Fast enough for which prediction?

A defensible reduction preserves a specified prediction under a specified preparation, input and observation.

CCBS · EST 5053 · Lecture 6 audience extension
Editorial synthesis of eight Week 3 student submissions · prepared with Codex · 22 September 2026

A reduced model can fit an endpoint and still miss the experiment. It may reach that endpoint too quickly, predict covalent capture before binding has occurred, or reproduce the mean while discarding most of the variance. Agreement must therefore be judged against a named observable.

The Week 3 contributions expose several ways a familiar summary can conceal failure. A final abundance leaves recovery speed undetermined. Affinity does not fix capture time. A good amplitude fit can hide phase lag, and an exact free-protein law can fail when fed a total-protein measurement. Even after those checks, a short molecular episode may leave a random packet that a downstream consumer can detect.

We will follow these failures from a hidden binding clock to a downstream output. At each step, the missing information becomes a concrete validation test. The core separates exact accounting from fast-state closure. The exposition supplies broader derivations, and the lineage puts the reduction programmes in context.

Names link to accepted submissions. Added connections and error criteria are editorial synthesis. They do not imply a classroom exchange or student approval of this prose. All numerical rates below are illustrative model choices, not fitted biological measurements.

1Equal levels leave a clock undetermined

A steady abundance determines a ratio, not a turnover rate. Changkai Li isolates this distinction with constant supply and linear removal. Let $x$ be abundance, $k$ the supply per unit time and $\gamma$ the per-capita removal rate. Starting from zero,

Fixing $k/\gamma=100$ makes every trajectory end at 100. The time to reach half that value is $t_{1/2}=\log(2)/\gamma$. Changkai's removal rates 0.2, 1 and 5 per chosen time unit give half-times 3.47, 0.693 and 0.139 time units. The shared endpoint leaves a 25-fold difference between the slowest and fastest recovery.

The extra measurement needed is a clock measurement. A time course after a controlled change can estimate $\gamma$. Combined with the final level, it determines $k$. An endpoint alone cannot distinguish those parameter pairs.

This example does not yet eliminate a state. It establishes why an endpoint fit cannot validate an elimination. We next place the hidden clock inside a process whose output is formed only after a binding event.

Separate the ratio fixed by a steady abundance from the rate fixed by a transient measurement.

2Equal affinity does not fix capture

A binding equilibrium predicts occupancy, not the speed of irreversible capture. Xinyu Wu compares two binders with the same affinity but different association and dissociation rates. His observable is the fraction captured by a finite time.

Consider one reactive molecule and a partner maintained in excess at concentration $[B]$. The molecule has three states: unbound, reversibly bound, and irreversibly captured. Let their probabilities be $p_U,p_C,p_X$. The unbound-to-bound transition has rate $a=k_\mathrm{on}[B]$, unbinding has rate $b=k_\mathrm{off}$, and capture from the bound state has rate $c$. All three rates are positive and have units of inverse time.

These are transitions between modeled states. They do not specify the elementary chemistry of a crosslinker. With no side reactions and an initially unbound molecule, the probability equations are

In the reversible binding submodel without capture, the bound probability is $p_\mathrm{bound}=a/(a+b)$. Holding the dissociation constant $K_d=k_\mathrm{off}/k_\mathrm{on}$ and partner concentration fixed preserves that probability. Multiplying both binding rates by a common factor leaves affinity unchanged while changing the binding relaxation time $1/(a+b)$.

The rapid-equilibrium approximation assumes that the uncaptured population continually has this bound fraction. It uses $p_C\approx p_\mathrm{bound}(1-p_X)$, giving

The model boundary matters for real crosslinking. Kammer and colleagues explicitly include hydrolysis, mono-link formation and competing crosslinks. Those pathways can change both yield and its dependence on rates. Xinyu's three-state construction isolates the binding-clock question. It does not claim to reproduce that full chemistry.¹

Define capture as a finite-time state probability, and distinguish equal binding occupancy from equal capture kinetics.

3Calculate the missing binding wait

The initial binding wait produces a measurable difference even at fixed affinity. Take Xinyu's $c=0.1\,\mathrm{s}^{-1}$ and $p_\mathrm{bound}=1/2$. Compare $a=b=0.01\,\mathrm{s}^{-1}$ with $a=b=1\,\mathrm{s}^{-1}$. At 60 seconds, the exact captured probabilities are 0.3663 and 0.9450. Rapid equilibrium predicts 0.9502 for both.

The mean capture time exposes the omitted wait without solving the full time course. Let $T_U$ and $T_C$ be the expected remaining times to capture from the unbound and bound states. First wait for the next state transition, then add the remaining time:

The first equation includes the initial association wait. The second includes a possible unbinding and restart. Solving gives

The rapid-equilibrium mean is 20 seconds in both cases. The exact means are 120 and 21 seconds. Their difference is not an affinity difference or an eventual-yield difference. It is time spent reaching the reactive state under the specified preparation.

Use first-step waiting equations to identify the delay that an occupancy-based approximation omits.

4Turn “fast” into a requirement

An error tolerance converts the clock comparison into a quantitative acceptance test. Building on Xinyu's exact mean, define relative error against the full initially unbound mean:

To require $e_\mathrm{mean}\leq\eta$, with $0<\eta<1$, it is necessary and sufficient in this model that

A 5% tolerance therefore requires a binding relaxation rate at least 19 times the capture rate. The fast example has ratio 20 and relative error 4.76%. The slow example has ratio 0.2 and error 83.33%. The criterion says exactly which approximation is acceptable for this mean-time question.

This criterion does not certify the whole capture curve. Relative error at very early times can remain large because the exact initial capture rate is zero. A finite-time yield tolerance needs its own curve comparison.

Starting from a pre-equilibrated mixture changes the mean-time criterion too. If its initial bound fraction is $p_\mathrm{bound}$, the extra wait above $T_\mathrm{RE}$ becomes $(1-p_\mathrm{bound})/a$, not $1/a$.

Attach every “fast enough” inequality to its observable, normalization and initial preparation.

5A moving target supplies another clock

A fast relaxation must be compared with the changing input, not just with another internal rate. Zixuan Wang gives this requirement an exact error budget. Let $z(t)$ track a target $u(t)$ with relaxation time $t_\mathrm{relax}$:

Here $z$ and $u$ are equally scaled deviations around positive reference states, not concentrations that become negative. For small perturbations of independent receptor binding, $t_\mathrm{relax}=1/(k_\mathrm{on}S_0+k_\mathrm{off})$, where $S_0$ is the reference free-ligand concentration. This connection is a local linearization with externally imposed free ligand.

After startup transients decay, define the dimensionless forcing speed $r=\omega t_\mathrm{relax}$. The exact gain is $G=1/\sqrt{1+r^2}$ and the phase lag is $\arctan r$. Quasi-steady tracking sets $z=u$ and loses both effects.

Zixuan measures waveform error over one settled period, normalized by the input's RMS amplitude. Denoting a full-period average by angle brackets,

The last inequality assumes $T=2\pi/\omega$ and $0<\eta<1$. A 10% tolerance requires a period at least 62.52 relaxation times. This is a result for the stated waveform metric and single-frequency linear model, not a universal “63 times faster” rule for enzymes.

Weihan Luo supplies a complementary moving-target example: $\dot s=1-s$ and $\epsilon\dot f=s-f$, starting at zero. Time is scaled to the slow relaxation time. At time one, $s=0.6321$. The fast state is 0.6128 for $\epsilon=0.05$, but only 0.2642 for $\epsilon=1$. Calling a variable “fast” does not establish that it follows its target.

The connection to fast-weight neural memory raises a further question: is the lag an error or the function? Ba and colleagues deliberately retain recent activity through rapidly changing associative weights. Their fast weights evolve more quickly than learned weights, but are not simply an instantaneous function of the current input. Weihan's tracking toy illustrates a condition for elimination. It does not justify deleting the history that a memory architecture was built to store.⁴

Test the input bandwidth and ask whether the fast state's retained history is negligible for the task, rather than eliminating it because of its name.

6Fast binding does not identify the measured pool

An accurate fast-state reduction can still fail at the observation map. Kehan Lang makes the course's free-versus-total warning quantitative. Let $R$ be free repressor monomer, $C_{RR}$ free dimer, $D$ free promoter and $C_{DRR}$ promoter-bound dimer. Assume rapid reversible equilibration and define dissociation scales by

Promoter conservation is $q_D=D+C_{DRR}$. Protein conservation counts two monomer constituents in every free or bound dimer. Together the constraints give

The promoter law is exactly degree two in free monomer within this reduced model. The total map is strictly increasing, so a measured $q_R$ determines one physical $R$. But using $q_R$ directly in the promoter law drops both storage terms. Fast binding does not make those terms small.

This comparison separates two approximations often bundled together. A fast-equilibrium assumption reconstructs occupancy from free species. A monomer-dominance assumption permits replacement of free protein by total protein. Failure of the second assumption does not refute the first or the dimer mechanism. Nor does a fitted Hill exponent alone identify that mechanism.

Reconstruct the free species from the conserved total before applying a free-species regulatory law, unless pool dominance has been checked.

7The mean is the easy observable

A reduction that preserves the mean may still discard substantial noise. Peng Su tests promoter averaging by its stationary Fano factor. His question is not whether arbitrarily fast switching can average out. It is how much speed a specified noise tolerance needs at the actual production amplitude.

Define a binary promoter state $G\in\{0,1\}$. It switches ON at rate $\alpha$ and OFF at rate $\beta$, independently of protein count. When ON, it adds single proteins at propensity $\rho$. Each protein is removed at rate $\gamma$. All four rates are positive. This effective state-and-count model is the switching construction studied by Peccoud and Ycart. It does not resolve transcription and translation separately.²

The stationary ON fraction is $p_\mathrm{on}=\alpha/(\alpha+\beta)$. If $N$ denotes protein count, stationary mean balance gives $\mu=\mathrm E[N]=\rho p_\mathrm{on}/\gamma$. Replacing the promoter by the constant propensity $\rho p_\mathrm{on}$ preserves this mean exactly at every switching speed.

The variance reveals the missing correlation. Let $C=\operatorname{Cov}(G,N)$ at stationarity. Applying the event generator to $GN$, using $G^2=G$, gives

Protein accumulation remembers recent promoter activity. Its variance balance retains that memory through $C$. Define the Fano factor $F=\operatorname{Var}(N)/\mu$. Then

The constant-rate reduction predicts a Poisson distribution and $F=1$. It matches the full mean while omitting the entire positive second term. The comparison is exact for this model, without a Gaussian approximation or a moment-closure assumption.

Check a fluctuation statistic when a reduction already preserves stationary mean balance.

8Amplitude sets the noise threshold

Peng's amplitude scan shows why a clock ratio alone is not a finite-error guarantee. Define the switching relaxation time $t_\mathrm{switch}=1/(\alpha+\beta)$, the protein lifetime $t_\mathrm{protein}=1/\gamma$, and the dimensionless production amplitude $R=\rho/\gamma$. Then

At $p_\mathrm{on}=1/21$ and $\epsilon=0.01$, switching relaxes 100 times faster than protein removal. Yet amplitudes $R=1,10,100,200$ give Fano factors 1.0094, 1.0943, 1.9430 and 2.8859. The same clock separation meets some noise requirements and misses others.

Specify an absolute Fano-excess tolerance $F-1\leq\tau$, with $\tau>0$. The exact acceptance condition is

If the right side is negative, every positive switching speed satisfies that particular tolerance. For Peng's $R=200$, $p_\mathrm{on}=1/21$ and $\tau=0.05$, the required ratio is 3808.5.

The required acceleration can be large even for an initially fast promoter. Peng's baseline uses $\rho=6\,\mathrm{s}^{-1}$, $\gamma=0.03\,\mathrm{s}^{-1}$, $\alpha=0.015\,\mathrm{s}^{-1}$ and $\beta=0.3\,\mathrm{s}^{-1}$. Their switching-to-removal ratio is 10.5. Reaching the chosen tolerance requires multiplying both switching rates by about 362.7 while preserving their ratio. The resulting switching relaxation time is about 8.75 milliseconds.

The error metric changes the conclusion's meaning. With equal means, the relative variance error of the Poisson reduction is $(F-1)/F$. An absolute Fano excess of 0.05 therefore means 4.76% relative variance error, not exactly 5%. A true 5% relative variance tolerance instead requires a rate ratio of 3618.0 in this example.

Increasing production also does not worsen every fluctuation metric. The squared coefficient of variation is variance divided by squared mean. Its excess above the Poisson value is

At fixed ON fraction and switching speed, this excess is independent of $\rho$. Peng's warning concerns finite accuracy for a named statistic. It does not say that production amplitude defeats fast switching under every normalization.

Choose the noise metric first, then solve for the switching rate that meets its tolerance at the declared production amplitude.

9A different limit retains packets

Fixed-amplitude averaging remains valid as switching becomes arbitrarily fast. If $R$ and $p_\mathrm{on}$ stay fixed while $\epsilon\to0$, Peng's formula gives $F\to1$. The finite-rate example does not contradict that limit. It shows why the convergence is not uniform over unrestricted production amplitudes.

A distinct joint scaling can preserve finite output from an ever-shorter active episode. During one ON episode, production events occur at rate $\rho$ and deactivation occurs at rate $\beta$. Let $B$ count proteins produced before deactivation. Competing exponential clocks give

Keep $\alpha,\gamma$ and $b_\mathrm{pkt}$ fixed while increasing $\rho$ and $\beta$ together. Active durations shrink to zero, but packet sizes remain geometrically distributed. This is not the previous family: the ON fraction now tends to zero while ON-state production diverges. The limit has packet arrivals at rate $\alpha$, mean count $\alpha b_\mathrm{pkt}/\gamma$ and Fano factor $1+b_\mathrm{pkt}$.

Runzi Yan moves the hidden clock from a promoter episode to an RNA lifetime. Translation at rate $k_p$ competes with RNA removal at rate $\gamma_m$. Counting translation events before removal gives a geometric packet with mean $\bar B=k_p/\gamma_m$ and variance $\bar B(1+\bar B)$. Speeding both clocks at fixed ratio shortens the episode without shrinking the packet.

Packet size and packet arrival rate are separate quantities. With independent RNA arrivals at rate $r_m$, linear protein removal at rate $\gamma_p$, and a short-RNA limit at fixed $\bar B$, proteins receive instantaneous geometric packets at rate $r_m$. Shorter RNA lifetime does not itself accelerate those arrivals. Shahrezaei and Swain's fast-RNA analysis retains the finite packet scale and excludes the unresolved early RNA transient.³

The count observed at one time is not the cumulative output of one RNA. In the full unregulated two-stage model, stationary protein moments give

Only when $\gamma_m\gg\gamma_p$ at fixed packet size does the count Fano approach $1+\bar B$. For Yan's illustrative $k_p=2\,\mathrm{min}^{-1}$ and $\gamma_m=0.2\,\mathrm{min}^{-1}$, the packet mean is 10 and packet variance is 110. The stationary protein Fano also needs $\gamma_p$. For example, $\gamma_p=0.02\,\mathrm{min}^{-1}$ gives 10.09, approaching 11 only in the stronger separation limit.

A different scaling erases the packets. Hold translation fixed, increase RNA removal, and increase the RNA source to preserve mean protein production. Individual packet sizes then collapse toward zero, with rare single-protein events. The resulting protein arrivals approach a Poisson process. This is not a geometric packet becoming a Poisson-distributed packet. It is a limiting arrival process assembled from many increasingly unproductive RNA episodes.

These alternatives show what a noise measurement can reject without uniquely identifying a cause. A stationary Fano incompatible with the specified two-stage model could reflect promoter switching, shared environmental variation, feedback or measurement effects. Matching the geometric packet law likewise cannot identify whether a promoter episode or RNA lifetime produced it.

State what is held fixed in a fast limit, and retain the random output of a short episode when that output remains finite.

10A downstream consumer can notice the packets

Preserving mean supply need not preserve mean function. Shunkang Xue tests this consequence with a noisy carrier pool feeding a nonlinear consumer. The question follows directly from the packet limit: if the reduced source retains only average input, what does the next process lose?

Let $E$ count available carrier molecules and $Y$ cumulative useful product. Unit additions arrive at rate $\bar J$, or fixed packets of $B$ carriers arrive at rate $\bar J/B$. Both have the same expected input. A work event consumes one carrier and adds one product at propensity $v(E)=V_\mathrm{max}E/(K_E+E)$. Leakage removes a carrier at propensity $\ell E$.

The mean output depends on the distribution of the carrier pool. The exact expectation equations are

The deterministic approximation replaces $\mathrm E[v(E)]$ by $v(\mathrm E[E])$. The saturating law is concave, so fluctuations lower average consumption relative to evaluating the same law at the same pool mean. Comparing two dynamical systems still requires solving their distributions: their pool means can change too. Concavity alone does not determine the difference between every pair of burst sources.

Shunkang includes the necessary unit-event control. It separates a burst-size effect from the difference between any discrete event source and a noiseless ODE. For $\bar J=4\,\mathrm{min}^{-1}$, $V_\mathrm{max}=6\,\mathrm{min}^{-1}$, $K_E=10$ carriers and $\ell=0.08\,\mathrm{min}^{-1}$, start every system at zero and observe for 60 minutes.

These values are independent editorial calculations from forward probability equations and an ODE, not reruns of the submitted simulation. The packet source spends more time near depletion and makes less useful product in this example. No cellular measurements enter the comparison, and a different consumer law could change the direction of the effect.

A linear consumer provides a sharper control. If $v(E)=kE$, matched mean input and initial mean give identical expected carrier and product trajectories for all packet sizes. This equality is exact because the mean equations close. No additional fast-clock or small-variance assumption is needed for that mean-output statement. Threshold risk and variance can still differ.

Test a reduced source through the downstream observable that matters, and retain packet statistics when a nonlinear or thresholded consumer is sensitive to them.

11Validate the prediction you need

A useful reduction test reaches beyond the summary used to fit the model. The clock comparisons hold steady levels or affinities fixed. The forcing comparison exposes phase error. The dimer example checks the measured coordinate. The packet and consumer examples test whether discarded fluctuations change an output. Together they specify what “accurate” must mean before a reduced model can earn that description.

The capture experiment requires controlled preparation and a clock. Start the two equal-affinity systems unbound and compare their capture fractions over time. A pre-equilibrated preparation tests a different prediction. Eventual complete capture cannot distinguish the clocks in the three-state model, so it is not a substitute for the time course.

The promoter experiment requires matched means and measured switching. Hold ON fraction, production and removal fixed while changing both switching rates together. The model predicts an unchanged mean but a declining Fano factor. Fit those rates independently if the purpose is to test the prediction rather than merely fit the noise curve.

A next experiment can then separate amplitude from time. Repeat the switching scan at a second production rate. The Fano excess should scale with production, while the excess squared coefficient of variation should not. Apply the detector's integration window before comparing either prediction with a recorded signal, as the Lecture 5 audience chapter shows. Failure of those paired predictions after observation and shared-variation controls would reject this independent binary-promoter description even if every mean still agreed.

Validate a reduction with a specified preparation, observation window and error metric, then choose a second observable that could falsify the retained mechanism.

Models and provenance

This chapter develops eight contributions from the accepted Week 3 snapshot checked on 22 September. The Lecture 5 audience chapter develops the other seven. Together they cover all 15 accepted essays at this cutoff. Attribution identifies intellectual inputs, not student coauthorship or approval of the editorial additions.

Independent calculations, source checksums, full-text passage records and the contribution ledger remain in the instructor's project under lectures/research/week03-extension-synthesis/. Matrix exponentials check capture curves. Backward equations and survival integration check capture means. Stationary generator equations check promoter moments. No student code was executed.

References

1. K.-M. Kammer, T. Eisgruber, P. Heid, R. Pellarin and F. Stengel, “Kinetic principles of chemical cross-link formation for protein–protein interactions,” PNAS 121, e2402040121 (2024). DOI. Pages 2–3 and Figure 1 specify hydrolysis, mono-links and competing crosslink pathways omitted from the three-state teaching model.
2. J. Peccoud and B. Ycart, “Markovian modelling of gene product synthesis,” Theoretical Population Biology 48, 222–234 (1995). DOI. Pages 223–224 define the binary gene state, protein transitions and master equations.
3. V. Shahrezaei and P. S. Swain, “Analytical distributions for stochastic gene expression,” PNAS 105, 17256–17261 (2008). DOI. The two-stage analysis on pp. 17256–17257 specifies the lifetime separation, finite burst parameter and transient-time qualification.
4. J. Ba, G. Hinton, V. Mnih, J. Z. Leibo and C. Ionescu, “Using fast weights to attend to the recent past,” Advances in Neural Information Processing Systems 29 (2016). Proceedings full text. Sections 1 and 3, especially equations 1–4: temporary associative history and its distinct role from slow learned weights.