Brownian Passage Time

Brownian Passage Time — renewal recurrence and the characteristic earthquake

A single-topic deep page. The Brownian Passage Time (BPT) model is a renewal model of earthquake recurrence: it treats a fault as a system that is steadily loaded toward failure and asks, given how long it has been since the last large event on that fault, how likely the next one is in the coming interval. Unlike the memoryless Poisson process — for which the answer is always the same regardless of elapsed time — BPT has a clock: the hazard climbs as elapsed time approaches the mean recurrence interval. It is the long-term, fault-specific layer that conditions the background over years to decades, and it travels together with the characteristic earthquake hypothesis about how large recurring events are distributed.

Honest framing. This page describes a forecast, not a prediction. BPT yields a conditional probability of recurrence over a stated window (e.g. "the 30-year probability of a characteristic event on this fault segment is 21 %"), never a date. Its skill over a plain Poisson baseline is often marginal when the recurrence data are thin — a limitation this page states plainly rather than hiding. It conditions long-term hazard; it does not issue alarms.

---

Table of contents

1. Intuition and history
2. The Brownian relaxation oscillator
3. The BPT (inverse-Gaussian) recurrence density
4. Hazard rate — the defining property
5. The conditional recurrence probability (the published number)
6. Parameter estimation: mean interval and aperiodicity
7. The characteristic-earthquake hypothesis
8. BPT versus Poisson — when time-dependence actually helps
9. Assumptions, strengths and limitations
10. A worked illustration
11. Role in operational earthquake forecasting
12. References

---

1. Intuition and history

The elastic-rebound picture of Reid (1910) is the physical seed: tectonic motion loads strain energy into a fault at a roughly steady rate; the fault holds until it reaches a failure threshold, ruptures in a large earthquake that releases the accumulated strain, and then begins re-loading from a lower stress state. If loading were perfectly steady and the threshold perfectly constant, large events on a single fault would be strictly periodic — a metronome. They are not: observed recurrence intervals on well-studied faults scatter substantially around their mean. So the right model is quasi-periodic: a clock with noise.

This motivates a renewal model — a model in which the time since the last event matters (the process "renews" at each event), in contrast to a Poisson model, which is memoryless (the time since the last event is irrelevant; the hazard is constant). Several renewal distributions have been used for recurrence times — the lognormal, the Weibull, the gamma, and the Brownian Passage Time. BPT, introduced for earthquakes by Matthews, Ellsworth & Reasenberg (2002), is distinguished because it is derived from an explicit, physically interpretable stochastic loading process rather than chosen for analytical convenience, and because its hazard function has exactly the shape elastic rebound predicts: near-zero immediately after an event, rising to a peak, then plateauing (not decaying) at long times.

BPT became the recurrence engine of major time-dependent national and regional hazard forecasts — the U.S. Working Group on California Earthquake Probabilities (WGCEP) forecasts, UCERF3-TD, and the Italian and Japanese national seismic-hazard models all use BPT (or close relatives) to push the long-term background up or down depending on where each modelled fault sits in its cycle.

flowchart LR
  A["Steady tectonic<br/>loading rate ρ"] --> B["State variable Y(t)<br/>drifts upward"]
  N["Brownian noise<br/>(perturbations)"] --> B
  B --> C{"Y reaches<br/>failure barrier?"}
  C -->|"yes"| D["Earthquake<br/>(rupture)"]
  D --> E["Reset Y → 0<br/>(elastic rebound)"]
  E --> B
  C -->|"recurrence interval T"| F["T ~ BPT(μ, α)<br/>inverse-Gaussian"]

Loading

---

2. The Brownian relaxation oscillator

BPT models the fault's internal state as a one-dimensional Brownian motion with positive drift. A "load state" variable $Y(t)$ starts at zero just after an earthquake and evolves as

$$dY(t) = \rho, dt + \sigma_B, dW(t),$$

where $\rho > 0$ is the steady drift (the deterministic tectonic loading rate), $\sigma_B$ is the amplitude of the Brownian perturbations (the noise that randomizes the timing), and $W(t)$ is a standard Wiener process. An earthquake occurs the first time $Y(t)$ reaches a fixed failure barrier $Y_f$; at that instant the state resets to zero (elastic rebound) and the cycle repeats. The recurrence interval $T$ is therefore the first-passage time of a drifting Brownian motion to a barrier — and that first-passage time has a known distribution: the inverse Gaussian (Wald) distribution. This is the entire content of the model: quasi-periodic loading toward a threshold, randomized by additive noise. The two physical knobs $(\rho, \sigma_B, Y_f)$ collapse, as shown next, into just two effective parameters — a mean and a measure of irregularity.

The drift $\rho$ sets the pace of the clock; the noise-to-drift ratio sets how good a clock it is. A strong drift with weak noise gives an almost perfect metronome; comparable noise and drift give something close to random (Poisson-like) timing. That competition is captured by a single dimensionless number, the aperiodicity $\alpha$, below.

---

3. The BPT (inverse-Gaussian) recurrence density

The probability density of the recurrence interval $T = t$ is the inverse-Gaussian / Wald density, written in the earthquake literature as

$$f(t;, \mu,, \alpha) = \sqrt{\frac{\mu}{2\pi,\alpha^2, t^3}}; \exp!\left(-\frac{(t - \mu)^2}{2,\mu,\alpha^2, t}\right), \qquad t > 0,$$

with just two parameters:

• $\mu$ — the mean recurrence interval (the expected time between successive characteristic events on the fault). It is set by the loading: $\mu = Y_f / \rho$.
• $\alpha$ — the aperiodicity, the coefficient of variation of the recurrence times ($\alpha = \mathrm{sd}(T)/\mathrm{mean}(T)$). It measures how irregular the clock is and is set by the noise-to-loading competition.

The mapping from the underlying oscillator to $(\mu, \alpha)$ is

$$\mu = \frac{Y_f}{\rho}, \qquad \alpha^2 = \frac{\sigma_B^2}{\rho, Y_f},$$

so the three physical quantities $(\rho, \sigma_B, Y_f)$ enter only through these two combinations — the model is genuinely two-parameter.

The limiting behaviour of $\alpha$ is the key to reading the model:

• $\alpha \to 0$ — quasi-periodic (the noise is negligible relative to the drift; the fault is an almost perfect metronome, recurrence tightly clustered around $\mu$).
• $\alpha \approx 0.5$ — moderately irregular; a common empirical value for well-recorded fault segments.
• $\alpha \approx 1$ — near-Poisson; the recurrence is so irregular that elapsed time carries little information, and BPT degenerates toward the memoryless exponential.

The mean and variance of $T$ are $\mathbb{E}[T] = \mu$ and $\mathrm{Var}(T) = \alpha^2 \mu^2$, which is why $\alpha$ is exactly the coefficient of variation.

---

4. Hazard rate — the defining property

The quantity that makes BPT physically attractive is its hazard rate (also called the failure rate or conditional intensity of recurrence),

$$h(t) = \frac{f(t)}{1 - F(t)}, \qquad F(t) = \int_0^t f(u), du,$$

the instantaneous probability density of rupture given that no rupture has yet occurred by elapsed time $t$. For the BPT distribution the hazard has a characteristic three-phase shape:

1. $h(t) \to 0$ as $t \to 0^+$. Immediately after an event the fault has just relaxed; it is physically implausible for it to rupture again instantly, and the model encodes this — the hazard starts at essentially zero. (This is the elastic-rebound "stress shadow" expressed as a hazard.)
2. $h(t)$ rises to a peak near the mean recurrence interval $\mu$, as the accumulating load carries the state variable toward the barrier.
3. $h(t)$ plateaus to a finite, constant asymptote at long elapsed times, rather than continuing to rise without bound.

This is a sharp, testable contrast with the two obvious alternatives:

Model	Hazard $h(t)$ behaviour
Poisson (exponential)	Constant for all $t$ — memoryless; elapsed time is irrelevant.
Lognormal	Rises, peaks, then decreases toward zero at long $t$ (implies a long-overdue fault becomes safer, which is physically odd).
BPT (inverse Gaussian)	Rises from ~0, peaks, then plateaus at a finite asymptote (a long-overdue fault stays dangerous).

The plateau is the physically sensible feature: a fault that is overdue relative to its mean does not become safe (lognormal) and is not unchanged (Poisson) — it remains at an elevated, bounded hazard. This is why BPT is preferred for time-dependent long-term hazard whenever the recurrence data are good enough to justify time-dependence at all.

flowchart LR
  subgraph H["Hazard h(t) vs elapsed time"]
    P["Poisson:<br/>flat (constant)"]
    L["Lognormal:<br/>rise → peak → decay → 0"]
    B["BPT:<br/>≈0 → rise → peak → PLATEAU"]
  end

Loading

---

5. The conditional recurrence probability (the published number)

The operationally useful quantity is not the unconditional density but the conditional probability: given that $T_e$ years have elapsed since the last event with no rupture yet, the probability that the next event occurs within the coming window $[T_e, T_e + \Delta T]$. By the definition of conditional probability for a renewal process,

$$P\big(\text{event in } [T_e,, T_e + \Delta T] ;\big|; \text{none by } T_e\big) = \frac{\displaystyle\int_{T_e}^{T_e + \Delta T} f(t), dt}{\displaystyle\int_{T_e}^{\infty} f(t), dt} = \frac{F(T_e + \Delta T) - F(T_e)}{1 - F(T_e)}.$$

The denominator $1 - F(T_e)$ is the survival probability to the present moment (the fault has not ruptured in the elapsed time $T_e$); the numerator is the probability mass in the coming window. The ratio re-normalizes "what could still happen" to "what is left to happen," which is the correct conditioning on the present state of the cycle.

Two limits make the behaviour concrete:

• If the fault is young in its cycle ($T_e \ll \mu$), the BPT hazard is still near zero, so the conditional probability is lower than the Poisson value — a recently-ruptured fault is, for a while, safer than average.
• If the fault is mature or overdue ($T_e \gtrsim \mu$), the hazard has risen to its plateau, so the conditional probability is higher than the Poisson value — the time-dependence "earns its keep" exactly here, by raising the long-term background where it should be raised.

This conditional probability is the number that flows into a time-dependent hazard forecast: it modulates the long-term background rate of the relevant fault segment up or down according to where the segment sits in its cycle, and that modulated background then enters the rest of the system (e.g. as a slowly-varying prior on the smoothed-seismicity background, or directly in a national hazard model).

---

6. Parameter estimation: mean interval and aperiodicity

Estimating $(\mu, \alpha)$ is where BPT meets the hard reality of data scarcity, because large characteristic earthquakes on a single fault recur on timescales of centuries to millennia, and the instrumental record is barely a century long.

• Mean recurrence interval $\mu$. Estimated from paleoseismology (trenching across a fault to date prehistoric ruptures from offset and disturbed strata), supplemented by historical catalogues, and/or from the slip-rate / slip-per-event ratio ($\mu \approx \text{(slip per event)} / \text{(long-term slip rate)}$). Each route carries large dating and correlation uncertainties.
• Aperiodicity $\alpha$. This is the genuinely difficult parameter. With only a handful of dated recurrence intervals, the coefficient of variation is estimated from a tiny sample and is itself highly uncertain. Because individual faults rarely yield enough events, $\alpha$ is often borrowed from a global or regional pooled estimate across many faults (commonly $\alpha \approx 0.5$), treating it as a generic property rather than fitting it per fault. This pooling is a documented assumption with real consequences for the forecast.

When several recurrence intervals are available, $(\mu, \alpha)$ are fit by maximum likelihood under the BPT density, ideally in a Bayesian framework that carries the (large) parameter uncertainty all the way through to the conditional probability — reporting a distribution of 30-year probabilities rather than a single deceptively precise number. Open-paleoseismic-record uncertainty, dating uncertainty, and the open interval since the last event (which is censored data — we know only that $T > T_e$) all belong in the likelihood, not swept under a point estimate.

Honesty point. With few observed cycles, $\alpha$ is poorly constrained, and the conditional probability is sensitive to $\alpha$. A model that reports a crisp "21 % in 30 years" while $\alpha$ is really known only to within a factor of two is overstating its own precision. Carry the uncertainty; report a band.

---

7. The characteristic-earthquake hypothesis

BPT is usually paired with the characteristic-earthquake hypothesis of Schwartz & Coppersmith (1984): a given fault (or fault segment) tends to rupture repeatedly in earthquakes of a similar, characteristic size — set by the segment's dimensions — rather than producing a full Gutenberg–Richter spectrum of sizes. The two ideas fit together naturally: renewal recurrence (BPT) describes the timing of the characteristic event, and the characteristic-earthquake hypothesis describes its size.

The empirical signature is a frequency–magnitude distribution that follows Gutenberg–Richter for small and moderate events but shows a bump (an excess) at the large, characteristic magnitude — more large events than the extrapolated GR line predicts, because the segment "saves up" for and repeatedly produces its signature rupture.

This is contested. Critics argue the apparent characteristic bump is partly an artifact of short catalogs (the instrumental window is too brief to sample the true large-event rate), spatial windowing (drawing the box tightly around one fault inflates its large events relative to the regional small-event rate), and magnitude-bin choices. A pure, self-similar Gutenberg–Richter distribution remains a viable competing description on many faults. The product therefore treats "characteristic vs. pure GR" as an open, region-specific modelling choice with its uncertainty reported, not a settled fact — and never lets a contested characteristic bump silently inflate a public tail probability.

---

8. BPT versus Poisson — when time-dependence actually helps

The Poisson process is the null: constant hazard, memoryless, no clock. BPT is the time-dependent challenger. The honest question is not "is BPT more sophisticated?" (it obviously is) but "does its extra structure buy real, demonstrable skill on this fault with this data?"

flowchart TB
  Q{"Is α well constrained<br/>by ≥ several dated cycles?"}
  Q -->|"no (1–2 intervals)"| POIS["Use Poisson background.<br/>BPT gain is marginal &<br/>α is guessed → false precision"]
  Q -->|"yes"| W{"Where is the fault<br/>in its cycle?"}
  W -->|"young: T_e ≪ μ"| LOW["BPT < Poisson<br/>(recently relaxed → safer)"]
  W -->|"mature/overdue: T_e ≳ μ"| HIGH["BPT > Poisson<br/>(time-dependence earns its keep)"]

Loading

The practical guidance the system follows:

• Use BPT only where the recurrence data genuinely justify it — a paleoseismic or historical record long enough to constrain both $\mu$ and (at least regionally) $\alpha$. On faults with one or two dated intervals, the gain over Poisson is marginal, and a confident-looking BPT number is mostly an artifact of a guessed $\alpha$.
• The gain is largest for mature/overdue, low-$\alpha$ (regular) faults, where the BPT hazard has climbed well above the Poisson constant and the model has something real to say.
• Report BPT next to its Poisson baseline, with the parameter-uncertainty band, so a reader sees both the time-dependent number and how much (or little) it differs from the memoryless null.

This mirrors the system-wide rule for the short-term models: a more elaborate model must demonstrate its advantage over the relevant null, not assume it. For BPT the null is Poisson; for the short-term layer it is smoothed seismicity and ETAS.

---

9. Assumptions, strengths and limitations

Core assumptions.

• Renewal / single-fault independence. Each fault segment is its own renewal process; the model conditions only on that segment's elapsed time, ignoring stress interaction with neighbouring faults (which static Coulomb stress transfer does address).
• Steady mean loading. The long-term loading rate $\rho$ (hence $\mu$) is constant over many cycles — questionable where loading is itself episodic (e.g. modulated by slow slip or post-seismic relaxation).
• Fixed failure barrier and full reset. Every characteristic event releases the accumulated load back to the same baseline — an idealization of the real, partial, variable stress drops.
• A characteristic size (when paired with the characteristic-earthquake hypothesis) — itself contested (§7).

Strengths.

• Physically motivated hazard shape. The rise-then-plateau hazard is exactly what elastic rebound implies and is more defensible than the lognormal's eventual decay.
• Only two parameters, both interpretable ($\mu$ = pace, $\alpha$ = regularity).
• Operationally proven as the recurrence engine of UCERF3-TD and the Italian/Japanese national hazard models.
• Cleanly nests the null: as $\alpha \to 1$ it degenerates toward Poisson, so the time-dependent and time-independent cases live on one continuum.

Limitations (stated honestly).

• Data starvation. Large recurring earthquakes are rare; $\mu$ and especially $\alpha$ are estimated from a handful of paleoseismic dates with large uncertainties. The forecast inherits that uncertainty and must report it.
• Marginal gain over Poisson when $\alpha$ is poorly constrained — the most important honesty point, repeated deliberately.
• Single-fault blindness. It ignores stress transfer between faults; a neighbour's rupture can load this segment in a way BPT's elapsed-time clock cannot see.
• Long-term only. It conditions the background over years to decades; it is not a short-horizon (days-to-weeks) forecaster and is irrelevant to the one-day-ahead probability except as a slow prior.

---

10. A worked illustration

Consider a fault segment whose paleoseismic record gives a mean recurrence interval $\mu = 200$ years and a (regionally pooled) aperiodicity $\alpha = 0.5$. The last characteristic rupture was $T_e = 150$ years ago. We want the 30-year conditional probability of the next characteristic event ($\Delta T = 30$ years).

1. Survival to the present. Compute $1 - F(150)$ under the BPT density with $(\mu, \alpha) = (200, 0.5)$ — the probability the segment has not yet ruptured in 150 years. The fault is at $T_e/\mu = 0.75$ of its mean cycle: past the low-hazard early phase, on the rising limb of the hazard curve.
2. Window mass. Compute $F(180) - F(150)$ — the probability mass between 150 and 180 years.
3. Conditional probability. $$P = \frac{F(180) - F(150)}{1 - F(150)}.$$ For these parameters this evaluates to roughly 0.18–0.22 (about a 1-in-5 chance in 30 years).
4. Compare to Poisson. A memoryless Poisson model with the same mean rate $1/\mu = 1/200$ per year gives $$P_{\text{Poisson}} = 1 - e^{-30/200} = 1 - e^{-0.15} \approx 0.14.$$

So at $T_e = 150$ years the BPT conditional probability (~0.20) is higher than Poisson (~0.14), because the segment is on the rising limb of its hazard. Had the segment ruptured only $T_e = 20$ years ago ($T_e/\mu = 0.10$), BPT would instead read below Poisson — the recently-relaxed fault is temporarily safer. The whole value of BPT is in this elapsed-time-dependent deviation from the flat Poisson line.

A vital caveat for this very example: the answer is sensitive to $\alpha$. Recomputing with $\alpha = 0.3$ (more regular) pushes the 30-year probability up; with $\alpha = 0.7$ (more irregular) it pulls it back toward — and eventually below — the Poisson value. Because $\alpha$ is usually the least-constrained parameter, the honest deliverable is a band of 30-year probabilities spanning the plausible $\alpha$ range, not a single figure.

---

11. Role in operational earthquake forecasting

BPT lives in the long-term, fault-specific layer of the system, far from the daily clock. Its job is to condition the background — to tell the slowly-varying part of the forecast that a given modelled fault is early, mature, or overdue in its cycle, nudging that segment's baseline rate down or up over years to decades. It is the renewal engine behind time-dependent national hazard models (UCERF3-TD, the Italian and Japanese models), and the product treats it the same way: a regional, optional conditioning layer applied only where paleoseismic recurrence is genuinely known.

What BPT is not, and what the product is careful never to let it pretend to be:

• It is not a short-term forecaster. It says nothing about whether tomorrow's probability is elevated; that is the work of ETAS and the aftershock models. BPT's output enters only as a slow prior on the background.
• It is not an alarm or a countdown. "Overdue" in a BPT sense means an elevated long-term conditional probability — it does not name a year, and a fault can remain "overdue" for a long time. Communicating this without inviting alarm is part of the product's honest-limits discipline.
• It is not to be used where the data do not support it. A confident BPT number on a fault with one dated interval is false precision; the system defaults to the Poisson background there and says so.

In short, BPT is the system's patient, decade-scale clock: it makes the long-term background honestly time-dependent where the recurrence record earns it, and it steps aside — back to the Poisson null — where it does not.

---

References

1. Matthews, M.V., Ellsworth, W.L. & Reasenberg, P.A. (2002). A Brownian model for recurrent earthquakes. Bulletin of the Seismological Society of America 92(6), 2233–2250. doi:10.1785/0120010267
2. Schwartz, D.P. & Coppersmith, K.J. (1984). Fault behavior and characteristic earthquakes: Examples from the Wasatch and San Andreas fault zones. Journal of Geophysical Research 89(B7), 5681–5698. doi:10.1029/JB089iB07p05681
3. Reid, H.F. (1910). The Mechanics of the Earthquake (The California Earthquake of April 18, 1906, Report of the State Earthquake Investigation Commission, Vol. 2). Carnegie Institution of Washington.
4. Ellsworth, W.L., Matthews, M.V., Nadeau, R.M., Nishenko, S.P., Reasenberg, P.A. & Simpson, R.W. (1999). A physically-based earthquake recurrence model for estimation of long-term earthquake probabilities. U.S. Geological Survey Open-File Report 99-522.
5. Field, E.H., Biasi, G.P., Bird, P., Dawson, T.E., Felzer, K.R., Jackson, D.D., et al. (2015). Long-term, time-dependent probabilities for the third Uniform California Earthquake Rupture Forecast (UCERF3). Bulletin of the Seismological Society of America 105(2A), 511–543. doi:10.1785/0120140093
6. Working Group on California Earthquake Probabilities (2003). Earthquake probabilities in the San Francisco Bay Region: 2002–2031. U.S. Geological Survey Open-File Report 03-214.
7. Jordan, T.H., Chen, Y.-T., Gasparini, P., Madariaga, R., Main, I., Marzocchi, W., Papadopoulos, G., Sobolev, G., Yamaoka, K. & Zschau, J. (2011). Operational Earthquake Forecasting: state of knowledge and guidelines for utilization. Annals of Geophysics 54(4), 315–391. doi:10.4401/ag-5350

---

See also: Models — Classical · Smoothed Seismicity · Rate-and-State & Coulomb · Evaluation & Tests · Honest Limits · Methodology & History. All equations are transcribed from the primary/authoritative sources cited above.