If you've ever worked with Weibull analysis — in reliability engineering, survival analysis, or manufacturing — you've encountered the characteristic life, often denoted \(\eta\). It's defined as the time by which 63.2% of units have failed (or converted, or experienced the event of interest). But where does that oddly specific number come from?
The answer is one of the most elegant results in applied statistics, and it connects the Weibull distribution to a constant that appears across all of mathematics and the natural sciences: Euler's number \(e\).
The Quick Answer
The cumulative distribution function (CDF) of the Weibull distribution is:
\[ F(t) = 1 - e^{-(t/\eta)^\beta} \]where \(\eta\) is the scale parameter (characteristic life) and \(\beta\) is the shape parameter. When we evaluate the CDF at \(t = \eta\), the shape parameter cancels entirely:
\[ F(\eta) = 1 - e^{-(\eta/\eta)^\beta} = 1 - e^{-1} \approx 0.6321 \]That's it. No matter what value \(\beta\) takes — whether you're modeling infant mortality (\(\beta < 1\)), constant hazard (\(\beta = 1\)), or wear-out failures (\(\beta > 1\)) — exactly 63.21% of the population will have failed by time \(\eta\).
Why \(e^{-1}\)? The Deeper Story
The number \(e \approx 2.71828\) is the base of the natural logarithm. Its reciprocal \(e^{-1} \approx 0.3679\) is the probability of surviving past the characteristic life, and \(1 - e^{-1} \approx 0.6321\) is the probability of having failed by that point.
This isn't a coincidence specific to the Weibull — it's a direct consequence of the natural exponential function sitting at the heart of the distribution. The Weibull CDF is built on the survival function \(S(t) = e^{-(t/\eta)^\beta}\), which generalizes the exponential decay that governs everything from radioactive half-lives to capacitor discharge.
The Exponential Distribution: The Special Case
When \(\beta = 1\), the Weibull reduces to the exponential distribution:
\[ F(t) = 1 - e^{-t/\eta} \]In this case the characteristic life \(\eta\) equals the mean lifetime \(1/\lambda\). The exponential is the only continuous memoryless distribution — "memoryless" meaning the probability of surviving an additional unit of time is the same regardless of how long you've already survived. Even in this simplest case, the 63.2% property holds: by the mean lifetime, 63.2% of units have failed.
This is often surprising to people encountering it for the first time. We might intuitively expect 50% to have failed by the "average" lifetime — but the exponential distribution is right-skewed, so the median (\(\eta \ln 2 \approx 0.693\eta\)) is less than the mean. Among the common parametric lifetime models, the characteristic life is deliberately chosen to be above the median to capture the bulk of the early-failure population.
63.2% Across Mathematics and Science
The constant \(1 - e^{-1}\) appears far beyond reliability engineering. Here are several prominent examples:
RC Circuits (Electrical Engineering)
When a capacitor charges through a resistor, the voltage reaches \(1 - e^{-1} \approx 63.2\%\) of its final value after one time constant \(\tau = RC\). This is the same mathematical structure: the charging curve \(V(t) = V_0 (1 - e^{-t/\tau})\) is identical in form to the Weibull CDF with \(\beta = 1\). Engineers routinely characterize circuit response by how many time constants it takes to reach steady state (roughly \(5\tau\) for 99.3%).
Radioactive Decay (Physics)
The fraction of atoms that have decayed by time \(t\) is \(1 - e^{-\lambda t}\). After one mean lifetime \(\tau = 1/\lambda\), exactly 63.2% of the atoms have decayed. Note that the half-life \(t_{1/2} = \tau \ln 2\) is the more publicly recognized metric, but in theoretical physics the mean lifetime (and its 63.2% threshold) is the more fundamental quantity.
The Poisson Distribution (Probability Theory)
For a Poisson process with rate \(\lambda\), the probability of observing at least one event in an interval of expected count \(\lambda t = 1\) is:
\[ P(X \geq 1) = 1 - P(X = 0) = 1 - e^{-1} \approx 0.6321 \]That is, if you expect exactly one event on average, there is a 63.2% chance that at least one event actually occurs. This is the discrete counting analogue of the Weibull characteristic life.
The Secretary Problem (Optimal Stopping Theory)
In the classical secretary problem, you interview \(n\) candidates sequentially and must hire one on the spot. The optimal strategy is to reject the first \(n/e \approx 36.8\%\) of candidates (observing but not hiring), then hire the next one who is better than all previously seen. The probability of selecting the best candidate converges to \(1/e \approx 36.8\%\) — the complement of our 63.2%.
This result from optimal stopping theory shows that \(e^{-1}\) isn't merely a feature of continuous distributions; it governs fundamental bounds on decision-making under uncertainty.
The Coupon Collector's Problem (Combinatorics)
To collect all \(n\) distinct coupon types, you need approximately \(n \ln n\) draws. After \(n\) draws, the expected fraction of distinct coupons seen is \(1 - (1 - 1/n)^n\), which converges to \(1 - e^{-1} \approx 63.2\%\) as \(n \to \infty\). The constant again emerges naturally from the exponential limit \((1 - 1/n)^n \to e^{-1}\).
Why Not Use the Median or the Mean?
A fair question: why define a special statistic at 63.2% instead of the more familiar 50th percentile (median) or the expected value (mean)?
- The median depends on \(\beta\). For the Weibull, the median is \(\eta (\ln 2)^{1/\beta}\), which varies with the shape parameter. This makes it harder to compare across different failure modes.
- The mean also depends on \(\beta\) through the gamma function: \(\mu = \eta \, \Gamma(1 + 1/\beta)\).
- The characteristic life \(\eta\) is the only quantile of the Weibull that is invariant to \(\beta\). This makes it the natural "shape-free" reference point for comparing Weibull fits across different products, processes, or failure mechanisms.
In practice this means you can compare the characteristic lives of two systems directly, even if their shape parameters differ dramatically (e.g., one with infant mortality \(\beta = 0.5\) and another with wear-out \(\beta = 3.5\)). The 63.2% quantile is always the scale parameter itself.
A Proof in Three Lines
For completeness, here is the derivation in full:
\[ \begin{aligned} F(\eta) &= 1 - \exp\!\Big[-\Big(\frac{\eta}{\eta}\Big)^\beta\Big] \\[6pt] &= 1 - \exp\!\big[-(1)^\beta\big] \\[6pt] &= 1 - e^{-1} \approx 0.6321 \end{aligned} \]The shape parameter \(\beta\) vanishes because \(1^\beta = 1\) for all real \(\beta\). This is the algebraic reason behind the invariance.
Summary
The Weibull characteristic life being 63.2% is not an arbitrary convention — it is a direct consequence of Euler's number \(e\) appearing in the exponential family of distributions. The same constant governs:
- Capacitor charging (63.2% of final voltage after one time constant)
- Radioactive decay (63.2% of atoms decayed after one mean lifetime)
- Poisson counting (63.2% chance of ≥ 1 event when the expected count is 1)
- Optimal stopping (36.8% = \(1/e\) rejection threshold in the secretary problem)
- Coupon collecting (63.2% of coupons seen after \(n\) of \(n\) draws)
Whenever a process is governed by exponential growth or decay, the constant \(1 - e^{-1}\) inevitably appears as the natural threshold after one characteristic unit of time, rate, or scale — making 63.2% one of the most universal constants in applied mathematics.