Introduction
Fouling, the slow accumulation of material at the interface between a flowing phase and a solid, is monitored in heat exchangers through the fouling resistance. Its growth costs fuel, design margin and production lost to cleaning [1 , 2 , 3 ] , and the operator’s question is when it will reach the limit at which cleaning becomes necessary [4 , 5 ] . In operation the resistance is not measured but reconstructed from temperatures and flow rates through the exchanger’s energy balance [6 , 7 ] , so that its change from one sampling period to the next is typically smaller than its measurement error. Decades of work have identified the contributing processes in detail, such as particle transport and attachment, crystallization, chemical reactions and ageing [3 , 8 , 9 , 10 ] . Fouling records are often still read against two main models, shown in Figure 1 . Kern and Seaton assume deposition at a constant rate and removal proportional to the deposit, giving an asymptotic rise to a plateau [11 ] ; Ebert and Panchal set a thermally activated deposition rate against a shear-dependent term, giving a threshold above which the resistance grows linearly [12 ] . Records are routinely sorted as linear, falling-rate or asymptotic [1 , 2 , 3 , 9 ] , while subsequent work has concentrated on what sets the rate: particle transport and attachment [13 , 14 , 15 , 16 , 17 , 18 ] , crystallization and wax deposition [19 , 20 ] , chemical-reaction fouling [21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 ] , ageing [31 , 32 , 33 , 34 , 35 , 36 , 37 ] and parameter estimation or data-driven prediction from plant data [38 , 39 , 40 , 41 ] . Replicate fouling runs under identical conditions scatter widely [42 ] , and probabilistic treatments attach this scatter to the deterministic laws as random growth parameters, from which they derive time-to-cleaning distributions and maintenance policies [43 , 44 , 45 , 46 ] ; other stochastic treatments [47 , 48 , 49 , 50 ] are mostly simulations. None derives the deterministic laws from the stochastic deposition, or asks what the distribution of the fouling measure allows a record to identify.
Figure 1. Prototypical fouling-resistance trajectories. Both laws show an induction period t i t_i t i before the resistance grows, after which the Kern–Seaton law saturates to a plateau R ∞ R_\infty R ∞ at a rate set by the removal process, while the Ebert–Panchal law grows approximately linearly once the deposition term dominates.
If fouling reflects many coupled processes, each with its own strong dependence on operating conditions, why do measured trajectories repeatedly collapse onto a small number of elementary shapes? The usual reading is that the shape identifies the dominant mechanism. In heat exchangers an asymptote is read as the signature of removal [51 , 52 ] , as the Kern–Seaton derivation suggests, and the saw-tooth pattern and scatter of plant data as evidence of it [53 ] ; yet Ebert and Panchal’s threshold concerns only the initial rate, and linear growth is not evidence that removal is absent [12 , 27 ] . The mechanistic readings have been debated—asymptotic data need not imply removal [1 ] , suppression and removal are indistinguishable during growth [37 ] , and likelihood comparison can select the wrong increment law [54 ] —but mostly through counterexamples and case studies, not a general statement of what a record can identify.
We show in this study that, under very general conditions, the classical laws are universal first-moment consequences of additive deposition processes. If deposition increments are stationary, independent and non-negative, the deposit is a Lévy subordinator and the two classical laws are first-moment identities: the mean depends only on the net deposition and removal rates, for every additive mechanism. The statement follows directly from the linearity of expectation, but its consequences are not obvious. The information absent from the mean lies in the fluctuations and is unequally accessible: removal mechanisms separate in a plateau variance that survives as the deposition becomes smooth, whereas the increment law separates only in higher cumulants that measurement noise attenuates, turning the noise-free sample size of model selection [54 ] into a record length that grows as
Nomenclature
caption:
:::
the fourth power of the noise. The dependence of the rate on operating conditions remains the proper object of mechanistic modelling, but it cannot be inferred from a single record at a single operating point. Section VII translates these results into what an operator can and cannot conclude from a monitored fouling resistance.
Framework and deterministic reference models
Fouling denotes the accumulation or growth of unwanted material on a surface [1 , 2 , 10 ] . We describe the state of a fouled surface by a nonnegative fouling measure R ( t ) R(t) R ( t ) , which is kept generic throughout the theory. Its principal application here is the thermal fouling resistance of a heat exchanger, R = 1 / U − 1 / U c l e a n R=1/U-1/U_{\mathrm{clean}} R = 1/ U − 1/ U clean where U U U is the overall heat-transfer coefficient. For a homogeneous deposit of thickness δ \delta δ and thermal conductivity k f k_f k f this is proportional to the deposit thickness R = δ / k f R=\delta/k_f R = δ / k f .
This framework describes deposition under approximately stationary operating conditions in which the deposit does not feed back on the process that creates it. Specifically, we assume that (i) the variables that set the deposition rate (flow, wall shear stress, film temperature, foulant concentration) are constant in time and uniform over the surface, so that R R R is an area-averaged scalar; (ii) the deposit does not appreciably change these variables, i.e. the loss of flow cross-section and the resulting changes in wall shear stress and interface temperature are negligible, such feedback being one route to an asymptotic record [55 ] ; (iii) R R R is additive in the deposited material, so that ageing, consolidation, and roughness effects that change R R R at fixed deposited mass are neglected [32 , 56 ] ; and (iv) removal, where present, is a deterministic function of the state. Removal triggered by the deposit state itself (e.g. thickness-dependent spalling) is not treated; Poisson-timed removal of a fixed fraction is (Sec. IV B). Additivity (iii) and the absence of feedback (ii) are what make the mean independent of mechanism; when (i)–(iv) fail, the net rate depends on the state and the mean laws of Table I no longer hold in general, as shown for nonlinear removal in Sec. IV C.
Deterministic fouling models
Fouling is conventionally modeled as a balance between the deposition and removal of material,
d R d t = R ˙ d − R ˙ r , (1) \frac{dR}{dt}=\dot{R}_{d}-\dot{R}_{r}, \tag*{(1)} d t d R = R ˙ d − R ˙ r , (1)
where R ˙ d \dot{R}_{d} R ˙ d and R ˙ r \dot{R}_{r} R ˙ r are the deposition and removal rates in units of R R R per unit time. Depending on the relative importance of deposition, adhesion, erosion, and detachment, observed fouling curves may be approximately linear, asymptotic, falling-rate, or irregular [2 , 10 ] . Two prototypical deterministic models used in the industry may provide useful reference limits: the asymptotic (Kern–Seaton) model [11 ] and the linear (Ebert–Panchal) model [12 ] .
Kern–Seaton. Deposition is proportional to the fluid velocity u u u and to the foulant concentration C C C , whereas removal grows with the wall shear stress τ w \tau_w τ w and with the amount of material already deposited. For particulate fouling, this reads as
R ˙ d = k d u C , R ˙ r = k r τ w R , (2) \dot{R}_{d}=k_d uC,\qquad\dot{R}_{r}=k_r\tau_w R, \tag*{(2)} R ˙ d = k d u C , R ˙ r = k r τ w R , (2)
where k d k_d k d and k r k_r k r are effective rate coefficients. The clean-surface deposition rate m = k d u C m=k_d uC m = k d u C may be interpreted as a mass-transfer-controlled capture rate of classical filtration theory under particulate-fouling assumptions, multiplied
by an attachment efficiency [13 , 14 ] ; deposition–removal models of this form for particulate and crystallisation fouling are developed in Refs. [15 , 16 ] and have been a staple of fouling modeling in the industry. The removal law assumes that R R R is proportional to the deposited mass, and in fully turbulent flow the wall shear stress is approximately proportional to u 2 u^2 u 2 , so that under stationary conditions
d R d t = m − ψ R , (3) \frac{dR}{dt}=m-\psi R, \tag*{(3)} d t d R = m − ψ R , (3)
with m = k d u C m=k_d uC m = k d u C and ψ = k r ′ u 2 \psi=k'_r u^2 ψ = k r ′ u 2 , where k r ′ k'_r k r ′ is k r k_r k r multiplied by the constant of proportionality between τ w \tau_w τ w and u 2 u^2 u 2 . The fouling resistance then rises monotonically to the plateau R ∞ = m / ψ R_\infty=m/\psi R ∞ = m / ψ with time constant ψ − 1 \psi^{-1} ψ − 1 [3 , 11 ] , and because removal is proportional to R R R the state cannot become negative.
Ebert–Panchal. In this family of models the net fouling rate is the difference between a thermally activated deposition contribution and a shear-dependent removal contribution [27 ] . Following the original derivation
d R d t = a R b exp ( − E a R g T f ) − γ τ w ≡ K E P . (4) \frac{dR}{dt}=aR^{b}\exp\left(-\frac{E_a}{R_gT_f}\right)-\gamma\tau_w\equiv K_{\mathrm{EP}}. \tag*{(4)} d t d R = a R b exp ( − R g T f E a ) − γ τ w ≡ K EP . (4)
Here E a E_a E a is the activation energy per mole, R g R_g R g the gas constant, T f T_f T f the film temperature, and a a a , b b b , γ \gamma γ are empirical parameters. The second term is interpreted as either a removal or suppression of the deposition rate, and it reduces the rate of sticking, i.e. the increase of R R R even without any removal mechanism [31 , 37 ] . The rate is accordingly clipped at zero, d R / d t = max ( K E P , 0 ) dR/dt=\max(K_{\mathrm{EP}},0) d R / d t = max ( K EP , 0 ) , so that R R R never decreases. In the stochastic description of Sec. III the two readings of the second term are no longer equivalent. Read as suppression , γ τ w \gamma\tau_w γ τ w lowers the fraction of arriving material that sticks; this is a thinning of the deposition process (Sec. V A, Eq. (27)). Read as erosion , it is a constant removal rate g ( R ) = γ τ w g(R)=\gamma\tau_w g ( R ) = γ τ w acting between deposition events, so that the path decreases during quiescent stretches and can reach R = 0 R=0 R = 0 , where it must be held. The two readings share the linear mean but not the variance, and the erosion reading is exact only away from the wall (Table I).
Several further models have been built on this basis [22 , 23 , 24 , 25 , 26 ] , and in all of them m m m and the suppression rate γ τ w \gamma\tau_w γ τ w are time-independent under stationary conditions, so that R R R grows linearly at the net rate K E P K_{\mathrm{EP}} K EP ; the condition K E P > 0 K_{\mathrm{EP}}>0 K EP > 0 defines the net-fouling regime, and below it no net accumulation is predicted. Both models are instances of d R / d t = m − g ( R ) dR/dt=m-g(R) d R / d t = m − g ( R ) with the same deposition rate m m m and different removal laws, g g g , and both mean trajectories are collected in Table I.
The deterministic models describe mean trajectories under fixed operating conditions. They do not determine the variability among nominally identical campaigns, the intermittency of deposition events, or the uncertainty in the time needed to reach a prescribed fouling threshold. The same mean trajectory may arise from frequent small deposition events or from rare, comparatively large bursts.
To represent this unresolved variability we keep the accumulation–removal structure of Eq. (1) but replace the smooth deposition increment by the increment of a non-negative stochastic process. Let D ( t ) D(t) D ( t ) be the cumulative deposition, with D ( 0 ) = 0 D(0)=0 D ( 0 ) = 0 , and let g ( R ) g(R) g ( R ) be a deterministic, state-dependent removal rate. The system is then governed by the Langevin-type equation:
d R ( t ) = d D ( t ) − g [ R ( t ) ] d t , R ( 0 ) = R 0 ≥ 0. (5) dR(t)=dD(t)-g[R(t)]\,dt,\qquad R(0)=R_0\ge0. \tag*{(5)} d R ( t ) = d D ( t ) − g [ R ( t )] d t , R ( 0 ) = R 0 ≥ 0. (5)
The deposition d D ( t ) dD(t) d D ( t ) acts as a non-negative noise, made of the deposition increments D ( t + Δ ) − D ( t ) D(t+\Delta)-D(t) D ( t + Δ ) − D ( t ) , i.e. the material accumulated in a window of duration Δ \Delta Δ , counteracted by the deterministic, state-dependent removal term g [ R ( t ) ] g[R(t)] g [ R ( t )] . Under stationary operating conditions we make three assumptions:
(a) Stationarity. The distribution of an increment depends on the window length Δ \Delta Δ , but not on the time at which the window begins.
(b) Independence. Deposition increments over non-overlapping windows are independent.
(c) Non-negativity. Deposition increments are nonnegative, so D ( t ) D(t) D ( t ) cannot decrease.
Together with D ( 0 ) = 0 D(0)=0 D ( 0 ) = 0 and stochastic continuity, assumptions (a)–(c) make D ( t ) D(t) D ( t ) a Lévy process. We take its deterministic drift to be zero, so that all deposition occurs through events, however small, making the process a subordinator [57 , 58 ] . Two kinds of subordinator exist, distinguished by the total mass of the Lévy measure, ν ( d x ) \nu(dx) ν ( d x ) , which we call the activity :
finite activity , r = ∫ ν ( d x ) < ∞ r=\int\nu(dx)<\infty r = ∫ ν ( d x ) < ∞ : a finite number of events per period, with possibly long quiet intervals in which nothing deposits (bursty, intermittent fouling);
infinite activity , r = ∫ ν ( d x ) = ∞ r=\int\nu(dx)=\infty r = ∫ ν ( d x ) = ∞ : infinitely many ever-smaller events in any interval, so the deposit strictly increases over every window almost surely; the sample paths remain discontinuous (pure-jump) despite this (gradual fouling).
The Lévy measure fixes the whole process through the Lévy–Khintchine formula [57 , 58 ] ; in particular, by Campbell’s theorem [59 ] , the cumulants of D ( t ) D(t) D ( t ) are proportional to t t t ,
κ n [ D ( t ) ] = q n t , q n = ∫ 0 ∞ x n ν ( d x ) , (6) \kappa_n[D(t)]=q_n t,\qquad q_n=\int_0^\infty x^n\nu(dx), \tag*{(6)} κ n [ D ( t )] = q n t , q n = ∫ 0 ∞ x n ν ( d x ) , (6)
with
κ 1 [ D ( t ) ] = E [ D ( t ) ] = m t , m = q 1 , (7) \kappa_1[D(t)]=\mathbb{E}[D(t)]=mt,\qquad m=q_1, \tag*{(7)} κ 1 [ D ( t )] = E [ D ( t )] = m t , m = q 1 , (7)
κ 2 [ D ( t ) ] = Var [ D ( t ) ] = s 2 t , s 2 = q 2 , (8) \kappa_2[D(t)]=\operatorname{Var}[D(t)]=s^2t,\qquad s^2=q_2, \tag*{(8)} κ 2 [ D ( t )] = Var [ D ( t )] = s 2 t , s 2 = q 2 , (8)
where m m m is the mean deposition rate and s 2 s^2 s 2 the variance rate. Taking expectations of Eq. (5) then gives d E [ R ] / d t = m − E [ g ( R ) ] d\mathbb{E}[R]/dt=m-\mathbb{E}[g(R)] d E [ R ] / d t = m − E [ g ( R )] , which for absent, constant or linear g g g closes into the mean laws of Table I whatever the Lévy measure. The variance rate and the mean rate have different dimensions, so the relative irregularity of the deposition must be referred to a time window. For a fixed reference window τ 0 \tau_0 τ 0 we define the coefficient of variation (c or CV) of the accumulated deposition as
c ≡ c ( τ 0 ) , c ( Δ ) 2 = Var [ D ( Δ ) ] E [ D ( Δ ) ] 2 = s 2 m 2 Δ , (9) c\equiv c(\tau_0),\qquad c(\Delta)^2=\frac{\operatorname{Var}[D(\Delta)]}{\mathbb{E}[D(\Delta)]^2}=\frac{s^2}{m^2\Delta}, \tag*{(9)} c ≡ c ( τ 0 ) , c ( Δ ) 2 = E [ D ( Δ ) ] 2 Var [ D ( Δ )] = m 2 Δ s 2 , (9)
so that
s 2 = c 2 m 2 τ 0 , Var [ D ( t ) ] = c 2 m 2 τ 0 t , s^2=c^2m^2\tau_0,\qquad\operatorname{Var}[D(t)]=c^2m^2\tau_0t, s 2 = c 2 m 2 τ 0 , Var [ D ( t )] = c 2 m 2 τ 0 t ,
c ( Δ ) = c τ 0 / Δ . (10) c(\Delta)=c\sqrt{\tau_0/\Delta}. \tag*{(10)} c ( Δ ) = c τ 0 /Δ . (10)
The CV therefore depends on the temporal resolution and decreases as windows are aggregated. Operationally, τ 0 \tau_0 τ 0 is set by the record itself—the native interval at which R ( t ) R(t) R ( t ) is sampled or reported—while Δ = k τ 0 \Delta=k\tau_0 Δ = k τ 0 for integer k k k is any coarser window obtained by aggregating k k k successive periods to suppress noise. The choice of τ 0 \tau_0 τ 0 is a convention; the physical content is the time t c = s 2 / m 2 = c 2 τ 0 t_c = s^2/m^2 = c^2\tau_0 t c = s 2 / m 2 = c 2 τ 0 , which is independent of τ 0 \tau_0 τ 0 and equals the mean time needed to deposit one size-biased event (Sec. III A). The parameter m m m sets how fast fouling accumulates on average, whereas c c c measures the relative irregularity of deposition over the chosen reference window. The value c = 2 c = 2 c = 2 used in most examples is illustrative, chosen large enough to make the differences between the deposition laws visible. Order-unity values require spatially coherent fluctuations of the deposition; granular arrivals alone give far smaller values (§ 3 B). The deterministic models of Sec. II are recovered in the limit t c → 0 t_c \to0 t c → 0 (equivalently c → 0 c \to0 c → 0 at fixed m m m ), in which fluctuations vanish relative to the mean and R ( t ) R(t) R ( t ) reduces to the smooth trajectories of Table I. The third and fourth cumulants give the skewness γ 1 = κ 3 / κ 2 3 / 2 \gamma_1 = \kappa_3/\kappa_2^{3/2} γ 1 = κ 3 / κ 2 3/2 and the excess kurtosis γ 2 = κ 4 / κ 2 2 \gamma_2 = \kappa_4/\kappa_2^2 γ 2 = κ 4 / κ 2 2 , all referring to the deposit in one sampling period. Since κ n [ D ( Δ ) ] = q n Δ \kappa_n[D(\Delta)] = q_n\Delta κ n [ D ( Δ )] = q n Δ , the standardized cumulant of order n n n decays as Δ 1 − n / 2 \Delta^{1-n/2} Δ 1 − n /2 under aggregation, with an exponent that does not depend on the Lévy measure.
Event spectrum and effective number of events
Under assumptions (a)–(c) the deposition is a superposition of effective events, each adding a positive amount of size x x x to R R R (the attachment of a particle of foulant), which need not be resolved individually. Events of each size occur as independent Poisson processes with rate density ν ( x ) \nu(x) ν ( x ) , so that a size class x x x contributes x ν ( x ) d x x\nu(x)\,dx xν ( x ) d x to the mean deposition rate and x 2 ν ( x ) d x x^2\nu(x)\,dx x 2 ν ( x ) d x to its variance rate. From Eq. (6) follows that the fluctuations (higher order cumulants) are dominated by the larger events.
The activity (Sec. III) separates two regimes. For finite r r r the process is a compound Poisson process (Supplemental Material, Sec. S2). Deposition occurs at random times and with random magnitude. It is bursty, with quiescent periods that are in principle observable. Only finitely many events occur in a finite interval, and the increment Δ D = D ( t + Δ ) − D ( t ) \Delta D = D(t + \Delta) - D(t) Δ D = D ( t + Δ ) − D ( t ) vanishes with probability Pr ( Δ D = 0 ) = e − r Δ \Pr(\Delta D = 0) = e^{-r\Delta} Pr ( Δ D = 0 ) = e − r Δ . For r = ∞ r = \infty r = ∞ instead, infinitely many arbitrarily small events occur in every interval and Δ D > 0 \Delta D > 0 Δ D > 0 almost surely for every Δ > 0 \Delta> 0 Δ > 0 . Deposition is then gradual, but it can still be intermittent, because most of the deposited mass may be carried by a few large events. Since D D D is nondecreasing, its paths have finite variation in both cases, and the model contains no Gaussian component. The distribution of the deposited mass over event sizes is described by the mass spectrum w ( x ) = x ν ( x ) / m w(x) = x\nu(x)/m w ( x ) = xν ( x ) / m , the fraction of the deposit carried by events of size x x x . Its mean,
ℓ = ∫ 0 ∞ x w ( x ) d x = q 2 q 1 = s 2 m = m t c , (11) \ell= \int_0^\infty x\,w(x)\,dx = \frac{q_2}{q_1} = \frac{s^2}{m} = mt_c, \tag*{(11)} ℓ = ∫ 0 ∞ x w ( x ) d x = q 1 q 2 = m s 2 = m t c , (11)
is the size-biased event size, i.e. the mean size of the event that deposits a given unit of R R R (denoted ℓ \ell ℓ to distinguish it from the process D ( t ) D(t) D ( t ) ). It has the dimension of R R R ,
does not depend on Δ \Delta Δ , and differs from the mean event size ∫ x ν d x / ∫ ν d x \int x\,\nu\,dx/\int\nu\,dx ∫ x ν d x / ∫ ν d x , which is dominated by the numerous small events. The squared coefficient of variation over a window Δ \Delta Δ is then
c ( Δ ) 2 = ℓ m Δ = 1 N e f f ( Δ ) , N e f f ( Δ ) = m Δ ℓ , (12) c(\Delta)^2 = \frac{\ell}{m\Delta} = \frac{1}{N_{\mathrm{eff}}(\Delta)}, \qquad N_{\mathrm{eff}}(\Delta) = \frac{m\Delta}{\ell}, \tag*{(12)} c ( Δ ) 2 = m Δ ℓ = N eff ( Δ ) 1 , N eff ( Δ ) = ℓ m Δ , (12)
where N e f f N_{\mathrm{eff}} N eff is the number of size-biased events contained in the mean deposit m Δ m\Delta m Δ , equivalently N e f f = Δ / t c N_{\mathrm{eff}} = \Delta/t_c N eff = Δ/ t c , the time t c t_c t c named in Sec. III and now given explicitly as t c = ℓ / m t_c = \ell/m t c = ℓ / m . The irregularity of the deposition is thus the inverse of the effective number of events. For N e f f ≲ 1 N_{\mathrm{eff}} \lesssim1 N eff ≲ 1 it is lumpy, with quiescent periods if r r r is finite and with intermittent gradual growth if r r r is infinite and a few large events carry most of the mass. Instead, for N e f f ≫ 1 N_{\mathrm{eff}} \gg1 N eff ≫ 1 the deposit in a window is Gaussian irrespective of the activity, as a consequence of the central limit theorem: provided that the cumulant rate q 2 q_2 q 2 is finite, the standardized higher cumulants γ i \gamma_i γ i tend to zero as N e f f N_{\mathrm{eff}} N eff increases.
Tempered-stable family of event spectra
A pure-jump deposition process is specified by the spectrum of its events. We use a tempered-stable Lévy subordinator form
ν ( d x ) = B x − 1 − α e − λ x d x , x > 0 , α < 1. (13) \nu(dx) = Bx^{-1-\alpha}e^{-\lambda x}\,dx, \qquad x > 0,\quad\alpha< 1. \tag*{(13)} ν ( d x ) = B x − 1 − α e − λ x d x , x > 0 , α < 1. (13)
The power-law factor controls the relative occurrence of small and large events, and the exponential factor imposes a maximum event size of order λ − 1 \lambda^{-1} λ − 1 , which makes all moments finite ([60 ] for the general theory of tempering). The cutoff may represent, for example, a maximum cluster or flake size that can form or attach and in the following α \alpha α is used as a statistical shape parameter. The condition α < 1 \alpha< 1 α < 1 ensures a finite mean deposition rate.
Given m m m and c c c , the two scale parameters B B B and λ \lambda λ are fixed for any α \alpha α ; in particular the cutoff is λ = ( 1 − α ) / ℓ \lambda= (1-\alpha)/\ell λ = ( 1 − α ) / ℓ (Supplemental Material, Sec. S3). The family therefore has the same mean and variance rates for all α \alpha α , and the standardized skewness and excess kurtosis over a window Δ \Delta Δ are
γ 1 ( Δ ) = 2 − α 1 − α c ( Δ ) , \gamma_1(\Delta) = \frac{2-\alpha}{1-\alpha}c(\Delta), γ 1 ( Δ ) = 1 − α 2 − α c ( Δ ) ,
γ 2 ( Δ ) = ( 2 − α ) ( 3 − α ) ( 1 − α ) 2 c ( Δ ) 2 , \gamma_2(\Delta) = \frac{(2-\alpha)(3-\alpha)}{(1-\alpha)^2}c(\Delta)^2, γ 2 ( Δ ) = ( 1 − α ) 2 ( 2 − α ) ( 3 − α ) c ( Δ ) 2 ,
c ( Δ ) = c τ 0 Δ . (14) c(\Delta) = c\sqrt{\frac{\tau_0}{\Delta}}. \tag*{(14)} c ( Δ ) = c Δ τ 0 . (14)
The prefactor ( 2 − α ) / ( 1 − α ) (2-\alpha)/(1-\alpha) ( 2 − α ) / ( 1 − α ) is the variance power of the Tweedie exponential-dispersion family, to which these laws belong [61 , 62 ] . The parameter α \alpha α has a direct physical reading. The mass-weighted spectrum of the family w ( x ) w(x) w ( x ) is a Gamma density of shape 1 − α 1-\alpha 1 − α and rate λ \lambda λ ,
w ( x ) = λ 1 − α Γ ( 1 − α ) x − α e − λ x , (15) w(x) = \frac{\lambda^{1-\alpha}}{\Gamma(1-\alpha)}x^{-\alpha}e^{-\lambda x}, \tag*{(15)} w ( x ) = Γ ( 1 − α ) λ 1 − α x − α e − λ x , (15)
with mean ℓ \ell ℓ for every α \alpha α and relative width ( 1 − α ) − 1 / 2 (1-\alpha)^{-1/2} ( 1 − α ) − 1/2 . At fixed m m m and c c c , α \alpha α therefore controls how the deposit is distributed over event sizes. For α → − ∞ \alpha\to-\infty α → − ∞ all events have the same size ℓ \ell ℓ , and deposition is a Poisson process of equal events; as α \alpha α increases, the deposit is carried by an increasingly broad range of sizes, with more of it in rare large events and, at the same time, in very small ones. Since the mean and the variance are unchanged, this broadening shows up only from the third cumulant onwards, and never lowers the skewness below c ( Δ ) c(\Delta) c ( Δ ) , a bound that holds for every subordinator (Supplemental Material, Sec. S3).
The three deposition laws
Three values of α \alpha α reproduce the standard processes used to model particulate deposition (Table II); their explicit increment distributions are given in the Supplemental Material, Sec. S3. They differ only in activity and in the breadth of the event-size spectrum. Compound Poisson (α = − 1 \alpha=-1 α = − 1 ) is the only finite-activity member, depositing discrete events of mean size ℓ / 2 \ell/2 ℓ /2 at rate 2 / t c 2/t_c 2/ t c and so admitting quiescent periods of probability e − r Δ e^{-r\Delta} e − r Δ ; any α < 0 \alpha<0 α < 0 is the same process with Gamma-distributed event sizes (Supplemental Material, Sec. S3). Gamma (α = 0 \alpha=0 α = 0 ) is its limit of increasingly frequent and increasingly weak events, in which attachment of fine particles, changes of contact area, pore filling and cluster growth cannot be told apart [63 , 64 , 65 ] , and inverse Gaussian (α = 1 / 2 \alpha=1/2 α = 1/2 ) differs from Gamma only in carrying a broader spectrum. All three are standard reliability models of monotone degradation, originating in shock-and-wear models [66 ] , developed in detail for the inverse-Gaussian case [67 , 68 , 69 , 70 , 71 ] , long since unified within the class of Lévy degradation processes [72 , 73 , 74 ] , and applied to heat-exchanger health [75].
The three laws should be viewed as alternative mesoscopic descriptions of deposition with identical mean and variance. As shown in Figure 2 , their event-size distributions differ, but the mean and the variance of the deposit are the same by construction at every window. Only the higher-order cumulants, skewness and excess kurtosis, are sensitive to the event-size distribution, and their sample estimates converge to the analytical values of Table II (Supplemental Material, Sec. S6).
Figure 2. Event spectrum, mass spectrum and sample paths for the three laws of Table II at matched m m m , c = 2 c = 2 c = 2 and τ 0 \tau_0 τ 0 . (a) Event-count spectrum ν ( x ) \nu(x) ν ( x ) of the compound-Poisson law, with its exponential jump law of mean β = λ − 1 = ℓ / 2 \beta= \lambda^{-1} = \ell/2 β = λ − 1 = ℓ /2 . (b) The same for the two infinite-activity laws; not normalizable. (c) Mass spectrum w ( x ) = x ν ( x ) / m w(x) = x\nu(x)/m w ( x ) = xν ( x ) / m , with the common size-biased mean ℓ \ell ℓ (dashed). (d)–(f) One path of each law with the mean m t mt m t and its ± 2 σ \pm2\sigma ± 2 σ band; only compound Poisson (d) shows quiescent stretches.
Aggregation and approach to Gaussian behavior
Equation (14) shows that aggregation acts on the three laws in the same way: the decay of the skewness and of the kurtosis with the window length has the same exponent for all of them, and α \alpha α only changes the prefactor. The crossover window to Gaussian behavior, defined by γ 1 ( Δ c ) = 1 \gamma_1(\Delta_c)=1 γ 1 ( Δ c ) = 1 , is
Δ c = t c ( 2 − α 1 − α ) 2 . (16) \Delta_c=t_c\left(\frac{2-\alpha}{1-\alpha}\right)^2. \tag*{(16)} Δ c = t c ( 1 − α 2 − α ) 2 . (16)
It is a crossover, not a sharp threshold, and it grows without bound as α → 1 \alpha\to1 α → 1 : the broader the spectrum of event sizes, the more events are needed before the deposit looks Gaussian. For Δ ≫ Δ c \Delta\gg\Delta_c Δ ≫ Δ c the three laws become increasingly difficult to distinguish on the basis of the shape of the increments alone, and the deposition is statistically a drift m m m with Gaussian fluctuations of variance rate s 2 s^2 s 2 : this is the central limit theorem for the underlying random walk, with N e f f = Δ / t c N_{\rm eff}=\Delta/t_c N eff = Δ/ t c playing the role of the number of summands. Eq. (14) quantifies the decay of the leading non-Gaussian cumulants.
Removal: fluctuations, intermittency and the limits of closure
Section III showed that, for absent, constant or linear removal, the mean fouling curve does not depend on the statistics of the deposition. Here we ask what removal does to the fluctuations, whether two removal mechanisms with the same mean trajectory can be told apart, and when the closure of the mean fails.
Linear removal: a Lévy-driven Ornstein–Uhlenbeck process
Solving Eq. (5) with g = ψ R g=\psi R g = ψ R gives
R ( t ) = R 0 e − ψ t + ∫ 0 t e − ψ ( t − u ) d U ( u ) , (17) R(t)=R_0e^{-\psi t}+\int_0^t e^{-\psi(t-u)}\,dU(u), \tag*{(17)} R ( t ) = R 0 e − ψ t + ∫ 0 t e − ψ ( t − u ) d U ( u ) , (17)
a superposition of deposition events, each decaying exponentially over the removal time ψ − 1 \psi^{-1} ψ − 1 . This is a shot-noise process and, in the language of stochastic processes, a Lévy-driven Ornstein–Uhlenbeck process [57 , 76 ] . Since the removal kernel is linear, the cumulants of R R R follow from those of D D D :
κ n [ R ( t ) ] = q n 1 − e − n ψ t n ψ ( n ≥ 2 ) , (18) \kappa_n[R(t)]=q_n\frac{1-e^{-n\psi t}}{n\psi}\qquad(n\ge2), \tag*{(18)} κ n [ R ( t )] = q n n ψ 1 − e − n ψ t ( n ≥ 2 ) , (18)
while the mean is given in Table I . All cumulants therefore saturate, and the plateau is characterized by:
Table I. Deterministic fouling laws, and the first two cumulants of the fouling state R ( t ) R(t) R ( t ) when the deposition is a subordinator with mean rate m m m and variance rate s 2 = c 2 m 2 τ 0 s^2=c^2m^2\tau_0 s 2 = c 2 m 2 τ 0 , starting from R ( 0 ) = R 0 R(0)=R_0 R ( 0 ) = R 0 .
E [ R ] ∞ = R ∞ = m ψ , Var [ R ] ∞ = s 2 2 ψ , \mathbb{E}[R]_{\infty}=R_{\infty}=\frac{m}{\psi},\qquad\operatorname{Var}[R]_{\infty}=\frac{s^2}{2\psi}, E [ R ] ∞ = R ∞ = ψ m , Var [ R ] ∞ = 2 ψ s 2 ,
CV ∞ 2 = γ / t c 2 = 1 2 N ∞ , N ∞ = R ∞ ℓ , (19) \operatorname{CV}_{\infty}^2=\frac{\gamma/t_c}{2}=\frac{1}{2N_{\infty}},\qquad N_{\infty}=\frac{R_{\infty}}{\ell}, \tag*{(19)} CV ∞ 2 = 2 γ / t c = 2 N ∞ 1 , N ∞ = ℓ R ∞ , (19)
where N ∞ N_{\infty} N ∞ is the number of size-biased events held in the plateau deposit.
Per-element detachment is a different mechanism: each deposited element survives an exponential time, then is removed entirely—the usual reading of the Kern–Seaton time constant as a mean residence time [51 ] . It shares the mean R ∞ = m / ψ R_{\infty}=m/\psi R ∞ = m / ψ but gives
κ n [ R ] ∞ = q n ψ , (20) \kappa_n[R]_{\infty}=\frac{q_n}{\psi}, \tag*{(20)} κ n [ R ] ∞ = ψ q n , (20)
Table 2. The three deposition laws as members of the tempered-stable family, Eq. (13) , at common m m m , c c c and τ 0 \tau_0 τ 0 ; ℓ = c 2 m τ 0 \ell= c^2m\tau_0 ℓ = c 2 m τ 0 . Skewness and excess kurtosis over a window Δ \Delta Δ follow from Eq. (14) : γ 1 = ( γ 1 / c ) c ( Δ ) \gamma_1 = (\gamma_1/c)c(\Delta) γ 1 = ( γ 1 / c ) c ( Δ ) and γ 2 = ( γ 2 / c 2 ) c ( Δ ) 2 \gamma_2 = (\gamma_2/c^2)c(\Delta)^2 γ 2 = ( γ 2 / c 2 ) c ( Δ ) 2 .
against q n / ( n ψ ) q_n/(n\psi) q n / ( n ψ ) under continuous removal, Eq. (18) . The difference lies in the fate of a single event of size x x x deposited at time u u u . Continuous removal erodes it steadily, so that it contributes x e − ψ ( t − u ) x e^{-\psi(t-u)} x e − ψ ( t − u ) ; per-element detachment leaves it intact until it goes, so that it contributes x x x with probability e − ψ ( t − u ) e^{-\psi(t-u)} e − ψ ( t − u ) and nothing otherwise. The two contributions have the same mean, but their n n n th powers average to x n e − n ψ ( t − u ) x^n e^{-n\psi(t-u)} x n e − n ψ ( t − u ) and x n e − ψ ( t − u ) x^n e^{-\psi(t-u)} x n e − ψ ( t − u ) respectively, and since the cumulants of a Poisson sum add these powers, every cumulant of order n n n is n n n times larger under detachment: the mean is unchanged and the variance doubled. Once standardized the factors nearly cancel: the plateau excess kurtosis is the same, γ 2 ( ψ − 1 ) \gamma_2(\psi^{-1}) γ 2 ( ψ − 1 ) , and
the skewness, γ 1 ( ψ − 1 ) \gamma_1(\psi^{-1}) γ 1 ( ψ − 1 ) , exceeds the continuous-removal value by only 3 / ( 2 2 ) ≈ 1.06 3/(2\sqrt{2}) \approx1.06 3/ ( 2 2 ) ≈ 1.06 (Supplemental Material, Sec. S4). The factor n n n requires each event to detach as a single unit: an event made of many sub-units detaching independently loses mass smoothly, as x e − ψ ( t − u ) x e^{-\psi(t-u)} x e − ψ ( t − u ) , and returns the continuous-removal cumulants.
The plateau is thus built from N ∞ = R ∞ / ℓ N_\infty= R_\infty/\ell N ∞ = R ∞ / ℓ size-biased events (Sec. III A), and its relative fluctuations are 1 / 2 N ∞ 1/\sqrt{2N_\infty} 1/ 2 N ∞ : a plateau that contains many events is regular, whereas one made of a few large events is not. The transient variance, ( s 2 / 2 ψ ) ( 1 − e − 2 ψ t ) (s^2/2\psi)(1-e^{-2\psi t}) ( s 2 /2 ψ ) ( 1 − e − 2 ψ t ) , grows as s 2 t s^2t s 2 t (pure deposition) for t ≪ ψ − 1 t \ll\psi^{-1} t ≪ ψ − 1 and saturates after a time of order ( 2 ψ ) − 1 (2\psi)^{-1} ( 2 ψ ) − 1 . For the tempered-stable family the stationary skewness and excess kurtosis under continuous linear removal are
γ 1 , ∞ = 4 3 2 − α 1 − α C V ∞ , γ 2 , ∞ = 2 ( 2 − α ) ( 3 − α ) ( 1 − α ) 2 C V ∞ 2 . (21) \gamma_{1,\infty} = \frac{4}{3}\frac{2-\alpha}{1-\alpha}\mathrm{CV}_{\infty},
\qquad
\gamma_{2,\infty} = 2\frac{(2-\alpha)(3-\alpha)}{(1-\alpha)^2}\mathrm{CV}_{\infty}^{2}. \tag*{(21)} γ 1 , ∞ = 3 4 1 − α 2 − α CV ∞ , γ 2 , ∞ = 2 ( 1 − α ) 2 ( 2 − α ) ( 3 − α ) CV ∞ 2 . (21)
equivalently
γ 1 , ∞ = 2 2 3 γ 1 ( ψ − 1 ) , γ 2 , ∞ = γ 2 ( ψ − 1 ) . (22) \gamma_{1,\infty} = \frac{2\sqrt{2}}{3}\gamma_1(\psi^{-1}),
\qquad
\gamma_{2,\infty} = \gamma_2(\psi^{-1}). \tag*{(22)} γ 1 , ∞ = 3 2 2 γ 1 ( ψ − 1 ) , γ 2 , ∞ = γ 2 ( ψ − 1 ) . (22)
The plateau has the shape of the deposit accumulated over one removal time, Δ = ψ − 1 \Delta=\psi^{-1} Δ = ψ − 1 (Supplemental Material, Sec. S4). The pattern of Sec. III B therefore repeats one level up: R ∞ R_\infty R ∞ and Var [ R ] ∞ \operatorname{Var}[R]_\infty Var [ R ] ∞ are independent of α \alpha α , while γ 1 , ∞ \gamma_{1,\infty} γ 1 , ∞ and γ 2 , ∞ \gamma_{2,\infty} γ 2 , ∞ are not.
What removal changes is how that distinguishing information behaves in time. Without removal the standardized cumulants decay indefinitely with the observation window, as Δ − 1 / 2 \Delta^{-1/2} Δ − 1/2 and Δ − 1 \Delta^{-1} Δ − 1 , so the longer the deposit accumulates before being inspected the harder the three laws are to tell apart. Under linear removal R ( t ) R(t) R ( t ) is built only from deposition of the last ψ − 1 \psi^{-1} ψ − 1 or so, older material having been removed, and the effective window cannot grow past Δ ∼ ψ − 1 \Delta\sim\psi^{-1} Δ ∼ ψ − 1 however long the process runs. The exact finite-time cumulants of Eq. (18) show this directly: γ 1 ( t ) \gamma_1(t) γ 1 ( t ) and γ 2 ( t ) \gamma_2(t) γ 2 ( t ) follow their no-removal decays while t ≪ ψ − 1 t\ll\psi^{-1} t ≪ ψ − 1 , then saturate at the values of Eq. (21) . Removal does not erase the shape information the way further aggregation does; it freezes it at the value reached after about one relaxation time. Figure 3 (a)–(Figure 3 b) illustrates this directly: sample paths with the common mean and ± 2 σ \pm2\sigma ± 2 σ band, and the exact skewness γ 1 ( t ) \gamma_1(t) γ 1 ( t ) for each law (solid), which tracks the no-removal t − 1 / 2 t^{-1/2} t − 1/2 decay (dashed) while t ≪ ψ − 1 t\ll\psi^{-1} t ≪ ψ − 1 and bends over to the law-dependent plateau γ 1 , ∞ \gamma_{1,\infty} γ 1 , ∞ once removal takes effect.
Figure 3. Fluctuations under removal (m = 1 m=1 m = 1 , c = 1 c=1 c = 1 , τ 0 = 1 \tau_{0}=1 τ 0 = 1 , ψ = 0.2 \psi=0.2 ψ = 0.2 , R ∞ = 5 R_{\infty}=5 R ∞ = 5 ). (a) Sample paths of the Lévy-driven Ornstein–Uhlenbeck process for the three laws, with the mean and the ± 2 σ ( t ) \pm2\sigma(t) ± 2 σ ( t ) band. (b) Exact skewness γ 1 ( t ) \gamma_{1}(t) γ 1 ( t ) under removal (solid, Eq. (18) ) and the no-removal t − 1 / 2 t^{-1/2} t − 1/2 reference (dashed). (c) One path each of continuous removal, per-element detachment, sloughing (η = 0.5 \eta=0.5 η = 0.5 , r s = ψ / η r_{\mathrm{s}}=\psi/\eta r s = ψ / η ) and independent-site blocking (equal events of size ℓ \ell ℓ ), all with the same mean (dashed). (d) Stationary variance against η \eta η at fixed ψ \psi ψ , Eq. (24) (lines) and Monte Carlo (markers), for c = 1 c=1 c = 1 and c = 0.1 c=0.1 c = 0.1 ; dashed: the continuous-removal baselines s 2 / 2 ψ s^{2}/2\psi s 2 /2 ψ (logarithmic scale).
Intermittent removal: sloughing
Continuous removal is not the only way to produce a plateau. Fouling layers may lose material in discrete events (sloughing, spalling, detachment of flakes) [77 , 78 ] . At the scale of a single particle the same structure arises from re-entrainment by turbulent bursts, each sweeping a fixed fraction of the wall and removing material that has not yet stabilised [47 ] . We model them as a Poisson process N s ( t ) N_s(t) N s ( t ) of rate r s r_s r s , each event removing a fixed fraction η ∈ ( 0 , 1 ] \eta\in(0,1] η ∈ ( 0 , 1 ] of the deposit,
d R ( t ) = d D ( t ) − η R ( t − ) d N s ( t ) . (23) dR(t)=dD(t)-\eta R(t^{-})dN_s(t). \tag*{(23)} d R ( t ) = d D ( t ) − η R ( t − ) d N s ( t ) . (23)
The rate and the fraction do not depend on the state of the deposit: removal is triggered by time, not by the thickness reached. (Thickness-triggered spalling is not treated here.) Averaging Eq. (23) gives d E [ R ] / d t = m − r s η E [ R ] d\mathop{\mathbb{E}}[R]/dt=m-r_s\eta\mathop{\mathbb{E}}[R] d E [ R ] / d t = m − r s η E [ R ] , the same mean equation as for continuous removal, with ψ = r s η \psi=r_s\eta ψ = r s η . The mean trajectory, the plateau R ∞ R_\infty R ∞ and the relaxation time cannot discriminate between the two mechanisms.
The relaxation time cannot discriminate between the two mechanisms. The same holds for the two-point correlation: E [ R ( t + τ ) ∣ R ( t ) ] \mathop{\mathbb{E}}[R(t+\tau)\mid R(t)] E [ R ( t + τ ) ∣ R ( t )] obeys the same linear equation in both cases, so the autocorrelation is again e − ψ τ e^{-\psi\tau} e − ψ τ .
The fluctuations, however, differ. At equal mean and relaxation rate, the stationary variance under sloughing is
Var [ R ] ∞ = s 2 ( 2 − η ) ψ + η 2 − η R ∞ 2 . (24) \operatorname{Var}[R]_\infty=\frac{s^2}{(2-\eta)\psi}+\frac{\eta}{2-\eta}R_\infty^2. \tag*{(24)} Var [ R ] ∞ = ( 2 − η ) ψ s 2 + 2 − η η R ∞ 2 . (24)
to be compared with s 2 / 2 ψ s^2/2\psi s 2 /2 ψ for continuous removal (Table III). The second term does not depend on the deposition statistics: it is the sawtooth generated by the random timing of the removal events, and is present even if the deposition is perfectly smooth (s = 0 s=0 s = 0 ). Its relative size is η / ( 2 − η ) \eta/(2-\eta) η / ( 2 − η ) . In the limit of frequent small events, η → 0 \eta\to0 η → 0 at fixed r s η r_s\eta r s η , Eq. (24) reduces to the continuous result, as expected for erosion by many small detachments. For complete cleaning events (η = 1 \eta=1 η = 1 ), C V 2 = ψ t c + 1 ≳ 1 \mathrm{CV}^2=\psi t_c+1\gtrsim1 CV 2 = ψ t c + 1 ≳ 1 : the fluctuations are of the order of the mean itself. Figure 3 (c)–(Figure 3 d) compares one realization of each mechanism at matched mean and validates the stationary variance against η \eta η , Eq. (24) , by Monte Carlo simulations. The moment comparison between the two mechanisms is collected in the Supplemental Material, Sec. S4.
Nonlinear removal
For a removal law g ( R ) g(R) g ( R ) that is neither constant nor linear, the mean equation, which is still exact,
d μ d t = m − E [ g ( R ) ] , μ = E [ R ] , (25) \frac{d\mu}{dt}=m-\mathbb{E}[g(R)],
\qquad
\mu=\mathbb{E}[R], \tag*{(25)} d t d μ = m − E [ g ( R )] , μ = E [ R ] , (25)
no longer closes: E [ g ( R ) ] ≠ g ( μ ) \mathbb{E}[g(R)]\ne g(\mu) E [ g ( R )] = g ( μ ) , and the mean depends on the fluctuations. Expanding g g g around μ \mu μ gives, to second order,
d μ d t ≃ m − g ( μ ) − 1 2 g ′ ′ ( μ ) V , d V d t ≃ s 2 − 2 g ′ ( μ ) V , (26) \frac{d\mu}{dt}\simeq m-g(\mu)-\frac{1}{2}g''(\mu)V,
\qquad
\frac{dV}{dt}\simeq s^2-2g'(\mu)V, \tag*{(26)} d t d μ ≃ m − g ( μ ) − 2 1 g ′′ ( μ ) V , d t d V ≃ s 2 − 2 g ′ ( μ ) V , (26)
with V = Var [ R ] V=\operatorname{Var}[R] V = Var [ R ] . The expansion is ordered in the relative event size 1 / N ∞ = ℓ / R ∞ 1/N_\infty=\ell/R_\infty 1/ N ∞ = ℓ / R ∞ ; the third-cumulant terms it neglects are smaller by a further factor 1 / N ∞ 1/N_\infty 1/ N ∞ (Supplemental Material, Sec. S4). To this order the deposition law enters only through s 2 s^2 s 2 , i.e. through c c c ; the shape parameter α \alpha α first appears through κ 3 [ R ] \kappa_3[R] κ 3 [ R ] at the next order in 1 / N ∞ 1/N_\infty 1/ N ∞ . If g g g is convex, fluctuations increase the mean removal rate and the mean fouling lies below the deterministic (mean-field) prediction; for concave g g g the opposite holds. The universality of the mean is therefore lost, and the mean curve now carries information on c c c .
The correction to the removal rate, 1 2 g ′ ′ V / g ∼ C V 2 ∼ 1 / N ∞ \frac{1}{2}g''V/g\sim\mathrm{CV}^2\sim1/N_\infty 2 1 g ′′ V / g ∼ CV 2 ∼ 1/ N ∞ , is small when the plateau holds many events, where the closure becomes asymptotically exact, and the closure fails when it holds few.
In real exchangers nonlinear dependence on the deposit arises, for instance, from the change of the flow cross-section and of the wall shear stress with the deposit thickness (condition (ii) of Sec. II); such feedback also makes the deposition itself non-stationary, and lies outside the scope of this work.
Table 3. Plateau statistics for three removal mechanisms reaching the same plateau R ∞ = m / ψ R_{\infty}=m/\psi R ∞ = m / ψ (for sloughing, ψ = r s η \psi=r_{\mathrm{s}}\eta ψ = r s η ); t c = s 2 / m 2 t_{c}=s^{2}/m^{2} t c = s 2 / m 2 . All three share the autocorrelation e − ψ ∣ τ ∣ e^{-\psi|\tau|} e − ψ ∣ τ ∣ . Only sloughing retains a non-vanishing plateau variance as the deposition becomes smooth, c → 0 c\to0 c → 0 , because its second term does not contain s 2 s^{2} s 2 .
Numerical verification. Every analytic result of this and the following sections is checked against Monte Carlo simulation of Eq. (5) . Deposition increments are drawn exactly from the three laws of Table II; continuous removal is applied as exponential decay over each step, sloughing as Poisson-timed events, per-element detachment exactly by thinning the alive deposit, and blocking by binomial filling of independent sites. Where a time step is needed it is 0.005 0.005 0.005 –0.02 τ 0 0.02\tau_{0} 0.02 τ 0 ; ensembles range from 10 3 10^{3} 1 0 3 records to 2 × 10 5 2\times10^{5} 2 × 1 0 5 paths per law, with fixed seeds, and the code is available (Data Availability). The simulations reproduce the transient and stationary moments of Table III to within sampling error, about 0.5 % 0.5\% 0.5% for the stationary variances (Figure S2 ), the plateau skewness of Eq. (21) , the record-length estimate of Sec. V D (Figure S5 ) and the first-passage statistics of Sec. VI (Figure S6 ).
Figure S2. Removal mechanisms with the same mean R ∞ = m / ψ R_{\infty}=m/\psi R ∞ = m / ψ and relaxation rate ψ \psi ψ (m = c = τ 0 = 1 m=c=\tau_{0}=1 m = c = τ 0 = 1 , ψ = 0.2 \psi=0.2 ψ = 0.2 , Gamma deposition; lines exact, markers Monte Carlo). (a) Transient variance, normalized by the continuous-removal plateau s 2 / 2 ψ s^{2}/2\psi s 2 /2 ψ : continuous removal [Eq. (18) of the main text]; per-element detachment, q 2 ( 1 − e − ψ t ) / ψ q_{2}(1-e^{-\psi t})/\psi q 2 ( 1 − e − ψ t ) / ψ , saturating at twice that value and half the rate; and sloughing (η = 1 2 \eta=\frac{1}{2} η = 2 1 ), Eqs. (69) –(70) , whose plateau carries a deposition-independent floor. (b) All three share the stationary autocorrelation e − ψ τ e^{-\psi\tau} e − ψ τ , so neither mean nor correlation time discriminates them. (c) Independent-site blocking vs. continuous removal at the same R ∞ R_{\infty} R ∞ , ψ \psi ψ , for equal events of size ℓ \ell ℓ : removal variance rises monotonically to ℓ R ∞ / 2 \ell R_{\infty}/2 ℓ R ∞ /2 , while blocking variance ℓ R ∞ θ ( 1 − θ ) \ell R_{\infty}\theta(1-\theta) ℓ R ∞ θ ( 1 − θ ) peaks at half coverage and vanishes at saturation (markers: binomial site filling).
Figure S5. Monte Carlo test of the record-length estimate, Eq. (84) , for the Gamma and inverse-Gaussian laws at c = 2 c=2 c = 2 . Each record of N N N base periods is observed with Gaussian noise set by ρ \rho ρ , Eq. (79) , and differenced over N / 2 k N/2k N /2 k windows of k = ρ / 2 k=\rho/2 k = ρ /2 periods; the law is assigned by comparing sample skewness with the midpoint of the two attenuated values, Eq. (80) . Markers: fraction of correct assignments in 1000 records per law. Line: asymptotic prediction Φ G ( N / N ∗ ) \Phi_{\mathrm{G}}(\sqrt{N/N^*}) Φ G ( N / N ∗ ) , with Φ G \Phi_{\mathrm{G}} Φ G the standard normal distribution function, i.e. d ′ = 2 d'=2 d ′ = 2 at N = N ∗ = 324 ρ 2 / c 2 N=N^*=324\rho^2/c^2 N = N ∗ = 324 ρ 2 / c 2 , which holds for ρ ≳ 10 2 \rho\gtrsim10^2 ρ ≳ 1 0 2 and is optimistic below (about 1.3 × 1.3\times 1.3 × longer needed at ρ = 40 \rho=40 ρ = 40 , about 3 × 3\times 3 × at ρ = 10 \rho=10 ρ = 10 , where ρ \rho ρ is no longer large compared with c 2 = 4 c^2=4 c 2 = 4 ).
Figure S6. Range of validity of the first-passage formulas, by Monte Carlo at c = 2 c = 2 c = 2 for the three matched laws. (a) E [ T x ] − n x \mathbb{E}[T_x] - n_x E [ T x ] − n x vs. the predicted ( 1 + c 2 ) / 2 (1+c^2)/2 ( 1 + c 2 ) /2 (dashed; error bars ± 1 \pm1 ± 1 Monte Carlo s.e.). (b) sd [ T x ] / n x \operatorname{sd}[T_x]/\sqrt{n_x} sd [ T x ] / n x vs. predicted c c c . (c) skew [ T x ] n x \operatorname{skew}[T_x]\sqrt{n_x} skew [ T x ] n x vs. predicted 3 c − γ 1 3c-\gamma_1 3 c − γ 1 (dashed; zero for the inverse Gaussian law). Accurate at n x = 220 n_x = 220 n x = 220 , degrading towards small n x n_x n x , the only regime where the deposition law would matter.
Identifiability of the deposition mechanism
What the mean cannot see
Linear removal and Poisson-timed sloughing, matched at ψ = r s η \psi= r_s\eta ψ = r s η , were shown in Sec. IV to share one mean curve while their fluctuations remain distinguishable (Table III).
Both mechanisms assume that every particle reaching the wall sticks, whereas attachment efficiencies well below unity are the rule in particulate deposition [13 , 14 ] , and the reaction-engineering treatment of fouling writes the deposition rate as a foulant production rate times a sticking probability [79 ] . Writing p p p for the sticking probability, two cases separate. A constant p p p thins the event-rate density,
ν → p ν , m → p m , c 2 → c 2 / p , (27) \nu\to p\nu,\qquad m \to pm,\qquad c^2 \to c^2/p, \tag*{(27)} ν → p ν , m → p m , c 2 → c 2 / p , (27)
leaving ℓ \ell ℓ and α \alpha α unchanged: the tempered-stable family is closed under the operation, and the only effect is a scaled net rate—a linear mean curve of reduced slope, or, where linear removal is also present, the same exponential saturation to a proportionally lower plateau. This is the suppression reading of the Ebert–Panchal term (Sec. II): with p = 1 − γ τ w / m p = 1 - \gamma\tau_w/m p = 1 − γ τ w / m the thinned process reproduces the linear mean of Table I exactly, without a boundary at R = 0 R = 0 R = 0 , and with a variance p s t 2 p s_t^2 p s t 2 smaller than that of the erosion reading. If instead p p p falls as the surface fills, p ( R ) = 1 − R / R ∞ p(R) = 1 - R/R_\infty p ( R ) = 1 − R / R ∞ , the Langmuir blocking law of site-limited adsorption, the mechanism is not removal at all: nothing leaves the surface and attachment becomes self-limiting. Each increment then depends on how full the surface already is, so the process leaves the Lévy class. We take the blocking to be independent-site filling, for which the mean closes exactly because p p p is linear in R R R and the variance vanishes at saturation; random sequential adsorption with excluded volume is a different process, approaching its jamming limit non-exponentially and retaining a variance proportional to the area, so the two are themselves distinguishable by the late-time scatter. The mean trajectory of independent-site blocking is nonetheless the same exponential saturation to R ∞ R_\infty R ∞ .
Four mechanisms with nothing physically in common—continuous shear removal, per-element detachment, intermittent sloughing and self-limiting Langmuir adsorption—therefore reach the same saturating mean at matched net rate [Figure 3 (c)], while constant attachment thinning rescales the linear branch without saturating it at all. Fitting an asymptotic curve returns R ∞ R_\infty R ∞ and a time constant and says nothing about which of the four produced it: a Kern–Seaton fit is not evidence of shear removal. Table III collects the three that settle to a stationary plateau; blocking is absent from it because its variance vanishes once the sites are full, leaving no stationary fluctuation.
The same difficulty of separating deposition, suppression, removal and consolidation in operating data, is known within the mechanistic literature [37 ] .
The degeneracy is not even confined to the mechanism. Stationarity and a finite first moment already give E [ D ( t ) ] = m t \mathbb{E}[D(t)] = mt E [ D ( t )] = m t , since expectations add whether or not the increments are independent, and for a single record rather than an ensemble it is enough that the increment sequence be ergodic; independence is needed only from the second cumulant onward. A straight or saturating mean curve is therefore not evidence even of the Lévy structure assumed in Sec. III. What such a curve identifies is the net mean rate and, under linear removal, the relaxation rate—and nothing beyond them.
Variance and the deposition law
The three laws of Sec. III C are matched in their first two cumulants by construction, so only the standardized third and fourth cumulants carry information about the increment law. What the variance can test is whether a granular picture of deposition is quantitatively consistent at all.
Suppose each attachment event is one particle of diameter d p d_p d p , depositing at solid fraction ϕ s \phi_s ϕ s , and that arrivals form a Poisson process. Averaged over the heat-transfer area A A A , a single attachment raises the fouling resistance by
J 1 = π d p 3 / 6 A ϕ s k f , whence c 2 = J 1 m τ 0 . (28) J_1 = \frac{\pi d_p^3/6}{A\phi_s k_f},\qquad\text{whence}\qquad c^2 = \frac{J_1}{m\tau_0}. \tag*{(28)} J 1 = A ϕ s k f π d p 3 /6 , whence c 2 = m τ 0 J 1 . (28)
Equation (28) is not an external estimate: it is Eq. (12) evaluated at Δ = τ 0 \Delta= \tau_0 Δ = τ 0 in the monodisperse limit α → ∞ \alpha\to\infty α → ∞ of the family, where every event deposits the same amount and the size-biased event size ℓ \ell ℓ collapses onto the jump J 1 J_1 J 1 itself. Equivalently c 2 = 1 / ( r τ 0 ) c^2 = 1/(r\tau_0) c 2 = 1/ ( r τ 0 ) —the classical shot-noise result for Poisson-timed pulses of equal weight [80 , 81 ] . No parameter is free: once the event size is fixed, the mean rate fixes the event frequency through r = m / J 1 r = m/J_1 r = m / J 1 . The derivation is given in the Supplemental Material, Sec. S5. The resulting magnitudes are extreme. For an exchanger of A = 200 m 2 A = 200\ \mathrm{m^2} A = 200 m 2 fouling at a rate of order 10 − 5 m 2 K W − 1 10^{-5}\ \mathrm{m^2\,K\,W^{-1}} 1 0 − 5 m 2 K W − 1 per sampling period, with deposit properties k f ∼ 0.5 W m − 1 K − 1 k_f \sim0.5\ \mathrm{W\,m^{-1}\,K^{-1}} k f ∼ 0.5 W m − 1 K − 1 and ϕ s ∼ 0.5 \phi_s \sim0.5 ϕ s ∼ 0.5 , Eq. (28) gives c ∼ 10 − 9 c \sim10^{-9} c ∼ 1 0 − 9 for 0.1 μ m 0.1\ \mu\mathrm{m} 0.1 μ m particles and c ∼ 3 × 10 − 5 c \sim3 \times10^{-5} c ∼ 3 × 1 0 − 5 at 100 μ m 100\ \mu\mathrm{m} 100 μ m . The transport mechanism is thus invisible in the mean, which sees only the net rate (Sec. V A), and invisible in the variance by smallness.
The scaling behind those numbers is the general statement. Because c 2 ∝ J 1 ∝ A − 1 c^2 \propto J_1 \propto A^{-1} c 2 ∝ J 1 ∝ A − 1 at fixed mean rate, granular arrival noise is suppressed as A − 1 / 2 A^{-1/2} A − 1/2 and attains order unity only on the scale of a single particle. The same square-root suppression governs the coverage of a deposited monolayer under random sequential adsorption, whose jamming count obeys a central limit theorem with variance proportional to the system size [82 , 83 ] . An area-averaged fouling resistance is therefore almost perfectly deterministic with respect to the granularity of deposition, and any observed coefficient of variation of order one must originate in a fluctuation that is coherent over a length comparable to the exchanger itself.
That bound bears on assumption (b) of Sec. III: independent granular arrivals cannot produce c c c of order unity without single events equivalent to a uniform layer several to tens of micrometres thick across the whole surface, which is not physical. A coherent fluctuation is required instead: operating-condition drift (c ≈ E a Δ T / ( R g T f 2 ) ≈ 0.1 c \approx E_a\Delta T/(R_gT_f^2) \approx0.1 c ≈ E a Δ T / ( R g T f 2 ) ≈ 0.1 for a 5 K film-temperature excursion at E a = 50 k J m o l − 1 E_a = 50\ \mathrm{kJ\,mol^{-1}} E a = 50 kJ mo l − 1 , T f = 500 K T_f = 500\ \mathrm{K} T f = 500 K ), turbulent-burst intermittency in near-wall transport [47 , 84 , 85 ] , or a coherent measurement error. The first two place the process outside the Lévy class; the third leaves the subordinator description intact and makes the observed c c c partly a sensor parameter.
Variance and the removal mechanism
For removal, linearising the variance equation gives
V ˙ ≃ s 2 − 2 g ′ ( μ ) V (29) \dot{V} \simeq s^2 - 2g'(\mu)V \tag*{(29)} V ˙ ≃ s 2 − 2 g ′ ( μ ) V (29)
so what matters is the slope of the removal law at the operating point, not whether removal is present. Pure deposition and Ebert–Panchal removal both have g ′ = 0 g' = 0 g ′ = 0 and therefore the same unbounded growth (Table I ; under the suppression reading the variance rate is reduced by the factor p p p , but the growth remains unbounded); only g ′ ≠ 0 g' \ne0 g ′ = 0 makes the variance informative.
A saturating record is exactly the regime where the variance separates mechanisms the mean cannot: sloughing retains a floor η R ∞ 2 / ( 2 − η ) \eta R_\infty^2/(2-\eta) η R ∞ 2 / ( 2 − η ) even as deposition becomes noise-free ((24) ), while every s 2 s^2 s 2 -proportional term—including per-element detachment’s ((20) )—vanishes with c c c . At realistic c c c the question is therefore not whether the plateau variance exceeds a baseline but whether it fluctuates at all; blocking is separated from continuous removal instead by its time course: with independent-site filling the variance is ℓ R ∞ θ ( 1 − θ ) \ell R_\infty\theta(1-\theta) ℓ R ∞ θ ( 1 − θ ) at coverage θ = E [ R ] / R ∞ = 1 − p \theta= \mathop{\rm E}[R]/R_\infty= 1-p θ = E [ R ] / R ∞ = 1 − p , rising to a maximum ℓ R ∞ / 4 \ell R_\infty/4 ℓ R ∞ /4 at half coverage and returning to zero as the sites fill. The binomial form holds for sites of a single capacity, each filled by one event of size J 1 J_1 J 1 , i.e. the monodisperse limit of Sec. V B with ℓ = J 1 \ell= J_1 ℓ = J 1 ; for polydisperse events the intermediate-time variance depends on how events of different size share the sites, but it still vanishes as the surface saturates. That maximum is set by the event size, so unlike the sloughing floor it is a granular quantity and inherits the smallness of Sec. V B.
What this test costs is worth setting beside Sec. V D. The sloughing floor is a fixed fraction of the plateau: at η = 1 2 \eta= \frac{1}{2} η = 2 1 it is R ∞ 2 / 3 R_\infty^2/3 R ∞ 2 /3 , so the discriminator asks only that the measurement noise stay below R ∞ / 3 R_\infty/\sqrt{3} R ∞ / 3 , under six-tenths of the plateau itself. What it does demand is duration, since the state decorrelates on the relaxation time and the plateau variance must therefore be estimated from a record spanning several multiples of ψ − 1 \psi^{-1} ψ − 1 . The contrast with the deposition law is not one of degree. Identifying the increment law needs both a better instrument and the record length of (32) ; identifying the removal mechanism needs a record that reaches its plateau and stays there.
Measurement noise: observability and record length
Temporal aggregation preserves the mean rate and suppresses every standardized higher cumulant, (14) : over a window Δ t \Delta t Δ t the coefficient of variation falls as c τ 0 / Δ t c\sqrt{\tau_0/\Delta t} c τ 0 /Δ t and the skewness as γ 1 τ 0 / Δ t \gamma_1\sqrt{\tau_0/\Delta t} γ 1 τ 0 /Δ t . A smooth macroscopic fouling curve is therefore compatible with strongly intermittent event-scale deposition, and the question for a real record is whether any observation window is short enough to retain that shape and long enough to clear the measurement noise.
Let the state be observed at intervals Δ t = k τ 0 \Delta t = k\tau_0 Δ t = k τ 0 as Y j = R ( j Δ t ) + ε j Y_j = R(j\Delta t) + \varepsilon_j Y j = R ( j Δ t ) + ε j with ε j ∼ i i d N ( 0 , σ 2 ) \varepsilon_j \overset{\mathrm{iid}}{\sim} \mathcal{N}(0,\sigma^2) ε j ∼ iid N ( 0 , σ 2 ) , and take disjoint increments Δ Y j = Y 2 j + 1 − Y 2 j \Delta Y_j = Y_{2j+1} - Y_{2j} Δ Y j = Y 2 j + 1 − Y 2 j so that their measurement errors are independent. Then Var [ Δ Y j ] = c 2 m 2 τ 0 Δ t + 2 σ 2 \operatorname{Var}[\Delta Y_j] = c^2m^2\tau_0\Delta t + 2\sigma^2 Var [ Δ Y j ] = c 2 m 2 τ 0 Δ t + 2 σ 2 : the deposition variance grows with the window, the differenced noise does not, and everything that follows is governed by the single dimensionless group formed from the two,
ρ = 2 σ 2 c 2 m 2 τ 0 2 , (30) \rho= \frac{2\sigma^2}{c^2m^2\tau_0^2}, \tag*{(30)} ρ = c 2 m 2 τ 0 2 2 σ 2 , (30)
the differenced noise variance divided by the deposition variance of one reference window; equivalently, ρ \rho ρ is the window length k k k at which the increment’s signal-to-noise ratio SNR ( k ) = k / ρ \operatorname{SNR}(k)=k/\rho SNR ( k ) = k / ρ reaches one, where deposition and noise contribute equally.
Let f = k / ( k + ρ ) = SNR / ( 1 + SNR ) f = k/(k+\rho) = \operatorname{SNR}/(1+\operatorname{SNR}) f = k / ( k + ρ ) = SNR / ( 1 + SNR ) be the share of the observed variance that is deposition. Gaussian noise leaves every cumulant above the second unchanged and inflates κ 2 \kappa_2 κ 2 by 1 / f 1/f 1/ f , so a standardized cumulant of order n n n , which carries κ 2 − n / 2 \kappa_2^{-n/2} κ 2 − n /2 , is attenuated by exactly f n / 2 f^{n/2} f n /2 :
γ 1 , k = γ 1 k f 3 / 2 = γ 1 k ( k + ρ ) 3 / 2 , \gamma_{1,k} = \frac{\gamma_1}{\sqrt{k}}f^{3/2} = \gamma_1\frac{k}{(k+\rho)^{3/2}}, γ 1 , k = k γ 1 f 3/2 = γ 1 ( k + ρ ) 3/2 k ,
γ 2 , k = γ 2 k f 2 = γ 2 k ( k + ρ ) 2 . (31) \gamma_{2,k} = \frac{\gamma_2}{k}f^2 = \gamma_2\frac{k}{(k+\rho)^2}. \tag*{(31)} γ 2 , k = k γ 2 f 2 = γ 2 ( k + ρ ) 2 k . (31)
the first factor in each being the noise-free value at that window. The two pull against each other: f n / 2 f^{n/2} f n /2 rises towards one as the window clears the noise (noise masking) and can go no further, while the aggregation factor (aggregation masking) falls without bound. A bounded gain against an unbounded loss forces an interior maximum, at k ∗ = 2 ρ k^* = 2\rho k ∗ = 2 ρ and k ∗ = ρ k^* = \rho k ∗ = ρ , with maxima ( 2 / ( 3 3 ) ) γ 1 / ρ (2/(3\sqrt{3}))\gamma_1/\sqrt{\rho} ( 2/ ( 3 3 )) γ 1 / ρ and γ 2 / ( 4 ρ ) \gamma_2/(4\rho) γ 2 / ( 4 ρ ) , so that once ρ ≫ 1 \rho\gg1 ρ ≫ 1 no window reveals the shape of the law: any window long enough to clear the noise has already aggregated the shape away (Fig. 4).
Detecting the shape is a different optimization from preserving it. Each disjoint increment consumes 2 k 2k 2 k base periods, so a record of N N N periods supplies n w = N / ( 2 k ) n_w=N/(2k) n w = N / ( 2 k ) independent windows, on which the sample skewness and sample excess kurtosis have standard errors 6 / n w \sqrt{6/n_w} 6/ n w and 24 / n w \sqrt{24/n_w} 24/ n w in the Gaussian limit ρ ≫ c 2 \rho\gg c^2 ρ ≫ c 2 . We define the separability index d ′ d' d ′ as the difference between the attenuated values of two laws, Δ γ 1 , k \Delta\gamma_{1,k} Δ γ 1 , k or Δ γ 2 , k \Delta\gamma_{2,k} Δ γ 2 , k , in units of that standard error: the distance, in standard errors, between the estimate obtained under one law and the value expected under the other. A longer window retains more of the cumulant but leaves fewer independent windows in a record of fixed length, and the two effects balance at k = ρ / 2 k=\rho/2 k = ρ /2 for the skewness and k = ρ / 3 k=\rho/3 k = ρ /3 for the excess kurtosis—both shorter than the windows that maximize the cumulants themselves, and both requiring ρ ≳ 3 \rho\gtrsim3 ρ ≳ 3 so that k ≥ 1 k \ge1 k ≥ 1 . Requiring d ′ = 2 d'=2 d ′ = 2 between the Gamma and inverse Gaussian laws (Δ γ 1 = c \Delta\gamma_1=c Δ γ 1 = c ) at the better of the two statistics then demands a record of
N ≈ 324 ρ 2 c 2 (32) N \approx324\frac{\rho^2}{c^2} \tag*{(32)} N ≈ 324 c 2 ρ 2 (32)
periods (Supplemental Material, Sec. S6). This pair is the easiest to separate; the Gamma and compound-Poisson laws differ by only Δ γ 1 = c / 2 \Delta\gamma_1=c/2 Δ γ 1 = c /2 and need four times as long a record. The accounting is also conservative: when readings are available at every base period, windows need only be separated by one period to keep their endpoint errors independent, n w ≃ N / k n_w \simeq N/k n w ≃ N / k , which halves the prefactor, and using all consecutive increments with their known first-order moving-average error structure would gain a further constant factor. Neither changes the ρ 2 \rho^2 ρ 2 scaling. The skewness is the cheaper statistic by a factor ρ / ( 14 c 2 ) \rho/(14c^2) ρ / ( 14 c 2 ) , so it, and not the excess kurtosis, sets the scaling of the requirement with the noise. Since the required record grows as the fourth power of the measurement noise, improving the instrument is by far the more effective route to the event-scale statistics. The estimate is asymptotic and specific to moment-based estimation, whereas the attenuation it rests on is not; its range of validity is given in the Supplemental Material, Sec. S6.
The numbers this implies can be estimated without reference to any particular record. A thermally reconstructed fouling resistance is inferred from temperature and flow measurements as a difference of reciprocals, R f = A ( 1 / U A − 1 / U A c l e a n ) R_f=A(1/UA-1/UA_{\mathrm{clean}}) R f = A ( 1/ U A − 1/ U A clean ) , so that a relative error ϵ \epsilon ϵ in the inferred conductance gives an error ϵ / U \epsilon/U ϵ / U in R f R_f R f , with U = U A / A U=UA/A U = U A / A . The deposition increment of one period changes the conductance by the relative amount ζ = U m τ 0 \zeta=Um\tau_0 ζ = U m τ 0 , and Eq. (30) becomes
ρ = 2 c 2 ( ϵ ζ ) 2 , N ≈ 1296 c 6 ( ϵ ζ ) 4 (33) \rho=\frac{2}{c^2}\left(\frac{\epsilon}{\zeta}\right)^2,\qquad N\approx\frac{1296}{c^6}\left(\frac{\epsilon}{\zeta}\right)^4 \tag*{(33)} ρ = c 2 2 ( ζ ϵ ) 2 , N ≈ c 6 1296 ( ζ ϵ ) 4 (33)
the second from Eq. (32) . For an exchanger with U ≈ 400 W m − 2 K − 1 U\approx400\ \mathrm{W\,m^{-2}\,K^{-1}} U ≈ 400 W m − 2 K − 1 fouling by 10 − 5 m 2 K W − 1 10^{-5}\ \mathrm{m^2\,K\,W^{-1}} 1 0 − 5 m 2 K W − 1 per three-day period, ζ ≈ 0.4 % \zeta\approx0.4\% ζ ≈ 0.4% : a conductance inferred to within 1 % 1\% 1% already carries a noise comparable to several periods of deposition (consistent with reported uncertainties of 4.6 % 4.6\% 4.6% for an industrial exchanger’s overall coefficient [52 ] and resistances uncertain below about 10 − 5 m 2 K W − 1 10^{-5}\ \mathrm{m^2\,K\,W^{-1}} 1 0 − 5 m 2 K W − 1 [7 ] ). Table IV evaluates Eq. (33) . Only the combination of a conductance accurate to about one per cent and strongly bursty deposition, c ≈ 2 c\approx2 c ≈ 2 , brings the requirement within years; at a few per cent, or for c ≲ 1 c\lesssim1 c ≲ 1 (the requirement scales as c − 6 c^{-6} c − 6 ), it runs to centuries and beyond, and for the granular values of c c c found in Sec. V B it is unbounded for practical purposes. The estimate is, moreover, favorable on two counts: it counts only the error of the inferred conductance, not the property and heat-loss closures behind the duty or the use of a design rather than a measured clean conductance, and it assumes errors that average away rather than correlate in time. For thermally reconstructed resistances, where a conductance error of a few per cent already exceeds the deposition of several sampling periods, the increment law is thus, in practice, unobservable.
Table IV. TABLE IV. Record length, in years, required to separate the Gamma and inverse-Gaussian laws, Eq. (33) , against the relative error ϵ \epsilon ϵ of the inferred conductance, for a relative conductance change ζ = 0.4 % \zeta=0.4\% ζ = 0.4% per sampling period τ 0 = 3 \tau_0=3 τ 0 = 3 days. The column ϵ / ζ \epsilon/\zeta ϵ / ζ expresses the noise in periods of deposition. Down a column the requirement grows as ϵ 4 \epsilon^4 ϵ 4 , across a row as c − 6 c^{-6} c − 6 . Entries marked † \dagger † have ρ < c 2 \rho<c^2 ρ < c 2 , outside the asymptotic regime of Eq. (32) , and are indicative only.
Measurement Uncertainty Effect. The error ϵ \epsilon ϵ follows from the reconstruction of the conductance. With U A = Q / Δ T l m UA=Q/\Delta T_{\mathrm{lm}} U A = Q /Δ T lm and the duty Q = m ˙ c p Δ T Q=\dot{m}c_p\Delta T Q = m ˙ c p Δ T taken from one stream, first-order propagation gives, neglecting correlations through shared sensors [6 , 7 ] ,
ϵ 2 ≃ ( δ m ˙ m ˙ ) 2 + ( δ ( Δ T ) Δ T ) 2 + ( δ ( Δ T l m ) Δ T l m ) 2 . (34) \epsilon^2 \simeq\left(\frac{\delta\dot{m}}{\dot{m}}\right)^2+\left(\frac{\delta(\Delta T)}{\Delta T}\right)^2+\left(\frac{\delta(\Delta T_{\mathrm{lm}})}{\Delta T_{\mathrm{lm}}}\right)^2. \tag*{(34)} ϵ 2 ≃ ( m ˙ δ m ˙ ) 2 + ( Δ T δ ( Δ T ) ) 2 + ( Δ T lm δ ( Δ T lm ) ) 2 . (34)
Only the random part of each error enters. A calibration offset of a thermocouple or a flow meter is constant over the record and cancels in the increments on which every statistic of this section is built, so that ϵ \epsilon ϵ is the precision error of a single reading: a better-calibrated instrument does not help, only a less noisy one does. For a stream heated by Δ T = 20 K \Delta T=20\ \mathrm{K} Δ T = 20 K , a log-mean temperature difference of 30 K 30\ \mathrm{K} 30 K , a random thermocouple error of 0.1 K 0.1\ \mathrm{K} 0.1 K and a random flow-meter error of 0.5 % 0.5\% 0.5% , the duty is uncertain to about 0.9 % 0.9\% 0.9% and the log-mean difference to about 0.3 % 0.3\% 0.3% , so that ϵ ≈ 1 % \epsilon\approx1\% ϵ ≈ 1% and ϵ / ζ ≈ 2.5 \epsilon/\zeta\approx2.5 ϵ / ζ ≈ 2.5 , the second row of Table IV . Averaging n n n readings within a sampling period reduces ϵ \epsilon ϵ by n \sqrt{n} n only if their errors are independent; plant noise is usually correlated in time, and ϵ \epsilon ϵ is then the error of the averaged reading.
The same competition between a noise floor at short lag and a degrading signal at long lag governs parameter estimation for noisy degradation and diffusion records generally [86 , 87 , 88 , 89 ] ; what is specific here is that the quantity estimated is a standardized higher cumulant, which aggregation itself destroys, so the upper cutoff is a property of the process rather than of the record length.
First-passage consequences
Consider deposition without removal, R ( t ) = R 0 + D ( t ) R(t)=R_0+D(t) R ( t ) = R 0 + D ( t ) , and a generic threshold x > R 0 x>R_0 x > R 0 . The continuous first-passage time is
T x = inf { t ≥ 0 : R ( t ) ≥ x } . (35) T_x=\inf\{t\ge0:R(t)\ge x\}. \tag*{(35)} T x = inf { t ≥ 0 : R ( t ) ≥ x } . (35)
Because the process is nondecreasing,
Pr ( T x ≤ t ) = Pr [ D ( t ) ≥ x − R 0 ] . (36) \Pr(T_x\le t)=\Pr[D(t)\ge x-R_0]. \tag*{(36)} Pr ( T x ≤ t ) = Pr [ D ( t ) ≥ x − R 0 ] . (36)
Figure 4. Observability is governed by the noise-to-signal ratio ρ \rho ρ alone. The vertical axis is the attenuated skewness γ 1 , k \gamma_{1,k} γ 1 , k , Eq. (31) , normalized by its noise-free value γ 1 \gamma_1 γ 1 , plotted against the aggregation window k k k (ρ = 0 \rho=0 ρ = 0 : the noise-free k − 1 / 2 k^{-1/2} k − 1/2 decay). The dot on each curve marks the best window k ∗ = 2 ρ k^*=2\rho k ∗ = 2 ρ , with peak value ( 2 / 3 3 ) γ 1 / ρ (2/3\sqrt{3})\gamma_1/\sqrt{\rho} ( 2/3 3 ) γ 1 / ρ : every curve has the same shape, rescaled by ρ \rho ρ , so only this peak height falls as ρ \rho ρ grows, clearing the dashed 1% line for ρ = 10 \rho=10 ρ = 10 , 10 2 10^2 1 0 2 , 10 3 10^3 1 0 3 but falling below it for any ρ ≳ 1.5 × 10 3 \rho\gtrsim1.5\times10^3 ρ ≳ 1.5 × 1 0 3 .
In an engineering application, setting x x x equal to a cleaning or performance limit makes T x T_x T x equivalent to a remaining useful life.
We define the dimensionless threshold distance
n x = x − R 0 m τ 0 . (37) n_x=\frac{x-R_0}{m\tau_0}. \tag*{(37)} n x = m τ 0 x − R 0 . (37)
For n x ≫ 1 n_x\gg1 n x ≫ 1 , and with time measured in sampling periods, the leading passage-time moments are
E [ T x ] τ 0 = n x + 1 + c 2 2 + o ( 1 ) , (38) \frac{\mathbb{E}[T_x]}{\tau_0}=n_x+\frac{1+c^2}{2}+o(1), \tag*{(38)} τ 0 E [ T x ] = n x + 2 1 + c 2 + o ( 1 ) , (38)
Var ( T x ) τ 0 2 = c 2 n x [ 1 + O ( n x − 1 ) ] , (39) \frac{\operatorname{Var}(T_x)}{\tau_0^2}=c^2n_x\left[1+O(n_x^{-1})\right], \tag*{(39)} τ 0 2 Var ( T x ) = c 2 n x [ 1 + O ( n x − 1 ) ] , (39)
the expansion being in the threshold distance n x n_x n x rather than in τ 0 \tau_0 τ 0 . Equation (37) is the mean for passage detected at the sampling instants; in continuous time the overshoot is the stationary overshoot of the subordinator, of mean ℓ / 2 \ell/2 ℓ /2 , and the correction is c 2 / 2 c^2/2 c 2 /2 instead: the extra half-period is the cost of discrete observation. To leading order,
CV ( T x ) = c n x + O ( n x − 1 ) . (40) \operatorname{CV}(T_x)=\frac{c}{\sqrt{n_x}}+O(n_x^{-1}). \tag*{(40)} CV ( T x ) = n x c + O ( n x − 1 ) . (40)
The mean rate fixes the leading passage time, while c c c fixes its leading relative uncertainty. Cleaning-schedule optimization under uncertainty treats the fouling law as deterministic with uncertain parameters or inputs [90 , 91 ] ; for a distant threshold that uncertainty dominates, the intrinsic variability adding only a relative spread c / n x c/\sqrt{n_x} c / n x . The deposition law enters neither the mean nor the spread, but only the third cumulant,
skew [ T x ] = 3 c − γ 1 n x , (41) \operatorname{skew}[T_x]=\frac{3c-\gamma_1}{\sqrt{n_x}}, \tag*{(41)} skew [ T x ] = n x 3 c − γ 1 , (41)
so that the remaining life is positively skewed for the compound Poisson and Gamma laws, whose increment skewness satisfies γ 1 < 3 c \gamma_1<3c γ 1 < 3 c , and symmetric to this order for the inverse Gaussian law, whose increment skewness is exactly 3 c 3c 3 c (Table II). Because the skew decays as n x − 1 / 2 n_x^{-1/2} n x − 1/2 (Figure 5 ), the law is visible in the life distribution only for a threshold a few increments away. At a threshold n x = 220 n_x=220 n x = 220 increments away, the three laws agree on the mean life to 0.2 sampling periods and on every percentile from the 5th to the 95th to within 0.9% (Supplemental Material, Sec. S7); at n x = 5 n_x=5 n x = 5 the same percentiles spread by up to three periods on a mean life of seven periods. For a distant threshold the choice of law is immaterial, and what is not immaterial is c c c , which sets the spread, and the uncertainty of m m m , which dominates it.
Figure 5. Distribution of the remaining useful life for compound-Poisson deposition at c = 2 c = 2 c = 2 (10 5 10^5 1 0 5 paths, passage detected at integer sampling periods), for thresholds n x = 5 , 20 , 40 , … , 200 n_x = 5, 20, 40, \ldots, 200 n x = 5 , 20 , 40 , … , 200 increments away. Each ridge is the probability mass of T x / τ 0 T_x/\tau_0 T x / τ 0 , scaled to a common peak height and drawn on a baseline at its n x n_x n x ; the black tick marks the mean life n x + ( 1 + c 2 ) / 2 n_x + (1 + c^2)/2 n x + ( 1 + c 2 ) /2 , Eq. (38) . The distribution drifts linearly with n x n_x n x , widens as c n x c\sqrt{n_x} c n x and loses its right skew as ( 3 c − γ 1 ) / n x (3c-\gamma_1)/\sqrt{n_x} ( 3 c − γ 1 ) / n x , Eq. (41) .
The asymptotic argument, overshoot qualification, and distinction between continuous passage and passage first detected at discrete observation times, together with the range of validity of Eqs. (38) –(41) at small n x n_x n x , are given in Supplemental Material, Sec. S7. With removal, the process is no longer monotone and Eq. (36) does not hold; the passage problem must then be treated separately. Exact results are nevertheless available in special cases: for constant-rate removal the moments of the time to clear a deposit are known in closed form for compound Poisson deposition [92 ] , and for infinite-activity deposition the passage event can be simulated exactly [93 , 94 ] .
The preceding results assume stationary operating conditions. A known, deterministic variation in deposition intensity can be represented by the operational clock
Λ ( t ) = ∫ 0 t χ ( u ) d u , (42) \Lambda(t)=\int_0^t \chi(u)\,\mathrm{d}u, \tag*{(42)} Λ ( t ) = ∫ 0 t χ ( u ) d u , (42)
where χ ( u ) ≥ 0 \chi(u) \ge0 χ ( u ) ≥ 0 is a relative deposition intensity. Replacing D ( t ) D(t) D ( t ) by D [ Λ ( t ) ] D[\Lambda(t)] D [ Λ ( t )] makes all cumulants scale with Λ ( t ) \Lambda(t) Λ ( t ) ; in particular,
E { D [ Λ ( t ) ] } = m Λ ( t ) . (43) \mathbb{E}\{D[\Lambda(t)]\}=m\Lambda(t). \tag*{(43)} E { D [ Λ ( t )]} = m Λ ( t ) . (43)
The stationary theory therefore remains valid on the operational clock. Prediction in calendar time additionally requires future operating conditions, whose uncertainty is distinct from the intrinsic deposition variability represented by c c c and α \alpha α .
Implications and conclusions
The classical linear and asymptotic fouling laws are universal first-moment laws of stationary additive deposition with a finite mean rate, and this fixes what a monitored fouling resistance can reveal. A linear or Kern–Seaton fit returns a net deposition rate m m m and, if the record saturates, a relaxation rate ψ \psi ψ , or equivalently R ∞ R_\infty R ∞ and ψ − 1 \psi^{-1} ψ − 1 ; nothing more is in the mean. A good fit is therefore not evidence that removal acts, nor of which removal mechanism acts, nor of any deposition model, since every additive mechanism with the same m m m and ψ \psi ψ fits equally well (§ 3 A); the same caution applies to a best-fitting blocking law in membrane filtration [95 , 96 ] and, once measurement noise is included, to an increment law selected by likelihood [54 ] . Even m m m is a single combination: writing deposition as transport to the wall followed by attachment [79 ] , it constrains the product of a transport coefficient, an attachment probability and the deposit density, not the factors individually. How m m m and ψ \psi ψ depend on velocity, temperature and composition remains the proper object of mechanistic modelling, but must be read across operating points, not from the shape of one curve.
The information the fit discards is in the scatter about it, and for a saturating record it can be read with three steps. First, subtract the fitted mean curve. Second, estimate the variance of the residuals on the plateau and subtract the measurement contribution (ϵ / U \epsilon/U ϵ / U )2 ^2 2 , with ϵ \epsilon ϵ from Eq. (34) . Third, compare the excess with the scales of Table III. Continuous removal and per-element detachment leave a plateau variance of order ℓ R ∞ \ell R_\infty ℓ R ∞ , a relative scatter C V ∞ 2 ∼ ℓ / R ∞ \mathrm{CV}_\infty^2 \sim\ell/R_\infty CV ∞ 2 ∼ ℓ / R ∞ that is small whenever the plateau holds many events. Sloughing leaves a floor of order η R ∞ 2 \eta R_\infty^2 η R ∞ 2 that does not shrink as the deposition becomes smooth, Eq. (24) . A plateau scatter well above the measurement noise, and a fair fraction of R ∞ R_\infty R ∞ , therefore points to intermittent removal; a scatter at the noise level is consistent with continuous removal or with blocking. Blocking is told apart in time rather than on the plateau: its variance peaks at half coverage and vanishes at saturation (§ 3 C). The autocorrelation of the residuals does not discriminate, since it decays as e − ψ τ e^{-\psi\tau} e − ψ τ for all three mechanisms of Table III.
The deposition law itself, beyond its coefficient of variation c c c , cannot be recovered from a thermally reconstructed resistance at realistic measurement accuracy (Table IV ); as § 3 D showed, a better instrument helps far more than a longer record, and a better-calibrated one not at all.
For a cleaning limit many sampling periods away, the time to reach it is set by the mean rate, its uncertainty, and c c c through the relative spread c / n x c/\sqrt{n_x} c / n x , Eq. (40) : the deposition law enters only through smoothness that fades with distance from the limit (Sec. VI). Schedules can therefore be planned on the fitted rate, with an uncertainty budget dominated by that of m m m , and the choice of law matters only when the limit is a few periods away.
Curve fitting and mechanism identification are thus different tasks: the first is governed by universal mean laws, the second requires fluctuation statistics, independent measurements, or observations across operating conditions. The formulation is restricted to additive, non-decreasing deposition with a removal rate that depends only on the current state; processes that change the thermal resistance without changing the deposited mass, such as ageing, consolidation or diffusion within wax gels [97 , 98 ] , and state-triggered detachment such as spalling fall outside it. Extending the theory to these effects, and testing the plateau-variance discriminator on records that visibly saturate, are the natural next steps.
Acknowledgments
This publication has been funded by the SFI NorwAI (Centre for Research-based Innovation, 309834). The author gratefully acknowledges the financial support from the Research Council of Norway and the partners of the SFI NorwAI.
Data Availability
Code producing figures and numerical checks is available at github.com/simonecasolo/fouling-first-moment-laws.
Notation
The deposition D ( t ) D(t) D ( t ) is a subordinator with Lévy measure ν ( d x ) \nu(dx) ν ( d x ) on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , activity r = ∫ ν ( d x ) r=\int\nu(dx) r = ∫ ν ( d x ) and cumulant rates q n = ∫ 0 ∞ x n ν ( d x ) q_n=\int_0^\infty x^n\nu(dx) q n = ∫ 0 ∞ x n ν ( d x ) , so that κ n [ D ( t ) ] = q n t \kappa_n[D(t)]=q_nt κ n [ D ( t )] = q n t . Its law is fixed by the Laplace exponent through E [ e − θ D ( t ) ] = e − t Φ ( θ ) \mathbb{E}[e^{-\theta D(t)}]=e^{-t\Phi(\theta)} E [ e − θ D ( t ) ] = e − t Φ ( θ ) with
Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) ν ( d x ) , θ ≥ 0. (44) \Phi(\theta)=\int_0^\infty(1-e^{-\theta x})\,\nu(dx),\qquad\theta\ge0. \tag*{(44)} Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) ν ( d x ) , θ ≥ 0. (44)
We write m = q 1 m=q_1 m = q 1 for the mean deposition rate, s 2 = q 2 s^2=q_2 s 2 = q 2 for the variance rate, c 2 = s 2 / ( m 2 τ 0 ) c^2=s^2/(m^2\tau_0) c 2 = s 2 / ( m 2 τ 0 ) for the squared coefficient of variation over a reference window τ 0 \tau_0 τ 0 , t c = s 2 / m 2 t_c=s^2/m^2 t c = s 2 / m 2 and ℓ = q 2 / q 1 = s 2 / m = m t c \ell=q_2/q_1=s^2/m=mt_c ℓ = q 2 / q 1 = s 2 / m = m t c for the size-biased event size.
Finite activity and the compound Poisson equivalence
This section proves the equivalence in Sec. 3.1: a subordinator has finite activity, r = ∫ ν ( d x ) < ∞ r=\int\nu(dx)<\infty r = ∫ ν ( d x ) < ∞ , if and only if it is a compound Poisson process [57 , 58 ] . It is worth separating the two directions: finite activity is a statement about the Lévy measure ν \nu ν , whereas being a compound Poisson process is a statement about how the sample path is assembled from discrete jumps.
From finite activity to a Poisson sum of jumps. Suppose r < ∞ r<\infty r < ∞ . Then ν \nu ν splits into an overall rate and a shape,
ν ( d x ) = r F ( d x ) , (45) \nu(dx)=rF(dx), \tag*{(45)} ν ( d x ) = r F ( d x ) , (45)
and F F F is a probability distribution on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , since ∫ F ( d x ) = r − 1 ∫ ν ( d x ) = 1 \int F(dx)=r^{-1}\int\nu(dx)=1 ∫ F ( d x ) = r − 1 ∫ ν ( d x ) = 1 . This normalization step is only possible because r r r is finite. Substituting ν = r F \nu=rF ν = r F into the Lévy–Khintchine exponent, Eq. (S1),
Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) r F ( d x ) = r [ 1 − F ^ ( θ ) ] . (46) \Phi(\theta)=\int_0^\infty(1-e^{-\theta x})\,rF(dx)=r[1-\widehat{F}(\theta)]. \tag*{(46)} Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) r F ( d x ) = r [ 1 − F ( θ )] . (46)
All dependence on the jump-size distribution is now carried by F ^ \widehat{F} F , where F ^ ( θ ) = ∫ 0 ∞ e − θ x F ( d x ) \widehat{F}(\theta)=\int_0^\infty e^{-\theta x}F(dx) F ( θ ) = ∫ 0 ∞ e − θ x F ( d x ) is the Laplace transform of F F F . Insert Eq. (S3) into E [ e − θ D ( t ) ] = e − t Φ ( θ ) \mathbb{E}[e^{-\theta D(t)}]=e^{-t\Phi(\theta)} E [ e − θ D ( t ) ] = e − t Φ ( θ ) and expand the exponential in powers of F ^ ( θ ) \widehat{F}(\theta) F ( θ ) :
E [ e − θ D ( t ) ] = e − r t e r t F ^ ( θ ) = ∑ n = 0 ∞ ( r t ) n e − r t n ! F ^ ( θ ) n , (47) \mathbb{E}[e^{-\theta D(t)}]=e^{-rt}e^{rt\widehat{F}(\theta)}
=\sum_{n=0}^{\infty}\frac{(rt)^n e^{-rt}}{n!}\widehat{F}(\theta)^n, \tag*{(47)} E [ e − θ D ( t ) ] = e − r t e r t F ( θ ) = n = 0 ∑ ∞ n ! ( r t ) n e − r t F ( θ ) n , (47)
where the coefficient of F ^ ( θ ) n \widehat{F}(\theta)^n F ( θ ) n is a Poisson( r t ) (rt) ( r t ) probability. This is exactly the Laplace transform one obtains by conditioning on the number of jumps: for N ( t ) ∼ Poisson ( r t ) N(t)\sim\operatorname{Poisson}(rt) N ( t ) ∼ Poisson ( r t ) and X i ∼ i i d F X_i\mathrel{\overset{\mathrm{iid}}{\sim}}F X i ∼ iid F independent of N ( t ) N(t) N ( t ) ,
E [ e − θ ∑ i = 1 N ( t ) X i ] = ∑ n = 0 ∞ Pr [ N ( t ) = n ] E [ e − θ ∑ i = 1 n X i ] = ∑ n = 0 ∞ Pr [ N ( t ) = n ] F ^ ( θ ) n , (48) \mathbb{E}\left[e^{-\theta\sum_{i=1}^{N(t)}X_i}\right]
=\sum_{n=0}^{\infty}\Pr[N(t)=n]\mathbb{E}\left[e^{-\theta\sum_{i=1}^{n}X_i}\right]
=\sum_{n=0}^{\infty}\Pr[N(t)=n]\widehat{F}(\theta)^n, \tag*{(48)} E [ e − θ ∑ i = 1 N ( t ) X i ] = n = 0 ∑ ∞ Pr [ N ( t ) = n ] E [ e − θ ∑ i = 1 n X i ] = n = 0 ∑ ∞ Pr [ N ( t ) = n ] F ( θ ) n , (48)
which is term-by-term identical to Eq. (S4). Since the Laplace transform determines the law, D ( t ) = ∑ i = 1 N ( t ) X i D(t)=\sum_{i=1}^{N(t)}X_i D ( t ) = ∑ i = 1 N ( t ) X i : D D D is, by definition, a compound Poisson process of rate r r r and jump law F = ν / r F=\nu/r F = ν / r .
From a Poisson sum of jumps back to finite activity. The converse needs no further calculation, only reading the definition in the other direction. A compound Poisson process of rate r r r and jump law F F F has a Lévy–Khintchine representation with Lévy measure ν = r F \nu=rF ν = r F ; by uniqueness of that representation, this is the Lévy measure of the process, not merely one that reproduces its law. Since F F F is a probability measure, ∫ ν ( d x ) = r ∫ F ( d x ) = r < ∞ \int\nu(dx)=r\int F(dx)=r<\infty ∫ ν ( d x ) = r ∫ F ( d x ) = r < ∞ : the process automatically has finite activity, equal to its jump rate.
Put together, the two directions say that the map ν ↔ ( r , F ) \nu\leftrightarrow(r,F) ν ↔ ( r , F ) is a bijection between Lévy measures with r < ∞ r<\infty r < ∞ and (rate, jump-law) pairs: every finite-activity subordinator is a compound Poisson process, and every compound Poisson process has finite activity equal to its own rate.
Tempered-stable family: derivations
This section carries out the five calculations that the tempered-stable spectrum, Eq. (13) of the main text, is built to support: fixing the two free scale parameters B B B and λ \lambda λ in terms of the physically meaningful triple ( m , c , τ 0 ) (m,c,\tau_0) ( m , c , τ 0 ) at any shape α \alpha α ; obtaining the cumulant rates q n q_n q n , and from them the aggregation scaling of the main text; obtaining the Laplace exponent in closed form; identifying the three named laws of Table 2 of the main text as three points on this family; and separating three notions—activity, path continuity, and observability—that are easy to conflate. Throughout, ℓ = q 2 / q 1 \ell=q_2/q_1 ℓ = q 2 / q 1 is the size-biased event size of the main text: the typical event size as seen by the mass it deposits, rather than by a plain count of events. Only the first calculation, the matching relation Eq. (S7), is needed before any of the others, so a reader interested only in one named law can read that paragraph and then skip ahead.
Cumulant rates, and matching to ( m , c , τ 0 ) (m,c,\tau_0) ( m , c , τ 0 ) . The spectrum ν ( d x ) \nu(dx) ν ( d x ) gives the rate at which events of size x x x occur; its moments q n = ∫ 0 ∞ x n ν ( d x ) q_n = \int_0^\infty x^n \nu(dx) q n = ∫ 0 ∞ x n ν ( d x ) are therefore exactly the cumulant rates of the subordinator, with q 1 = m q_1 = m q 1 = m the mean deposition rate and q 2 q_2 q 2 the variance rate (s 2 = q 2 s^2 = q_2 s 2 = q 2 in the notation of the main text). For the tempered-stable spectrum every such moment reduces to a single Gamma-function integral:
q n = ∫ 0 ∞ x n ν ( d x ) = B ∫ 0 ∞ x n − 1 − α e − λ x d x = B Γ ( n − α ) λ α − n , (49) q_n = \int_0^\infty x^n \nu(dx) = B \int_0^\infty x^{n-1-\alpha} e^{-\lambda x}\,dx
= B\Gamma(n-\alpha)\lambda^{\alpha-n}, \tag*{(49)} q n = ∫ 0 ∞ x n ν ( d x ) = B ∫ 0 ∞ x n − 1 − α e − λ x d x = B Γ ( n − α ) λ α − n , (49)
valid for every n ≥ 1 n \ge1 n ≥ 1 : the exponential tempering factor e − λ x e^{-\lambda x} e − λ x ensures the integral converges at large x x x regardless of n n n , and the restriction α < 1 \alpha< 1 α < 1 ensures convergence at x → 0 x \to0 x → 0 , which is what keeps even the first moment, q 1 = m q_1 = m q 1 = m , finite.
Two of these moments are enough to fix both free parameters of the family, B B B and λ \lambda λ . Imposing q 1 = m q_1 = m q 1 = m sets the overall deposition rate; imposing q 2 / q 1 = ℓ q_2/q_1 = \ell q 2 / q 1 = ℓ sets the size-biased event size rather than q 2 q_2 q 2 itself, because ℓ \ell ℓ carries the same units as a jump size and is the more physically transparent quantity to specify. Using Γ ( 2 − α ) = ( 1 − α ) Γ ( 1 − α ) \Gamma(2-\alpha) = (1-\alpha)\Gamma(1-\alpha) Γ ( 2 − α ) = ( 1 − α ) Γ ( 1 − α ) to simplify the resulting ratio, these two conditions solve to
λ = 1 − α ℓ = 1 − α c 2 m τ 0 , B = m λ 1 − α Γ ( 1 − α ) . (50) \lambda= \frac{1-\alpha}{\ell} = \frac{1-\alpha}{c^2m\tau_0}, \qquad B = \frac{m\lambda^{1-\alpha}}{\Gamma(1-\alpha)}. \tag*{(50)} λ = ℓ 1 − α = c 2 m τ 0 1 − α , B = Γ ( 1 − α ) m λ 1 − α . (50)
Only two parameters are free in the family, so once λ \lambda λ and B B B are fixed by m m m and ℓ \ell ℓ at a chosen α \alpha α , every higher q n q_n q n is determined by Eq. (49) ; no further fitting is possible or needed, and α \alpha α alone governs how the higher cumulants compare to the first two. This dependence can be made explicit without re-evaluating Gamma functions at every n n n . Taking the ratio of consecutive cumulant rates in Eq. (49) gives q n / q n − 1 = ( n − 1 − α ) / λ q_n/q_{n-1} = (n-1-\alpha)/\lambda q n / q n − 1 = ( n − 1 − α ) / λ , and substituting λ = ( 1 − α ) / ℓ \lambda= (1-\alpha)/\ell λ = ( 1 − α ) / ℓ from Eq. (50) turns this into a one-step recursion, q n / q n − 1 = ( n − 1 − α ) ℓ / ( 1 − α ) q_n/q_{n-1} = (n-1-\alpha)\ell/(1-\alpha) q n / q n − 1 = ( n − 1 − α ) ℓ / ( 1 − α ) . Iterating the recursion down from q 1 = m q_1 = m q 1 = m telescopes it into the closed form
q n = m ℓ n − 1 ∏ k = 2 n − 1 k − α 1 − α ( n ≥ 2 ) ; (51) q_n = m\ell^{n-1}\prod_{k=2}^{n-1}\frac{k-\alpha}{1-\alpha}\qquad(n\ge2); \tag*{(51)} q n = m ℓ n − 1 k = 2 ∏ n − 1 1 − α k − α ( n ≥ 2 ) ; (51)
the product is empty (equal to 1) for n = 2 n = 2 n = 2 , consistent with q 2 = m ℓ q_2 = m\ell q 2 = m ℓ by the definition of ℓ \ell ℓ itself—a useful check on the algebra before using the recursion for n = 3 , 4 n = 3,4 n = 3 , 4 below.
Skewness and kurtosis of the increments. The third and fourth cumulant rates are the ones most directly tied to data, since they set the skewness and excess kurtosis of the increment distribution over any observation window. Setting n = 3 , 4 n = 3,4 n = 3 , 4 in Equation (51) gives
q 3 q 2 = 2 − α 1 − α ℓ , q 4 q 2 = ( 2 − α ) ( 3 − α ) ( 1 − α ) 2 ℓ 2 . (52) \frac{q_3}{q_2} = \frac{2-\alpha}{1-\alpha}\ell,\qquad
\frac{q_4}{q_2} = \frac{(2-\alpha)(3-\alpha)}{(1-\alpha)^2}\ell^2. \tag*{(52)} q 2 q 3 = 1 − α 2 − α ℓ , q 2 q 4 = ( 1 − α ) 2 ( 2 − α ) ( 3 − α ) ℓ 2 . (52)
These enter the standardized cumulants of a window Δ \Delta Δ , γ 1 ( Δ ) = q 3 q 2 − 3 / 2 Δ − 1 / 2 \gamma_1(\Delta) = q_3q_2^{-3/2}\Delta^{-1/2} γ 1 ( Δ ) = q 3 q 2 − 3/2 Δ − 1/2 and γ 2 ( Δ ) = q 4 q 2 − 2 Δ − 1 \gamma_2(\Delta) = q_4q_2^{-2}\Delta^{-1} γ 2 ( Δ ) = q 4 q 2 − 2 Δ − 1 , only through the combination q 2 Δ = s 2 Δ q_2\Delta= s^2\Delta q 2 Δ = s 2 Δ , the accumulated variance over the window; this is expected, since γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 are dimensionless cumulant ratios and Δ \Delta Δ can only enter through how much variance has built up by the end of the window. Writing this out and using c ( Δ ) 2 = ℓ / ( m Δ ) = s 2 / ( m 2 Δ ) c(\Delta)^2 = \ell/(m\Delta) = s^2/(m^2\Delta) c ( Δ ) 2 = ℓ / ( m Δ ) = s 2 / ( m 2 Δ ) to trade ℓ \ell ℓ and Δ \Delta Δ for the single window-dependent coefficient of variation c ( Δ ) c(\Delta) c ( Δ ) reproduces the aggregation scaling of the main text directly, i.e. it is this substitution that the main text’s statement of how γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 shrink with the aggregation window rests on. A useful check that survives this substitution is the ratio
γ 2 ( Δ ) γ 1 ( Δ ) 2 = q 4 q 2 q 3 2 = 3 − α 2 − α , (53) \frac{\gamma_2(\Delta)}{\gamma_1(\Delta)^2} = \frac{q_4q_2}{q_3^2} = \frac{3-\alpha}{2-\alpha}, \tag*{(53)} γ 1 ( Δ ) 2 γ 2 ( Δ ) = q 3 2 q 4 q 2 = 2 − α 3 − α , (53)
which depends on α \alpha α alone: c c c and Δ \Delta Δ cancel because both γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 carry the same power of c ( Δ ) c(\Delta) c ( Δ ) through the aggregation scaling of the main text (γ 1 ∝ c ( Δ ) \gamma_1 \propto c(\Delta) γ 1 ∝ c ( Δ ) , γ 2 ∝ c ( Δ ) 2 \gamma_2 \propto c(\Delta)^2 γ 2 ∝ c ( Δ ) 2 ), so this ratio is a pure shape diagnostic, usable in principle to estimate α \alpha α from data without first knowing c c c , since it depends only on the shape of the increment distribution and not on its overall scale.
Laplace exponent. The Laplace exponent
Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) ν ( d x ) , (54) \Phi(\theta) = \int_0^\infty(1-e^{-\theta x})\nu(dx), \tag*{(54)} Φ ( θ ) = ∫ 0 ∞ ( 1 − e − θ x ) ν ( d x ) , (54)
already defined in the Lévy–Khintchine formula, Eq. (44) , is the generating function of the subordinator: it determines Δ D \Delta D Δ D ’s distribution over any window through E [ e − θ Δ D ] = e − Δ Φ ( θ ) \mathbb{E}[e^{-\theta\Delta D}] = e^{-\Delta\Phi(\theta)} E [ e − θ Δ D ] = e − ΔΦ ( θ ) , and is the object matched below to identify the increment law of each named process. It cannot be obtained by integrating x − 1 − α e − λ x x^{-1-\alpha}e^{-\lambda x} x − 1 − α e − λ x alone: for α ≥ 0 \alpha\ge0 α ≥ 0 that integral diverges at x = 0 x = 0 x = 0 , since a single power-law tail with α ≥ 0 \alpha\ge0 α ≥ 0 carries infinite activity (infinitely many arbitrarily small events, Sec. 3.1 of the main text). What converges is the difference 1 − e − θ x 1-e^{-\theta x} 1 − e − θ x against the spectrum, because this difference vanishes linearly as x → 0 x \to0 x → 0 and tames the singularity that defeats direct integration. Since ( 1 − e − θ x ) e − λ x = e − λ x − e − ( λ + θ ) x (1-e^{-\theta x})e^{-\lambda x} = e^{-\lambda x}-e^{-(\lambda+\theta)x} ( 1 − e − θ x ) e − λ x = e − λ x − e − ( λ + θ ) x , the tempering already present in ν \nu ν turns the integrand into a difference of two exponentials, each divergent against x − 1 − α x^{-1-\alpha} x − 1 − α at x = 0 x = 0 x = 0 but with a convergent difference. This reduces Φ \Phi Φ to the one nontrivial integral
∫ 0 ∞ ( e − λ x − e − ( λ + θ ) x ) x − 1 − α d x = Γ ( − α ) [ λ α − ( λ + θ ) α ] , (55) \int_0^\infty\left(e^{-\lambda x}-e^{-(\lambda+\theta)x}\right)x^{-1-\alpha}\,dx
= \Gamma(-\alpha)\left[\lambda^\alpha-(\lambda+\theta)^\alpha\right], \tag*{(55)} ∫ 0 ∞ ( e − λ x − e − ( λ + θ ) x ) x − 1 − α d x = Γ ( − α ) [ λ α − ( λ + θ ) α ] , (55)
a standard tempered-stable identity [60 ] , a generalization of Frullani’s integral to a power-law weight. Multiplying Eq. (55) by B B B gives Φ ( θ ) \Phi(\theta) Φ ( θ ) itself, so the only remaining step is to trade B B B and Γ ( − α ) \Gamma(-\alpha) Γ ( − α ) —both of which are awkward to evaluate directly for general α \alpha α , the latter diverging at every nonnegative integer—for the physically meaningful m m m and λ \lambda λ already fixed in Eq. (50) . Using Γ ( 1 − α ) = − α Γ ( − α ) \Gamma(1-\alpha) = -\alpha\Gamma(-\alpha) Γ ( 1 − α ) = − α Γ ( − α ) together with the matching relation B Γ ( 1 − B\Gamma(1- B Γ ( 1 − α ) = m λ 1 − α \alpha)=m\lambda^{1-\alpha} α ) = m λ 1 − α from Eq. (50) eliminates B B B and Γ ( − α ) \Gamma(-\alpha) Γ ( − α ) in favor of m m m and λ \lambda λ alone:
Φ ( θ ) = m λ α [ ( 1 + θ λ ) α − 1 ] , α ≠ 0 , (56) \Phi(\theta)=\frac{m\lambda}{\alpha}\left[\left(1+\frac{\theta}{\lambda}\right)^\alpha-1\right],\qquad\alpha\ne0, \tag*{(56)} Φ ( θ ) = α mλ [ ( 1 + λ θ ) α − 1 ] , α = 0 , (56)
Φ ( θ ) = m λ ln ( 1 + θ / λ ) , α = 0 , (57) \Phi(\theta)=m\lambda\ln(1+\theta/\lambda),\qquad\alpha=0, \tag*{(57)} Φ ( θ ) = mλ ln ( 1 + θ / λ ) , α = 0 , (57)
the α = 0 \alpha=0 α = 0 case following from the α ≠ 0 \alpha\ne0 α = 0 one by L’Hôpital’s rule as α → 0 \alpha\to0 α → 0 , since [ ( 1 + θ / λ ) α − 1 ] / α → ln ( 1 + θ / λ ) [(1+\theta/\lambda)^\alpha-1]/\alpha\to\ln(1+\theta/\lambda) [( 1 + θ / λ ) α − 1 ] / α → ln ( 1 + θ / λ ) in that limit.
Mass-weighted spectrum. The spectrum ν ( x ) \nu(x) ν ( x ) weights every event by how often it occurs, not by how much mass it carries, so a large rare event and a small common one can have the same ν ( x ) \nu(x) ν ( x ) while contributing very differently to the deposit. To see the spectrum from the standpoint of mass rather than event count, multiply the event-rate density by the size x x x and normalize by the mean rate m m m : this turns ν \nu ν into a genuine probability density, w ( x ) = x ν ( x ) / m w(x)=x\nu(x)/m w ( x ) = xν ( x ) / m , giving the fraction of deposited mass that arrives in events of size near x x x :
w ( x ) = λ 1 − α Γ ( 1 − α ) x − α e − λ x , (58) w(x)=\frac{\lambda^{1-\alpha}}{\Gamma(1-\alpha)}x^{-\alpha}e^{-\lambda x}, \tag*{(58)} w ( x ) = Γ ( 1 − α ) λ 1 − α x − α e − λ x , (58)
which is recognizable as the Gamma( 1 − α , λ ) (1-\alpha,\lambda) ( 1 − α , λ ) density — shape 1 − α 1-\alpha 1 − α , rate λ \lambda λ — with mean ( 1 − α ) / λ = ℓ (1-\alpha)/\lambda=\ell ( 1 − α ) / λ = ℓ by construction (consistent with ℓ \ell ℓ being defined as the size-biased, i.e. mass-weighted, event size) and variance ( 1 − α ) / λ 2 (1-\alpha)/\lambda^2 ( 1 − α ) / λ 2 , i.e. relative width (coefficient of variation) ( 1 − α ) − 1 / 2 (1-\alpha)^{-1/2} ( 1 − α ) − 1/2 . This relative width is governed by α \alpha α alone, independently of m m m and ℓ \ell ℓ : the mass spectrum narrows as α → − ∞ \alpha\to-\infty α → − ∞ , concentrating the deposit into a few near-identical events, and broadens as α → 1 \alpha\to1 α → 1 , spreading it over an increasingly wide range of event sizes. This is the precise sense in which α \alpha α “controls how the deposit is distributed over event sizes” in Sec. 3.2 of the main text.
The three named laws. The tempered-stable family is a one-parameter continuum in α \alpha α , but three particular values recover subordinators with well-known, independently derived increment distributions. Evaluating Eqs. (50) –(57) at these three values of α \alpha α recovers the three named processes, identified by matching Φ \Phi Φ above against their known Laplace exponents term by term; once Φ \Phi Φ matches, the increment law over any window Δ \Delta Δ follows immediately, since the Laplace exponent determines the distribution uniquely. Beyond the algebra, the three laws correspond to three qualitatively different pictures of how the deposit actually accumulates, set by how far α \alpha α has pushed the event-size spectrum toward few-and-large or many-and-small events (Mass-weighted spectrum, above); it is worth stating each picture in words before the matching calculation that confirms it:
α = − 1 \alpha=-1 α = − 1 (compound Poisson). Physically, this is the law of a surface that sits idle most of the time and then, at random, unrelated instants, gains a lump of deposit of random size: nothing happens between events, and when something does happen, its size is drawn independently of everything else. Concretely, the waiting time between successive events is exponential (events occur at a constant rate, with no memory of how long it has already been quiet), and each event’s size is itself exponentially distributed, independently of the waiting times and of all other events. It is this combination—a finite, constant rate of occurrence together with jumps that occur strictly one at a time—that produces the quiescent stretches visible in Figure 2 (d) of the main text: genuine flat intervals of the process, not merely small increments. Algebraically, substituting α = − 1 \alpha=-1 α = − 1 into Eq. (50) gives λ = 2 / ℓ \lambda=2/\ell λ = 2/ ℓ and B = m λ 2 B=m\lambda^2 B = m λ 2 , so by the finite-activity criterion of Sec. 3.1 of the main text the total event rate is r = ∫ ν ( x ) d x = B / λ = 2 / ℓ = 2 / t c r=\int\nu(x)\,\mathrm{d}x=B/\lambda=2/\ell=2/t_c r = ∫ ν ( x ) d x = B / λ = 2/ ℓ = 2/ t c : events of all sizes occur at a finite combined rate, as required for a list of discrete events to exist at all. This lets Sec. S2 apply directly: the normalized jump law is F ( d x ) = ν ( d x ) / r = λ e − λ x d x F(\mathrm{d}x)=\nu(\mathrm{d}x)/r=\lambda e^{-\lambda x}\,\mathrm{d}x F ( d x ) = ν ( d x ) / r = λ e − λ x d x , exponential of mean 1 / λ = ℓ / 2 1/\lambda=\ell/2 1/ λ = ℓ /2 , so Δ D = ∑ i = 1 N J i \Delta D=\sum_{i=1}^{N}J_i Δ D = ∑ i = 1 N J i with N ∼ Poisson ( r Δ ) N\sim\operatorname{Poisson}(r\Delta) N ∼ Poisson ( r Δ ) and J i ∼ i i d Exp ( mean = ℓ / 2 ) J_i\mathrel{\stackrel{\mathrm{iid}}{\sim}}\operatorname{Exp}(\text{mean}=\ell/2) J i ∼ iid Exp ( mean = ℓ /2 ) , confirming in symbols the picture above: a Poisson-distributed number of events, each an independent draw from an exponential jump-size distribution. More generally, any α < 0 \alpha<0 α < 0 —not only α = − 1 \alpha=-1 α = − 1 —keeps r < ∞ r<\infty r < ∞ and the same idle-then-jump picture, only with the jump size now F F F a Gamma(shape = − α =-\alpha = − α , rate = λ =\lambda = λ )-distributed draw rather than a plain exponential; writing a = − α a=-\alpha a = − α for the shape, the aggregation scaling of the main text gives γ 1 / c = ( a + 2 ) / ( a + 1 ) \gamma_1/c=(a+2)/(a+1) γ 1 / c = ( a + 2 ) / ( a + 1 ) , which recovers 3 / 2 3/2 3/2 at a = 1 a=1 a = 1 (α = − 1 \alpha=-1 α = − 1 ) above.
α = 0 \alpha=0 α = 0 (Gamma process). Physically, this is the opposite extreme from compound Poisson: instead of occasional large jumps separated by quiescent gaps, the surface receives a continuous rain of events so frequent and so small that no individual event can be pointed to, and no stretch of time is ever truly flat. The record still only ever increases (deposition cannot run backward), but it does so through an uninterrupted accumulation of infinitesimal contributions rather than through discrete, separately identifiable arrivals. Formally, λ = 1 / ℓ \lambda=1/\ell λ = 1/ ℓ and B = m λ B=m\lambda B = mλ ; the activity is infinite (0 ≤ α < 1 0\le\alpha<1 0 ≤ α < 1 ), so there is no finite list of discrete events to thin or enumerate—the process deposits through a dense set of arbitrarily small jumps rather than a countable sequence of them, which is the precise sense of “continuous rain” above. Even so, Φ ( θ ) = m λ ln ( 1 + θ / λ ) \Phi(\theta)=m\lambda\ln(1+\theta/\lambda) Φ ( θ ) = mλ ln ( 1 + θ / λ ) , Eq. (57) , is exactly the Laplace exponent of a Gamma process of shape rate B B B and scale 1 / λ 1/\lambda 1/ λ , for which Δ D ∼ Gamma ( shape = B Δ , scale = 1 / λ ) \Delta D\sim\operatorname{Gamma}(\text{shape}=B\Delta,\text{scale}=1/\lambda) Δ D ∼ Gamma ( shape = B Δ , scale = 1/ λ ) : the match is read off by comparing this expression term by term against the textbook Laplace exponent of a Gamma process. Since B Δ = m Δ / ℓ = Δ / t c B\Delta=m\Delta/\ell=\Delta/t_c B Δ = m Δ/ ℓ = Δ/ t c and 1 / λ = ℓ 1/\lambda=\ell 1/ λ = ℓ , this is Δ D ∼ Gamma ( shape = Δ / t c , scale = ℓ ) \Delta D\sim\operatorname{Gamma}(\text{shape}=\Delta/t_c,\text{scale}=\ell) Δ D ∼ Gamma ( shape = Δ/ t c , scale = ℓ ) .
finite activity, so there are no isolated events or quiescent gaps either, but its mass spectrum is broader (Mass-weighted spectrum, above), so a somewhat larger share of the deposit is carried by comparatively rarer, comparatively larger contributions riding on top of the same dense background of infinitesimal ones—a bursty, rather than perfectly smooth, continuous accumulation. (The inverse Gaussian process also has an independent, well-known construction as the first-passage time of a drifting Brownian motion through a fixed level, which is a separate fact about the law and not how it is derived here.) Formally, λ = 1 / ( 2 ℓ ) \lambda=1/(2\ell) λ = 1/ ( 2 ℓ ) and, from Eq. (50) with Γ ( 1 / 2 ) = π \Gamma(1/2)=\sqrt{\pi} Γ ( 1/2 ) = π , B = m λ / π B=m\sqrt{\lambda/\pi} B = m λ / π ; by the same term-by-term matching, Φ ( θ ) = 2 m λ [ ( 1 + θ / λ ) 1 / 2 − 1 ] \Phi(\theta)=2m\lambda[(1+\theta/\lambda)^{1/2}-1] Φ ( θ ) = 2 mλ [( 1 + θ / λ ) 1/2 − 1 ] is exactly the Laplace exponent of an inverse Gaussian process, for which Δ D ∼ IG ( μ Δ = m Δ , λ I G = m Δ 2 / t c ) \Delta D\sim\operatorname{IG}(\mu_\Delta=m\Delta,\lambda_{\mathrm{IG}}=m\Delta^2/t_c) Δ D ∼ IG ( μ Δ = m Δ , λ IG = m Δ 2 / t c ) . As a consistency check, its variance μ Δ 3 / λ I G = c 2 m 2 τ 0 Δ \mu_\Delta^3/\lambda_{\mathrm{IG}}=c^2m^2\tau_0\Delta μ Δ 3 / λ IG = c 2 m 2 τ 0 Δ agrees with the variance rate of the main text, as it must since both were fixed by the same matching condition q 2 / q 1 = ℓ q_2/q_1=\ell q 2 / q 1 = ℓ .
Figure S1 places the three laws on one continuous family, showing both how their skewness and kurtosis compare (panel a) and how the underlying mass spectrum w ( x ) w(x) w ( x ) reshapes itself as α \alpha α sweeps between them (panel b). Panel (a) shows γ 1 / c \gamma_1/c γ 1 / c approaching, but never crossing, the value 1 as the mass spectrum degenerates toward equal-sized events (u → 0 u\to0 u → 0 , i.e. α → − ∞ \alpha\to-\infty α → − ∞ , where every event carries the same size). This floor is not special to the tempered-stable family: it is a general property of every subordinator, as Sec. S3 A shows.
Figure S1. The tempered-stable family as a continuum, at common m m m , c c c and τ 0 \tau_0 τ 0 . (a) Standardized skewness and excess kurtosis, in units of c ( Δ ) c(\Delta) c ( Δ ) and c ( Δ ) 2 c(\Delta)^2 c ( Δ ) 2 , against u = 1 / ( 1 − α ) u=1/(1-\alpha) u = 1/ ( 1 − α ) (top axis: α \alpha α ): γ 1 / c = 1 + u \gamma_1/c=1+u γ 1 / c = 1 + u , γ 2 / c 2 = ( 1 + u ) ( 1 + 2 u ) \gamma_2/c^2=(1+u)(1+2u) γ 2 / c 2 = ( 1 + u ) ( 1 + 2 u ) . Compound Poisson, Gamma and inverse Gaussian (u = 1 2 , 1 , 2 u=\frac{1}{2},1,2 u = 2 1 , 1 , 2 ; markers) lie on these curves, and u → 0 u\to0 u → 0 approaches the bound γ 1 ≥ c \gamma_1\ge c γ 1 ≥ c . (b) Mass spectrum w ( x ) w(x) w ( x ) for five values of α \alpha α , all with mean ℓ \ell ℓ (dotted): α \alpha α redistributes the deposit between few large and many small events at fixed mean and variance rate.
Activity, path continuity and observability. The results above are easy to over-interpret, because they blur together three properties of the family that are logically independent and should be kept apart.
Activity is the event count: finite for α < 0 \alpha<0 α < 0 , where the process can be written as a genuine list of discrete jumps as in the compound Poisson case above, and infinite for α ≥ 0 \alpha\ge0 α ≥ 0 , where it cannot. Path continuity is a separate matter and does not track activity: every member of the family, finite or infinite activity alike, is a pure-jump process, so the Gamma and inverse-Gaussian paths—despite having infinite activity—are not continuous but strictly increasing through a dense set of arbitrarily small jumps. In particular, the Gamma process is a continuum limit only in this measure-theoretic sense of infinitely many infinitesimal jumps, and not in the sense of a deposit thickening smoothly and uniformly the way a continuous function will; no member of the family has continuous sample paths. Observability is a third, and independent, question again: whether any of the first two properties can actually be resolved from a measured record depends on the sampling interval and the measurement noise, and for an area-averaged thermal resistance it cannot (main text, Sec. 5), regardless of the true activity or path structure of the underlying process. The quiescent stretches that distinguish compound Poisson in Figure 2 (d) of the main text illustrate this: they are a statement about the model’s event count (activity), not a prediction that a measured record will show matching flat intervals, since the measurement itself may not resolve them.
Skewness bound
For any subordinator, the Cauchy–Schwarz inequality q 2 2 ≤ q 1 q 3 q_2^2\le q_1q_3 q 2 2 ≤ q 1 q 3 gives γ 1 ( Δ ) ≥ c ( Δ ) \gamma_1(\Delta)\ge c(\Delta) γ 1 ( Δ ) ≥ c ( Δ ) , with equality only if all events have the same size (approached as α → − ∞ \alpha\to-\infty α → − ∞ in the family above): no matter how the event-size spectrum is shaped, the skewness of the increments can never fall below the coefficient of variation itself.
Proof. With the window cumulants κ n = q n Δ \kappa_n=q_n\Delta κ n = q n Δ ,
c ( Δ ) = q 2 Δ q 1 Δ , γ 1 ( Δ ) = q 3 Δ ( q 2 Δ ) 3 / 2 , c(\Delta)=\frac{\sqrt{q_2\Delta}}{q_1\Delta},\qquad\gamma_1(\Delta)=\frac{q_3\Delta}{(q_2\Delta)^{3/2}}, c ( Δ ) = q 1 Δ q 2 Δ , γ 1 ( Δ ) = ( q 2 Δ ) 3/2 q 3 Δ ,
γ 1 ( Δ ) c ( Δ ) = q 1 q 3 q 2 2 , (59) \frac{\gamma_1(\Delta)}{c(\Delta)}=\frac{q_1q_3}{q_2^2}, \tag*{(59)} c ( Δ ) γ 1 ( Δ ) = q 2 2 q 1 q 3 , (59)
so the ratio is independent of the window and depends on the event-size spectrum alone. Writing x 2 = x 1 / 2 x 3 / 2 x^2=x^{1/2}x^{3/2} x 2 = x 1/2 x 3/2 and applying the Cauchy–Schwarz inequality with respect to ν \nu ν ,
q 2 = ∫ 0 ∞ x 1 / 2 x 3 / 2 ν ( d x ) ≤ ( ∫ 0 ∞ x ν ( d x ) ) 1 / 2 ( ∫ 0 ∞ x 3 ν ( d x ) ) 1 / 2 = q 1 q 3 , (60) q_2=\int_0^\infty x^{1/2}x^{3/2}\,\nu(dx)
\leq\left(\int_0^\infty x\,\nu(dx)\right)^{1/2}\left(\int_0^\infty x^3\,\nu(dx)\right)^{1/2}=\sqrt{q_1q_3}, \tag*{(60)} q 2 = ∫ 0 ∞ x 1/2 x 3/2 ν ( d x ) ≤ ( ∫ 0 ∞ x ν ( d x ) ) 1/2 ( ∫ 0 ∞ x 3 ν ( d x ) ) 1/2 = q 1 q 3 , (60)
which is γ 1 ≥ c \gamma_1\geq c γ 1 ≥ c . Equality requires x 1 / 2 ∝ x 3 / 2 x^{1/2}\propto x^{3/2} x 1/2 ∝ x 3/2 on the support of ν \nu ν , i.e. a single event size: compound Poisson with fixed jumps attains γ 1 = c \gamma_1=c γ 1 = c , and any spread of sizes raises γ 1 / c \gamma_1/c γ 1 / c above 1. In the tempered-stable family γ 1 / c = 1 + u \gamma_1/c=1+u γ 1 / c = 1 + u , i.e. 1.5, 2 and 3 for the compound Poisson, Gamma and inverse Gaussian laws, and the bound is approached only as u → 0 u\to0 u → 0 [Fig. S1(a)].
Physically, a deposit that grows only by adding nonnegative pieces must have increments at least as right-skewed as they are dispersed. A Gaussian increment model, with zero skewness and finite spread, is therefore never an exact additive law, only its large-window limit, in which both γ 1 \gamma_1 γ 1 and c c c vanish as Δ − 1 / 2 \Delta^{-1/2} Δ − 1/2 . Conversely, measured increments whose skewness falls clearly below their coefficient of variation cannot come from additive deposition alone: measurement noise, removal or a non-additive mechanism must contribute.
Removal: derivations
Cumulants of the Lévy-driven Ornstein–Uhlenbeck process. For a Lévy process D D D and a deterministic kernel f f f , the cumulants of a stochastic integral against D D D are
κ n [ ∫ 0 t f ( u ) d D ( u ) ] = q n ∫ 0 t f ( u ) n d u . (61) \kappa_n\left[\int_0^t f(u)\,dD(u)\right]=q_n\int_0^t f(u)^n\,du. \tag*{(61)} κ n [ ∫ 0 t f ( u ) d D ( u ) ] = q n ∫ 0 t f ( u ) n d u . (61)
With f ( u ) = e − ψ ( t − u ) f(u)=e^{-\psi(t-u)} f ( u ) = e − ψ ( t − u ) this gives
κ n [ R ( t ) ] = q n ∫ 0 t e − n ψ ( t − u ) d u = q n 1 − e − n ψ t n ψ , (62) \kappa_n[R(t)]=q_n\int_0^t e^{-n\psi(t-u)}\,du=q_n\frac{1-e^{-n\psi t}}{n\psi}, \tag*{(62)} κ n [ R ( t )] = q n ∫ 0 t e − n ψ ( t − u ) d u = q n n ψ 1 − e − n ψ t , (62)
Eq. (18) of the main text; the deterministic initial condition contributes only to the mean. At stationarity, κ n = q n / ( n ψ ) \kappa_n=q_n/(n\psi) κ n = q n / ( n ψ ) .
Stationary plateau. The mean obeys d μ / d t = m − ψ μ d\mu/dt=m-\psi\mu d μ / d t = m − ψ μ (Table 1 of the main text), so μ ∞ = R ∞ = m / ψ \mu_\infty=R_\infty=m/\psi μ ∞ = R ∞ = m / ψ . Taking n = 2 n=2 n = 2 above,
Var [ R ] ∞ = κ 2 = q 2 2 ψ = s 2 2 ψ , (63) \operatorname{Var}[R]_\infty=\kappa_2=\frac{q_2}{2\psi}=\frac{s^2}{2\psi}, \tag*{(63)} Var [ R ] ∞ = κ 2 = 2 ψ q 2 = 2 ψ s 2 , (63)
since q 2 = s 2 q_2=s^2 q 2 = s 2 . The plateau accordingly contains
N ∞ = R ∞ ℓ = m / ψ s 2 / m = m 2 ψ s 2 = 1 ψ t c (64) N_\infty=\frac{R_\infty}{\ell}=\frac{m/\psi}{s^2/m}=\frac{m^2}{\psi s^2}=\frac{1}{\psi t_c} \tag*{(64)} N ∞ = ℓ R ∞ = s 2 / m m / ψ = ψ s 2 m 2 = ψ t c 1 (64)
size-biased events (using t c = s 2 / m 2 = ℓ / m t_c=s^2/m^2=\ell/m t c = s 2 / m 2 = ℓ / m ), and
C V ∞ 2 = Var [ R ] ∞ R ∞ 2 = s 2 / 2 ψ ( m / ψ ) 2 = ψ s 2 2 m 2 = ψ t c 2 = 1 2 N ∞ , \mathrm{CV}_\infty^2=\frac{\operatorname{Var}[R]_\infty}{R_\infty^2}=\frac{s^2/2\psi}{(m/\psi)^2}=\frac{\psi s^2}{2m^2}=\frac{\psi t_c}{2}=\frac{1}{2N_\infty}, CV ∞ 2 = R ∞ 2 Var [ R ] ∞ = ( m / ψ ) 2 s 2 /2 ψ = 2 m 2 ψ s 2 = 2 ψ t c = 2 N ∞ 1 ,
the plateau relations of the main text. The shape of the plateau follows the same route:
γ 1 , ∞ = κ 3 κ 2 3 / 2 = q 3 / 3 ψ ( q 2 / 2 ψ ) 3 / 2 = 2 2 3 ψ 1 / 2 q 3 q 2 − 3 / 2 = 2 2 3 γ 1 ( ψ − 1 ) , (65) \begin{aligned}
\gamma_{1,\infty}=\frac{\kappa_3}{\kappa_2^{3/2}}=\frac{q_3/3\psi}{(q_2/2\psi)^{3/2}}=\frac{2\sqrt{2}}{3}\psi^{1/2}q_3q_2^{-3/2} \\
&=\frac{2\sqrt{2}}{3}\gamma_1(\psi^{-1}), \tag*{(65)}
\end{aligned} γ 1 , ∞ = κ 2 3/2 κ 3 = ( q 2 /2 ψ ) 3/2 q 3 /3 ψ = 3 2 2 ψ 1/2 q 3 q 2 − 3/2 = 3 2 2 γ 1 ( ψ − 1 ) , (65)
and γ 2 , ∞ = κ 4 / κ 2 2 = ψ q 4 q 2 − 2 = γ 2 ( ψ − 1 ) \gamma_{2,\infty}=\kappa_4/\kappa_2^2=\psi q_4q_2^{-2}=\gamma_2(\psi^{-1}) γ 2 , ∞ = κ 4 / κ 2 2 = ψ q 4 q 2 − 2 = γ 2 ( ψ − 1 ) . Using γ 1 ( Δ ) = [ ( 2 − α ) / ( 1 − α ) ] c ( Δ ) \gamma_1(\Delta)=[(2-\alpha)/(1-\alpha)]c(\Delta) γ 1 ( Δ ) = [( 2 − α ) / ( 1 − α )] c ( Δ ) , γ 2 ( Δ ) = [ ( 2 − α ) ( 3 − α ) / ( 1 − α ) 2 ] c ( Δ ) 2 \gamma_2(\Delta)=[(2-\alpha)(3-\alpha)/(1-\alpha)^2]c(\Delta)^2 γ 2 ( Δ ) = [( 2 − α ) ( 3 − α ) / ( 1 − α ) 2 ] c ( Δ ) 2 and c ( ψ − 1 ) = 2 C V ∞ c(\psi^{-1})=\sqrt{2}\,\mathrm{CV}_\infty c ( ψ − 1 ) = 2 CV ∞ for the tempered-stable family gives
γ 1 , ∞ = 4 3 2 − α 1 − α C V ∞ , \gamma_{1,\infty}=\frac{4}{3}\frac{2-\alpha}{1-\alpha}\mathrm{CV}_\infty, γ 1 , ∞ = 3 4 1 − α 2 − α CV ∞ ,
γ 2 , ∞ = 2 ( 2 − α ) ( 3 − α ) ( 1 − α ) 2 C V ∞ 2 , \gamma_{2,\infty}=2\frac{(2-\alpha)(3-\alpha)}{(1-\alpha)^2}\mathrm{CV}_\infty^2, γ 2 , ∞ = 2 ( 1 − α ) 2 ( 2 − α ) ( 3 − α ) CV ∞ 2 ,
the plateau skewness and kurtosis of the main text.
Per-element detachment. Write the deposition as D ( t ) = ∫ 0 t ∫ 0 ∞ x N D ( d u , d x ) D(t)=\int_0^t\int_0^\infty x\,N_D(du,dx) D ( t ) = ∫ 0 t ∫ 0 ∞ x N D ( d u , d x ) , with N D N_D N D a Poisson random measure of intensity d u ν ( d x ) du\,\nu(dx) d u ν ( d x ) , and mark each atom with an independent lifetime τ ∼ Exp ( ψ ) \tau\sim\operatorname{Exp}(\psi) τ ∼ Exp ( ψ ) . The marked measure N ( d u , d x , d τ ) N(du,dx,d\tau) N ( d u , d x , d τ ) is again Poisson, with intensity d u ν ( d x ) ψ e − ψ τ d τ du\,\nu(dx)\,\psi e^{-\psi\tau}d\tau d u ν ( d x ) ψ e − ψ τ d τ , and the deposit present at time t t t is R ( t ) = ∫ ∫ ∫ x 1 { t − u < τ } N ( d u , d x , d τ ) R(t)=\int\!\!\int\!\!\int x\,\mathbf{1}_{\{t-u<\tau\}}N(du,dx,d\tau) R ( t ) = ∫ ∫ ∫ x 1 { t − u < τ } N ( d u , d x , d τ ) : each event leaves as a single intact unit at the end of its lifetime. By Campbell’s theorem [59 ] the cumulants of a Poisson integral are the integrals of the n n n th power of the integrand against the intensity,
κ n [ R ( t ) ] = ∫ 0 t d u ∫ 0 ∞ x n ν ( d x ) Pr ( τ > t − u ) = q n ∫ 0 t e − ψ ( t − u ) d u = q n 1 − e − ψ t ψ , \begin{aligned}
\kappa_n[R(t)]=\int_0^t du\int_0^\infty x^n\nu(dx)\Pr(\tau>t-u) \\
&=q_n\int_0^t e^{-\psi(t-u)}du=q_n\frac{1-e^{-\psi t}}{\psi},
\end{aligned} κ n [ R ( t )] = ∫ 0 t d u ∫ 0 ∞ x n ν ( d x ) Pr ( τ > t − u ) = q n ∫ 0 t e − ψ ( t − u ) d u = q n ψ 1 − e − ψ t ,
which holds for finite and infinite activity alike, since only the moments q n = ∫ x n ν ( d x ) q_n=\int x^n\nu(dx) q n = ∫ x n ν ( d x ) , n ≥ 1 n\geq1 n ≥ 1 , enter. At stationarity κ n = q n / ψ \kappa_n=q_n/\psi κ n = q n / ψ , n n n times larger than the Ornstein–Uhlenbeck value q n / ( n ψ ) q_n/(n\psi) q n / ( n ψ ) . The difference lies in the integrand: under proportional erosion an event of size x x x contributes x e − ψ ( t − u ) xe^{-\psi(t-u)} x e − ψ ( t − u ) , whose n n n th power integrates to 1 / ( n ψ ) 1/(n\psi) 1/ ( n ψ ) , whereas under per-element detachment it contributes x x x or nothing, and 1 n = 1 1^n=1 1 n = 1 integrates to 1 / ψ 1/\psi 1/ ψ . An event made of many sub-units detaching independently contributes, by the law of large numbers, its surviving fraction x e − ψ ( t − u ) x e^{-\psi(t-u)} x e − ψ ( t − u ) , and so returns the Ornstein–Uhlenbeck cumulants. The plateau shape follows as before:
γ 1 , ∞ = q 3 / ψ ( q 2 / ψ ) 3 / 2 = ψ 1 / 2 q 3 q 2 − 3 / 2 = γ 1 ( ψ − 1 ) , \gamma_{1,\infty}=\frac{q_3/\psi}{(q_2/\psi)^{3/2}}=\psi^{1/2}q_3q_2^{-3/2}=\gamma_1(\psi^{-1}), γ 1 , ∞ = ( q 2 / ψ ) 3/2 q 3 / ψ = ψ 1/2 q 3 q 2 − 3/2 = γ 1 ( ψ − 1 ) ,
γ 2 , ∞ = ψ q 4 q 2 − 2 = γ 2 ( ψ − 1 ) ; (66) \gamma_{2,\infty}=\psi q_4q_2^{-2}=\gamma_2(\psi^{-1}); \tag*{(66)} γ 2 , ∞ = ψ q 4 q 2 − 2 = γ 2 ( ψ − 1 ) ; (66)
with C V ∞ 2 = ψ q 2 / m 2 = ψ t c = c ( ψ − 1 ) 2 \mathrm{CV}_\infty^2=\psi q_2/m^2=\psi t_c=c(\psi^{-1})^2 CV ∞ 2 = ψ q 2 / m 2 = ψ t c = c ( ψ − 1 ) 2 these are
γ 1 , ∞ = 2 − α 1 − α C V ∞ , \gamma_{1,\infty}=\frac{2-\alpha}{1-\alpha}\mathrm{CV}_\infty, γ 1 , ∞ = 1 − α 2 − α CV ∞ ,
γ 2 , ∞ = ( 2 − α ) ( 3 − α ) ( 1 − α ) 2 C V ∞ 2 . \gamma_{2,\infty}=\frac{(2-\alpha)(3-\alpha)}{(1-\alpha)^2}\mathrm{CV}_\infty^2. γ 2 , ∞ = ( 1 − α ) 2 ( 2 − α ) ( 3 − α ) CV ∞ 2 .
Autocorrelation. For continuous, per-element and intermittent removal alike the mean equation closes, and it holds conditionally on the state at an earlier time:
E [ R ( t + τ ) ∣ R ( t ) ] = R ∞ + ( R ( t ) − R ∞ ) e − ψ τ , τ > 0. (67) \mathbb{E}[R(t+\tau)\mid R(t)]=R_\infty+(R(t)-R_\infty)e^{-\psi\tau},\qquad\tau>0. \tag*{(67)} E [ R ( t + τ ) ∣ R ( t )] = R ∞ + ( R ( t ) − R ∞ ) e − ψ τ , τ > 0. (67)
Multiplying by R ( t ) R(t) R ( t ) and averaging gives
Cov [ R ( t ) , R ( t + τ ) ] = Var [ R ] e − ψ τ . (68) \operatorname{Cov}[R(t),R(t+\tau)]=\operatorname{Var}[R]e^{-\psi\tau}. \tag*{(68)} Cov [ R ( t ) , R ( t + τ )] = Var [ R ] e − ψ τ . (68)
Moments under sloughing. Applying the generator of the sloughing equation of the main text to f ( R ) = R f(R)=R f ( R ) = R and f ( R ) = R 2 f(R)=R^2 f ( R ) = R 2 ,
d μ 1 d t = m − r s η μ 1 , (69) \frac{d\mu_1}{dt}=m-r_s\eta\mu_1, \tag*{(69)} d t d μ 1 = m − r s η μ 1 , (69)
d μ 2 d t = 2 m μ 1 + s 2 − r s η ( 2 − η ) μ 2 , (70) \frac{d\mu_2}{dt}=2m\mu_1+s^2-r_s\eta(2-\eta)\mu_2, \tag*{(70)} d t d μ 2 = 2 m μ 1 + s 2 − r s η ( 2 − η ) μ 2 , (70)
where the deposition contributes 2 R m + s 2 2Rm+s^2 2 R m + s 2 and an event changes R 2 R^2 R 2 by − η ( 2 − η ) R 2 -\eta(2-\eta)R^2 − η ( 2 − η ) R 2 . At stationarity μ 1 = m / ( r s η ) = R ∞ \mu_1=m/(r_s\eta)=R_\infty μ 1 = m / ( r s η ) = R ∞ and μ 2 = ( 2 m μ 1 + s 2 ) / [ r s η ( 2 − η ) ] \mu_2=(2m\mu_1+s^2)/[r_s\eta(2-\eta)] μ 2 = ( 2 m μ 1 + s 2 ) / [ r s η ( 2 − η )] , so that
Var = μ 2 − μ 1 2 = 2 R ∞ 2 2 − η + s 2 ( 2 − η ) ψ − R ∞ 2 = s 2 ( 2 − η ) ψ + η 2 − η R ∞ 2 . (71) \begin{aligned}
\operatorname{Var}=\mu_2-\mu_1^2=\frac{2R_\infty^2}{2-\eta}+\frac{s^2}{(2-\eta)\psi}-R_\infty^2 \\
&=\frac{s^2}{(2-\eta)\psi}+\frac{\eta}{2-\eta}R_\infty^2. \tag*{(71)}
\end{aligned} Var = μ 2 − μ 1 2 = 2 − η 2 R ∞ 2 + ( 2 − η ) ψ s 2 − R ∞ 2 = ( 2 − η ) ψ s 2 + 2 − η η R ∞ 2 . (71)
with ψ = r s η \psi=r_s\eta ψ = r s η , which is the sloughing variance of the main text. The transient moments of the three removal mechanisms, and of site blocking, are compared in Figure S2 .
Nonlinear removal. Starting from d R = d D − g ( R ) d t dR=dD-g(R)\,dt d R = d D − g ( R ) d t , the mean and the variance obey
d μ d t = m − E [ g ( R ) ] , d V d t = s 2 − 2 Cov [ R , g ( R ) ] . (72) \frac{d\mu}{dt}=m-\mathbb{E}[g(R)],\qquad\frac{dV}{dt}=s^2-2\operatorname{Cov}[R,g(R)]. \tag*{(72)} d t d μ = m − E [ g ( R )] , d t d V = s 2 − 2 Cov [ R , g ( R )] . (72)
Expanding g ( R ) = g ( μ ) + g ′ ( μ ) δ R + 1 2 g ′ ′ ( μ ) δ R 2 + ⋯ g(R)=g(\mu)+g'(\mu)\delta R+\frac{1}{2}g''(\mu)\delta R^2+\cdots g ( R ) = g ( μ ) + g ′ ( μ ) δ R + 2 1 g ′′ ( μ ) δ R 2 + ⋯ around the mean gives, to the order kept by the main text’s second-order closure,
E [ g ] = g ( μ ) + 1 2 g ′ ′ ( μ ) V + O ( κ 3 ) , \mathbb{E}[g]=g(\mu)+\frac{1}{2}g''(\mu)V+O(\kappa_3), E [ g ] = g ( μ ) + 2 1 g ′′ ( μ ) V + O ( κ 3 ) ,
Cov [ R , g ( R ) ] = g ′ ( μ ) V + O ( κ 3 ) ; (73) \operatorname{Cov}[R,g(R)]=g'(\mu)V+O(\kappa_3); \tag*{(73)} Cov [ R , g ( R )] = g ′ ( μ ) V + O ( κ 3 ) ; (73)
the neglected terms contain the third cumulant and carry the first dependence on α \alpha α . Explicitly, the next order is
Cov [ R , g ( R ) ] = g ′ ( μ ) V + 1 2 g ′ ′ ( μ ) κ 3 [ R ] + ⋯ , \operatorname{Cov}[R,g(R)]=g'(\mu)V+\frac{1}{2}g''(\mu)\kappa_3[R]+\cdots, Cov [ R , g ( R )] = g ′ ( μ ) V + 2 1 g ′′ ( μ ) κ 3 [ R ] + ⋯ ,
E [ g ] = g ( μ ) + 1 2 g ′ ′ ( μ ) V + 1 6 g ′ ′ ′ ( μ ) κ 3 [ R ] + ⋯ . (74) \mathbb{E}[g]=g(\mu)+\frac{1}{2}g''(\mu)V+\frac{1}{6}g'''(\mu)\kappa_3[R]+\cdots. \tag*{(74)} E [ g ] = g ( μ ) + 2 1 g ′′ ( μ ) V + 6 1 g ′′′ ( μ ) κ 3 [ R ] + ⋯ . (74)
The closure is ordered in 1 / N ∞ = ℓ / R ∞ 1/N_\infty=\ell/R_\infty 1/ N ∞ = ℓ / R ∞ : near the plateau V = O ( ℓ R ∞ ) V=O(\ell R_\infty) V = O ( ℓ R ∞ ) and κ 3 [ R ] = O ( ℓ 2 R ∞ ) \kappa_3[R]=O(\ell^2R_\infty) κ 3 [ R ] = O ( ℓ 2 R ∞ ) , i.e. κ n [ R ] / μ n = O ( N ∞ 1 − n ) \kappa_n[R]/\mu^n=O(N_\infty^{1-n}) κ n [ R ] / μ n = O ( N ∞ 1 − n ) , and with g ( j ) ∼ g / μ j g^{(j)}\sim g/\mu^j g ( j ) ∼ g / μ j each κ 3 \kappa_3 κ 3 term is smaller than the retained V V V term of the same equation by a factor O ( 1 / N ∞ ) O(1/N_\infty) O ( 1/ N ∞ ) . Dropping g ′ ′ κ 3 g''\kappa_3 g ′′ κ 3 in the variance equation while keeping g ′ ′ V g''V g ′′ V in the mean equation is therefore consistent at leading order in 1 / N ∞ 1/N_\infty 1/ N ∞ , although both terms carry g ′ ′ g'' g ′′ . Figure S3 compares the closure with simulation.
Figure S3. Nonlinear removal, g ( R ) = ψ R + ε R 2 g(R)=\psi R+\varepsilon R^{2} g ( R ) = ψ R + ε R 2 (m = τ 0 = 1 m=\tau_{0}=1 m = τ 0 = 1 , ψ = 0.2 \psi=0.2 ψ = 0.2 , ε = 0.05 \varepsilon=0.05 ε = 0.05 , Gamma deposition). (a) Shift of the stationary mean from the mean-field plateau m = g ( μ ) m=g(\mu) m = g ( μ ) , against the relative event size ℓ / μ \ell/\mu ℓ / μ , varied through c c c : second-order closure, Eq. (26) of the main text (line), vs. Monte Carlo (circles). The closure captures the shift while the plateau holds many events, and overestimates it once a single event is comparable to the deposit (ℓ / μ ≳ 0.3 \ell/\mu\gtrsim0.3 ℓ / μ ≳ 0.3 ), having dropped − g ′ ′ κ 3 [ R ] -g^{\prime\prime}\kappa_{3}[R] − g ′′ κ 3 [ R ] . Under linear removal (ε = 0 \varepsilon=0 ε = 0 , squares) the plateau does not move: only there is the mean universal. (b) Mean trajectory at c = 1 c=1 c = 1 : fluctuations lower the plateau from 2.90 (mean field, dashed) to 2.81 (Monte Carlo), against 2.79 from the closure (solid).
Numerical checks. Monte Carlo simulations (m = c = τ 0 = 1 m=c=\tau_0=1 m = c = τ 0 = 1 , ψ = 0.2 \psi=0.2 ψ = 0.2 , 2 × 10 4 2\times10^4 2 × 1 0 4 paths, run for T = 60 = 12 ψ − 1 T=60=12\psi^{-1} T = 60 = 12 ψ − 1 to reach the stationary state) reproduce the stationary skewness of the plateau of the main text for the three laws ( 0.63 , 0.82 , 1.27 (0.63,\,0.82,\,1.27 ( 0.63 , 0.82 , 1.27 against 0.63 , 0.84 , 1.27 ) 0.63,\,0.84,\,1.27) 0.63 , 0.84 , 1.27 ) , the stationary variances of continuous removal, per-element detachment and sloughing ( 2.52 , 5.02 (2.52,\,5.02 ( 2.52 , 5.02 and 11.67 11.67 11.67 for η = 0.5 \eta=0.5 η = 0.5 , Gamma deposition, Figure S2 ; per-element detachment is simulated exactly by thinning the alive deposit, and the residuals are within sampling error), the exponential autocorrelation ( 0.366 (0.366 ( 0.366 against e − 1 = 0.368 e^{-1}=0.368 e − 1 = 0.368 at τ = 5 ) \tau=5) τ = 5 ) , and the nonlinear plateau of Figure S3 .
Granular shot noise and the coefficient of variation
This section derives the granular estimate of the coefficient of variation quoted in Sec. 5.2 in the main text, and the area scaling that follows from it. The monodisperse limit of the family. In the tempered-stable family, Eq. (13) of the main text, the mass-weighted spectrum has relative width ( 1 − α ) − 1 / 2 (1-\alpha)^{-1/2} ( 1 − α ) − 1/2 , so as α → − ∞ \alpha\to-\infty α → − ∞ it collapses onto a single size: deposition becomes a compound Poisson process in which every event deposits the same amount J 1 J_1 J 1 . Writing r r r for the event rate, the cumulant rates are q n = r J 1 n q_n=rJ_1^n q n = r J 1 n , so that
m = q 1 = r J 1 , s 2 = q 2 = r J 1 2 , ℓ = q 2 q 1 = J 1 . (75) m=q_1=rJ_1,\qquad s^2=q_2=rJ_1^2,\qquad\ell=\frac{q_2}{q_1}=J_1. \tag*{(75)} m = q 1 = r J 1 , s 2 = q 2 = r J 1 2 , ℓ = q 1 q 2 = J 1 . (75)
the size-biased event size coinciding with the jump itself, as it must when all jumps are equal. Substituting ℓ = J 1 \ell=J_1 ℓ = J 1 into the squared coefficient of variation over a window Δ \Delta Δ , c ( Δ ) 2 = ℓ / ( m Δ ) c(\Delta)^2=\ell/(m\Delta) c ( Δ ) 2 = ℓ / ( m Δ ) , and evaluating it at the reference window Δ = τ 0 \Delta=\tau_0 Δ = τ 0 ,
c 2 = J 1 m τ 0 = 1 r τ 0 , (76) c^2=\frac{J_1}{m\tau_0}=\frac{1}{r\tau_0}, \tag*{(76)} c 2 = m τ 0 J 1 = r τ 0 1 , (76)
one over the mean number of attachments per reference window. This is the classical shot-noise result for a superposition of Poisson-timed pulses of equal weight [80 , 81 ] ; here it is not imported but obtained as the α → − ∞ \alpha\to-\infty α → − ∞ member of the family of the main text.
resistance by
J 1 = δ 1 k f = π d p 3 / 6 A ϕ s k f . (77) J_{1}=\frac{\delta_{1}}{k_{f}}=\frac{\pi d_{p}^{3}/6}{A\phi_{s}k_{f}} . \tag*{(77)} J 1 = k f δ 1 = A ϕ s k f π d p 3 /6 . (77)
The size of a single attachment. It remains to express J 1 J_{1} J 1 for one attached particle. A sphere of diameter d p d_{p} d p has volume π d p 3 / 6 \pi d_{p}^{3}/6 π d p 3 /6 . Deposited into a porous layer of solid fraction ϕ s \phi_{s} ϕ s it occupies a layer volume ( π d p 3 / 6 ) / ϕ s (\pi d_{p}^{3}/6)/\phi_{s} ( π d p 3 /6 ) / ϕ s ; spread over the heat-transfer area A A A this raises the deposit thickness by δ 1 = ( π d p 3 / 6 ) / ( A ϕ s ) \delta_{1}=(\pi d_{p}^{3}/6)/(A\phi_{s}) δ 1 = ( π d p 3 /6 ) / ( A ϕ s ) , and with R = δ / k f R=\delta/k_{f} R = δ / k f the fouling
Equations (S33) and (S34) together are the estimate used in the main text. Nothing in it is adjustable: once d p d_{p} d p and the deposit properties are fixed, the observed mean rate fixes the event frequency through r = m / J 1 r=m/J_{1} r = m / J 1 .
Area scaling. Because J 1 ∝ A − 1 J_{1}\propto A^{-1} J 1 ∝ A − 1 at fixed particle size, Eq. (76) gives c ∝ A − 1 / 2 c\propto A^{-1/2} c ∝ A − 1/2 at fixed mean rate: granular arrival noise in an area-averaged observable is suppressed by the square root of the averaging area. Setting c = 1 c = 1 c = 1 identifies the area on which one sampling period deposits a single event,
A 1 = π d p 3 / 6 ϕ s k f m τ 0 . (78) A_1 = \frac{\pi d_p^3/6}{\phi_s k_f m \tau_0}. \tag*{(78)} A 1 = ϕ s k f m τ 0 π d p 3 /6 . (78)
For d p = 1 μ m d_p = 1\ \mu\mathrm{m} d p = 1 μ m , k f = 0.5 W m − 1 K − 1 k_f = 0.5\ \mathrm{W\,m^{-1}K^{-1}} k f = 0.5 W m − 1 K − 1 , ϕ s = 0.5 \phi_s = 0.5 ϕ s = 0.5 and a mean rate of order 10 − 5 m 2 K W − 1 10^{-5}\ \mathrm{m^2\,K\,W^{-1}} 1 0 − 5 m 2 K W − 1 per sampling period, A 1 ≈ 2 × 10 − 13 m 2 A_1 \approx2 \times10^{-13}\ \mathrm{m^2} A 1 ≈ 2 × 1 0 − 13 m 2 , about a quarter of the cross-section of the particle itself. That is the sense in which an area-averaged fouling resistance is deterministic with respect to the granularity of deposition: order-unity fluctuations would require averaging over a surface no larger than one grain.
Record length for law identification
This section derives the noise attenuation of the standardized cumulants and the record-length estimate quoted in the main text. Write
ρ = 2 σ 2 c 2 m 2 τ 0 2 (79) \rho= \frac{2\sigma^2}{c^2m^2\tau_0^2} \tag*{(79)} ρ = c 2 m 2 τ 0 2 2 σ 2 (79)
for the noise-to-signal ratio. The observed increment over k k k periods is the sum of the deposition increment, with cumulants κ n [ Δ D ] = q n k τ 0 \kappa_n[\Delta D] = q_n k\tau_0 κ n [ Δ D ] = q n k τ 0 , and two independent Gaussian measurement errors of variance σ 2 \sigma^2 σ 2 each, which contribute only to κ 2 \kappa_2 κ 2 . Writing κ 2 [ Δ y ] = κ 2 [ Δ D ] ( 1 + ρ / k ) \kappa_2[\Delta y] = \kappa_2[\Delta D](1+\rho/k) κ 2 [ Δ y ] = κ 2 [ Δ D ] ( 1 + ρ / k ) and substituting into the standardized cumulants gives
γ 1 , k = γ 1 k ( k + ρ ) 3 / 2 , γ 2 , k = γ 2 k ( k + ρ ) 2 . (80) \gamma_{1,k} = \gamma_1 \frac{k}{(k+\rho)^{3/2}}, \qquad\gamma_{2,k} = \gamma_2 \frac{k}{(k+\rho)^2}. \tag*{(80)} γ 1 , k = γ 1 ( k + ρ ) 3/2 k , γ 2 , k = γ 2 ( k + ρ ) 2 k . (80)
The windows that best retain the cumulants maximize Eq. (80) over k k k . For the skewness,
d d k k ( k + ρ ) 3 / 2 = ( k + ρ ) − 3 2 k ( k + ρ ) 5 / 2 = 0 ⇒ k ∗ = 2 ρ , (81) \frac{d}{dk}\frac{k}{(k+\rho)^{3/2}} = \frac{(k+\rho)-\frac{3}{2}k}{(k+\rho)^{5/2}} = 0 \quad\Rightarrow\quad k^* = 2\rho, \tag*{(81)} d k d ( k + ρ ) 3/2 k = ( k + ρ ) 5/2 ( k + ρ ) − 2 3 k = 0 ⇒ k ∗ = 2 ρ , (81)
max k γ 1 , k = γ 1 2 ρ ( 3 ρ ) 3 / 2 = 2 3 3 γ 1 ρ . (82) \max_k \gamma_{1,k} = \gamma_1\frac{2\rho}{(3\rho)^{3/2}} = \frac{2}{3\sqrt{3}}\frac{\gamma_1}{\sqrt{\rho}}. \tag*{(82)} k max γ 1 , k = γ 1 ( 3 ρ ) 3/2 2 ρ = 3 3 2 ρ γ 1 . (82)
For the excess kurtosis, ( k + ρ ) − 2 k = 0 (k+\rho)-2k=0 ( k + ρ ) − 2 k = 0 gives k ∗ = ρ k^*=\rho k ∗ = ρ and max k γ 2 , k = γ 2 / ( 4 ρ ) \max_k\gamma_{2,k}=\gamma_2/(4\rho) max k γ 2 , k = γ 2 / ( 4 ρ ) . Noise-free, the sample estimators numerically converge on the law-specific values (Figure S4 ). Detection requires a second ingredient—the sampling error of the estimator—and therefore selects a different window. Let two laws differ, before attenuation, by Δ γ 1 \Delta\gamma_1 Δ γ 1 or Δ γ 2 \Delta\gamma_2 Δ γ 2 . Each window is formed from a disjoint pair of observations k k k periods apart and so consumes 2 k 2k 2 k periods, giving n w = N / ( 2 k ) n_w=N/(2k) n w = N / ( 2 k ) independent windows; the standard errors of the sample skewness and sample excess kurtosis under approximate normality are 6 / n w \sqrt{6/n_w} 6/ n w and 24 / n w \sqrt{24/n_w} 24/ n w .
Figure S4. Recovering the microscopic law from skewness and kurtosis: finite-record convergence of the sample estimators, at m m m , c = 2 c=2 c = 2 , τ 0 \tau_0 τ 0 , noise-free. For N N N independent periods (300 Monte Carlo records per N N N ), markers/lines are the median sample skewness (a) and excess kurtosis (b), shaded bands the 5–95% range, dashed lines the exact values of Table 2 of the main text. Both converge onto the correct, law-specific value within a feasible record length (N ∼ 10 2 N\sim10^2 N ∼ 1 0 2 –10 3 10^3 1 0 3 ), given clean data; for an area-averaged thermal measurement, however, the same estimators are attenuated by Eq. (80) , so the required record grows to the lengths estimated in Sec. 5.4 of the main text.
We measure the separation of two laws in units of the standard error of a single estimate: the record is tested against the value expected under the alternative law, so d ′ = Δ γ 1 , k / 6 / n w d'=\Delta\gamma_{1,k}/\sqrt{6/n_w} d ′ = Δ γ 1 , k / 6/ n w and d ′ = Δ γ 2 , k / 24 / n w d'=\Delta\gamma_{2,k}/\sqrt{24/n_w} d ′ = Δ γ 2 , k / 24/ n w . Comparing two independent records, the standard error of their difference is larger by 2 \sqrt{2} 2 and N N N doubles. With the attenuated differences Δ γ 1 , k = Δ γ 1 k / ( k + ρ ) 3 / 2 \Delta\gamma_{1,k}=\Delta\gamma_1 k/(k+\rho)^{3/2} Δ γ 1 , k = Δ γ 1 k / ( k + ρ ) 3/2 and Δ γ 2 , k = Δ γ 2 k / ( k + ρ ) 2 \Delta\gamma_{2,k}=\Delta\gamma_2 k/(k+\rho)^2 Δ γ 2 , k = Δ γ 2 k / ( k + ρ ) 2 of Eq. (80) , the separability indices are
d 1 ′ = Δ γ 1 N 12 k 1 / 2 ( k + ρ ) 3 / 2 , d'_1 = \Delta\gamma_1\sqrt{\frac{N}{12}}\frac{k^{1/2}}{(k+\rho)^{3/2}}, d 1 ′ = Δ γ 1 12 N ( k + ρ ) 3/2 k 1/2 ,
d 2 ′ = Δ γ 2 N 48 k 1 / 2 ( k + ρ ) 2 . (83) d'_2 = \Delta\gamma_2\sqrt{\frac{N}{48}}\frac{k^{1/2}}{(k+\rho)^2}. \tag*{(83)} d 2 ′ = Δ γ 2 48 N ( k + ρ ) 2 k 1/2 . (83)
Setting the derivative of the k k k -dependent factor to zero gives 1 2 ( k + ρ ) − 3 2 k = 0 \frac{1}{2}(k+\rho)-\frac{3}{2}k=0 2 1 ( k + ρ ) − 2 3 k = 0 , i.e. k = ρ / 2 k=\rho/2 k = ρ /2 , for the skewness, and 1 2 ( k + ρ ) − 2 k = 0 \frac{1}{2}(k+\rho)-2k=0 2 1 ( k + ρ ) − 2 k = 0 , i.e. k = ρ / 3 k=\rho/3 k = ρ /3 , for the excess kurtosis—both shorter than k ∗ k^* k ∗ , because a longer window is also a rarer one. At these windows the factors equal ( 2 / 3 3 ) ρ − 1 (2/3\sqrt{3})\rho^{-1} ( 2/3 3 ) ρ − 1 and ( 3 3 / 16 ) ρ − 3 / 2 (3\sqrt{3}/16)\rho^{-3/2} ( 3 3 /16 ) ρ − 3/2 , so that d ′ = 2 d'=2 d ′ = 2 requires N = 12 ⋅ 4 ⋅ ( 27 / 4 ) ρ 2 / Δ γ 1 2 N=12\cdot4\cdot(27/4)\rho^2/\Delta\gamma_1^2 N = 12 ⋅ 4 ⋅ ( 27/4 ) ρ 2 /Δ γ 1 2 and N = 48 ⋅ 4 ⋅ ( 256 / 27 ) ρ 3 / Δ γ 2 2 N=48\cdot4\cdot(256/27)\rho^3/\Delta\gamma_2^2 N = 48 ⋅ 4 ⋅ ( 256/27 ) ρ 3 /Δ γ 2 2 , that is,
N ≳ 324 ρ 2 Δ γ 1 2 , N ≳ 1820 ρ 3 Δ γ 2 2 . (84) N\gtrsim324\frac{\rho^2}{\Delta\gamma_1^2},\qquad N\gtrsim1820\frac{\rho^3}{\Delta\gamma_2^2}. \tag*{(84)} N ≳ 324 Δ γ 1 2 ρ 2 , N ≳ 1820 Δ γ 2 2 ρ 3 . (84)
For the Gamma/inverse-Gaussian pair, Δ γ 1 = c \Delta\gamma_1=c Δ γ 1 = c and Δ γ 2 = 9 c 2 \Delta\gamma_2=9c^2 Δ γ 2 = 9 c 2 (Sec. S3), giving N ≈ 324 ρ 2 / c 2 N\approx324\rho^2/c^2 N ≈ 324 ρ 2 / c 2 and N ≈ 22 ρ 3 / c 4 N\approx22\rho^3/c^4 N ≈ 22 ρ 3 / c 4 ; the skewness is therefore the cheaper statistic by a factor ρ / ( 14 c 2 ) \rho/(14c^2) ρ / ( 14 c 2 ) , and it sets the scaling of the requirement with the noise. Separating windows by one period instead of k k k gives n w ≃ N / k n_w\simeq N/k n w ≃ N / k and halves both prefactors without changing the powers of ρ \rho ρ . The estimates are asymptotic in ρ \rho ρ : Figure S5 tests them by simulation and finds them accurate to within a factor 1–2 for ρ ≳ 10 2 \rho\gtrsim10^2 ρ ≳ 1 0 2 and optimistic below. They are also specific to moment-based estimation, whereas the attenuation, Eq. (80) , is a property of the measurement and applies to any estimator.
First-passage time: exact distribution and asymptotics
Exact distribution. Because R ( t ) = R 0 + D ( t ) R(t)=R_0+D(t) R ( t ) = R 0 + D ( t ) is nondecreasing, { T x ≤ t } = { D ( t ) ≥ x − R 0 } \{T_x\le t\}=\{D(t)\ge x-R_0\} { T x ≤ t } = { D ( t ) ≥ x − R 0 } , and Pr ( T x ≤ t ) \Pr(T_x\le t) Pr ( T x ≤ t ) follows directly from the law of D ( t ) D(t) D ( t ) given in the main text. For the Gamma law, D ( t ) ∼ Gamma ( t / ( c 2 τ 0 ) , c 2 m τ 0 ) D(t)\sim\operatorname{Gamma}(t/(c^2\tau_0),c^2m\tau_0) D ( t ) ∼ Gamma ( t / ( c 2 τ 0 ) , c 2 m τ 0 ) and Pr ( T x ≤ t ) = Q ( t / ( c 2 τ 0 ) , ( x − R 0 ) / ( c 2 m τ 0 ) ) \Pr(T_x\le t)=Q(t/(c^2\tau_0),(x-R_0)/(c^2m\tau_0)) Pr ( T x ≤ t ) = Q ( t / ( c 2 τ 0 ) , ( x − R 0 ) / ( c 2 m τ 0 )) , the regularized upper incomplete gamma function. For the compound Poisson law, conditioning on the number of events N ( t ) = n N(t)=n N ( t ) = n gives a sum of n n n exponential jumps, i.e. an Erlang( n , ℓ / 2 ) (n,\ell/2) ( n , ℓ /2 ) variable, so Pr ( D ( t ) ≥ x − R 0 ∣ N ( t ) = n ) = Q ( n , 2 ( x − R 0 ) / ℓ ) \Pr(D(t)\ge x-R_0\mid N(t)=n)=Q(n,2(x-R_0)/\ell) Pr ( D ( t ) ≥ x − R 0 ∣ N ( t ) = n ) = Q ( n , 2 ( x − R 0 ) / ℓ ) ; weighting by the Poisson probabilities of N ( t ) N(t) N ( t ) and summing over n ≥ 1 n\ge1 n ≥ 1 gives Pr ( T x ≤ t ) \Pr(T_x\le t) Pr ( T x ≤ t ) in closed form. For the inverse Gaussian law, D ( t ) ∼ IG ( μ t , λ t ) D(t)\sim\operatorname{IG}(\mu_t,\lambda_t) D ( t ) ∼ IG ( μ t , λ t ) with μ t = m t \mu_t=mt μ t = m t , λ t = m t 2 / ( c 2 τ 0 ) \lambda_t=mt^2/(c^2\tau_0) λ t = m t 2 / ( c 2 τ 0 ) , and
the standard inverse-Gaussian distribution function gives Pr ( T x ≤ t ) = 1 − IG ( x − R 0 ; μ t , λ t ) \Pr(T_x\le t)=1-\operatorname{IG}(x-R_0;\mu_t,\lambda_t) Pr ( T x ≤ t ) = 1 − IG ( x − R 0 ; μ t , λ t ) . All three reduce, for x − R 0 x-R_0 x − R 0 large compared with the event scale ℓ \ell ℓ , to the normal approximation Φ G ( [ m t − ( x − R 0 ) ] / ( c m τ 0 t ) ) \Phi_{\mathrm{G}}([mt-(x-R_0)]/(cm\sqrt{\tau_0t})) Φ G ([ m t − ( x − R 0 )] / ( c m τ 0 t )) .
Asymptotic cumulants. Write the deposition as a sum of one-period increments X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … with cumulant generating function K ( θ ) = log E e θ X 1 K(\theta)=\log\mathbb{E}e^{\theta X_1} K ( θ ) = log E e θ X 1 , mean κ 1 = m τ 0 \kappa_1=m\tau_0 κ 1 = m τ 0 , variance κ 2 = c 2 m 2 τ 0 2 \kappa_2=c^2m^2\tau_0^2 κ 2 = c 2 m 2 τ 0 2 and third cumulant κ 3 = γ 1 c 3 m 3 τ 0 3 \kappa_3=\gamma_1c^3m^3\tau_0^3 κ 3 = γ 1 c 3 m 3 τ 0 3 , and let T = T x T=T_x T = T x be the number of periods at which the partial sum S T = ∑ i < T X i S_T=\sum_{i<T}X_i S T = ∑ i < T X i first reaches x − R 0 x-R_0 x − R 0 , with overshoot O = S T − ( x − R 0 ) > 0 O=S_T-(x-R_0)>0 O = S T − ( x − R 0 ) > 0 . For θ \theta θ in a neighbourhood of 0 0 0 where K K K is finite, M n = exp [ θ S n − n K ( θ ) ] M_n=\exp[\theta S_n-nK(\theta)] M n = exp [ θ S n − n K ( θ )] is a martingale: since the increments are independent, E [ M n + 1 ∣ X 1 , … , X n ] = M n E [ e θ X n + 1 ] e − K ( θ ) = M n \mathbb{E}[M_{n+1}\mid X_1,\ldots,X_n]=M_n\mathbb{E}[e^{\theta X_{n+1}}]e^{-K(\theta)}=M_n E [ M n + 1 ∣ X 1 , … , X n ] = M n E [ e θ X n + 1 ] e − K ( θ ) = M n , and E M n = 1 \mathbb{E}M_n=1 E M n = 1 . The passage time T T T is finite with exponential moments for a subordinator of positive mean, so optional stopping applies and E M T = 1 \mathbb{E}M_T=1 E M T = 1 ; with S T = ( x − R 0 ) + O S_T=(x-R_0)+O S T = ( x − R 0 ) + O this reads
E [ e θ ( x − R 0 ) + θ O − T K ( θ ) ] = 1. (85) \mathbb{E}\left[e^{\theta(x-R_0)+\theta O-TK(\theta)}\right]=1. \tag*{(85)} E [ e θ ( x − R 0 ) + θ O − T K ( θ ) ] = 1. (85)
Choosing θ = θ ( u ) \theta=\theta(u) θ = θ ( u ) to solve K ( θ ) = − u K(\theta)=-u K ( θ ) = − u turns − T K ( θ ) -TK(\theta) − T K ( θ ) into u T uT u T , and moving the constant factor e θ ( u ) ( x − R 0 ) e^{\theta(u)(x-R_0)} e θ ( u ) ( x − R 0 ) to the right-hand side gives Wald’s identity
E [ e u T + θ ( u ) O ] = e − θ ( u ) ( x − R 0 ) . (86) \mathbb{E}\left[e^{uT+\theta(u)O}\right]=e^{-\theta(u)(x-R_0)}. \tag*{(86)} E [ e u T + θ ( u ) O ] = e − θ ( u ) ( x − R 0 ) . (86)
For a distant threshold the overshoot becomes asymptotically independent of T T T and converges to the stationary overshoot of renewal theory, with mean E [ O ∞ ] = E [ X 1 2 ] / ( 2 κ 1 ) = ( κ 2 + κ 1 2 ) / ( 2 κ 1 ) \mathbb{E}[O_\infty]=\mathbb{E}[X_1^2]/(2\kappa_1)=(\kappa_2+\kappa_1^2)/(2\kappa_1) E [ O ∞ ] = E [ X 1 2 ] / ( 2 κ 1 ) = ( κ 2 + κ 1 2 ) / ( 2 κ 1 ) . The expectation in Eq. (86) then factorizes, and its logarithm is log E e u T = − θ ( u ) ( x − R 0 ) − log E e θ ( u ) O ∞ \log\mathbb{E}e^{uT}=-\theta(u)(x-R_0)-\log\mathbb{E}e^{\theta(u)O_\infty} log E e u T = − θ ( u ) ( x − R 0 ) − log E e θ ( u ) O ∞ . Since θ ( u ) = − u / κ 1 + O ( u 2 ) \theta(u)=-u/\kappa_1+O(u^2) θ ( u ) = − u / κ 1 + O ( u 2 ) , the last term is − θ ( u ) E [ O ∞ ] + O ( u 2 ) = u E [ O ∞ ] / κ 1 + O ( u 2 ) -\theta(u)\mathbb{E}[O_\infty]+O(u^2)=u\mathbb{E}[O_\infty]/\kappa_1+O(u^2) − θ ( u ) E [ O ∞ ] + O ( u 2 ) = u E [ O ∞ ] / κ 1 + O ( u 2 ) , so that
log E e u T = − θ ( u ) ( x − R 0 ) + u E [ O ∞ ] κ 1 + O ( u 2 ) . (87) \log\mathbb{E}e^{uT}=-\theta(u)(x-R_0)+\frac{u\mathbb{E}[O_\infty]}{\kappa_1}+O(u^2). \tag*{(87)} log E e u T = − θ ( u ) ( x − R 0 ) + κ 1 u E [ O ∞ ] + O ( u 2 ) . (87)
Inverting K ( θ ) = − u K(\theta)=-u K ( θ ) = − u order by order in u u u (matching K ( θ ) = κ 1 θ + 1 2 κ 2 θ 2 + 1 6 κ 3 θ 3 + ⋯ = − u K(\theta)=\kappa_1\theta+\frac{1}{2}\kappa_2\theta^2+\frac{1}{6}\kappa_3\theta^3+\cdots=-u K ( θ ) = κ 1 θ + 2 1 κ 2 θ 2 + 6 1 κ 3 θ 3 + ⋯ = − u ) and reading off the coefficients of u n / n ! u^n/n! u n / n ! in − θ ( u ) ( x − R 0 ) -\theta(u)(x-R_0) − θ ( u ) ( x − R 0 ) gives the cumulants of T T T to leading order in x − R 0 x-R_0 x − R 0 :
κ 1 ( T ) = x − R 0 κ 1 + O ( 1 ) , \kappa_1(T)=\frac{x-R_0}{\kappa_1}+O(1), κ 1 ( T ) = κ 1 x − R 0 + O ( 1 ) ,
κ 2 ( T ) = ( x − R 0 ) κ 2 κ 1 3 + O ( 1 ) . (88) \kappa_2(T)=\frac{(x-R_0)\kappa_2}{\kappa_1^3}+O(1). \tag*{(88)} κ 2 ( T ) = κ 1 3 ( x − R 0 ) κ 2 + O ( 1 ) . (88)
κ 3 ( T ) = ( x − R 0 ) ( 3 κ 2 2 κ 1 5 − κ 3 κ 1 4 ) + O ( 1 ) . \kappa_3(T)=(x-R_0)\left(\frac{3\kappa_2^2}{\kappa_1^5}-\frac{\kappa_3}{\kappa_1^4}\right)+O(1). κ 3 ( T ) = ( x − R 0 ) ( κ 1 5 3 κ 2 2 − κ 1 4 κ 3 ) + O ( 1 ) .
Substituting κ 1 = m τ 0 \kappa_1=m\tau_0 κ 1 = m τ 0 , κ 2 = c 2 m 2 τ 0 2 \kappa_2=c^2m^2\tau_0^2 κ 2 = c 2 m 2 τ 0 2 , κ 3 = γ 1 c 3 m 3 τ 0 3 \kappa_3=\gamma_1c^3m^3\tau_0^3 κ 3 = γ 1 c 3 m 3 τ 0 3 and n x = ( x − R 0 ) / ( m τ 0 ) n_x=(x-R_0)/(m\tau_0) n x = ( x − R 0 ) / ( m τ 0 ) reproduces the expressions quoted in the main text for the mean and the coefficient of variation of T x T_x T x (in units of τ 0 \tau_0 τ 0 ), and gives for the skewness
skew [ T x ] = 3 c − γ 1 n x , (89) \operatorname{skew}[T_x] = \frac{3c-\gamma_1}{\sqrt{n_x}}, \tag*{(89)} skew [ T x ] = n x 3 c − γ 1 , (89)
which is positive whenever γ 1 < 3 c \gamma_1<3c γ 1 < 3 c (true for the compound Poisson and Gamma laws at matched ( m , c ) (m,c) ( m , c ) , Sec. S3) and vanishes for the inverse Gaussian law, whose skewness is exactly 3 c 3c 3 c . The O ( 1 ) O(1) O ( 1 ) remainders in Eq. (88) come from the overshoot and do not grow with the distance to the threshold, which is why the leading-order formulas improve as n x n_x n x increases and are least reliable for n x n_x n x of order 10 or smaller, the regime flagged in the main text.
References
Range of validity. Monte Carlo over n x ∈ { 5 , 10 , 25 , 50 , 220 } n_x \in\{5, 10, 25, 50, 220\} n x ∈ { 5 , 10 , 25 , 50 , 220 } (Figure S6 ) recovers the mean to within 0.09 sampling periods for n x ≥ 10 n_x \ge10 n x ≥ 10 , but at n x = 5 n_x=5 n x = 5 the inverse Gaussian law is biased by − 0.23 -0.23 − 0.23 periods, its spread is 18 % below c n x c\sqrt{n_x} c n x , and its predicted zero skewness is still 0.45. At n x = 220 n_x=220 n x = 220 the three laws agree on the mean life to 0.2 periods and on the 5th–95th percentiles to 0.9 %. For a threshold within a handful of increments the life distribution should be simulated rather than taken from Eqs. (88) –(89) .