Recovering Exponential Rates from Quantiles

Why Recover a Rate from a Quantile?

An exponential model is often specified by its rate, but a waiting-time description may instead give a median or another percentile. For example, a hypothetical service system might be modeled so that 90% of requests are resolved within a stated number of hours. If the exponential assumption is appropriate, that information determines the model’s rate.

The calculation is an inversion: begin with the exponential cumulative distribution function (CDF), set it equal to the stated percentile probability, and solve for the rate. The main risks are interpreting a survival percentage as a lower-tail percentile, losing track of time units, and rounding the rate before checking the result.

Definition: Let \(T\) be an exponential waiting time with rate \(\lambda>0\), measured in inverse units of time. For \(t\geq 0\), its CDF is \(F_T(t)=P(T\leq t)=1-e^{-\lambda t}\), and its survival function is \(P(T>t)=e^{-\lambda t}\). The \(p\)-quantile, for \(0<p<1\), is the time \(q_p\) satisfying \(F_T(q_p)=p\).

The CDF gives the probability that the waiting time is at most a specified time. Thus, a statement that “the 80th percentile is 10 hours” means \(P(T\leq 10)=0.80\), not \(P(T>10)=0.80\). The exponential CDF is continuous and strictly increasing for \(t\geq0\), so each percentile strictly between 0 and 1 corresponds to exactly one positive quantile when \(\lambda>0\).

Deriving the Rate Formula

Suppose a model specifies that the \(p\)-quantile is \(q_p>0\). Substituting \(q_p\) into the CDF and isolating \(\lambda\) gives the rate directly.

\begin{aligned} p &= F_T(q_p)=1-e^{-\lambda q_p},\\ e^{-\lambda q_p} &= 1-p,\\ -\lambda q_p &= \ln(1-p),\\ \lambda &= -\frac{\ln(1-p)}{q_p}. \end{aligned}

Theorem: Under an exponential model, if \(q_p>0\) is the \(p\)-quantile for \(0<p<1\), then the unique rate consistent with it is \(\lambda=-\ln(1-p)/q_p\). The rate has units reciprocal to those of \(q_p\).

The inversion is valid because \(1-p\) lies strictly between 0 and 1, so its natural logarithm is negative. Therefore \(-\ln(1-p)>0\), consistent with the required positive exponential rate. A time of zero cannot be a positive-probability quantile for this continuous model, while \(p=0\) and \(p=1\) do not produce a finite positive rate from this formula.

Translate the wording.
Write the given percentile as \(P(T\leq q_p)=p\). If the statement describes the proportion still waiting, first convert it to a lower-tail probability.

Record the time unit.
For example, if the quantile is in hours, the rate obtained is per hour. Do not silently mix minutes, days, or other units.

Substitute and solve.
Use \(\lambda=-\ln(1-p)/q_p\), retaining the logarithmic expression as long as possible.

Check in the original model.
Substitute the unrounded rate into \(F_T(q_p)=1-e^{-\lambda q_p}\), or check that the survival probability is \(1-p\).

Worked Rate-Recovery Problems

Worked Example: A 90th-Percentile Resolution Time

In a hypothetical support system, request resolution times are modeled as exponential with an unknown rate measured per day. The model’s 90th percentile is 14 days. Find the rate and verify that the stated percentile is reproduced.

Method. The 90th percentile means \(P(T\leq14)=0.90\), so \(p=0.90\) and \(q_p=14\) days. Applying the quantile formula,

\lambda=-\frac{\ln(1-0.90)}{14} =-\frac{\ln(0.10)}{14} =\frac{\ln(10)}{14} =\frac{2.302585092994\ldots}{14} =0.1644703638\ldots\ \text{per day}.

The rate is approximately \(0.1645\) per day to four significant figures. For the check, use the unrounded rate expression, not the displayed rounded decimal:

F_T(14) =1-e^{-(\ln(10)/14)14} =1-e^{-\ln(10)} =1-0.10 =0.90.

The stated 90th percentile is recovered exactly by the unrounded expression. Common mistake avoided: rounding the rate and then treating the resulting exponent as exact. The rounded rate \(0.1644704\) gives \(14(0.1644704)=2.3025856\), approximately, rather than exactly \(\ln(10)\). The exact check uses \(\ln(10)/14\); a check using the rounded rate must use approximate equality throughout.

Worked Example: Recovering a Rate from a Median

A hypothetical sensor is monitored until its first alert. The exponential model assigns a median waiting time of 6 hours. Find the rate per hour and check the median condition.

A median is the 50th percentile, so \(p=0.50\) and \(q_{0.50}=6\) hours. Therefore,

\lambda=-\frac{\ln(1-0.50)}{6} =-\frac{\ln(0.50)}{6} =\frac{\ln(2)}{6} =\frac{0.6931471806\ldots}{6} =0.1155245301\ldots\ \text{per hour}.

To four significant figures, the rate is \(0.1155\) per hour. Checking with the unrounded rate,

P(T\leq 6) =1-e^{-(\ln(2)/6)6} =1-e^{-\ln(2)} =1-\frac12 =0.50.

Thus 6 hours is indeed the model median. Common mistake avoided: using \(\ln(0.50)/6\) without the leading negative sign. That expression is negative, so it cannot be an exponential rate. The rate must be positive.

Worked Example: Converting a Survival Statement into a Percentile

In a hypothetical queue, the exponential model says that 30% of jobs are still waiting after 5 hours. Find the rate per hour and identify the percentile represented by 5 hours.

“Still waiting after 5 hours” describes the survival probability \(P(T>5)=0.30\). The corresponding lower-tail probability is

P(T\leq 5)=1-0.30=0.70.

Therefore 5 hours is the 70th percentile, and

\lambda =-\frac{\ln(1-0.70)}{5} =-\frac{\ln(0.30)}{5} =\frac{1.2039728043\ldots}{5} =0.2407945609\ldots\ \text{per hour}.

To four significant figures, the rate is \(0.2408\) per hour. Check the original survival statement using the unrounded rate:

P(T>5) =e^{-( -\ln(0.30)/5)5} =e^{\ln(0.30)} =0.30.

The check agrees with the given proportion still waiting. Common mistake avoided: setting \(F_T(5)=0.30\). That would model 30% finished by 5 hours and 70% still waiting, reversing the stated condition.

Interpreting and Checking the Result

A recovered rate is a model parameter, not a time percentile. Its units must match the time scale used in the quantile. If the quantile is in days, the formula returns a rate per day; if the same time is expressed in hours, the numerical rate changes accordingly. State both the number and its time unit in a final answer.

The formula uses the population or model percentile supplied by the problem. It does not by itself quantify uncertainty in a percentile estimated from a finite sample. The calculation also presumes an exponential waiting-time model; a single quantile alone does not establish that the model is suitable.

A reliable verification returns to the original probability statement. If the problem gives a lower-tail percentile, check \(F_T(q_p)=p\). If it gives the fraction still waiting, check \(P(T>q_p)=1-p\). Keep the exact logarithmic expression until the check is complete, and round the reported rate only at the end.

Key takeaway: For a stated exponential \(p\)-quantile \(q_p\), recover the rate with \(\lambda=-\ln(1-p)/q_p\). First identify whether the wording gives \(P(T\leq q_p)\) or \(P(T>q_p)\), then preserve the time unit and verify using the unrounded rate.

Check Your Understanding

For each problem, assume the stated waiting time follows an exponential model. Keep the time unit attached to the rate, and use the final answers to check your calculations.

A hypothetical dispatch system has a 25th-percentile response time of 4 minutes. Find the exponential rate per minute. (Answer: \(-\ln(0.75)/4=0.07192\) per minute, to four significant figures.)
A hypothetical device has a 60th-percentile time to its first warning of 8 hours. Find the rate per hour and state the survival probability at 8 hours. (Answer: \(-\ln(0.40)/8=0.1145\) per hour; survival probability \(0.4000\).)
A hypothetical repair duration has a median of 9 days under an exponential model. Recover the rate per day and verify the median condition. (Answer: \(\ln(2)/9=0.07702\) per day; \(P(T\leq9)=0.5000\).)
In a hypothetical inspection process, 85% of items are inspected within 12 months. Find the rate per month and identify the percentile represented by 12 months. (Answer: \(-\ln(0.15)/12=0.1581\) per month; 12 months is the 85th percentile.)
A hypothetical claim review is still in progress for 20% of claims after 3.2 years. Find the rate per year and state the lower-tail percentile at 3.2 years. (Answer: \(-\ln(0.20)/3.2=0.5029\) per year; 3.2 years is the 80th percentile.)

What You'll Learn