The measurement is part of the stochastic model

Pairing, phase, exposure time and survival determine what a stochastic observation can tell us. CCBS · EST 5053 · Lecture 5 audience extensionEditorial synthesis of seven Week 3 student submissions · prepared with Codex · 22 September 2026 A broad molecule-count histogram does not tell us why the counts differ. Independent molecular events can broaden it. So can mixing cells with different expression rates, or measuring the same driven process at different phases. Those explanations make different predictions once we retain more information than one histogram. The Week 3 submissions turn this ambiguity into a sequence of experiments. A reporter pair identifies shared variation. A phase label exposes a hidden mixture. A time series recovers a clock, but its exposure window can hide fluctuations. A stopped history answers a decision or extinction question that no stationary histogram can settle. The common argument is that observation is part of the model. We must specify what is paired, pooled, integrated or conditioned on before interpreting the resulting statistic. More accurate simulation cannot repair a mismatch between the biological question and the quantity extracted from its paths. Names below link to the accepted source essays. Added derivations are editorial synthesis, not student quotations or an invented conversation. The core supplies the event model. The exposition develops its probability equations, and the lineage traces the measurements that motivated them. Every numerical example here is illustrative, not fitted to cellular data.

1A histogram does not name its cause

Pairing adds information that separate histograms discard. Álvaro Bátrez makes this visible by comparing differences within reporter pairs with differences among pair averages. Two reporters can rise together across cells while remaining close within each cell. Alternatively, the two readings can differ substantially within the same cell. The interpretation requires a model of what the reporters share. Let

X

and

Y

be calibrated readings from two matched reporters. Let

Z

specify their shared condition or relevant environmental history. Assume that, conditional on

Z

, the reporters are independent and have the same mean

m(Z)

and variance

v(Z)

These assumptions make “intrinsic” and “extrinsic” operational labels. Reporter-specific variation remains when the shared condition is fixed. Shared variation changes the conditional mean of both reporters. A fluctuating transcript shared by both reporters belongs to the second category for this construction. The same transcript would not be shared by reporters with independent transcripts.

Elowitz and colleagues implemented the two-reporter idea with distinguishable fluorescent proteins under matched regulatory control in E. coli. Swain, Elowitz and Siggia formulated the conditional decomposition that explains what such a measurement identifies. Neither result makes every correlated fluorescence signal a unique biochemical diagnosis.¹²

Specify the shared condition and reporter-matching assumptions before assigning a noise source to a joint measurement.

2What pairing identifies

The variance identities calibrate Álvaro's geometric picture. Define the reporter-specific contribution $V_\mathrm{int}=\mathrm E[v(Z)]$ and the shared contribution $V_\mathrm{ext}=\operatorname{Var}(m(Z))$ . The law of total variance and conditional independence give

\begin{aligned} \operatorname{Var}(X)&=V_\mathrm{int}+V_\mathrm{ext},\\ \operatorname{Cov}(X,Y)&=V_\mathrm{ext},\\ \frac12\mathrm E[(X-Y)^2]&=V_\mathrm{int}. \end{aligned}

The covariance follows because the conditional product mean is $m(Z)^2$ . The difference identity follows because the conditional mean difference is zero and its conditional variance is $2v(Z)$ . No Gaussian assumption is needed.

A pair average still contains reporter-specific noise. With $M=(X+Y)/2$ , the same identities give

\operatorname{Var}(M)=V_\mathrm{ext}+\frac12V_\mathrm{int}.

Álvaro already includes the corresponding subtraction in his calculation and caveat. Treating the entire pair-mean variance as shared noise would miss that correction. His absolute-distance comparisons are useful descriptive summaries, but the exact variance decomposition uses squared deviations.

Reporter mismatch can break the identification even when every variance is measured precisely. Unequal maturation, different regulation or shared detector noise can change the covariance. Independent detector noise contributes to the difference alongside reporter-specific chemistry. Pairing identifies conditional statistical components under the stated assumptions. Further controls are needed to name their molecular origin.

Use covariance and half the mean squared reporter difference to recover shared and conditional variance, and correct the pair-mean variance before interpreting it.

3Observation phase changes the ensemble

A mixture need not come from permanent differences among cells. Haobo Zheng asks what happens when a periodically driven process is observed at different phases. A complete-cycle time average and a fixed-phase ensemble answer different questions.

Let $N(t)$ be a molecule count. Single molecules enter at propensity $b(t)=b_0[1+\epsilon\sin(\omega t)]$ . Each molecule is independently removed at rate $\gamma$ . Here $b_0$ is an event rate, $0\leq\epsilon<1$ is the drive depth, and $\omega$ is the angular frequency. Assume $b_0,\gamma,\omega>0$ . These are effective count events, not a claim about elementary synthesis chemistry.

The mean satisfies the exact equation $\dot m=b(t)-\gamma m$ . Its attracting periodic solution is

\begin{aligned} m_\mathrm{per}(t)&=\mu+A\sin(\omega t-\delta),\\ \mu&=\frac{b_0}{\gamma},\\ A&=\frac{b_0\epsilon}{\sqrt{\gamma^2+\omega^2}},\qquad \delta=\arctan\frac{\omega}{\gamma}. \end{aligned}

The count is Poisson at each phase in the periodic regime. To see why, separate molecules already present from later arrivals. Independent survival thins a Poisson initial count into another Poisson count. Surviving arrivals from the prescribed Poisson input are independently Poisson. Their sum remains Poisson, with mean given by the equation above. An arbitrary initial distribution need not be Poisson during its transient, but its surviving initial population vanishes at long times.

A fixed-phase threshold probability therefore has a precise meaning. Define phase $\phi=\omega t$ modulo $2\pi$ and write its periodic mean as $m_\mathrm{per}(\phi)=\mu+A\sin(\phi-\delta)$ . The probability of at least four molecules at that phase is

p_4(\phi)=\Pr(N\geq4\mid\phi) =1-e^{-m_\mathrm{per}(\phi)}\sum_{n=0}^{3}\frac{m_\mathrm{per}(\phi)^n}{n!}.

Uniformly pooling phases gives $\bar p_4=(2\pi)^{-1}\int_0^{2\pi}p_4(\phi)\,d\phi$ . It is a single number, not the phase-dependent curve. Isolated phases may happen to equal that number.

Distinguish a fixed-phase ensemble from a phase mixture before comparing a time average with a population measurement.

4Pooling can create extra noise

Haobo's phase example connects directly to Álvaro's conditional decomposition. This editorial connection uses phase as the shared variable. At each phase, variance equals the Poisson mean. Define the Fano factor $F$ as variance divided by mean. Across uniformly sampled phases, total variance gives

\mathrm E[N]=\mu,\qquad \operatorname{Var}(N)=\mu+\frac{A^2}{2},\qquad F_\mathrm{pooled}=1+\frac{A^2}{2\mu}.

Pooling creates overdispersion even though every phase-conditioned distribution is Poisson. Its Fano factor is one at each phase, yet exceeds one in the pooled distribution whenever the periodic mean varies. Overdispersion alone therefore does not establish burst-like production. Mixing ordinary Poisson distributions is enough.

Haobo's illustrative rates make the size of this effect concrete. Take $b_0=2\,\mathrm{min}^{-1}$ , $\gamma=1\,\mathrm{min}^{-1}$ , $\epsilon=0.6$ and an eight-minute period. The mean count is 2 and the response amplitude is 0.9437. Uniform phase pooling gives $F=1.2227$ .

Figure 1. A threshold probability depends on when the count is read. The fixed-phase probability ranges from 0.0226 to 0.3402. Uniform phase pooling gives 0.1624. Replacing the periodic input by its average instead gives 0.1429. These are independently calculated model probabilities, not simulation estimates.

Averaging the input before calculating a nonlinear threshold probability is a different operation from averaging the resulting probabilities. The larger pooled value in this example is a calculated result. It is not a universal inequality for every threshold, mean and drive amplitude.

A long, residence-weighted trajectory samples the phase mixture, not an arbitrarily selected fixed phase. If the process is already in its periodic law, the expected occupation fraction over any integer number of complete cycles equals the phase average. Convergence of a single long record also relies on the process losing memory. The present addition–removal model does so. The difference between the two observables is not itself a failure of ergodicity.

Check whether unrecorded observation conditions can account for excess variance before assigning that excess to an additional molecular mechanism.

5Keep the input clock

A simulation of the driven process must preserve the changing input between events. Haobo makes this requirement explicit. When no event occurs, the count is fixed but the input propensity still changes. Freezing every propensity at its current value until the next jump would simulate a different process.

Suppose the current count is $n$ at time $t$ . The probability of no event during the next $s$ time units is $\exp[-H_t(s)]$ , where the integrated total hazard is

H_t(s)=(b_0+\gamma n)s+ \frac{b_0\epsilon}{\omega}[\cos(\omega t)-\cos(\omega(t+s))].

Draw a uniform random variable $U$ on $(0,1)$ and solve $H_t(s)=-\log U$ . At the resulting event time, select an addition with probability $b(t+s)/[b(t+s)+\gamma n]$ . The complementary probability selects removal. Anderson's time-dependent formulation derives these integrated-hazard and future-event channel rules.³

The observation rule must also survive the simulation. An occupation fraction weights each state by the time spent there. Counting each visited state once instead weights rapidly visited states differently. For a phase-conditioned analysis, split residence intervals at the relevant phase-bin boundaries. Exact event timing does not by itself guarantee the right estimator.

Preserve the prescribed time dependence in event sampling and the intended time weighting in the statistic extracted from the path.

6A stationary histogram erases the clock

Stationary occupancy compares residence-weighted inputs, not association speed alone. Fei Wang makes this distinction with two buffered signals competing for identical targets. A free target enters the A-bound state at rate $a_A$ or the B-bound state at rate $a_B$ . Clearance of the bound signal recycles the target at rates $d_A$ and $d_B$ . All rates have units of inverse time.

Balance of each occupied state gives its occupancy probability relative to the free state. If those probabilities are $p_A,p_B,p_0$ , then

a_Ap_0=d_Ap_A,\qquad a_Bp_0=d_Bp_B,\qquad \frac{p_A}{p_B}=\frac{a_A/d_A}{a_B/d_B}.

A twofold association advantage can therefore coexist with a twofold occupancy disadvantage if clearance is fourfold faster. This is a driven replacement model, not a measurement of reversible binding affinity.

Occupancy is also different from throughput. The steady clearance flux ratio is $d_Ap_A/(d_Bp_B)=a_A/a_B$ . A target can spend less time in a state while processing more events through it.

Even an entire stationary distribution can leave the absolute clock unidentified. Siheng Wang holds the ratio of constant addition rate $b$ to per-molecule removal rate $d$ fixed. The stationary count is Poisson with mean $\mu=b/d$ . Scaling both rates leaves every histogram bin unchanged.

A lagged measurement restores the missing information. Given the current count, the conditional future mean is $\mathrm E[N(t+u)\mid N(t)]=\mu+[N(t)-\mu]e^{-du}$ . Multiply its centered value by the present fluctuation and average at stationarity:

C_N(u)=\operatorname{Cov}[N(t),N(t+u)]=\mu e^{-du},\qquad u\geq0.

Siheng's three systems all have mean 20, but removal rates 0.2, 1 and 5 per minute give correlation half-times 3.47, 0.693 and 0.139 minutes. A time-lapse record can identify $d$ , then recover $b=\mu d$ . Hidden promoter states or detector dynamics could produce additional clocks, so a single-exponential fit is also a test of this particular model.

Use lagged correlations to identify a clock that a stationary histogram cannot contain, and do not confuse occupancy with event throughput.

7The detector changes the statistic

A recorded time series need not consist of instantaneous counts. Xiaomeng Zhu keeps the molecular dynamics fixed and changes only the measurement window. Let $P(t)$ be a stationary protein count with mean $\mu$ and covariance $C_P(u)$ . An ideal detector reports the trailing average

Y_T(t)=\frac1T\int_{t-T}^{t}P(s)\,ds.

Averaging preserves the mean but integrates correlations between all pairs of times inside the exposure. Stationarity makes the covariance depend only on their separation. Grouping those pairs by separation gives

\mathrm E[Y_T]=\mu,\qquad \operatorname{Var}(Y_T)=\frac{2}{T^2}\int_0^T(T-u)C_P(u)\,du.

Xiaomeng's example uses an independently switching binary promoter. ON and OFF rates are each $0.5\gamma$ , ON-state protein addition is $20\gamma$ , and each protein is removed at rate $\gamma$ . The stationary mean is 10 and variance is 60. With time measured in units of $1/\gamma$ , the covariance is $C_P(u)=(60+50u)e^{-u}$ . The extra term records promoter memory.

Figure 2. The same process looks quieter through a longer exposure. Independently integrating Xiaomeng's covariance gives measured variances 59.211, 31.421 and 10.200 at windows of 0.2, 5 and 20 protein lifetimes. The underlying instantaneous variance stays 60. These are model expectations, not estimates from the student's simulated trajectories.

This effect is different from sampling instantaneous values less frequently. Under stationarity, sparse instantaneous sampling preserves their marginal variance but loses temporal resolution. Exposure integration changes the marginal variance of the reported value itself. A smooth record can therefore reflect a detector's integration window without any reduction in molecular noise.

The kernel also connects Xiaomeng's example to Siheng's hidden clocks. A longer correlation time gives more strongly correlated count pairs inside the same exposure and hence less variance suppression. Once the observation window is known, measurements at several windows can test a proposed covariance model. They do not establish a unique biochemical mechanism without further controls.

Compare a measured variance with the model after applying the actual observation window, not with the instantaneous count variance by default.

8A decision is a stopped history

A first-arrival decision cannot be read from a snapshot histogram. Chengwei Jin asks whether a slower molecular signal can win a race. The answer depends on the path up to the decision, not merely on which source has the larger mean.

Let two independent Poisson sources add molecules at constant rates $0<a<b$ . Both counts start at zero, neither has removal, and both use the same positive integer threshold $K$ . The decision time $T$ is the first time either count reaches $K$ .

The merged event stream has rate $a+b$ . Each event belongs to the slower source with probability $p=a/(a+b)$ , independently of event time, and to the faster source with probability $q=1-p$ . For the slower source to win while the other count is $j<K$ , the final event must be its own. Counting the preceding event orders gives

\Pr(\text{slower wins})= \sum_{j=0}^{K-1}\binom{K+j-1}{j}p^Kq^j.

The same construction gives the waiting cost. Define $w_j=\binom{K+j-1}{j}(p^Kq^j+q^Kp^j)$ , the probability that either source wins after exactly $K+j$ events. Each event interval has mean $1/(a+b)$ , so

\mathrm E[T]=\frac{1}{a+b}\sum_{j=0}^{K-1}(K+j)w_j.

Figure 3. More evidence takes longer to accumulate. At threshold 1, the slower source wins with probability 0.4 after a mean of 0.4 minutes. At threshold 20, its winning probability is 0.1021 and the mean decision time is 13.09 minutes. Exact sums were checked independently by backward recursion on the absorbing count states.

The mean curves predict only that the faster source accumulates more molecules on average. They do not rule out the slower winner in Chengwei's experiment. Increasing the common threshold makes that outcome less likely in this model, at a cost in decision time.

This is a specified threshold readout, not a complete biochemical decision circuit. Removal would allow counts to decrease. Feedback, coupled sources or unequal thresholds would change the race. The calculation isolates a speed–reliability tradeoff under explicit assumptions without claiming a universal biological law.

Use a first-passage calculation for a first-arrival decision, and state the accumulation and stopping rules that make the calculation valid.

9Surviving is not the same as being stationary

A population that appears stable can still have a finite extinction time. Leo Sin asks what becomes of a stable logistic fixed point when the count can hit zero. His question changes the stopping event from reaching a high threshold to losing the final copy.

Let the count $N$ increase by one at rate $\lambda_n=bn$ and decrease by one at rate $\mu_n=dn+cn^2$ , with $b>d>0$ and $c>0$ . The deterministic equation $\dot x=(b-d)x-cx^2$ has an attracting positive level $K=(b-d)/c$ . These are declared population-event laws, not an elementary reaction mechanism.

Zero is absorbing because no addition can occur without an existing individual. In this density-regulated chain, removal eventually dominates addition strongly enough to prevent indefinite escape to larger counts. Extinction occurs with probability one. Absorbing-state reachability alone would not establish that conclusion for an arbitrary infinite-state process.

The useful prediction is the lifetime, not simply the eventual outcome. Define $T_n$ as the mean time to first reach zero from count $n$ . Conditioning on the next event gives

-1=\lambda_n(T_{n+1}-T_n)+\mu_n(T_{n-1}-T_n),\qquad T_0=0.

This backward equation is the extinction counterpart of Chengwei's race. Doering, Sargsyan and Sander derive discrete extinction-time expressions and their large-population asymptotics. Their analysis also explains why a diffusion approximation accurate near a stable level need not predict the rare route to extinction accurately.⁴

Figure 4. Stability can buy a long lifetime without preventing extinction. For

b=2

d=1

and

c=1/K

, independently computed mean lifetimes from

N(0)=K

are 25.70, 5354.66 and

5.55\times10^{12}

K=10,30,100

. Time is measured in inverse baseline-removal units. Positive backward recurrences agree at two high-count cutoffs and with 60-digit arithmetic. The curve is a numerical model result, not a measured population lifetime.

A histogram restricted to survivors answers a different question from the unconditional distribution. The surviving population can settle into a nearly time-independent conditional shape near $K$ while its total surviving probability declines. This is quasi-stationarity. The unconditional long-time probability instead accumulates at zero. Retaining only surviving trajectories can therefore hide the event that matters most.

Adding immigration changes the boundary and the conclusion. Zero then ceases to be absorbing. A constitutively transcribed RNA also need not require an existing transcript to restart production. The extinction claim belongs to the specified self-replicating, density-regulated process, not every low-copy molecular species.

Report survival probability or first-passage time when extinction matters, and label a survivor-conditioned histogram as conditional.

10Choose a discriminating observation

The useful next measurement separates plausible explanations. Pairing and phase labels test shared variation. Lagged correlations identify clocks. Exposure duration determines how those clocks enter the recorded variance. First-passage records keep decisions and extinctions visible. Increasing precision on a stationary histogram cannot recover information that the observation discarded.

The first two ideas combine into a concrete experiment. Record matched reporter pairs under the periodic input and retain each observation's phase. If the reporters are conditionally independent Poisson counts with the mean used above, uniform phase pooling predicts reporter covariance $A^2/2=0.4453$ . At an exactly fixed phase, that covariance is zero. The expected half squared difference remains $\mu=2$ after uniform phase averaging.

The detector must resolve the time dependence this test uses. Record exposure durations and compare the model after the same temporal integration. Finite phase bins also retain mixture variance, so narrowing the bins supplies a separate control. If shared covariance persists after those corrections, the phase-only explanation is insufficient. If it disappears as predicted, the experiment has identified an observation-induced component without inventing an extra burst mechanism.

Keep the pairing, phase, exposure and survival information that distinguishes competing models, then test the variation left after the predicted contribution is removed.

Models and provenance

This chapter develops seven contributions from the accepted Week 3 snapshot checked on 22 September. The Lecture 6 audience chapter develops the other eight. Together they cover all 15 accepted essays at this cutoff. Attribution identifies intellectual inputs, not student coauthorship or approval of the editorial prose.

The instructor's project retains source checksums, a complete contribution ledger, inspected literature and independent calculations under lectures/research/week03-extension-synthesis/. Forward probability equations check the periodic Poisson law. Quadrature checks phase mixing and integrated hazards. Absorbing-state recursion checks the race. No student code was executed.

References

M. B. Elowitz, A. J. Levine, E. D. Siggia and P. S. Swain, “Stochastic gene expression in a single cell,” Science 297, 1183–1186 (2002). DOI. Matched two-reporter construction and noise measurements, especially p. 1184.
P. S. Swain, M. B. Elowitz and E. D. Siggia, “Intrinsic and extrinsic contributions to stochasticity in gene expression,” PNAS 99, 12795–12800 (2002). DOI. Conditional decomposition and paired measurements, p. 12796, equations 2–9.
D. F. Anderson, “A modified next reaction method for simulating chemical systems with time dependent propensities and delays,” Journal of Chemical Physics 127, 214107 (2007). DOI. Section V, especially equations 7–8 and the event-channel probabilities at the sampled future time.
C. R. Doering, K. V. Sargsyan and L. M. Sander, “Extinction times for birth-death processes: exact results, continuum asymptotics, and the failure of the Fokker–Planck approximation,” Multiscale Modeling & Simulation 3, 283–299 (2005). DOI. Inspected 2004 preprint, pp. 3–7: logistic rates, backward equation, discrete solution and superthreshold lifetime scaling.