Equilibrium Gain and Stability in Homeostatic Feedback

Why Feedback Changes the Sustained Drug Response

A drug may shift a biomarker away from its untreated baseline, while homeostatic feedback acts to oppose that shift. The sustained response is therefore not determined by the drug input alone: it also depends on how strongly the system restores the biomarker toward baseline. A useful first model separates these contributions and makes their effects on the equilibrium explicit.

Let \(x(t)\) be the biomarker deviation from its untreated baseline, measured in biomarker units. Suppose the drug supplies a constant perturbing input \(D\), while passive restoration and instantaneous homeostatic feedback each act in proportion to \(x\). The model is

\frac{dx}{dt}=D-kx-gx=D-(k+g)x.

Here \(D\) has units of biomarker units per time, and \(k\) and \(g\) have units of inverse time. Assume initially that \(k>0\), \(g\geq 0\), and \(D\) is constant. The term \(-kx\) is passive turnover or restoration; \(-gx\) is negative feedback. When \(x>0\), feedback lowers the rate of change, and when \(x<0\), it raises the rate of change toward baseline.

Definition: An equilibrium \(x^*\) is a value at which \(dx/dt=0\) under the fixed input \(D\). It is locally asymptotically stable if sufficiently small perturbations from it decay back toward it over time. In this linear model, the perturbation dynamics can be solved exactly, so the stability conclusion applies to any initial displacement, not only a small one.

Equilibrium, Gain, and Stability

At equilibrium, the drug input is balanced by passive restoration and feedback. If \(k+g\ne 0\), setting the rate of change to zero gives a unique equilibrium. The stability result follows by subtracting this equilibrium from the dynamic equation.

0=D-(k+g)x^* \quad\Longrightarrow\quad x^*=\frac{D}{k+g}, \qquad e(t)=x(t)-x^*, \qquad \frac{de}{dt}=-(k+g)e.

The error equation has solution \(e(t)=e(0)\exp[-(k+g)t]\). Thus, if \(k+g>0\), every initial error decays exponentially and the equilibrium is asymptotically stable. Its relaxation time is \(1/(k+g)\): increasing the total restoration rate makes the return to equilibrium faster. If \(k+g<0\), the equilibrium exists but perturbations grow exponentially, so it is unstable.

The boundary \(k+g=0\) must be handled separately. In that case the model reduces to \(dx/dt=D\). If \(D\ne0\), there is no equilibrium. If \(D=0\), every value of \(x\) is an equilibrium and a displacement remains constant; perturbations do not decay, so none of these equilibria is asymptotically stable.

Key result: For \(k+g\ne0\), the unique equilibrium is \(x^*=D/(k+g)\), and it is asymptotically stable exactly when \(k+g>0\). At \(k+g=0\), an equilibrium exists only if \(D=0\); then perturbations remain constant rather than decaying.

To quantify attenuation, compare the equilibrium with the same constant input when there is no feedback (\(g=0\)). For \(k>0\), the no-feedback response is \(x^*_0=D/k\). With \(g\geq0\), the ratio of the signed responses is

\frac{x^*}{x^*_0} = \frac{D/(k+g)}{D/k} = \frac{k}{k+g} \qquad (D\ne0).

The ratio is between zero and one under the stated assumptions. Feedback therefore reduces the magnitude of the sustained response by the fraction \(1-k/(k+g)=g/(k+g)\), relative to this no-feedback model. The result concerns the equilibrium response to a constant input; it does not say that feedback changes the drug input itself.

Worked Example: Calculating Attenuation and Approach to Equilibrium

Suppose a hypothetical biomarker model has a constant drug input \(D=12\) biomarker units per hour, passive restoration rate \(k=0.40\ \text{h}^{-1}\), and feedback gain \(g=0.60\ \text{h}^{-1}\). Assume the biomarker begins at baseline, so \(x(0)=0\). Find the equilibrium, the no-feedback equilibrium, and the response after two hours.

The total restoration rate is \(k+g=0.40+0.60=1.00\ \text{h}^{-1}\), which is positive, so the equilibrium is stable. Substitution gives

x^*=\frac{12\ \text{units}/\text{h}}{1.00\ \text{h}^{-1}} =12.0\ \text{units}.

Without feedback, the equilibrium would be \(x^*_0=D/k\):

x^*_0=\frac{12\ \text{units}/\text{h}}{0.40\ \text{h}^{-1}} =30.0\ \text{units}, \qquad \frac{x^*}{x^*_0}=\frac{0.40}{1.00}=0.400.

Thus feedback reduces the sustained response by \(1-0.400=0.600\), or 60.0%, compared with the no-feedback equilibrium. Since \(x(0)=0\), the exact solution is \(x(t)=x^*[1-\exp(-(k+g)t)]\). At two hours,

x(2)=12.0[1-\exp(-1.00\times2)] =12.0(1-0.1353) =10.38\ \text{units}.

This value is below the equilibrium, as expected for an increasing response starting at zero, and the remaining error is \(12.0-10.38=1.62\) units.

Interpreting Gain and Checking the Model

The ratio \(k/(k+g)\) shows why the absolute size of \(g\) is not sufficient to describe attenuation: feedback acts relative to passive restoration. For fixed \(k\), larger nonnegative \(g\) lowers the equilibrium magnitude and increases the rate of return. At \(g=k\), the response is half the no-feedback response; if \(g\) is much smaller than \(k\), attenuation is modest.

This equilibrium comparison relies on specific assumptions: the input is constant long enough for equilibrium to be relevant; the deviation is represented by one state variable; restoration and feedback are linear in \(x\); and feedback is instantaneous. The model also treats \(D\), \(k\), and \(g\) as fixed. If the input varies over time, \(D\) cannot be replaced by one constant when calculating a single equilibrium response.

Worked Example: Inferring Gain From a Hypothetical Equilibrium

In a hypothetical experiment, a constant input of \(D=8.0\) biomarker units per day is maintained, passive restoration is known to be \(k=0.50\ \text{day}^{-1}\), and the measured equilibrium deviation is \(x^*=8.0\) units. Under the linear model, calculate the implied feedback gain and the relative attenuation.

The equilibrium relation is \(x^*=D/(k+g)\). Rearranging gives \(g=D/x^*-k\). Substituting the given values,

g=\frac{8.0\ \text{units}/\text{day}}{8.0\ \text{units}} -0.50\ \text{day}^{-1} =1.00-0.50 =0.50\ \text{day}^{-1}.

The total restoration rate is \(k+g=1.00\ \text{day}^{-1}>0\), so the equilibrium is stable. The no-feedback equilibrium is \(D/k=(8.0\ \text{units}/\text{day})/(0.50\ \text{day}^{-1})=16.0\) units. The observed equilibrium is half of that value, consistent with \(k/(k+g)=0.50/1.00=0.500\). This inference depends on \(k\) being known and the model being appropriate; an equilibrium measurement alone does not establish that the response was generated by feedback.

A more general instantaneous negative-feedback term can be written as \(-h(x)\), giving \(dx/dt=D-kx-h(x)\). If \(x^*\) is an equilibrium and \(h\) is differentiable there, a small perturbation \(e=x-x^*\) obeys approximately \(de/dt=-(k+h'(x^*))e\). Therefore \(k+h'(x^*)>0\) is the local stability condition. This is a local result: it uses the slope of the feedback function at the operating equilibrium, not its value across the entire response range. The linear model is the special case \(h(x)=gx\), for which \(h'(x^*)=g\).

Worked Example: The Zero-Rate Boundary and an Unstable Equilibrium

Consider the same equation \(dx/dt=D-(k+g)x\), but now allow \(g\) to be negative as a mathematical representation of feedback opposing restoration rather than reinforcing it. Let \(k=0.30\ \text{h}^{-1}\) and \(g=-0.30\ \text{h}^{-1}\). Examine the cases \(D=0\) and \(D=2.0\) units per hour, then consider \(g=-0.50\ \text{h}^{-1}\) with \(D=1.0\) unit per hour.

For the first two cases, \(k+g=0\), so \(dx/dt=D\). When \(D=0\), \(dx/dt=0\) for every \(x\): every state is an equilibrium and a perturbation stays at its displaced value. It does not decay. When \(D=2.0\ \text{units}/\text{h}\), \(x(t)=x(0)+(2.0\ \text{units}/\text{h})t\), so no equilibrium exists.

In the last case, \(k+g=0.30-0.50=-0.20\ \text{h}^{-1}\), which is nonzero. The equilibrium exists and is

x^*=\frac{1.0\ \text{unit}/\text{h}}{-0.20\ \text{h}^{-1}} =-5.00\ \text{units}.

The perturbation evolves as \(e(t)=e(0)\exp[0.20t]\), with \(t\) measured in hours. Its magnitude grows, so the equilibrium is unstable. The negative gain values in this example are outside the assumed homeostatic negative-feedback case \(g\geq0\); they illustrate why equilibrium existence and stability must be checked separately.

Diagnostics, Alternatives, and Limitations

A model should be questioned when its predicted rate of return or equilibrium is inconsistent with observations. Under constant input, the linear model predicts a single stable equilibrium for \(k+g>0\), with deviations decaying exponentially at rate \(k+g\). Persistent drift after the input has been held constant may indicate that equilibrium has not been reached, that the input or parameters are changing, or that the one-state linear model is inadequate. A return that is not approximately exponential can also signal nonlinear feedback or additional response processes.

A particularly important limitation is parameter confounding. If only \(D\) and the equilibrium \(x^*\) are observed, the equation identifies the sum \(k+g=D/x^*\), not \(k\) and \(g\) separately. Independent information on passive restoration, or informative time-course data together with a justified model, is needed to separate their contributions. Even time-course data do not guarantee identifiability if the sampling design or model is inadequate.

Approach	What it represents	Main trade-off
No-feedback model	Sets \(g=0\), so \(x^*=D/k\)	Simple reference, but cannot describe attenuation attributed to feedback
Linear instantaneous feedback	Adds \(-gx\) to the rate equation	Gives a transparent gain and stability condition, but assumes proportional, immediate feedback
Nonlinear instantaneous feedback	Uses a general opposing function \(h(x)\)	Can represent response-dependent gain, but requires its form and local slope to be supported by data

The linear model is useful because it makes equilibrium gain and stability easy to interpret, not because biological feedback must be linear. It does not by itself represent saturation, a changing drug input, multiple interacting states, or a lag between deviation and feedback. Those features require a different model structure, and the simple exponential stability result should not be assumed to apply unchanged.

Common Mistakes

Confusing a smaller equilibrium with a slower response. The equilibrium magnitude is \(D/(k+g)\), while the return rate is \(k+g\). With positive rates, stronger feedback both attenuates the equilibrium and speeds exponential relaxation in this model.
Calling every zero-net-rate case marginally stable. If \(k+g=0\) and \(D\ne0\), no equilibrium exists. Only when \(D=0\) are all states equilibria, with perturbations remaining constant.
Interpreting gain as a probability or percentage by itself. \(g\) has units of inverse time. Attenuation relative to no feedback is the dimensionless ratio \(g/(k+g)\), under \(k>0\) and \(g\geq0\).
Claiming that an equilibrium identifies feedback. The equilibrium identifies \(k+g\) only when \(D\) is known. Separate evidence is needed to attribute part of that total to feedback.

Check Your Understanding

Answer each question using the linear model \(dx/dt=D-(k+g)x\). Brief solutions follow the questions.

For \(D=6.0\) units per hour, \(k=0.30\ \text{h}^{-1}\), and \(g=0.70\ \text{h}^{-1}\), find the stable equilibrium and the attenuation relative to no feedback.
If \(k+g=-0.40\ \text{day}^{-1}\) and \(D\ne0\), does an equilibrium exist, and is it stable?
What happens when \(k+g=0\) and \(D=0\)?
With constant \(D\), \(k\), and \(g\), what combination of \(k\) and \(g\) is identified by one equilibrium measurement?

Solutions. (1) \(x^*=6.0/(0.30+0.70)=6.00\) units. The no-feedback value is \(6.0/0.30=20.0\) units, so the response ratio is \(0.30/1.00=0.300\), a 70.0% attenuation. Since \(k+g=1.00\ \text{h}^{-1}>0\), it is stable. (2) Because \(k+g\ne0\), an equilibrium \(D/(k+g)\) exists, but the negative net rate makes perturbations grow, so it is unstable. (3) Every \(x\) is an equilibrium and displacements remain constant; there is no asymptotic return. (4) The equilibrium identifies the sum \(k+g=D/x^*\), not the two rates separately.