How a dynamical claim earns its evidence

Governors, chemical networks and synthetic gene switches were built to solve different problems. Their intersecting histories explain why a good feedback sign, a stable matrix and a functioning memory are different achievements.

CCBS · EST 5053 · Lecture 4 lineage · lecturer companion
Prepared with Codex · 15 September 2026 · read beside the core and exposition

A diagram can show a sensible correction while the machine it describes oscillates itself out of useful operation. A mathematical model can contain two stable states while a constructed circuit realizes only one. A population can remain stationary while every individual molecule continues to enter and leave.

Each mismatch changes what counts as evidence. The history behind Lecture 4 is not a procession of classification techniques. It is a series of attempts to make a stronger claim than the available picture initially justified.

We will follow three intersecting traditions. Mechanical regulation demanded recovery after disturbances. Chemical kinetics added conservation and structure to the state space. Mathematical biology and synthetic circuits asked which structures permit memory and what measurements demonstrate it. The connections below are a retrospective comparison of problems, not a claim that every later author directly inherited a method from the preceding one.

1The intended correction is not the motion

James Clerk Maxwell's 1868 paper On Governors begins with a practical task: keep a machine's speed nearly uniform when driving power or resistance changes. A centrifugal mechanism responds to excess speed by increasing resistance or restricting supply. That description gives the intended feedback sign, not the resulting motion.¹

The relevant observation was that adjustment could improve regulation and then produce growing oscillation. Maxwell describes Fleeming Jenkin's adjustable governor developing a dancing motion and a jerking main shaft. The machine was responding to its error, but its response did not die away.¹

This separates two design questions. Does the mechanism oppose the disturbance? Does the coupled system return to operation after the disturbance? A wiring sketch can answer the first while leaving the second entirely open.

Separate the sign of an intended correction from the stability of the machine that implements it.

2Recovery becomes a root condition

Maxwell classified disturbances by whether they grew or decayed, with or without oscillation. He translated acceptable recovery into a condition on characteristic roots. He also stated the limits of his result: he had not completely determined the conditions beyond degree three.¹

In modern notation, a linear displacement model is $\dot z=Az$. If $Aw=sw$, the displacement along that mode is proportional to $e^{st}w$. The real part of $s$ determines whether its envelope grows or decays. The imaginary part determines rotation.

A Hurwitz matrix has every root strictly in the left half-plane. The condition tests all linear displacement modes, including those a designer did not intend to excite. It therefore answers a family of initial-condition questions rather than describing one successful run.

Read characteristic roots as disturbance modes, and distinguish a test covering all small displacements from one observed recovery trace.

3More feedback can make things worse

A three-state illustration shows why feedback strength cannot be judged from its sign alone. Let $z_1,z_2,z_3$ be signed deviations, not chemical concentrations, and take

The loop is negative: an increase propagated through the last two states returns as a decrease in the first. Its characteristic equation is $(s+1)^3=-g$. The conjugate pair has real part $-1+g^{1/3}/2$, so it crosses the imaginary axis at $g=8$.

The extra states give the correction its dynamics. Removing them would change the question being answered. A scalar equation with a restoring slope cannot reproduce this instability merely by turning its negative slope more negative.

The lesson transfers to biological circuits without assuming that a promoter is mechanically a governor. A regulatory sign describes a direction of influence. The times and intermediate states through which that influence travels determine how it acts on a changing system.

Test feedback together with its intermediate states and timescales, and avoid identifying a linear instability with a completed nonlinear oscillation mechanism.

4A perturbation must stay in its class

Horn and Jackson's 1972 treatment made chemical kinetics a geometric problem as well as a list of rate equations. The stoichiometric subspace records all directions in which reactions can change composition. A trajectory stays in the affine translate determined by its initial composition, restricted to nonnegative concentrations.²

For $\dot x=\Gamma v(x)$, any constant vector $\ell$ satisfying $\ell^{\mathrm T}\Gamma=0$ gives a conserved quantity $\ell^{\mathrm T}x$. These identities do not depend on the values of the reaction rates.

Stability must respect that geometry. A reversible binding system can have a unique attracting equilibrium for each pair of conserved totals, while the full species-space Jacobian has zero modes associated with changing those totals. Returning to the old equilibrium after adding material is a different requirement from recovering after redistributing it.

Identify the stoichiometric compatibility class before asking whether a chemical equilibrium attracts nearby states.

5A chemical analogue of energy

Horn and Jackson also developed a scalar function with the structure of a relative entropy. For positive concentrations $x$ and a positive reference equilibrium $x^*$, a shifted form is

Complex balance supplies conditions under which this function decreases along mass-action dynamics. The reference state and balance assumptions matter. An arbitrary open biochemical network does not acquire this property merely because its concentrations are positive.²

The mechanism can be seen directly in a reversible elementary conversion $A\xrightleftharpoons[k]{k}B$ with fixed total 1. Let $x$ be the fraction in state $A$. Then $\dot x=k(1-2x)$, and

Here the two factors in the derivative have opposite signs away from equilibrium. The system cannot move uphill in $V$. A mechanical energy and this chemical function have different physical origins, but both turn a trajectory question into a sign calculation on the full vector field.

Define a candidate scalar and compute its derivative along the actual reactions before using it as a stability certificate.

6A global claim needs a boundary argument

A decreasing scalar is not a substitute for stating the domain of the conclusion. In the reversible example, the interval is invariant and the endpoints point inward. Its explicit solution approaches 1/2 from every allowed initial fraction. The global conclusion is therefore straightforward.

For a larger chemical network, a sublevel set can approach the boundary where one or more concentrations vanish. A function that decreases in the positive interior does not, by that fact alone, exclude every limiting boundary behavior. Conservation, boundedness and the largest invariant set where the derivative vanishes must be examined.

The same issue appears in the damped spring. Its energy derivative is $-cv^2$, which vanishes whenever velocity is zero. A displaced spring cannot remain at zero velocity, so the only invariant part of that set is equilibrium. Compact energy sublevels supply the remaining confinement argument.

This is why the lecture's exposition separates local stability, attraction and a domain-wide conclusion. A certificate earns the stronger statement only when its hypotheses cover the entire region to which the statement refers.

Check confinement, boundary behavior and invariant zero-derivative sets before upgrading a local stability argument to a global one.

7A matrix can represent a family

Qualitative matrix theory asks what can be proved when interaction signs are known more confidently than their magnitudes. Jeffries, Klee and van den Driessche's 1977 paper gives a general characterization of sign stability and supplies full proofs, with applications spanning economics, ecology and chemistry.³

The essential change is a quantifier. An ordinary stability test asks whether this matrix is Hurwitz. A sign-stability test asks whether every matrix with these signs is Hurwitz. A collection of stable numerical samples cannot establish the second statement.

A structural guarantee is stronger within its stated class and silent outside it. Coupling magnitudes constrained by chemistry may occupy only a subset of all sign-compatible matrices. Conversely, changing an omitted state can change the matrix dimension and leave the original family entirely.

Write the family quantified over by a structural theorem, and distinguish an exact family guarantee from parameter sampling.

8A positive circuit is permission

The question of multiple states led to a different graph criterion. René Thomas proposed that positive feedback is necessary for multistationarity. Soulé's 2003 treatment proves a precise version for a differentiable vector field on an open product of intervals with at least two nondegenerate zeros: some interaction graph in the domain must contain a positive circuit.⁴

The direction of the implication matters. Multiple nondegenerate equilibria require suitable positive circuitry under those hypotheses. Finding a positive circuit does not prove that the equilibria exist, that two are stable or that a cell can be switched between them.

The domain also matters. Soulé explicitly examines what breaks at a closed boundary and when zeros are degenerate. These qualifications are part of the result, not optional cautions attached after applying it.

This graph is an interaction graph of derivatives, not simply the list of chemical reaction arrows. A reaction may affect several derivatives, and their signs can depend on state. Confusing the two graphs can make a biological cartoon appear to satisfy a theorem that has never actually been tested.

Read a positive-circuit condition as a necessary constraint with explicit domain and nondegeneracy assumptions, not as a sufficient recipe for bistability.

9Geometry can exclude a clock

A parameter-free conclusion can come from the geometry of state space rather than an interaction graph. For an autonomous scalar equation with unique solutions, a trajectory moves monotonically between equilibria. It cannot complete a nonconstant periodic orbit.

In a smooth planar system, closed curves separate an inside from an outside. The Poincaré–Bendixson theorem gives a strong consequence: a nonempty compact limit set with no equilibrium is a periodic orbit. The compactness, absence of equilibria and planar autonomous setting are all part of the statement.⁵

These results do not imply that any two-dimensional system oscillates. The reduced two-protein toggle has constant negative divergence, which excludes closed orbits by the planar flux argument. Its two stable states are compatible with that exclusion.

The lecture's center–cycle distinction addresses another ambiguity. Classical Lotka–Volterra dynamics has a conserved function and a family of closed orbits. A stable limit cycle instead attracts nearby trajectories toward one isolated orbit. A repeated oscillation on a graph is not enough to distinguish the two.⁵

Match a claim about cycling to the system's effective dimension, invariant region and attraction properties, not merely to a repeated time trace.

10The toggle joins theory to construction

Gardner, Cantor and Collins's 2000 genetic toggle connected a compact model to a constructed circuit in Escherichia coli. Two repressors inhibit one another's expression. Transient chemical or thermal induction can favor one side and leave the circuit in a different expression state.⁶

The construction changed the evidential task. A feedback diagram and nullclines could guide a design, but the resulting cells still had to exhibit the predicted behavior. The same architecture could be assembled from components with strengths that landed outside the bistable region.

The useful achievement was therefore a connection between levels: physical regulatory parts, an effective model, predicted parameter regimes and a protocol that tested the resulting cells. No one of those levels replaced the others.

Use a circuit model to propose an experimentally testable regime rather than treating its diagram as proof that a constructed device has the desired function.

11The diagram needs a parameter region

Mutual repression does not select one behavior independently of strength. The lecture's illustrative toggle writes that dependency explicitly:

At equilibrium, eliminating the second variable gives $y=4/(1+x^2)$ and $a_x=x[1+y^2]$. Parameterizing this relation by $x$ reveals the full equilibrium branch without losing the turning points to a coarse parameter scan.

A design margin is a distance from failure within a specified perturbation family. Moving a production capacity changes that margin. Changing cooperativity, coupling through growth or adding promoter states changes the family itself. The word “robust” needs to say which change is under discussion.

Locate a design within its bifurcation diagram and distinguish a parameter margin from robustness to changes in model structure.

12A memory test removes the inducer

To demonstrate memory, the intervention must end while its effect persists. The original toggle experiments compared induced and uninduced preparations and then removed the inducer. In one long-term test, differently prepared cultures remained in their respective high and low states for a further 22 hours without inducer.⁶

A high signal during induction would answer a weaker question. It could be an ordinary driven response. Persistence after removal asks whether the system's own dynamics retains the result.

A histogram with two peaks also leaves alternatives. Separate subpopulations with different parameters can be bimodal without either cell being bistable. A memory protocol must therefore control the conditions and relate the observed state to preparation history. The stronger claim is about what the same system can retain, not only what a mixed sample contains.

Finite observation also sets the scope of the conclusion. Persistence for a measured interval establishes performance on that interval. It does not demonstrate infinite retention or zero stochastic switching.

Test memory by removing the driving intervention, controlling preparation history and stating the observation interval.

13The host enters the circuit

Tan, Marguet and You's 2009 circuit showed why the host cannot always remain a passive background. A mutant T7 RNA polymerase activated its own expression noncooperatively, yet the cells showed bistable behavior. Circuit activation also slowed growth, reducing dilution of the expressed protein.⁷

The feedback diagram had acquired another path. More expression altered host growth, and altered growth changed expression loss. Omitting that path made the circuit's behavior appear inconsistent with the simpler model.

The evidence combined expression measurements, growth measurements, history-dependent single-cell behavior and model comparisons. This was not simply permission to explain any surprising curve by adding “context.” The proposed context supplied a measurable variable and a specific feedback route.

When a circuit changes its host, make the proposed host feedback explicit and test the physiological variable that carries it.

14Nonlinearity changes its address

Later modeling asked how far growth feedback could extend the circuit's repertoire. Melendez-Alvarez and Tian's 2022 analysis considered ultrasensitive growth suppression and showed how a constant-production circuit can acquire a nonmonotone removal curve and bistability.⁸

This is not a claim that bistability needs no feedback. The feedback is supplied by protein burden affecting growth, and growth affecting protein dilution. Nor does every slowing of growth suffice. The removal curve must turn, and the production level must intersect it in the appropriate range.

The distinction refines the lecture's cooperative-production example. That example is a valid construction of memory, not a proof that transcriptional cooperativity is the only construction. The invariant requirement lies in the full dynamical balance, while the biological location of its nonlinearity can change.

The student-inspired growth example works this out numerically and derives a slope criterion. The lineage's point is why the question changed: host physiology can move a dynamical mechanism outside the boundary of the original circuit diagram.

Separate a sufficient molecular construction from a universal dynamical requirement, and look for feedback across the chosen model boundary.

15Matching a curve leaves a clock free

An equilibrium relation can be unchanged while the time needed to reach it changes. For a two-state occupancy fraction $p$, write

The positive parameters $a,b$ are effective transition rates under fixed external conditions. Scaling both by the same factor preserves the occupancy curve and changes its clock. This elementary calculation shows why an equilibrium fit alone cannot validate instantaneous occupancy.

A finite writing pulse can end before the hidden promoter state catches up. A phase portrait in protein coordinates may then retain the correct equilibria while mispredicting switching time. This is an observable-specific failure of reduction, not a reason to reject all reduced models.

The distinction has an earlier counterpart in enzyme kinetics. Briggs and Haldane showed that a familiar saturation curve can arise without assuming binding equilibrium. A successful effective curve need not uniquely identify the microscopic process that generated it.⁹

Validate a reduction against the time-dependent observable and protocol of interest, not only its steady-state curve.

16Stationary does not mean motionless

Gillespie's 1977 stochastic simulation paper treated chemical evolution as discrete random reaction events under a spatially homogeneous model. A deterministic trajectory suppresses those event fluctuations. An exact stochastic simulation is exact for its declared event model, not for every feature of a real cell.¹⁰

Consider constant molecular births at rate $\alpha$ and independent deaths at rate $\gamma n$. Its stationary count is Poisson with mean $\bar n=\alpha/\gamma$. Conditioning on the present count gives

Multiplying by the present centered count and averaging shows that the stationary autocovariance is $\bar n e^{-\gamma\tau}$. The distribution stays the same while individual counts fluctuate and lose correlation.

For a memory, those fluctuations raise a different question: how long before a random trajectory leaves the intended state? That requires the event rates and an escape criterion. The location of deterministic attractors is an important input, but it is not the retention-time answer.

Distinguish a stationary distribution from an unchanging individual state, and ask for escape statistics when the claim concerns memory retention.

17Evidence belongs to a claim

The historical comparisons reveal different, non-interchangeable kinds of evidence. They are easiest to keep separate by stating the conclusion each one supports.

This table is not a ranking in which mathematics always outranks experiment. A theorem about the wrong model does not establish the behavior of a cell. An experiment can reveal a missing state that invalidates the model's biological interpretation. The useful relation is reciprocal: theory makes the test precise, and observation tests whether the description deserves to represent the system.

Attach each conclusion to the evidence that actually supports it, and state the untested step between the model and the physical system.

18The inherited method

The durable method is to strengthen the question before strengthening the claim. Ask for recovery, not merely corrective wiring. Ask for a state-space domain, not merely a list of species. Ask for a family, not merely a stable parameter choice. Ask for a completed writing and retention protocol, not merely two peaks or two fixed points.

Lecture 4's tools fit together because they answer those different questions. The phase line gives an ordering argument. The Jacobian gives a local mode test. A Lyapunov function gives a trajectory constraint. Network structure can replace uncertain parameter values with a theorem's hypotheses. Experiments and stochastic models decide what a cell actually retains.

The Week 2 audience chapter makes this inheritance operational. The students vary conserved totals, host growth, loop strengths, basin measures, pulse durations and hidden molecular clocks. Each variation names a possible gap between a phase portrait and a functioning device. Closing that gap is the work the portrait was meant to organize.

Turn a qualitative dynamical claim into a sequence of explicit state, mechanism, robustness and measurement tests.

References and reading basis

Important full texts were inspected in the course's root literature archive. The source log records the passages used. Figures are independently calculated teaching examples unless stated otherwise. Historical episodes are compared without asserting undocumented direct influence between their authors.

1. J. C. Maxwell, “On governors,” Proceedings of the Royal Society of London 16, 270–283 (1868). DOI. Opening discussion of disturbances, characteristic roots and Jenkin's governor. Local copy is a modern typeset transcription.
2. F. Horn and R. Jackson, “General mass action kinetics,” Archive for Rational Mechanics and Analysis 47, 81–116 (1972). DOI. Stoichiometric geometry, the pseudo-Helmholtz function and complex balance. No unrestricted global convergence theorem is inferred here from positivity alone.
3. C. Jeffries, V. Klee and P. van den Driessche, “When is a matrix sign stable?” Canadian Journal of Mathematics 29, 315–326 (1977). DOI. Introduction and characterization of sign stability.
4. C. Soulé, “Graphic requirements for multistationarity,” Complexus 1, 123–133 (2003). DOI. Theorem 1 and the discussion of open domains, nondegenerate zeros and boundary counterexamples.
5. M. W. Hirsch, S. Smale and R. L. Devaney, Differential Equations, Dynamical Systems, and an Introduction to Chaos, 3rd ed. (Academic Press, 2013). Publisher. Sections 10.5 and 11.2 supply the modern planar theorem and predator–prey comparison, not a primary historical account of their invention.
6. T. S. Gardner, C. R. Cantor and J. J. Collins, “Construction of a genetic toggle switch in Escherichia coli,” Nature 403, 339–342 (2000). DOI. Box 1 and the construction and memory experiments.
7. C. Tan, P. Marguet and L. You, “Emergent bistability by a growth-modulating positive feedback circuit,” Nature Chemical Biology 5, 842–848 (2009). DOI. Full author manuscript, including growth measurements, single-cell history tests and model limitations.
8. J. R. Melendez-Alvarez and X.-J. Tian, “Emergence of qualitative states in synthetic circuits driven by ultrasensitive growth feedback,” PLOS Computational Biology 18, e1010518 (2022). DOI. Constitutive-promoter analysis, pages 4–5.
9. G. E. Briggs and J. B. S. Haldane, “A note on the kinetics of enzyme action,” Biochemical Journal 19, 338–339 (1925). DOI. Both pages of the original note.
10. D. T. Gillespie, “Exact stochastic simulation of coupled chemical reactions,” Journal of Physical Chemistry 81, 2340–2361 (1977). DOI. Event-level description, assumptions and algorithm. The birth–death calculation here is an independent analytic illustration.