Why Bistability Requires More Than a Suggestive Trace
A steep signaling response, two apparent response levels, or a long-lived state can each look like evidence for bistability. None is sufficient by itself. A steep but continuous input-output curve can arise in a system with one equilibrium at every input level. Two different measurements at the same nominal input can reflect transient lag, unmeasured changes in the system, or noise. And a response that changes very slowly may still be relaxing toward a single equilibrium.
The central diagnostic question is therefore not whether a plot looks switch-like. It is whether, under the same fixed conditions, the system has at least two distinct attracting steady states, and whether perturbations provide evidence for their separate basins of attraction. Experiments can support that claim, but finite observations cannot establish it merely by showing two plateaus.
This definition concerns the dynamics at fixed conditions. It is stronger than observing two different responses during an input sweep. In many bistable systems an unstable steady state separates the basins of attraction of the stable states, but that pattern is not part of the definition in every possible model. A bistable system may also have additional equilibria or more complicated dynamics.
What a Stability Calculation Can and Cannot Show
For a one-state model, local stability can often be assessed by examining the derivative of the rate function at an equilibrium. Let a small displacement from \(x^*\) be \(\delta=x-x^*\). A Taylor expansion gives
When \(f_x(x^*,u)<0\), the linear term drives small displacements back toward zero; if the equilibrium is hyperbolic (the derivative is nonzero), this establishes local asymptotic stability. When \(f_x(x^*,u)>0\), small displacements grow in at least one direction, so the equilibrium is unstable. If \(f_x(x^*,u)=0\), this linear test is inconclusive and higher-order terms must be examined. For a multidimensional system, the corresponding local test uses the Jacobian matrix: if all eigenvalues have negative real parts, the equilibrium is locally asymptotically stable; an eigenvalue with positive real part indicates instability. Zero real parts require further analysis.
A stability calculation for a proposed model is not the same as evidence that the model describes the experiment. In data, researchers estimate whether measured states have settled, whether the condition truly remained fixed, and whether perturbations lead back to the same state. A finite observation can only provide evidence over its measured time window and resolution.
Worked Example: A Steep Response With Only One Equilibrium
Suppose a hypothetical pathway has normalized activity \(x\) that follows
where \(u\) is a fixed, dimensionless input. The fourth power can make the steady-state response steep around \(u=1\). To check whether steepness implies bistability, set the rate to zero:
For each fixed \(u\), this equation gives exactly one steady state. At \(u=1\), the equilibrium is \(x^*=1/(1+1)=0.500\). The derivative of the rate with respect to \(x\) is \(f_x=-1\), so this equilibrium is locally asymptotically stable. Thus, even if the measured input-output curve looks switch-like, this model has one attracting steady state at each fixed input. Steepness is not a count of equilibria.
Why Apparent Hysteresis May Be a Transient
An up-sweep and a down-sweep may produce different measured responses at the same input even in a monostable system. If the input changes faster than the state can track its equilibrium, the state lags behind. The resulting loop depends on sweep rate and can shrink as the experiment is slowed or as measurements are taken after longer settling periods. This differs from equilibrium hysteresis, which reflects distinct stable states available under the same fixed conditions over an interval of input values.
Worked Example: A Sweep-Rate Loop From a Single Equilibrium
Consider a hypothetical first-order response \( \dot{x}=(u-x)/\tau \), where \(u\) is the input and \( \tau=10 \) minutes. At every fixed \(u\), the only steady state is \(x^*=u\), and the derivative with respect to \(x\) is \(-1/\tau=-0.100\) per minute, so that equilibrium is locally asymptotically stable.
During an upward ramp \(u=vt\), starting at \(x=0\) when \(u=0\), the solution is
Take a hypothetical ramp rate \(v=0.0100\) input units per minute. After several time constants, the exponential term is small, and the upward-sweep response is approximately \(x_{\mathrm{up}}\approx u-v\tau=u-0.100\). On the downward ramp, after its transient has similarly diminished, \(x_{\mathrm{down}}\approx u+v\tau=u+0.100\). At a common input such as \(u=0.500\), the responses are therefore approximately \(0.400\) and \(0.600\), a difference of \(0.200\). Yet if the input is held at \(0.500\), both trajectories converge to \(x=0.500\). The apparent loop is caused by lag, not by two steady states.
A useful diagnostic is to repeat the sweep at slower rates and to hold the input at selected values until the response settles. A loop that narrows with slower sweeps or disappears during fixed-input holds is not, by itself, evidence for equilibrium bistability.
Slow Relaxation Is Not a Second Stable State
A response can appear nearly flat during an experiment because its relaxation time is long, not because it has reached an alternative equilibrium. For a stable first-order deviation \(z\) from the unique equilibrium, \( \dot{z}=-kz \), the residual deviation is \(z(t)=z(0)e^{-kt}\). A small rate constant \(k\) makes this decay difficult to detect over a short observation window. The distinction between Lyapunov stability and local asymptotic stability matters here: an observed state that stays close for a while has not thereby been shown to be an equilibrium, or even to be locally asymptotically stable.
Worked Example: A Long-Lived Transient
Suppose a hypothetical pathway relaxes toward one equilibrium according to \( \dot{x}=-0.0100(x-x^*) \), with time in minutes. Starting \(0.800\) activity units above \(x^*\), its remaining displacement after 30 minutes is
A measurement after 30 minutes could therefore show a substantial, apparently persistent elevation. After 300 minutes, however, the displacement is \(0.800e^{-3.00}=0.0398\) activity units. This trajectory is a slow transient toward a single equilibrium, not proof of a second stable state. Longer observation, repeated measurements, and perturbation tests can help distinguish these explanations; they cannot make an insufficiently long experiment conclusive.
Steady-State and Perturbation Tests
A strong experimental case for bistability combines steady-state evidence with controlled perturbations. The conditions must be matched: input, cell state, measurement protocol, and other relevant environmental variables should be held constant or measured. Researchers should approach the same fixed input from distinct initial states, allow sufficient time for relaxation, and test whether each putative state persists. They should then perturb each state by small amounts in both directions. For locally asymptotically stable states, small perturbations should decay back toward the corresponding state. Larger perturbations may cross a basin boundary and take the system to another attractor.
These tests support a dynamical interpretation; they do not prove that every alternative mechanism has been excluded. For instance, unmeasured slow changes in receptor abundance or cell condition can make two groups appear to have different equilibria. Population averages can also conceal cells in different states, while noise can cause transitions between states. The measurement unit and variability therefore matter when interpreting apparent plateaus.
Worked Example: Testing Two Candidate Attractors at Fixed Input
Consider a hypothetical normalized activity \(x\), held between 0 and 1, with a fixed input \(u_0\). Its dynamics are
The steady states are \(x^*=0\), \(0.400\), and \(1\). The derivative of the rate function at an equilibrium is
Thus \(f_x(0)=-0.400\), \(f_x(0.400)=0.240\), and \(f_x(1)=-0.600\). The two endpoint equilibria are locally asymptotically stable, while the middle equilibrium is unstable. For example, at \(x=0.0200\), \(x(1-x)(x-0.400)<0\), so activity moves downward toward 0. At \(x=0.980\), the rate is positive, so activity moves upward toward 1. Small displacements from either endpoint still move back toward that endpoint.
In an experiment, the corresponding evidence would be two distinct steady responses at the same fixed \(u_0\), reached from different initial conditions, with small perturbations around each response returning toward it. Merely measuring one low response and one high response at different inputs would not demonstrate this property. Nor would two short-lived plateaus without evidence that they persist under fixed conditions.
Comparing Evidence and Recognizing Failure Modes
| Observation or test | What it can support | What it cannot establish alone |
|---|---|---|
| Steep input-output curve | Strong dependence of response on input | More than one steady state at a fixed input |
| Different responses on up- and down-sweeps | Path dependence during the protocol | Equilibrium hysteresis; lag or drift may explain the loop |
| Long-lived response after a change | Slow dynamics over the observation window | A second equilibrium or local asymptotic stability |
| Distinct fixed-input plateaus and return after small perturbations | Evidence for separate attracting states under those conditions | Exclusion of all unmeasured slow processes or noise effects |
A frequent error is to reason, “The response stayed high for a long time, so the high state is stable.” The correction is that persistence over a finite interval only bounds how much change was observed in that interval. It does not show that the system has reached a steady state or that nearby perturbations return to it. Likewise, a loop is not an equilibrium property until transient lag, changing conditions, and other plausible sources of path dependence have been considered.
The main alternatives answer different questions. A response curve describes how measured output varies with input, but does not by itself identify the number or stability of steady states. A time-course after a perturbation reveals relaxation and possible transitions, but its interpretation depends on the duration and conditions of observation. A model-based equilibrium and stability analysis can count and classify steady states for specified equations and parameters, but its conclusions are conditional on that model. The strongest support combines model analysis with fixed-input measurements and perturbation experiments.
When data are noisy, do not require an exactly constant measured value to identify a candidate steady state. Instead, prespecify a practical criterion for settling, account for measurement variability, and check that the conclusion is robust to the observation window and perturbation size. Conversely, do not label random switching between high and low observations as deterministic bistability without checking whether noise, cell-to-cell heterogeneity, or changing conditions could explain it. Noise-driven switching can occur in a system with two attractors, but switching observations alone do not establish that deterministic structure.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A system has one steady state at each fixed input, but its response curve is very steep. Does that establish bistability? Why?
- In a scalar model, an equilibrium has \(f_x(x^*)=0\). What can the linear stability test conclude?
- An up-sweep and down-sweep differ at the same input, but the responses converge to one value after a long hold. What is the most direct interpretation?
- State the distinction between Lyapunov stability and local asymptotic stability.
- What experimental pattern would provide stronger evidence for bistability than two different responses observed at different inputs?
Solutions. 1. No. Steepness describes sensitivity to input; bistability requires at least two locally asymptotically stable steady states at the same fixed input. 2. The linear test is inconclusive because its first-order term vanishes; higher-order terms or another analysis are needed. 3. The sweep difference is consistent with transient lag or another protocol-dependent effect, rather than two equilibrium responses at that input. 4. Lyapunov stability requires nearby trajectories to remain nearby; local asymptotic stability additionally requires them to converge back to the equilibrium. 5. Hold conditions fixed, approach from distinct initial states, demonstrate distinct persistent steady responses, and show small perturbations around each response decay back toward it.