故事 · 不会变老的灯泡
Origin Story · The Bulb That Refuses to Age
指数分布回答一个最朴素的问题:"下一件随机事件还要等多久?"它的密度公式只有一个参数 λ:
f(t) = λe−λt。1909 年,哥本哈根电话公司的工程师 A.K. Erlang 发现来电的到达间隔正服从这个分布,由此奠基了排队论。
它最迷人的性质叫无记忆性:已经等待了 s 时间之后,再多等 t 时间的概率,和从零开始等 t 时间完全相同 ——
P(T > s+t | T > s) = P(T > t)。换句话说,一只指数寿命的元件"不积累损耗",它只是在每一刻都有同样的概率随机失效。
The exponential distribution answers one of the most natural questions in probability: "how long until the next random event?" Its density has a single parameter, λ:
f(t) = λe−λt. In 1909, the Copenhagen telephone engineer A.K. Erlang discovered that inter-arrival times between calls obeyed exactly this law — the founding observation of queueing theory.
Its most enchanting property is memorylessness: after waiting time s, the probability of waiting another t is identical to waiting t from the very beginning —
P(T > s+t | T > s) = P(T > t). Put plainly: an exponentially distributed component does not accumulate damage. At every instant it carries the same constant chance of failing at random.
1 无记忆:已经等过的时间,白等了 Memoryless: The Time You've Already Waited Buys You Nothing
曲线纹丝不动curve unchanged2 随机失效:间隔忽长忽短,毫无规律 Random Failures: Inter-arrival Gaps Stretch and Shrink Without Warning
每条竖线是一次随机失效。间隔服从指数分布:短间隔最常见,偶尔出现很长的"空窗"。这正是泊松过程(上一页)间隔的另一面。 Each tick is one random failure. The gaps follow the exponential law: short gaps dominate, with the occasional long quiet stretch. This is just the inter-arrival side of the Poisson process from the previous page.
3 现实里的指数分布 The Exponential in the Real World
可靠性 MTBF:无磨损期的随机失效,失效率 λ 恒定,平均无故障时间正是 1/λ。
Reliability & MTBF: pure random failure with no wear-out — the hazard rate λ is constant and the mean time between failures is exactly 1/λ.
随机失效期:浴盆曲线的平底段(电子元件),失效与已用时长无关,纯属偶然。
Random-failure region: the flat bottom of the bathtub curve (typical of electronics) — failures are independent of accumulated run-time, just plain chance.
排队等待:呼叫到达间隔、服务台空闲时长,Erlang 排队论与座席配置的输入。
Queueing: time between inbound calls, idle stretches at a service desk — the core input to Erlang's queueing theory and staffing models.
泊松过程间隔:放射性衰变、网络请求到达,泊松计数的相邻事件间隔即指数分布。
Poisson-process gaps: radioactive decay, network request arrivals — the gap between any two consecutive Poisson events is exponentially distributed.
一句话In One Line
泊松数"一段时间里发生几次",指数问"下一次还要等多久" —— 同一枚硬币的两面。
它唯一的参数 λ 同时定均值(1/λ)和失效率,是排队论、可靠性(MTBF)、维修策略的基石。
但记住它的致命前提:失效率恒定、不会老化。会磨损的机械零件不归它管,那是下一页的威布尔。无记忆性既是它的魔法,也是它的局限。
Poisson counts how many events arrive in a window; the exponential asks how long until the next one — two sides of the same coin.
Its single parameter λ fixes both the mean life (1/λ) and the hazard rate, making it the bedrock of queueing theory, reliability (MTBF), and maintenance policy.
But remember its critical precondition: the hazard rate must be constant — no aging allowed. Components that wear out (bearings, cutting tools) belong to the Weibull on the next page. Memorylessness is the exponential's magic and its hard ceiling.
常见误用Common Mistakes
给会磨损的零件套指数(轴承、刀具)。有老化用威布尔;指数只适合恒定失效率。
Applying the exponential to parts that wear out (bearings, cutting tools). Use Weibull when there's aging — the exponential only fits a constant hazard rate.
把 λ 当成平均寿命。λ 是失效率,平均寿命是它的倒数 1/λ。
Reading λ as the mean life. λ is the hazard rate; the mean life is its reciprocal, 1/λ.
误以为"等久了就快到了"。无记忆:已等多久都不改变剩余等待分布。
Believing "I've waited a long time, so I must be due". Memoryless: no matter how long you've waited, the distribution of the remaining wait is unchanged.