Normal Approximation to Poisson Counts at Large Rates

Why a Large-Rate Poisson Count Becomes Approximately Normal

A Poisson model is useful for counting events in a fixed interval when events occur independently at a stable average rate. Its mean and variance are both equal to its rate. When that rate is large, the count’s distribution is often approximated by a normal distribution with the same mean and variance. This can turn a complicated count probability into a standard-normal calculation.

The phrase large rate matters. A normal distribution is continuous and symmetric, while a Poisson count takes nonnegative integer values and can be strongly right-skewed at low rates. The approximation is therefore an asymptotic method, not an identity. A useful calculation begins by specifying the interval and rate, then checking whether the resulting count distribution is plausibly close enough to normal for the question being asked.

Definition: If \(X\sim\operatorname{Poisson}(\lambda)\), then \(X\) is a count with probability mass function \(P(X=k)=e^{-\lambda}\lambda^k/k!\) for \(k=0,1,2,\ldots\), where \(\lambda>0\) is the expected count in the specified interval. Its mean and variance are \(E[X]=\lambda\) and \(\operatorname{Var}(X)=\lambda\). For a sufficiently large \(\lambda\), the normal approximation is \(X\approx Y\), where \(Y\sim N(\lambda,\lambda)\); the second parameter is the variance, so the normal standard deviation is \(\sqrt{\lambda}\).

The parameter \(\lambda\) is the expected count for the whole interval in the model, not necessarily a rate per unit of time. If the expected rate is \(r\) events per hour and the observation lasts \(t\) hours, then the count has parameter \(\lambda=rt\), assuming the rate is stable and the Poisson model applies over that interval. A rate in events per day cannot be used unchanged for a count over several weeks.

Why the Matching Normal Distribution Works

The Poisson mean and variance determine the normal model’s center and scale. The approximation is justified more precisely by standardizing the Poisson variable and examining its moment-generating function. For \(X\sim\operatorname{Poisson}(\lambda)\), its moment-generating function is \(M_X(s)=\exp(\lambda(e^s-1))\). Set \(Z=(X-\lambda)/\sqrt{\lambda}\). Then, for fixed real \(t\), the moment-generating function of \(Z\) is

M_Z(t) =E\left[e^{t(X-\lambda)/\sqrt{\lambda}}\right] =\exp\left(\lambda\left(e^{t/\sqrt{\lambda}}-1\right)-t\sqrt{\lambda}\right).

As \(\lambda\) grows, expand the exponential in the exponent:

e^{t/\sqrt{\lambda}} =1+\frac{t}{\sqrt{\lambda}}+\frac{t^2}{2\lambda} +O\left(\lambda^{-3/2}\right), $$ $$ \lambda\left(e^{t/\sqrt{\lambda}}-1\right)-t\sqrt{\lambda} =\frac{t^2}{2}+O\left(\lambda^{-1/2}\right).

Thus \(M_Z(t)\) tends to \(e^{t^2/2}\), the moment-generating function of a standard normal variable. This gives the large-rate approximation \(Z\approx N(0,1)\), or equivalently \(X\approx N(\lambda,\lambda)\). The argument explains why both the center \(\lambda\) and variance \(\lambda\) appear. It does not say that every finite rate gives an accurate approximation.

Practical guideline: There is no universal cutoff that guarantees a good Poisson-to-normal approximation for every event. Rates in the few units are generally too small to presume a normal shape. Rates in the tens or above may be useful for many central calculations, but accuracy still depends on the event and the required precision. Check the model, the rate, and whether the event lies in a region where the count distribution is reasonably represented by a symmetric curve.

Standardizing a Poisson Count

Once the normal approximation is judged reasonable, use the normal standardization formula. If \(X\sim\operatorname{Poisson}(\lambda)\) is approximated by \(Y\sim N(\lambda,\lambda)\), then the corresponding standard-normal value for a count threshold \(c\) is

z=\frac{c-\lambda}{\sqrt{\lambda}}.

This tutorial uses raw CLT standardization in its numerical probability examples: an integer threshold is inserted directly into the displayed formula. This makes the approximation being calculated explicit. A half-unit boundary adjustment is a separate refinement for discrete counts; it is not applied in these examples.

The direction of the event still matters. A count below a threshold uses the lower standard-normal tail, while a count above it uses the upper tail. For an interval, subtract the lower-tail probability at the lower standardized endpoint from the lower-tail probability at the upper endpoint. The discrete event’s integer limits must be read correctly before using any approximation.

Worked Example: A Central Range of Weekly Service Requests

A hypothetical support desk models the number \(X\) of requests in one week as Poisson with parameter \(\lambda=64\). Estimate \(P(55\le X\le73)\) using raw CLT standardization. No interest rate or mortality assumption is relevant; the model is \(X\sim\operatorname{Poisson}(64)\).

Method and model check. The mean count is 64 and the variance is 64, so the matching normal model is \(Y\sim N(64,64)\), with standard deviation \(\sqrt{64}=8\). A rate of 64 is large enough for a normal approximation to be plausible for a central range, though it remains an approximation.

Standardize both endpoints:

z_{\text{lower}}=\frac{55-64}{8}=-1.125, \qquad z_{\text{upper}}=\frac{73-64}{8}=1.125.

Using standard-normal probabilities, \(\Phi(1.125)\approx0.8697\) and \(\Phi(-1.125)=1-\Phi(1.125)\approx0.1303\). Therefore,

P(55\le X\le73) \approx \Phi(1.125)-\Phi(-1.125) =0.8697-0.1303 =0.7394.

The estimate is 0.7394, to four decimal places. The interval is symmetric around the mean, as seen from its endpoints being 9 below and 9 above 64; the equal-magnitude z-scores provide a check on the calculation. The method avoids treating the variance as the standard deviation: the standard deviation is 8, not 64.

Worked Example: Combining Rates Before Approximating

A hypothetical monitoring system records independent event counts from four sources during the same month. The respective counts are Poisson with parameters 24, 26, 18, and 32. Estimate the probability that the total count \(T\) is at least 115 using raw CLT standardization. No interest rate or mortality assumption is relevant.

Method and model check. The sum of independent Poisson variables is Poisson with parameter equal to the sum of their parameters. Therefore,

\lambda_T=24+26+18+32=100, \qquad T\sim\operatorname{Poisson}(100).

The matching normal model is \(Y\sim N(100,100)\), so its standard deviation is \(\sqrt{100}=10\). The raw standardized value at 115 is

z=\frac{115-100}{10}=\frac{15}{10}=1.5.

The desired event is an upper tail. Since \(\Phi(1.5)\approx0.9332\),

P(T\ge115)\approx P(Z\ge1.5) =1-\Phi(1.5) =1-0.9332 =0.0668.

The estimate is 0.0668, to four decimal places. The calculation uses the monthly total rate, not an individual source’s rate. A common error is to use 100 as the normal standard deviation, producing \(z=15/100=0.15\) and an entirely different tail probability. The correction is to use the Poisson standard deviation \(\sqrt{100}=10\).

When a Small Rate Is a Warning

A low Poisson rate can make the count distribution visibly unlike a normal curve. The count has no probability below zero and places probability only at integers. A normal variable is continuous, symmetric, and can be negative. For the matching normal variable \(Y\sim N(\lambda,\lambda)\), the probability of a negative value is \(\Phi(-\sqrt{\lambda})\). This probability decreases as \(\lambda\) increases, but it is only one diagnostic: a small negative probability does not guarantee accurate tail probabilities or probabilities at individual counts.

At a low rate, the shape and event location both deserve attention. A calculation near the center may behave differently from one near zero or far into a tail. If a problem calls for a reliable result and the rate is small, use the Poisson model directly when feasible rather than assuming that the normal approximation is adequate.

Worked Example: Diagnosing a Low-Rate Approximation

A hypothetical small facility models the number \(X\) of equipment alerts in a shift as Poisson with parameter \(\lambda=4\). Assess a raw normal approximation to \(P(X\le2)\), and compare it with the exact Poisson probability. No interest rate or mortality assumption is relevant; \(X\sim\operatorname{Poisson}(4)\).

Raw normal estimate. The matching model is \(Y\sim N(4,4)\), with standard deviation \(2\). Standardizing the threshold 2 gives

z=\frac{2-4}{2}=-1, \qquad P(X\le2)\approx\Phi(-1)=0.1587.

Exact Poisson check. The event includes counts 0, 1, and 2, so

P(X\le2) =P(X=0)+P(X=1)+P(X=2) =e^{-4}\left(1+4+\frac{4^2}{2!}\right) =13e^{-4} \approx13(0.0183156) =0.2381.

The raw normal estimate, 0.1587, is materially below the exact value 0.2381. The gap is about \(0.2381-0.1587=0.0794\). Also, the matching normal model assigns probability \(\Phi(-2)=0.0228\) to negative values, which are impossible for a count. These are warnings that \(\lambda=4\) is too small to rely on the normal approximation for this event. The method avoids assuming that matching the mean and variance alone ensures a good approximation.

Common Traps and a Reliable Workflow

The normal approximation is straightforward once the model and scale have been set correctly. Most errors come from misreading the interval or the rate rather than from the standard-normal calculation itself.

Using the rate as the standard deviation. For a Poisson parameter \(\lambda\), variance is \(\lambda\), but standard deviation is \(\sqrt{\lambda}\).
Forgetting to scale the observation interval. Convert a per-hour or per-day rate to the full interval before defining \(\lambda\).
Summing dependent counts as if independent. The Poisson sum rule used above requires independence. If the counts are dependent, their total is not justified as Poisson merely by adding the stated rates.
Assuming a large mean guarantees every tail is accurate. The normal approximation is asymptotic. A stringent accuracy requirement or an extreme event can call for a direct Poisson calculation.
Reversing an upper tail. For a standardized threshold \(z\), the upper-tail probability is \(1-\Phi(z)\), not \(\Phi(z)\).

A disciplined solution follows a short sequence: identify the count and its observation interval; state the Poisson parameter and assumptions; calculate the matching normal mean, variance, and standard deviation; assess whether the rate supports approximation; then standardize the event and choose the correct tail or interval operation. This keeps the model decision separate from the arithmetic and makes the approximation’s limitations visible.

Key takeaway: For a large-rate Poisson count \(X\sim\operatorname{Poisson}(\lambda)\), use the approximation \(X\approx N(\lambda,\lambda)\), with standard deviation \(\sqrt{\lambda}\). Treat it as a model approximation whose adequacy depends on the rate and the event—not as an exact replacement for the Poisson distribution.

Check Your Understanding

Use raw CLT standardization as specified in each question; do not apply a half-unit boundary adjustment. Give probability answers to four decimal places.

A hypothetical kiosk has a Poisson count of 121 transactions per day. State the matching normal mean, variance, and standard deviation, and explain whether a normal approximation is plausible. (Answer: mean 121, variance 121, standard deviation 11; plausible for many central calculations, but not guaranteed for every tail.)
A hypothetical laboratory models its daily sample count as \(X\sim\operatorname{Poisson}(169)\). Estimate \(P(X\le156)\) using raw CLT standardization. (Answer: \(z=-1\), so \(\Phi(-1)=0.1587\).)
Three independent hypothetical counters have Poisson rates 30, 45, and 25 per shift. Approximate the probability that their total is between 90 and 110, inclusive, using raw CLT standardization. (Answer: total rate 100, standard deviation 10, z-scores \(-1\) and \(1\), so \(0.6827\).)
A hypothetical repair depot has \(X\sim\operatorname{Poisson}(3)\) alerts in a shift. Find the exact value of \(P(X\le1)\), and compare it with the raw normal approximation. (Answer: exact \(e^{-3}(1+3)=0.1991\); raw normal estimate \(\Phi((1-3)/\sqrt3)=\Phi(-1.1547)\approx0.1241\), indicating a poor approximation for this low rate.)
A hypothetical network models hourly failures as \(X\sim\operatorname{Poisson}(400)\). Estimate \(P(X\ge440)\) by raw CLT standardization. (Answer: \(z=(440-400)/20=2\), so \(1-\Phi(2)=0.0228\).)

What You'll Learn