Why the Geometric Distribution Is Memoryless

What “Memoryless” Means for a Geometric Waiting Time

A geometric waiting time describes the number of trials up to and including the first success. Its memoryless property says that, after observing failures, the distribution of the additional wait is the same as it was at the start. This is a precise statement about conditional probabilities, not a claim that the elapsed trials literally have no effect in every waiting-time model.

The assumptions matter. The geometric model uses independent Bernoulli trials with the same success probability on every trial. Independence means earlier outcomes do not change the probabilities of future outcomes; the constant probability means each future trial has the same chance of success. The memoryless identity follows from both features.

Definition: Let \(X_1,X_2,\ldots\) be independent Bernoulli trials, each with success probability \(p\), where \(0<p<1\). Let \(T\) be the trial number of the first success, so \(T\in\{1,2,\ldots\}\), and write \(q=1-p\). Then \(T\) has the geometric distribution with probability mass function \(P(T=k)=q^{k-1}p\), for \(k=1,2,\ldots\). The geometric waiting time is memoryless if, for integers \(m,n\geq 0\), \(P(T>m+n\mid T>m)=P(T>n)\), whenever the conditioning event has positive probability.

With this convention, \(T>m\) means that the first \(m\) trials were all failures. It does not mean that the first success occurs on trial \(m\). Being precise about this event is essential: the identity compares the wait for more than \(n\) additional trials, given \(m\) failures, with the original probability of waiting more than \(n\) trials.

Why the Identity Holds

The event \(T>k\) occurs exactly when the first \(k\) trials are failures. By independence and the common failure probability \(q\), its probability is \(q^k\). We can therefore calculate the conditional probability directly.

P(T>m+n\mid T>m) =\frac{P(T>m+n)}{P(T>m)} =\frac{q^{m+n}}{q^m} =q^n =P(T>n).

The denominator is positive because \(0<p<1\), so \(q^m>0\) for every finite \(m\). The cancellation works because the probability of surviving the first \(m+n\) trials factors into the probability of \(m\) failures and the probability of another \(n\) failures. Under the independent, constant-\(p\) assumptions, observing the first \(m\) failures does not alter the probability of the next \(n\) failures.

The same reasoning gives the full conditional distribution of the additional wait. For an integer \(r\geq 1\), the event \(T-m=r\), given \(T>m\), means failures on the next \(r-1\) trials followed by a success. Thus

P(T-m=r\mid T>m) =\frac{q^{m+r-1}p}{q^m} =q^{r-1}p =P(T=r).

This is a stronger way to express the result: conditional on having had \(m\) failures, the number of additional trials until the first success has the same geometric distribution as \(T\). The result concerns the additional wait, not the original total \(T\), which is already known to exceed \(m\).

Using the Property in Conditional Calculations

Once the model and conditioning event are stated clearly, the memoryless identity reduces many calculations to ordinary geometric probabilities. For example, the probability of a success within the next \(n\) trials, given \(m\ initial failures, is the complement of surviving all \(n\) further trials: \(1-q^n\). The probability that the first success comes exactly \(r\) trials after those failures is \(q^{r-1}p\).

Worked Example: Waiting for a Hypothetical Inspection Alert

A hypothetical device is checked once per hour. On each check, it independently produces an alert with probability \(p=0.25\), unchanged from hour to hour. Let \(T\) be the hour of the first alert. Given that there was no alert in the first 2 hours, find (a) the probability of no alert in the next 3 hours and (b) the probability that the first alert occurs exactly 3 hours after the observed failures. Compare the first result with the unconditional probability that \(T>5\).

Method and model. This is a geometric waiting time in trials-until-first-success form, with \(p=0.25\) and \(q=1-p=0.75\). The condition \(T>2\) says the first two checks failed.

(a) No alert in the next 3 hours. This is the event \(T>5\) conditional on \(T>2\). Applying the identity with \(m=2\) and \(n=3\),

P(T>5\mid T>2)=P(T>3)=q^3=(0.75)^3=0.421875.

Rounded to four decimal places, the conditional probability is \(0.4219\).

(b) First alert exactly 3 hours later. The next two checks must fail and the third must succeed. Therefore,

P(T-2=3\mid T>2)=q^2p=(0.75)^2(0.25) =0.5625(0.25)=0.140625.

Rounded to four decimal places, this probability is \(0.1406\). As a check using conditional probability directly, \(P(T=5)=(0.75)^4(0.25)=0.0791015625\) and \(P(T>2)=(0.75)^2=0.5625\). Their ratio is \(0.0791015625/0.5625=0.140625\), as above.

Comparison. Unconditionally, \(P(T>5)=(0.75)^5=0.2373046875\), or \(0.2373\) to four decimals. This is not the conditional probability in part (a): the conditioning on two failures removes those already-observed failures from the additional wait.

Common mistake avoided: Replacing \(P(T>5\mid T>2)\) with \(P(T>5)\) gives \(0.2373\), but ignores the condition. The correct calculation divides \(P(T>5)\) by \(P(T>2)\), or uses the identity to obtain \(P(T>3)=0.4219\).

Conditional Ranges and Exact Future Trials

The property also handles a range of future outcomes. Given \(T>m\), the probability of a first success within the next \(n\) trials is one minus the probability of no success throughout those trials. For an exact future trial, count the failures before that success. These event translations keep the time origin clear: “within the next \(n\)” includes the next trial and the \(n\)th one after the observed failures.

Worked Example: A Hypothetical Weekly Claim Report

In a hypothetical portfolio-monitoring exercise, a weekly review independently reports a new claim with probability \(p=0.10\) each week, with the probability fixed across weeks. Let \(T\) be the week of the first report. Given that no report occurred in the first 6 weeks, calculate (a) the probability that the first report occurs within the next 4 weeks and (b) the probability that it occurs exactly 3 weeks after week 6.

Method and model. The stated independent, constant-probability assumptions give a geometric model with \(p=0.10\) and \(q=0.90\). The condition is \(T>6\).

(a) Within the next 4 weeks. The complement is no report in all four additional weeks. By memorylessness,

P(T\leq 10\mid T>6) =1-P(T>10\mid T>6) =1-P(T>4) =1-(0.90)^4.

The intermediate power is \((0.90)^2=0.81\), so \((0.90)^4=(0.81)^2=0.6561\). Thus the requested probability is \(1-0.6561=0.3439\).

(b) Exactly 3 weeks after week 6. The first two additional weeks must have no report, followed by a report in the third. Hence

P(T=9\mid T>6) =(0.90)^2(0.10) =0.81(0.10) =0.0810.

As a direct check, \(P(T=9)=(0.90)^8(0.10)\) and \(P(T>6)=(0.90)^6\); dividing cancels the first six failure factors and leaves \((0.90)^2(0.10)=0.0810\).

Common mistake avoided: The condition \(T>6\) does not make “within the next 4 weeks” the event \(T\leq4\). In the original time scale it is \(T\leq10\); equivalently, the additional wait is at most 4 weeks.

Why a Constant Success Probability Matters

If trials are independent but their success probabilities change over time, the probability of surviving a block of trials is a product of their individual failure probabilities. The probability of surviving the next block need not match the probability of surviving an equally long block at the start. Thus independence by itself does not guarantee the geometric memoryless identity; a common success probability is also needed.

For independent trials with success probabilities \(p_1,p_2,\ldots\), the first \(k\) trials all fail with probability \(\prod_{j=1}^{k}(1-p_j)\). Consequently, the conditional probability of no success in trials \(m+1\) through \(m+n\), given no success in the first \(m\), is

\prod_{j=m+1}^{m+n}(1-p_j).

This product is not generally equal to the probability of no success in the first \(n\) trials, \(\prod_{j=1}^{n}(1-p_j)\). They are equal under a constant success probability, but need not be equal when the probabilities vary.

Worked Example: Changing Alert Probabilities Break the Identity

A hypothetical four-check process has independent checks, but the success probabilities are not constant: \(p_1=0.20\), \(p_2=0.40\), \(p_3=0.50\), and \(p_4=0.25\). Let \(T\) be the first successful check. Compare the probability of no success in checks 3 and 4 given failures in checks 1 and 2 with the unconditional probability of no success in the first two checks.

Method and assumptions. The checks are independent, so the relevant failure probabilities multiply. However, this is not a geometric model because the success probabilities differ by check.

Conditional probability for checks 3 and 4. Given failures in the first two checks, the probability of no success in the next two is

P(T>4\mid T>2) =(1-p_3)(1-p_4) =(1-0.50)(1-0.25) =(0.50)(0.75)=0.3750.

Unconditional probability for the first two checks. The probability that both fail is

$$ P(T>2)=(1-p_1)(1-p_2) =(1-0.20)(1-0.40) =(0.80)(0.60)=0.4800. $$

The two probabilities are \(0.3750\) and \(0.4800\), so the geometric identity would fail here: \(P(T>4\mid T>2)\ne P(T>2)\). The earlier failures do not change the specified probabilities of future independent checks, but the future checks have different probabilities from the first two.

Common mistake avoided: Independence does not make these checks identically distributed. Applying the geometric formula \(q^2\) would require the same \(q\) on every check, which this process does not have.

There is a separate reason independence is part of the model. Even if each trial has the same marginal success probability, dependence can make future outcomes change in probability after failures have been observed. In that case, the cancellation in the geometric derivation is not justified. To use the memoryless result, verify both independence and a fixed success probability, not just one of them.

Common Interpretations and Exam Traps

Confusing elapsed time with additional time. Given \(T>m\), the identity concerns the next \(n\) trials. In original trial numbering, that event is \(T>m+n\).
Using the wrong geometric convention. Here \(T\) counts trials up to and including success, so \(T>m\) means \(m\) failures. If a problem instead counts failures before success, translate its event carefully before using a formula.
Assuming the property from independence alone. Independent trials with changing success probabilities generally do not satisfy the geometric identity.
Treating a conditional probability as unconditional. \(P(T>m+n\mid T>m)\) is not usually \(P(T>m+n)\); the denominator \(P(T>m)\) matters.
Using an impossible conditioning event. The identity as stated requires \(P(T>m)>0\). For the geometric model with \(0<p<1\), this holds for every finite \(m\).

In actuarial applications, the result can simplify calculations for repeated opportunities to observe a first event, provided the trials genuinely follow the assumed model. It should not be applied automatically to durations or sequences whose success probabilities vary with time or whose outcomes depend on earlier results. The mathematical convenience of memorylessness is a consequence of the model assumptions, not a general feature of waiting.

Key takeaway: For an independent sequence of Bernoulli trials with the same success probability \(p\), the geometric waiting time satisfies \(P(T>m+n\mid T>m)=P(T>n)\). The constant probability makes the future block have the same failure probabilities as the initial block; independence permits the conditional cancellation.

Check Your Understanding

For each problem, use the trials-until-first-success convention and check the conditioning event before calculating.

A hypothetical sensor succeeds independently on each check with probability \(0.30\), constant across checks. Given that the first 4 checks failed, find the probability that the next 3 checks also fail. (Answer: \(0.7^3=0.3430\).)
A hypothetical weekly test independently detects a condition with fixed probability \(0.20\). Given no detection in the first 3 weeks, find the probability of at least one detection in the next 2 weeks. (Answer: \(1-0.8^2=0.3600\).)
A repeated independent process has fixed success probability \(0.25\). Given \(T>2\), find the probability that the first success occurs exactly 3 trials later. (Answer: \(0.75^2(0.25)=0.1406\), to four decimal places.)
For a geometric waiting time with \(p=0.10\), calculate \(P(T>8\mid T>5)\), then compare it with \(P(T>3)\). (Answer: Both equal \(0.9^3=0.7290\).)
Independent checks have success probabilities \(p_1=0.10\), \(p_2=0.20\), \(p_3=0.40\), \(p_4=0.50\), and \(p_5=0.50\). Calculate \(P(T>5\mid T>3)\) and \(P(T>2)\), and decide whether the geometric identity holds for these blocks. (Answer: \(P(T>5\mid T>3)=0.5(0.5)=0.2500\), while \(P(T>2)=0.9(0.8)=0.7200\); they are not equal.)

What You'll Learn