How separate questions became one account of adaptation

An amplifier designer, a physicist, control theorists, biochemists, chemotaxis modellers, network searchers and synthetic biologists each answered a different question. This page follows what each answer meant when it was asked, and what it came to mean once the answers were joined.

CCBS · Lecture 10 · lineage · 13 October 2026

Robust adaptation joins questions that were first asked separately. Can feedback reduce sensitivity to changing parts? Can a disturbed output return exactly? Can chemistry implement the required memory? What happens to sensitivity and noise when that memory closes a loop? The history below follows how these questions were joined, and why their assumptions still matter.

The core asks how an adaptation biomachine implements its function with binding and catalysis. The exposition derives a method for checking an adaptation claim. Here the organizing evidence is what each source asked, measured or proved.

The biomachine perspective is the course’s organizing choice. Functional roles let us compare a governor, receptor modification and a synthetic controller. The comparison is useful because it asks what each system senses, stores and changes. It does not imply that the historical authors used this vocabulary or built interchangeable mechanisms.

The sources below did not set out to build one account. Each asked its own question, with its own kind of evidence, and several of them changed meaning when a later source joined them to another. The page is ordered by those questions, not by date, and each section keeps three things apart: what the source claimed, how it was later read, and how the course now builds on it.

Lecture 9's lineage already tells how Black came to his amplifier, from drifting telephone repeaters to his recollection of a ferry crossing. This page recalls that story in two sentences and starts from the factor his amplifier made famous.

1Each piece of today's account of adaptation answered its own question

The organizing question is which questions, and which kinds of evidence, produced the pieces now used to explain robust adaptation. A second question follows: what did each piece come to mean once it was joined to the others? The table answers both in one line per piece, and the sections below give each answer its evidence.

Six kinds of evidence appear, and the page keeps them apart. A theorem holds exactly under its hypotheses. A mathematical construction produces a mechanism satisfying stated constraints. A model calculation shows what a stated model does at stated parameters. A sampled model ensemble shows how often a property survives random parameter changes inside one model. A choice among fitted candidate models ranks the candidates that were fitted, and only those. A measurement shows what cells or circuits did under one protocol, within its precision.

Four changes of meaning carry the page. A necessity theorem for designed regulators became a guide for reading biological networks. Saturated modification and sequestration became memory constructions after being studied as sources of ultrasensitivity. Black’s question about amplifier gain became a question about the constants that set a biological reference. Frequency sensitivity then separated a correct steady mean from the quality of its regulation.

Before weighing any source on adaptation, name the question it asked and the kind of evidence it offers: a theorem, a mathematical construction, a model calculation, a sampled ensemble, a choice among fitted models, or a measurement.

2Black's feedback bought insensitivity to the amplifier's own parts

Black's problem was a gain that drifted with its tubes. He built amplifiers with about 40 dB more gain than needed and fed the output back on the input "so as to throw away the excess gain" (pp. 1–2).¹ Lecture 9's lineage tells the rest of that story. What this page needs is the factor.

Black's own algebra already joined two results. He wrote the output as signal plus noise plus distortion, with the noise and distortion "assumed to be generated within the amplifier" and the distortion "a function only of the signal output" (p. 3). With feedback, all three are divided by one factor at a fixed input (eq. (2), p. 5). Compared at the same signal output, with the same amplifier and its input raised, the internally generated noise and distortion are smaller by that factor. The amplifier's own variations are stabilized by it, while those of the feedback circuit are "neither appreciably improved nor degraded" (p. 6). In the limit of a loop gain far above 1, the amplification stops depending on the amplifier and is set by the feedback circuit (p. 3).¹

Bode later gave the two ideas separate names. The usual conception of feedback, he wrote, "includes two distinct ideas": a loop return, which he called the return difference, and "a reduction in the effects of variations in the tube characteristics", which he called the sensitivity (p. 47).² In normal circuits, he added, the two are related by simple mathematical laws, so that "feedback" can name both (p. 47). His sensitivity is a ratio of percentage changes, so an amplifier with 40 dB of feedback has sensitivity 100 for its forward elements (pp. 52–53). Åström and Murray's textbook calls the reciprocal of Bode's quantity the sensitivity function, and reads it as the factor between a relative error in the process and the relative error of the closed loop (eqs. (12.2) and (13.23), pp. 12-3 and 13-17).³

Pointwise insensitivity raises a further design question. A dynamic circuit must remain stable across frequencies. Section 2a follows that question before the history turns to exact return.

A static construction shows the connection, with Black's quantities under the course's names. Call the amplifier's gain without feedback $G_\mathrm{amp}$ and the fraction of the output fed back $\rho$, so the loop gain is $H=G_\mathrm{amp}\rho$, the dimensionless product that sets the static feedback factor. Add a disturbance $\chi_\mathrm{out}$ at the amplifier's output and a disturbance $\chi_\mathrm{in}$ to the command $V_\mathrm{in}$ before the comparison. This is the course's explanation, not Black's derivation:

One factor answers both questions. The closed-loop gain's relative sensitivity to $G_\mathrm{amp}$ is $1/(1+H)$, and to $\rho$ it is $-H/(1+H)$, so as $H$ grows the dependence moves from the amplifier to the feedback path. The disturbance added at the output reaches the output divided by the same $1+H$. The disturbance added to the command reaches it multiplied by $G_\mathrm{cl}$, exactly as the command does (Figure 1).

A share needs a stated baseline, and the page uses two. On Black's basis, the same amplifier with its input raised and the noise voltages fixed, both disturbances keep $1/(1+H)$ of their shares. Figure 1(b) uses baseline (B): an amplifier without feedback whose gain equals $G_\mathrm{cl}$, driven by the same $V_\mathrm{in}$ and carrying the same noise voltages. Against (B) the disturbance added at the output keeps $1/(1+H)$ of its share, and the disturbance added to the command keeps all of it. So the first result holds against both baselines used here, and the second depends on the baseline. Neither holds on every basis. With the same amplifier at the same input, feedback lowers the signal and the disturbance added at the output together, so their ratio does not change. These are comparisons inside the course's model, not a claim that two physical amplifier designs generate the same internal noise.

The factor works at the scale of Black's measurement. Forty decibels of feedback means $1+H=100$. A change of 0.7 dB in an amplifier's own gain then becomes about 0.007 dB in the closed loop, 0.0067 dB exactly for that finite change. The 0.7 dB was measured on a conventional amplifier of comparable size, not on the open loop of Black's feedback amplifier, so the arithmetic illustrates the factor at the scale of his measurement and does not predict it.

Before calling a feedback loop robust, name the parameter, the disturbance's entry point and the baseline. The factor $1/(1+H)$ shrinks the amplifier's own variations, and against baseline (B) it shrinks a disturbance added at the output but not one added to the command.

2aFrequency sensitivity made feedback a tradeoff across frequencies

Bode asked what realizable networks permit over a whole frequency range. A useful reduction of sensitivity at one frequency is only part of an amplifier design. His 1945 book uses contour integration to relate real-frequency characteristics to physical realizability (p. 144). It also asks how available phase margin should be divided between interstage networks and the rest of a feedback circuit (pp. 487–488). The question is about the attainable response, not only a feedback factor at one operating point.²

The modern Bode sensitivity integral expresses one such constraint. Let P(s) describe the plant, C(s) its negative-feedback controller, and $L(s)=P(s)C(s)$ the loop transfer function. The complex frequency s has units of inverse time. Define $S(s)=1/[1+L(s)]$. An additive disturbance at the plant input reaches its output through PS instead of P. Thus S compares the controlled response with the uncontrolled response to the same disturbance.

The constraint needs stable closed-loop dynamics and sufficient high-frequency decay. For a real-rational single-input, single-output loop, internal asymptotic stability and $L(s)=O(s^{-2})$ give

Here $\omega$ is angular frequency and the sum counts the right-half-plane poles of L with multiplicity. With none, the integral is zero. An origin integrator gives an integrable logarithmic singularity at zero frequency. The signed area uses linear frequency. Less sensitivity over one band requires more over another. If $a=\lim_{s\to\infty}sL(s)\ne0$, the right side instead includes $-\pi a/2$. Åström and Murray state these results in Theorem 14.3 and Exercise 14.3, printed pp. 14-6 and 14-31.³

The integral does not specify an individual stability margin. Its value constrains a frequency integral. The maximum sensitivity, phase margin and gain margin still depend on the response shape. The antithetic performance question will reuse this distinction. The core’s A20 works out a matched comparison rather than assuming that the presence of an integral orders the margins.

A claim of robustness needs its frequencies as well as its disturbance. The Bode sensitivity integral constrains redistribution across frequency under stated stability and high-frequency assumptions.

3Exact return was a different question, answered by a part that integrates

Black's factor shrinks an error without removing it, and an older paper had already separated the two. Maxwell opened his 1868 paper on governors by defining a governor as the part of a machine that keeps its velocity "nearly uniform", and then split regulators into two kinds (transcription p. 1).⁴ A historian of control reads the opening the same way: Maxwell begins by "carefully distinguishing between regulators which give offset, which, following Siemens, he terms moderators, and governors which have no offset" (p. 68).⁵ The archived copy of Maxwell's paper is a modern typeset transcription that does not mark the original page numbers, so this page cites the transcription's pages.

A moderator reduces the error and keeps it. In a moderator, "the resistance is increased by a quantity depending on the velocity", so an increase of driving power still increases the velocity, "though a much smaller increase than would be produced without the moderator" (transcription p. 1). His equation (1) puts the final velocity above the normal velocity by the change of load divided by the moderator's force constant (transcription p. 4). Following Siemens, he would not call this a governor.

A governor returns exactly, because a separate piece accumulates the error. If the centrifugal part "sets in motion a contrivance which continually increases the resistance as long as the velocity is above its normal value", the governor brings the velocity "to the same normal value whatever variation (within the working limits of the machine)" is made in the load (transcription p. 1). For an idealized piece with no resistance of its own, integrating its equation shows its momentum equal to the force constant times the accumulated excess distance, so when the piece rests even the machine's position is restored (transcription p. 4).

Jenkin's governor keeps the speed exact but not the position. Its moving piece meets a viscous resistance, and then "the position of the machine is affected by that of the governor, but the final velocity is constant" (eq. (8), transcription p. 5). The box writes this governor as a state model from his equations (6) and (10).

The integral is a combination of the piece's momentum and its displacement. By the last line of the box, $J_\mathrm{g}v_\mathrm{g}+c\,\varphi$ changes at a rate proportional to the speed deviation, so the governor can rest only at the normal speed. The displacement alone does not integrate. At rest it is the displacement whose brake force balances the load, $\varphi=d/s$, so the displacement holds the load (Figure 2). Integrating the last line from one rest to the next gives the machine's extra distance, $c\,d/(\kappa s)$, which is not zero. This is the course's calculation behind Maxwell's remark that the position is affected.

Maxwell then asked a separate question: does the motion settle? For Jenkin's governor he derived a condition, his (14), on the piece's viscous resistance against the brake's strength. In the box's names it reads $(\kappa/J+c/J_\mathrm{g})\,c/J_\mathrm{g}>s/J_\mathrm{g}$. If it fails, "there will be a dancing motion of the governor", so the weight is hung in a viscous liquid to damp it (transcription p. 6).⁴

A governor establishes that accumulated error can remove an offset, but it does not establish when such a mechanism is necessary. Maxwell analyzed a chosen device and tested whether it settled. The next question reverses that reasoning: what must every regulator contain if exact return is to survive changes of the plant?

Tell a moderator from a governor by asking whether the corrective action depends on the present error or on its accumulated history, and test stability separately in either case.

4The internal model principle asked for exact regulation that survives changes of the plant

Francis and Wonham asked a converse question. Others had given procedures for designing regulators whose stability and regulation survive small changes of parameters, among them Davison, Staats and Pearson, and Wonham. Francis and Wonham asked instead: "What controller structure is necessary for structural stability?" (p. 457).⁶ Their paper summarizes and extends their own 1975 results, so it is neither the first appearance of these results nor the invention of integral control.

Structural stability joins Black's question to Maxwell's. A regulator synthesis is structurally stable when closed-loop stability and output regulation both survive the specified small changes of parameters (p. 457). Maxwell's governor contributes exact return. Black contributes the question of insensitivity to parameters, which his amplifier answered only in part: at finite loop gain its sensitivity stays $1/(1+H)$. Structural stability asks for both at once.

The answer is an internal model of the signals, not of the plant. For their finite-dimensional linear, time-invariant regulators, readability of the regulated variable and a stated technical condition imply that a structurally stable synthesis needs a compensator that contains an internal model of the exogenous signal generator, controllable by the regulated variable and observable by the actuator (Theorem 2, p. 460). For constant signals that model is a mode at zero frequency, an integrator. The internal model supplies closed-loop zeros that cancel the unstable poles of the exogenous signals (p. 460).

The theorem carries its scope in its own words. "A smaller class of parameter variations in general requires less reduplication in the internal model" (p. 460). For constant signals, uncertain plant parameters leave no offset, but a slight perturbation of the internal model itself leaves "a small steady state offset error", because "the accuracy of regulation depends on the fidelity of the internal model" (p. 461). The weakly nonlinear extension is local, for constant signals and a smooth plant designed on its linearization (pp. 462, 464). The conclusion calls perfect asymptotic rejection an idealization (p. 464).⁶

The internal model idea already had a biological echo. Francis and Wonham credit plausibility arguments for an internal model to Kelley and to Conant and Ashby, and note that internal models had a role in theories of visual perception and of brain function (p. 460). Sontag later wrote that the principle "originates in the biological cybernetics literature", and that it is "not a theorem but rather a 'mold' for many possible theorems" (p. 120).⁷ The core's §3a states the classroom theorem with its hypotheses. Its A10 compares Sontag’s different setting.

State the internal model principle with its hypotheses, and keep uncertainty in the plant apart from changes to the reference machinery: for constant signals the first leaves no offset, while the second can.

5Goldbeter and Koshland read a saturated modification cycle as a switch

Goldbeter and Koshland asked how a pathway could respond sharply without cooperative enzymes. Cooperativity and a ligand acting at several steps were known answers. They examined a protein that one enzyme modifies and another demodifies, and asked what happens "in the absence of allosteric cooperativity and multiple inputs" (p. 6840).⁸ In the core's names, the target is $S$, its modified form $S^*$, and the two enzymes $E$ and $F$, with capacities $k_Eq_E$ and $k_Fq_F$ (their $V_1$ and $V_2$). The conserved substrate inventory is $S_0=S+S^*+C_{ES}+C_{FS^*}$. These are the names used in the core’s §6.

Their answer was a steady-state switch. When both converter enzymes are saturated, the transition from unmodified to modified protein becomes far more abrupt (p. 6841). Their response coefficient, the ratio of capacity ratios between 90% and 10% modification, is 81 for a Michaelis-Menten response and falls toward 1 as both enzymes saturate (Eq. 10, p. 6841). With both Michaelis constants at a hundredth of the target total, they compare the steepness to "a Hill coefficient greater than 13". Their Eq. 10 gives about 26 (Figure 3a).

Three ingredients and one rate explain it. The kinetics must be at least partly zero order, the target's total must be conserved, and the system must reach a steady state rather than an equilibrium (pp. 6843–6844). When both enzymes are saturated, the rate of modification is the difference of the two capacities, $k_Eq_E-k_Fq_F$, whatever the amounts of the two forms, and it continues until one form falls below saturation (p. 6844). That rate is the fact a later reading would use differently.

They also asked what enzyme complexes do. With enzyme totals at 10% of the target or less, the switch barely changes. Larger totals damp it, and how much depends on whether the target stays active while bound, so the readout may count bound forms (Fig. 3 and p. 6842). Figure 3(b) reconstructs that calculation and adds the course's question: what a readout sees if it counts the bound modified form.

Adaptation appears in the paper only as a reason to want sensitivity. The authors name "an adaptive sensory system in which one pathway must be turned on and another pathway turned off" as a place where zero-order ultrasensitivity could supply the needed sensitivity (p. 6844). They add that "the data are not yet available" to say the device is used in cells (p. 6844). Their curves are model calculations.

Read zero-order ultrasensitivity as a statement about a steady state: two saturated enzymes, a conserved target, and a capacity ratio that crosses 1.

5aA kinase cascade connected the switch calculation to biological evidence

Huang and Ferrell asked whether a phosphorylation cascade could create a sharp biological response. In 1996 they modeled the mitogen-activated protein kinase cascade and tested Xenopus oocyte extracts. Added Mos kinase was the input. Steady p42 kinase activity was the output. Their measurements agreed with a steep model response whose effective Hill coefficient was about 4.9, without assigning cooperative rate laws to the enzymes (pp. 10081–10083, Fig. 3).²⁴

They separated saturation from the cascade’s other source of ultrasensitivity. Changing substrate amounts and Michaelis constants changed the predicted steepness. Two successive phosphorylation steps also mattered. Their note added in proof reported distributive phosphorylation and partially saturated phosphatase activity (pp. 10081–10083). This provides biological support for saturation as one contributor. It does not reduce the measured cascade to the course’s single modification cycle or turn a static switch into an adaptation result.²⁴

The course uses the single modification cycle to isolate that contributor. The core’s §6 varies opposing catalytic capacities and reads the modified fraction. It then gives the same modification chemistry a different task by making its removal depend on a sensed output. The biological cascade motivates the components. The following feedback calculations establish what those components do in the course’s chosen architecture.

The extract experiment supports a sharp kinase response. Saturation and successive phosphorylation remain distinct mechanisms, and adaptation needs an additional argument.

6Chemotaxis turned the precision of adaptation into a question about parameters

Barkai and Leibler asked Black's question of a cell. They addressed "the sensitivity of the networks to variations in their biochemical parameters" (p. 913).⁹ Bacterial chemotaxis was their case: after a step of attractant, the steady tumbling frequency returns to its old value. Earlier models, they wrote, "assumed that the biochemical parameters are fine-tuned" to keep that return (p. 913). Their examples include a 1986 model of exact sensory adaptation by Segel, Goldbeter, Devreotes and Knox (their ref. 17, p. 917).

Their answer was a structure, not a tuning. Adaptation is robust when the methylating enzyme CheR works at saturation and the demethylating enzyme CheB acts only on active receptors, so that "the rate of reverse modification is determined by the system's output" (Fig. 1b legend, p. 913; Box 1, p. 914). They describe it as "a feed-back system" in which the activity is "continuously compared to a reference" (pp. 915–916). They do not use the word integral.

Their evidence was a sampled model ensemble. Starting from a perfectly adaptive reference model, they changed every biochemical parameter at random by a factor of two. Over 80% of 6,157 such altered models stayed within 15% of perfect adaptation, while the adaptation time varied widely (p. 915, Fig. 3). They also predicted which property would move: "the steady-state activity itself, which is not a robust property of the network, depends on the enzyme concentrations" (p. 916).

They proposed the experiment that would test it. Vary the expression of the chemotaxis enzymes, then measure the precision of adaptation, the adaptation time and the tumbling frequency (p. 915).⁹

Four objects stay distinct here. The pathway, in which CheA also phosphorylates CheB. Barkai and Leibler's reduced module, their eq. (1), written for the modified-enzyme concentration and the activity (p. 916). Yi and colleagues' weighted methylation coordinate (§8). And the course's teaching model, which adopts the activity-only rate form for the mean methylation (core §5).

Ask whether a claim of robustness to parameters was made on a sampled model ensemble or measured in cells, and which property it predicts will move.

7Protein titrations measured which property was robust

Alon and colleagues turned the model's claim into a measurement. To tell the fine-tuned scenario from the robust one, they systematically varied the concentrations of the chemotaxis proteins and measured the response (p. 169).¹⁰ Their observable was the tumbling frequency of a population, before and after a saturating step of attractant.

The precision held and the reference moved. As CheR varied, the adaptation time and the steady tumbling frequency changed, while the precision of adaptation stayed within experimental error (p. 169). The measured return is consistent with exact adaptation within experimental uncertainty. Finite data do not establish mathematical exactness.

A mutant removed one candidate mechanism, within that test. A form of CheB lacking the domain that CheA phosphorylates still adapted, which "seems to exclude CheB phosphorylation as the mechanism of exact adaptation" (p. 170). The test shows that phosphorylation was not required in that assay. It does not exclude phosphorylation from physiological regulation.

The molecular measurements answer narrower questions than the adaptation experiment. Li and Hazelbauer’s receptor-to-enzyme ratios support enzyme scarcity. Lai and colleagues distinguish recruitment to the receptor tether from binding to catalytic substrate sites. The exposition compares these measurements with the conditions of the reduced law. The measurements constrain a mechanism. Neither abundance nor recruitment alone establishes an activity-only methylation rate.^{22, 23}

The course reads the same split. The core's teaching model reproduces it: cutting CheR slows the return and lowers the adapted activity, and the return stays exact (core A12). The core's A16 names the split: an adaptive property kept across parameter changes, and a reference value that moves.

Design a robustness test as Alon and colleagues did: vary one component's level, then measure both the property claimed robust and a property expected to move.

8Integral control turned the internal model principle into a way to read networks

Yi, Huang, Simon and Doyle read the chemotaxis model as a controller. They proposed that the robustness of perfect adaptation "is the result of this system possessing the property of integral feedback control", and showed that integral control is "structurally inherent" in Barkai and Leibler's model (p. 4649).¹¹ They traced the robustness theory behind that reading to electrical network design, citing Bode's book (p. 4649). Black's line of work met chemotaxis there.

Their necessity argument was linear. They argued that "integral control in some form is necessary" (p. 4649). The proof in their appendix assumes a linearization about an equilibrium whose matrix has "all its eigenvalues in the open left half of the complex plane", and for a nonlinear system concludes only that "the linearization must have integral control". They called the result "a special case of the internal model principle" (p. 4653).

Then they turned the theorem into a way of reading. The necessity "is important to biologists, because they must reverse engineer systems 'designed' by evolution" (p. 4652). They proposed "to catalog the types of basic biochemical networks that can implement integral control" (p. 4652). A theorem about what a designed regulator must contain became a guide to what an evolved network might contain.

Sontag rebuilt the theorem for that use. Citing Yi and colleagues, he wrote that the principle "could help guide experimentalists and modelers" and "help distinguish among mathematical models which do or do not contain internal models" (p. 120).⁷ For biology he wanted a theorem without a split into plant and controller and without the robustness assumption. He also named how biology changes the objective: control design would "not even notice disturbances", while in biology "signal detection is often an objective, to be followed by a return to default values" (p. 120).

Sontag's theorem is nonlinear, and it trades robustness for detection. It treats a smooth system with one input and one output, affine in the input (p. 122). It assumes adaptation in a strong sense: for every input in the class and every initial state, the solution exists for all time and stays bounded, and the output tends to zero (p. 122). It assumes signal detection, a uniform relative degree: the lower derivatives of the output do not depend on the input's present value, and one derivative does, through a factor that vanishes nowhere (Assumption 1, p. 123). It adds two technical conditions on certain vector fields, completeness and commutation, which hold automatically for linear systems and are "quite strong" for nonlinear ones (Assumptions 2 and 3, p. 123; Remark 3.2, p. 125). And the inputs must come from a Poisson-stable exosystem, whose trajectories keep returning near their start (pp. 123–124). Under these hypotheses the system contains an output-driven internal model of its inputs (Theorem 1, p. 124).⁷

The course keeps the two theorems apart and looks for coordinates. The core's optional A8 bookkeeping is stated for an output that responds to the step, which is Sontag's setting rather than a design that never notices. Its §3a and optional A1 set the hypotheses side by side, and its §10 finds an incoherent feedforward loop's integral as a coordinate that no molecule carries. Optional A1 writes that coordinate in Sontag's output-driven form and says that this alone does not check every hypothesis of his global theorem.

Use an internal model theorem as a reading guide only after naming which hypotheses the network meets, and look for the integral as a coordinate, which need not be one molecule.

9A search from function to topology found the switch working as an integrator

Ma and colleagues asked a new kind of question. Instead of explaining one pathway, they asked "What are all network topologies that are capable of robust adaptation?" (p. 761).¹² Their kinetic model is part of the question: each of three nodes has a fixed total and an active and inactive form, links are Michaelis-Menten conversions, regulations of one sign add, and basal enzymes supply a missing direction (pp. 762, 772).

Their counts carry their criteria. They searched the 16,038 three-node topologies with a path from input to output, each with 10,000 sampled parameter sets. A sampled circuit counted as adapting when its sensitivity exceeded 1 and its precision exceeded 10. For their usual 20% input step the final output then lies within 2% of its start, not exactly at it (p. 762). The 395 robust topologies are those that adapted in more than 10 of their 10,000 samples, and every one contains a negative feedback loop with a buffer node or an incoherent feedforward loop with a proportioner node (pp. 766–767).

Here the switch's saturation became an integrator in print. When the two enzymes acting on the buffer node are saturated, "the buffer node B integrates the difference between the output activity C and its input-independent steady-state value", and the motif "implements integral control" (p. 764). Precision then follows the flatness of the buffer's nullcline, which zero-order ultrasensitivity supplies, and here they cite Goldbeter and Koshland's 1984 paper (pp. 765–766). The rate that Goldbeter and Koshland used to explain a steep steady state is here proportional to the output's error, and the buffer's level is its integral.

The same constant sets both readings. Figure 4 puts them on one axis. As the saturating enzymes' Michaelis constant falls, the switch's equivalent Hill coefficient rises (Eq. 10 of 1981), and the buffer circuit's precision rises in proportion to its inverse.

The scaling is derived, not assumed. At rest the buffer's two fluxes balance, and the balance fixes the output's steady activity $C$ in terms of the buffer's $B$, both as active fractions. The factor 0.5 is the output's saturated limit (their Eq. 3). The second line holds while both buffer enzymes are saturated, $K\ll B$ and $K\ll 1-B$, at both rests:

The precision follows at leading order. Write $g(B)=1/(1-B)-1/B$ for the coefficient of $K$. The input step moves the buffer's rest from $B_1$ to $B_2$, so the output changes by the fraction $K\,|g(B_2)-g(B_1)|$, and the precision is $0.2/(K\,|g(B_2)-g(B_1)|)$, inversely proportional to $K$. The dashed line in Figure 4 takes the rests in the limit $K\to0$, $B_1=0.5$ and $B_2\approx0.589$, which makes it about $0.27/K$. The computed precision is about 28 at $K=0.01$ and 271 at $K=0.001$. It exceeds the leading order by a nearly constant 0.93, so the relative gap shrinks in proportion to $K$: 3.4% at $K=0.01$ and 0.34% at $K=0.001$. The precision crosses Ma and colleagues' threshold of 10 near $K=0.03$.

The feedforward class came with an open question. The proportioner node tracks the input when its activation is saturated and its inactivation linear (p. 766). In their models the feedforward class adapted better, yet "there have so far not been clear cases where IFFLP is implemented in any biological systems to achieve good adaptation" (p. 771).¹² The core's §6 first derives the modification switch, and its §7 builds the two-node reduction that Jeynes-Smith and Araujo discuss, which is not a design printed by Ma and colleagues.

Read a topology count with its model and criteria beside it: 395 is a count under one enzymatic model, one sampling rule and a 2% tolerance, not a census of biology.

9aCoupled phosphorylation and feedback buffered a biological output

Fritsche-Guenther and colleagues asked why active Erk changed less than total Erk. In colon cancer cell lines, they reduced Erk protein levels and measured doubly phosphorylated Erk. They linked the compensation to inhibitory feedback from Erk to Raf-1. Raf variants that bypassed the relevant feedback lost the robustness. The compensating response did not require new transcription or translation (Results, Figs. 3–6).²⁵

The experiment supports coupled modification and feedback with a narrower claim than exact adaptation. The perturbed quantity was protein abundance. The observed quantity was phosphorylated Erk. The authors also rejected Mek saturation as the explanation of this robustness. The course’s two-cycle circuit therefore illustrates a related design question. It is not a literal fit of this pathway or a demonstration that these cells occupy its selected saturation regime.²⁵

Use the biological pathway to motivate feedback among modified proteins. Use its actual perturbation and observable to delimit what the experiment establishes.

10A molecular feedforward adapter was identified by choosing between fitted models

Takeda and colleagues took up the gap that Ma and colleagues named. They began from it: "no clear examples of biological networks that use the incoherent feed-forward strategy have been identified" (Introduction, paragraph 1).¹³ The archived full text is the PubMed Central author manuscript, which carries no page numbers, so this section cites section and paragraph.

They measured near-perfect adaptation of activated Ras through a reporter's location. Cells of the amoeba Dictyostelium received sudden uniform steps of the chemoattractant cAMP in a microfluidic device. The reporter, the Ras-binding domain of human Raf1 fused to GFP (RBD-GFP), preferentially reports activated RasG, while inactive RasG already sits along the plasma membrane (Results, paragraph 1). RBD-GFP moved from the cytosol to the membrane when it bound activated Ras. Its cytosolic signal fell as membrane recruitment rose, then returned (Results, paragraph 2). The measurement is the reporter's location, and Ras activity is inferred from it.

Their evidence was a choice between fitted models. They fitted the two simple topologies that can adapt perfectly, an incoherent feedforward loop and an integral-feedback loop with RasGAP as the buffer, to the same 21 data points, with nine free parameters each. "The integral control mechanism was incompatible with our experimental results", because its adaptation kinetics depended strongly on the size of the step (Results, paragraph 9). This is the kind of evidence that §1 called a choice among fitted candidate models: it ranks the candidates that were fitted.

The fitted candidates do not exhaust the possible mechanisms. Further integral-control variants were tested in the supplement, which this page has not read. The authors also leave open an adapting module between the receptors and RasGEF (Discussion, paragraph 2).

An actin perturbation tested one feedback route rather than an output load. Blocking actin left adaptation intact (Results, paragraph 6). The core uses this biological case to motivate its feedforward example. Its A16 compares the response to a separate output load, which supplies another possible way to discriminate candidate mechanisms.

Treat a model-selection result as a ranking of the candidates that were fitted, and ask which perturbation would separate the ones that remain.

11Writing out the complexes reopened a question the switch had already raised

Jeynes-Smith and Araujo asked whether the switch-based integrator survives its own chemistry. Ma and colleagues had used Michaelis-Menten kinetics, which leave the enzyme-substrate complexes out (p. 2).¹⁴ Jeynes-Smith and Araujo asked whether "the prevalent use of simplified models" had given "a false hope for ultrasensitivity-driven RPA mechanisms" (p. 4). They wrote every complex out and sampled parameter sets.

They stated the join between switch and integrator directly. Embedded in a negative feedback loop, "the narrow transition zone for the ultrasensitive switch is converted into a narrow tolerance around an RPA setpoint" (Fig. 1 legend, p. 2). Here RPA is robust perfect adaptation.

With every complex written out, the free output did not track the set point in the cases they studied. Their free active output $A^*$ "never tracks the estimated setpoint", nor any other nontrivial set point, at any rate constants or abundances they examined (p. 5). Where they found tracking over a range of inputs, it was by $C_3$, the complex of $A^*$ bound to the unmodified opposer protein $B$, while free $A^*$ fell below their tolerance. They therefore named the pool $A^*+C_3$ as the adapting variable, and noted that it "corresponds to the location of the ultrasensitive switch in the Goldbeter–Koshland model" (p. 5).¹⁴

The question of which form counts is as old as the switch. Goldbeter and Koshland had already shown that appreciable complexes damp the switch, and that the right readout depends on whether bound target stays active (p. 6842; Figure 3b).⁸ The question did not begin in 2023. What changed is the job: in 1981 the readout decided how sharp a switch looks, and in 2023 it decided which quantity adapts.

Reference tracking and input invariance ask different questions. Reference tracking asks whether a quantity sits at a specified set point. Input invariance asks whether that quantity returns to the same value as the persistent input changes.

The course’s selected regime identifies a pool that approximately tracks the reference. The core's §7 defines the totals and chooses one joint dominance regime. Exact flux balance then pins the complex $C_3$ near the set point $\sigma$. At illustrative parameter set PT, the partial pool $A^*+C_3$ stays within 5% of that reference over a broad input range. The full modified total also counts the complex containing the memory and leaves that band.

A free concentration can be nearly input invariant far from the reference. At the same parameter set, free $A^*$ varies by at most 5% over one reported input interval. Yet its value sits 99.6% below $\sigma$. That concentration nearly holds a constant value without tracking the specified reference.

The regime calculation is a sufficient example belonging to the course. The exposition’s observable audit and two-cycle reduction derive the exact flux balance and the finite-pool errors. This is not a classification of every adaptive regime or a regime survey attributed to Jeynes-Smith and Araujo.

Before naming the adapting quantity, ask which complexes the readout counts, and test reference tracking and input invariance separately.

12Positivity turned "does it integrate?" into "where can the integral live?"

Xiao and Doyle asked where a reaction network can keep an integral. Their 2018 construction begins from the contrast with electrical and computer implementations. A voltage or signed charge can cross zero. A species concentration cannot. Their analysis treats this physical constraint as an input to controller design, rather than searching for a molecule that already satisfies a desired signed equation (PDF pp. 3–5).¹⁵

The obstacle is a boundary condition on removal. A process cannot remove a species that is absent. Elementary mass-action removal therefore carries the species concentration as a factor. Suppose a desired memory concentration Z obeyed $\dot Z=\mu-\theta X$, with output X, reference production flux $\mu$ and sensing frequency $\theta$. At $Z=0$ and $\theta X>\mu$, that equation asks Z to become negative. The physical boundary rules out this unrestricted implementation (PDF pp. 3–4).¹⁵

They separated three construction methods. A species can satisfy a constrained error balance that ceases to guarantee zero error at a boundary. A change of coordinates can place the integral in a combination of species. An approximation can make a physical rate nearly independent of the stored amount within a regime. Their population-control and feedforward examples use the first two methods. Saturated enzymes illustrate the third (PDF pp. 4–5).¹⁵

The feedforward construction separates the molecular graph from the integral coordinate. In the course’s notation, the reduced equations are $\dot x_1=k_1w-\delta x_1x_2$ and $\dot x_2=k_2w-\gamma x_2$. The paper’s linear change of coordinates cancels the shared input. Restoring the two production constants gives

The coordinate integrates the output error with a positive, state-dependent gain when $x_2>0$. No output-to-degrader edge has been added to the reaction graph. This explicit construction explains the core’s reversal without treating feedforward adaptation as an exception to nonlinear control theory.¹⁵

The course asks what binding does to the feedforward output being measured. Its §9 derives bilinear removal when both totals are mainly free. Its A19 then compares free and total output at finite binding affinity, under a declared degrader-turnover convention. That calculation belongs to the course model. The measured quantity and turnover convention must be supplied before transferring the result to a cell.

The course turns these methods into three concrete chemical memories. Its §3b compares them before the examples.

These are useful constructions rather than a complete classification. The feedforward coordinate shows why the molecule that changes most after a disturbance need not itself be the integral.

Ask where a proposed integral can live with nonnegative amounts: a saturated removal, a difference of two species, or the logarithm of a species that grows in proportion to itself.

13Concentration robustness was also found from network structure alone

Shinar and Feinberg asked which structures fix a concentration, with no controller in view. They sought "structural attributes that impart concentration robustness to any mass-action network possessing them" (p. 1389).¹⁶ Their examples were two bacterial modules in which approximate robustness had been measured, the osmoregulation system EnvZ-OmpR and the isocitrate dehydrogenase kinase-phosphatase system (p. 1389). No error, controller or integral appears in their question.

Absolute concentration robustness is a property of steady states. A network has it for a species when that species takes the same value in every positive steady state the network admits, at fixed rate constants, so the species is protected even against large changes in the total supply of the network's components (p. 1389). The property says nothing by itself about whether that steady state is reached, which one is reached, or how a load on the species is handled.

Later work placed it among the adaptive properties, with care. Xiao and Doyle wrote that it "is a special case" of perfect adaptation, because perfect adaptation may reject disturbances other than the total (PDF p. 5).¹⁵ The exposition's §9 audits Shinar and Feinberg's toy module and, for that network, constructs an exact identity in a logarithm. It derives the threshold total, the return and the stability separately, because the theorem does not supply them.

Read absolute concentration robustness as a statement about steady states at fixed rate constants, and derive return, reachability and stability separately before calling a network adaptive.

14Sequestration was designed as an exact comparator for noisy cells

Briat, Gupta and Khammash asked what integral feedback means when reactions are noisy. Negative feedback was understood for deterministic networks, but "the effect of noise on their regulatory function is not understood" (p. 15).¹⁷ Engineered circuits with perfect adaptation were rare, and they called for simple synthetic circuits with integral feedback (p. 17).

Their comparator was a pair that annihilates one to one. In the concentration version of their eq. 1, a reference species $Z_1$ is made at a constant rate $\mu$ and a sensing species $Z_2$ at the rate $\theta X$, in proportion to the output X, and the two annihilate each other (p. 19). Annihilation need not destroy the pair, since an inert dimer of high affinity would do, and they offered a sigma factor and its anti-sigma factor in E. coli as "a hypothetical, concrete example" (p. 17).

The mean count difference integrates the mean error. Count the molecules in a fixed volume, as Lecture 5 does: $N_X$, $N_{Z_1}$ and $N_{Z_2}$, with $\Omega$ the count-to-concentration factor, so that $N_X/\Omega$ is the concentration of X. The reference then fires with propensity $\Omega\mu$, and sensing with $\theta N_X$. Annihilation removes one of each controller, so it leaves $N_{Z_1}-N_{Z_2}$ unchanged on every path. A reference event raises the difference by one and a sensing event lowers it by one, so the difference is not a conserved inventory. A path of the difference is constant between events and jumps at event times. It is not a solution of the deterministic concentration equation in the core’s §8. Briat and colleagues state the consequence for the mean: "the mean difference between the two controller species is proportional to the integral of the mean tracking error" (eq. 3, p. 20). In Lecture 5's angle brackets, with finite means, the identity is exact:

Their conclusion keeps its conditions. Tracking needs a closed loop that is ergodic, with bounded first and second moments (p. 19). Under those conditions their Theorem 1 gives tracking of the mean, whatever the initial conditions and the other parameters: the stationary mean count of X is $\Omega\mu/\theta$, the set point $\mu/\theta$ as a mean concentration (p. 20). Their Theorem 2 gives conditions on the open-loop network under which the closed loop is ergodic (p. 21). A stationary distribution is not a claim that each noisy trajectory converges. They also found that noise can stabilize a loop whose deterministic version is unstable (p. 18).¹⁷

The course defines its totals before using their difference. In the deterministic model with reversible binding, the core's §8 writes $q_{Z_1}=Z_1+C_{Z_1Z_2}$ and $q_{Z_2}=Z_2+C_{Z_1Z_2}$, then shows that their difference changes at $\mu-\theta X$ at every binding allocation. It tests reachability, dilution and stability separately, and applies the stochastic theorem in its optional A4 to the annihilation form only.

Read a designed integrator by its identity and its hypotheses separately: name the difference that integrates, then treat ergodicity, reachability and growth as separate claims.

15A sequestration switch became a controller that adapts nearly but not exactly in growing cells

The pair that built the controller had first been used in this setting to build a switch. Chen and Arkin put the Bacillus subtilis sigma factor SigW and its cognate anti-sigma factor RsiW into E. coli. SigW activated its own promoter, arabinose controlled the anti-sigma factor, and a fluorescent reporter read the SigW-responsive promoter. They found that "ultrasensitivity from sequestration combined with positive feedback is sufficient to build a bistable switch" (article p. 1; Fig. 2, p. 2).¹⁸ Their experiment tested switching and hysteresis, not adaptation.

Aoki and colleagues reused the same pair as a comparator. They reused the B. subtilis pair SigW and RsiW from that earlier report, citing Chen and Arkin (p. 535).¹⁹ In their circuit SigW drives the controlled protein AraC and the reporter sfGFP, and AraC closes the loop by regulating the anti-sigma factor's production (p. 536). The reporter is meant to track the controlled protein, and the two are different molecules (Figure 5).

The same sequestration did a different job because the loop closed differently. Closed on SigW itself with positive feedback, sequestration gave a threshold and two stable states. Closed through the controlled protein, the same one-to-one removal compared a reference with a measured output. The pair follows the modification cycle of §9: a source of ultrasensitivity read again as a memory.

In growing cells the adaptation is near-perfect, not exact. Their theory assumed negligible degradation of the two controller proteins, and they note that "in living cells, these assumptions are not satisfied" (p. 535). The core's A14 shows why. Growth dilutes free and bound controllers alike, so the complex cancels from the difference of the totals and leaves a leak, $-\lambda(q_{Z_1}-q_{Z_2})=-\lambda(Z_1-Z_2)$, where $\lambda$ is the growth rate. The leak generally leaves a steady error.

Expect a molecular pair to change its job with its circuit: the same sequestration gives a threshold inside positive feedback and a comparator inside an integral loop.

15aAntithetic performance reopened the frequency-sensitivity question

Olsman and colleagues asked what an adaptive antithetic controller gives up in its transient response. Their 2019 Cell Systems study considered regulated processes built from chains of production and removal reactions. They linearized the plant and controller around an equilibrium, then related biochemical parameters to stability, speed, steady-state error and sensitivity to small disturbances (pp. 49–56).²⁶

They organized the tradeoffs through the Bode sensitivity integral. Low-frequency rejection and amplification elsewhere were different parts of one sensitivity curve. Controller degradation could change the low-frequency error and the curve’s maximum. The comparison was about a declared model and local frequency response, rather than a universal ranking of biological controllers (pp. 53–57).²⁶

The course’s actuation comparison asks whether the integral distinguishes implementations. Its A20 holds the plant, positive equilibrium and low-frequency integral coefficient fixed. The reference species can increase production, the sensor species can suppress it, or both can act. For the specified dynamic plants and stable closed loops, every route has the same zero Bode sensitivity integral. Their maximum sensitivities and margins differ. The result concerns frequency redistribution. It is not an identity of their full dynamics.

The high-frequency plant model can change that answer. A20 also states the correction for an instantaneous plant. Sensor-containing actuation can then change the integral itself because its high-frequency decay differs. The equality for the dynamic plants is conditional. It does not follow from adaptation alone.

A local frequency argument cannot replace the chemical operating limits. The core’s A14 and A20 distinguish a stable equilibrium, an unstable equilibrium and an unreachable target. The Bode sensitivity integral applies to the stable local response. A sustained nonlinear cycle needs a nonlinear dynamical argument. Windup needs the actuator ceiling and the accumulating inventory. Equal dilution can remove exact constant-disturbance rejection while leaving the stable loop’s Bode sensitivity integral zero.

The antithetic mean-error identity fixes a possible steady mean. Its performance still depends on frequency response, dilution and actuator feasibility.

15bActuation choices separated mean adaptation from molecular noise

Kell, Ripsman and Hilfinger asked whether perfect adaptation necessarily raises noise. Their 2023 analysis supported that penalty for the commonly studied ideal controller with reference-only actuation. They then examined sensor-only actuation, in which the output-dependent species represses output production. Approximate analysis and exact stochastic simulations found noise suppression below the corresponding open-loop level even with the ideal mean-error identity (pp. 2–5).²⁷

The noise result for sensor-only actuation kept an unresolved stochastic condition. The earlier ergodicity theorem applied to reference-only actuation. Kell and colleagues supported ergodicity of their cases with sensor-only actuation by comparing time and ensemble averages in simulations, rather than proving that theorem for every network with sensor-only actuation (p. 5). Sensor-and-reference actuation was another possible architecture, not an equivalent name for sensor-only actuation.²⁷

Filo and colleagues explained a dynamical contribution of sensor-only actuation. Their 2026 study identified a filtered proportional contribution within the controller with sensor-only actuation. Their theory and simulations connected that contribution to transient and noise performance (pp. 1–8). It did not require adding a new species to the ideal antithetic motif.²⁸

Their bacterial implementation tested a specific comparison. Split inteins removed the repressive activity of TetR in E. coli. Anhydrotetracycline was the imposed disturbance. AraC fused to mScarlet-I supplied the fluorescent output. The built circuit with sensor-only actuation was compared with filtered proportional control and open-loop circuits. They did not build an intein circuit with reference-only actuation for a direct experimental comparison. Noise measurements included cellular variability beyond intrinsic reaction noise (pp. 9–11, Fig. 7).²⁸

Frequency sensitivity and stochastic variability need different evidence. The following comparison prevents the same word, robustness, from collapsing their questions.

The next experiment should match the question being disputed. A small sinusoidal perturbation tests frequency sensitivity. Single-cell distributions at matched means test variability. A large persistent load tests reachability. The research essay uses these as distinct ways to discriminate mechanisms that share a step response.

The same adaptive mean does not determine the noise or the transient response. Name the actuation route and the experiment before transferring a performance claim.

16Black's question returned as which constants the set point may depend on

Black's amplifier made its gain depend on the feedback path in the high-gain limit. At finite loop gain its sensitivity to the amplifier's gain is $1/(1+H)$. As $H$ grows, the closed-loop gain stops depending on $G_\mathrm{amp}$ and depends on $\rho$ instead (p. 3).¹ The value the loop holds is set by the part the designer controls.

Francis and Wonham drew the same line between the plant and the reference machinery. Uncertain plant parameters leave no offset, while a perturbed internal model leaves a small one (p. 461).⁶

The set point of Briat and colleagues' controller depends on two constants. Under their conditions, Briat and colleagues' mean output tracks the set point $\mu/\theta$, as a mean concentration, whatever the initial conditions and the other model parameters (p. 20).¹⁷ Here $\mu$ and $\theta$ are the reference and sensing constants, the core's names.

Gupta and Khammash made the line a definition. A robust perfect adaptation claim "presupposes many specifications", among them "how the set point for the output species is encoded in terms of network parameters" (p. 1).²⁰ They define maximal robust perfect adaptation as adaptation with respect to perturbations in virtually all network parameters, and note that earlier topological conditions "only guarantee insensitivity of the output set point to the designated input parameter" (pp. 1–2).

The course holds the reference constants fixed when it compares designs. The core's A16 asks "robust to what?" with $\mu$ and $\theta$ fixed, and finds that the incoherent feedforward loop keeps its input invariance while its set point moves with its four rate constants. The exposition's §7 tests input invariance, a fixed reference, a load and growth dilution as separate changes.

Before calling a set point robust, list the constants it depends on, and hold the reference constants fixed when comparing designs.

17Function, chemical implementation and operating limits connect the histories

The record establishes several claims at different levels. Internal model theorems give necessity under named hypotheses. Chemical constructions locate exact or approximate error balances. Experiments show adaptation, switching or compensation under particular perturbations and measurement precision. Frequency and noise analyses explain why matching a steady mean does not match the quality of regulation. A claim must remain at the level its evidence supports.

Reaction orders turn the historical comparison into a new analytical question. Saturation removes the stored concentration from a leading removal law. One-to-one pair removal cancels from a difference exactly. Bilinear removal lets a feedforward circuit cancel its common input. Sensor-only actuation can also add a proportional contribution to a local controller. Which dynamical conclusions follow from these rate dependencies, and which conditions make the dependencies chemically attainable? Lecture 11 takes the order-to-dynamics question first, then derives attainable reaction orders from binding conservation.

Five readings of the record deserve distrust. A necessity claim stated without its hypotheses. A count stated without its criteria, such as 395 without its tolerance. A sampled model ensemble read as a measurement. A choice among fitted models read as a unique wiring. And a claim of priority that the sources do not make.

The biological transfer remains an open argument. The Koshland group’s 1982 amplification-and-adaptation paper remains unread in this account, so this page makes no priority claim for joining those ideas.²¹ Takeda’s additional model variants lie in an unread supplement. More generally, a biological reporter need not read the pool that a model regulates. Natural controllers need not satisfy a selected saturation regime or outrun dilution. A Bode sensitivity integral does not determine chemical noise, and a matched local sensitivity curve does not determine a large-load response. Those gaps identify the measurements and hypotheses a new argument needs.

The exposition's six audit questions each come from one of these pieces. The table traces them. It is the reverse of this page's first table: there each source led to a question, and here each question leads back to its sources.

Place a new adaptation claim by naming its function, chemical implementation, measured output, operating conditions and kind of evidence. The research essay then asks which perturbation distinguishes the candidate explanations.

References

Every number on this page is either quoted from a source at the page or paragraph given, or marked as a present-day calculation. The calculations behind Figures 1 to 4, and the checks that compare each plotted curve with its formula, are kept with the Lecture 10 research records.

Lecture 10: core · exposition · lineage · research essay