Why Delay Changes a Feedback Response
Homeostatic feedback counteracts a drug-induced change, but the compensatory response may depend on a signal that reflects the system’s state some time earlier. During that lag, the drug’s effect can continue to develop even though compensation is already being generated. The result may be an overshoot followed by oscillation. Whether those oscillations decay or grow depends on both feedback strength and delay.
A useful starting point is a simple model for a compensatory state \(x(t)\). Suppose a persistent perturbation drives the state at rate \(u\), while a delayed increase in \(x\) suppresses that drive:
Here \(u>0\) has units of state per unit time, \(b>0\) has units of inverse state, and \(\tau\geq0\) is the feedback delay in units of time. For a step perturbation, use the history \(x(t)=0\) for \(t\leq0\). The equilibrium is \(x^*=1/b\): at that value the delayed feedback exactly balances the constant drive. This is a deliberately simplified model of compensation, not a complete physiological description.
The equation for \(y(t)\) describes how perturbations around equilibrium evolve. If they decay, the equilibrium is locally asymptotically stable. If perturbations grow, the equilibrium is unstable. This stability question is distinct from whether a particular step response temporarily exceeds its equilibrium.
Stability and the Delay Threshold
To determine stability, consider an exponential deviation \(y(t)=e^{\lambda t}\), where \(\lambda\) may be complex. Substitution into \(y'(t)=-a y(t-\tau)\) gives the characteristic equation
A characteristic root with negative real part represents a decaying mode; a root with positive real part represents a growing mode. The stability boundary occurs when a root lies on the imaginary axis. Write \(\lambda=i\omega\), with \(\omega>0\). Separating real and imaginary parts gives
Because \(a>0\), the first equation requires \(\cos(\omega\tau)=0\). The second requires \(\sin(\omega\tau)=\omega/a\). Since \(|\sin(\omega\tau)|\leq1\), the first crossing has \(\omega=a\), \(\sin(\omega\tau)=1\), and \(a\tau=\pi/2\). The subsequent imaginary-axis crossings occur at \(a\tau=\pi/2+2k\pi\), for nonnegative integers \(k\).
At zero delay, the characteristic equation reduces to \(\lambda=-a\), which is stable. At the first crossing, the root moves into the right half-plane as the delay increases. To check the direction, implicitly differentiate the characteristic equation \(F(\lambda,\tau)=\lambda+a e^{-\lambda\tau}=0\):
At the first crossing, \(\lambda=ia\), \(a\tau=\pi/2\), and \(e^{-i\pi/2}=-i\), so
Thus the crossing is from negative to positive real part. The standard stability result for this scalar delayed-feedback equation is:
“Unstable” means that small deviations grow in this linear model; it does not mean that every observed trajectory must settle into a sustained cycle. Nonlinear physiology, bounds on the state, and other processes can change what happens after deviations become large. Conversely, when the equilibrium is stable, the trajectory can still overshoot and oscillate transiently before settling.
Worked Example: Checking Whether Feedback Is Stable
Suppose a hypothetical compensatory state has \(u=0.80\) state units per hour, \(b=0.050\) per state unit, and delay \(\tau=20\) hours. Determine whether small deviations around equilibrium decay.
First calculate the feedback gain:
The dimensionless gain-delay product is
Because \(0.800<\pi/2\approx1.571\), the equilibrium is asymptotically stable. Equivalently, the critical delay at this gain is
A delay of 20 hours is below that threshold. This conclusion concerns the long-run stability of small deviations; it does not rule out a transient peak above equilibrium.
Finding a Transient Peak Without Extrapolating a Segment
For a step perturbation with \(x(t)=0\) before time zero, the delayed equation can be solved successively on intervals of length \(\tau\). On the first interval, \(0\leq t\leq\tau\), the delayed state is zero, so \(x'(t)=u\) and \(x(t)=ut\). On the next interval, \(\tau\leq t\leq2\tau\), the delayed state is \(u(t-\tau)\). Integrating the equation on that interval yields
Within this interval, the slope is \(x'(t)=u[1-bu(t-\tau)]\). Its candidate zero is \(t=\tau+1/(bu)\), but that candidate belongs to this segment only if \(1/(bu)\leq\tau\), or \(bu\tau\geq1\). If \(bu\tau<1\), the slope remains positive through \(2\tau\). The response must then be continued using the next delayed interval before its first peak can be located.
Worked Example: A Peak That Occurs After the Second Delay Interval
Consider the hypothetical step response \(u=1.00\) state unit per hour, \(b=0.100\) per state unit, and \(\tau=8.00\) hours, with \(x(t)=0\) for \(t\leq0\). Find the first peak rather than extending the second-interval formula beyond its validity.
For \(0\leq t\leq8\), \(x(t)=t\), so \(x(8)=8.00\) state units. For \(8\leq t\leq16\), the segment formula becomes
The candidate slope zero on this segment would be at \(t=8+1/(bu)=18.0\) hours. But \(bu\tau=(0.100)(1.00)(8.00)=0.800<1\), so that time is outside the segment, which ends at 16 hours. At 16 hours the state is
For the next interval, write \(t=16+s\), where \(0\leq s\leq8\). Then \(t-\tau=8+s\), which lies in the second interval, and
The first zero of this slope is the smaller root:
This value is within \(0\leq s\leq8\), so the first peak occurs at \(t\approx18.254\) hours. Integrating the slope from 16 hours to that peak gives
The equilibrium is \(x^*=1/b=10.0\) state units, so this first peak exceeds equilibrium by about \(3.016\) state units. The gain-delay product is \(ub\tau=0.800<\pi/2\), so this overshoot is compatible with a stable equilibrium: the transient peak does not imply growing oscillations.
Comparing Delayed Feedback With Simpler Alternatives
The explicit-delay model isolates the effect of a lag. Its predictions should be compared with models that represent the same system in other ways, because an observed oscillation alone does not identify the mechanism.
| Model | What it represents | Implication for oscillation |
|---|---|---|
| Instantaneous negative feedback | Feedback responds to the current state, with no delay. | In this first-order model, deviations decay monotonically; delay-driven oscillation is absent. |
| Explicit-delay feedback | Feedback depends on the state \(\tau\) time units earlier. | Can produce overshoot; stability changes at \(a\tau=\pi/2\). |
| Transit-compartment approximation | A finite chain of states represents a distributed lag. | Can produce phase lag and overshoot, but its dynamics are not identical to a single fixed delay. |
An instantaneous model is a useful baseline: if it cannot reproduce the timing and shape of the response, a lagged mechanism may be worth considering. A transit-compartment approximation is useful when a single fixed lag is implausible or when the delay is spread across several processes. The two representations should not be treated as interchangeable: their stability properties and predicted response shapes can differ.
The delayed model has important limitations. A single \(\tau\) compresses potentially diverse biological processes into one lag. The linear feedback term does not impose physiological bounds, and the constant drive \(u\) may be unrealistic over long periods. The model also predicts the equilibrium \(1/b\) independently of \(u\), a consequence of this particular formulation rather than a general feature of homeostasis. These assumptions should be checked against the biological system and the observed time course.
Diagnostics and Common Mistakes
A model is less convincing when it predicts the wrong timing or direction of successive peaks, when residuals show repeating structure, or when fitted delay and gain change substantially under modest changes in the observation window. Sparse sampling can miss peaks and make an apparent delay difficult to distinguish from a slower response. Estimating \(a\) and \(\tau\) separately may also be difficult when the data mainly constrain their product \(a\tau\); informative observations around the onset and turning points are particularly valuable.
- Confusing transient overshoot with instability. A stable system can temporarily exceed equilibrium. Stability concerns the long-run response to small deviations, not the height of one peak.
- Using a segment outside its interval. The formula for \(\tau\leq t\leq2\tau\) cannot be used to locate a peak after \(2\tau\). Continue the delay equation interval by interval.
- Calling every oscillation delay-driven. Several coupled processes or a time-varying input can also produce oscillatory patterns. Compare plausible alternatives and inspect whether the model reproduces the timing of the observations.
- Interpreting an unstable linear model as a sustained biological cycle. Linear instability predicts growing deviations near equilibrium; it does not specify the eventual bounded behavior.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For \(a=0.0300\ \text{h}^{-1}\) and \(\tau=40.0\ \text{h}\), is the equilibrium asymptotically stable?
- If \(a=0.0500\ \text{h}^{-1}\), what is the critical delay at the first stability boundary?
- In the step-response model, why is the candidate peak \(t=\tau+1/(bu)\) not valid on the interval \([\tau,2\tau]\) when \(bu\tau<1\)?
- Does a first peak above \(x^*\) prove that feedback is unstable? Explain.
- At the first stability boundary, what is the oscillation period when \(a=0.0400\ \text{h}^{-1}\)?
Solutions. (1) \(a\tau=(0.0300)(40.0)=1.20<\pi/2\), so the equilibrium is asymptotically stable. (2) \(\tau_{\mathrm{crit}}=\pi/(2a)=1.571/(0.0500\ \text{h}^{-1})=31.42\ \text{h}\). (3) The candidate requires \(1/(bu)\leq\tau\), equivalent to \(bu\tau\geq1\); if \(bu\tau<1\), it occurs after \(2\tau\), so the delayed state must be evaluated on a later interval. (4) No. Overshoot can occur in a stable transient; instability requires growing deviations near equilibrium. (5) At the boundary, \(\omega=a\), so the period is \(2\pi/a=2\pi/(0.0400\ \text{h}^{-1})\approx157.1\ \text{h}\).