical effect of more infecteds. Under the same condition, lowering transmissibility \(\beta\) can marginally increase current infection flow, and lowering infection cost \(C\) can raise short-run health losses. Those reversals are local and state dependent, not generic claims about the entire epidemic trajectory.

Several BF-SIR subclasses retain a unimodal infection curve despite the added feedback. The rank-1 network model of [2507.03852] proves unimodality for the weighted aggregate
\[
\bar y(t)=\sum_{j=1}^n f_j(y_j(t))y_j(t)
\]
under positivity, monotonicity, and concavity assumptions. The scalar BF-SIR model of [2509.12257] likewise proves that \(y(t)\) is either strictly decreasing or single-peaked, because \(R(x,y)\) is increasing in \(x\) and non-increasing in \(y\). These results do not preclude multimodality at node level or under different feedback structures; they identify subclasses where aggregate one-peak behavior survives.

A recurring misconception concerns the word “feedback” itself. "Spurious self-feedback of mean-field predictions inflates infection curves" [2312.14851] does not study adaptive behavior. Its “feedback” is a closure artifact created by mean-field factorization in network SIR and SIRS, where paths of the form \(i\to j\to i\) spuriously re-enter the marginal infection hazard of node \(i\). The paper’s second-order Plefka/TAP correction subtracts that artifact. This is conceptually important for behavioral-feedback SIR modeling because mean-field closures can manufacture amplification loops that are distinct from genuine endogenous behavioral feedback.

## 6. Control, intervention, and normative benchmarks

Behavioral-feedback SIR models have also become a control-theoretic platform. In the controlled BF-SIR model of [2512.08544],
\[
\dot x=-(1-u)\beta(x,y)xy,\qquad
\dot y=(1-u)\beta(x,y)xy-\gamma y,
\]
with
\[
R(x,y)=\frac{x\beta(x,y)}{\gamma},
\]
the planner minimizes
\[
\int_0^\infty u(t)\,dt
\]
subject to \(y(t)\le \bar y\). Under the monotonicity assumptions \(x\beta_x(x,y)+\beta(x,y)>0\) and \(\beta_y(x,y)\le 0\), the unique optimal policy is the filling-the-box strategy: no intervention while \(y<\bar y\), then the minimal control
\[
u=\rho(x,\bar y)=1-\frac{\gamma}{x\beta(x,\bar y)}
\]
that keeps \(\dot y=0\) on the threshold, and finally no control once \(R(x,\bar y)\le 1\).

The related model in [2509.12257] derives the same threshold-hugging feedback for
\[
\beta(x,y)=\frac{b(x)}{1+a y},
\]
namely
\[
u=\rho(x,\bar y)=1-\frac{\gamma(1+a\bar y)}{x b(x)},
\]
and computes the exact cost of that feasible strategy as
\[
J(u^*|x_0,y_0)=
\begin{cases}
0,& (x_0,y_0)\in D_{\bar y}^{-},\\[4pt]
\dfrac{1}{\gamma \bar y}\int_{\bar x}^{h(x_0,y_0)}\rho(x,\bar y)\,dx,& (x_0,y_0)\in D_{\bar y}^{+}.
\end{cases}
\]
However, that paper explicitly leaves global optimality open.

A classical benchmark for these results is the peak-minimization problem in the controlled SIR model under an \(L^1\) budget constraint [2203.05800]. There the explicit optimal feedback is null–singular–null: no intervention until \(I\) reaches a target plateau, then
\[
u=1-\frac{1}{\mathcal R_0 S}
\]
to keep \(\mathcal R_e=1\), and no intervention after \(S\) reaches the herd-immunity threshold. This is not itself a behavioral model, but it provides a normative reference for prevalence-triggered transmission attenuation.

Control can also act on the behavioral subsystem itself. In [2509.09075], the control vector \((\alpha,\eta,\mu,\nu)\) simultaneously targets disease prevention, treatment, slowdown of noncompliance contagion, and re-compliance campaigns. In [2308.01038], vaccination and protocol adherence jointly determine whether the system converges to disease-free equilibrium, an endemic state, or complex positive-feedback dynamics. Taken together, these models suggest that behavioral-feedback SIR theory has evolved from a descriptive modification of incidence terms into a framework for state-constrained intervention design, threshold management, and the analysis of how voluntary adaptation interacts with formal control.

Source: https://www.emergentmind.com/topics/behavioral-feedback-sir-model