LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
For daily counts modeled as Poisson, the maximum likelihood estimate of the rate is the sample average. A Poisson law is the discrete distribution used for counts, with one parameter, often called lambda, that is also the mean. In a simulated hospital with a true rate of 85 patients a day, 100 beds, and 3,650 days, the average of the simulated counts was about 84.8. The chance of more than 100 patients that day was about 5 percent using the estimate, and about 5 percent using the true rate, a couple of basis points apart.
Why it matters
The story is beds, not tickers, so the probability has a plain meaning: the share of days the ward cannot take everyone. The lecture’s simplifications both make the risk look smaller than a real ward. Patients are assumed to leave the same day they arrive, so nobody occupies a bed overnight. Days are assumed independent, so a contagious week cannot cluster arrivals. The lecture says you should notice the direction of those mistakes even while the algebra stays clean. Refine the model later. Do not pretend the 5 percent is a conservative figure.
The idea in plain words
Each day is a count that cannot be a fraction of a person. The lecture writes the Poisson probability of a count as e to the minus lambda, times lambda to that count, divided by the count’s factorial. The data are ten years of days. Lambda is the unknown population rate.
Maximum likelihood chooses the lambda that makes the observed sample as probable as possible. Because the days are modeled as independent and identical, the joint probability is the product of the daily probabilities. The log of a product is a sum, and the lambda that maximizes the probability also maximizes the log. Substitute the Poisson formula, use the log rules to split products and quotients, and the factorial term does not depend on lambda.
Differentiate that sum with respect to lambda and set the derivative to zero. The lecture skips a second-derivative test and treats the critical point as the maximum. The terms simplify to one statement: lambda equals the average of the counts. That is the maximum likelihood estimator. The lecture lists properties it wants you to remember in large samples: the bias goes to zero, the variance is as small as the theory allows in the asymptotic sense they state, and the estimate’s distribution becomes normal even though the counts are Poisson. “Asymptotic” here means as the number of days grows.
The simulation draws 3,650 Poisson counts with lambda 85, the stand-in for ten years of 365 days. The mean of one such draw was given as 84.83, and a nearby remark says 84.6. The lecture’s point is “near 85,” not a single sacred rounding. Resamples would scatter, and that scatter is itself approximately normal by the central limit theorem. SciPy’s Poisson object, parameterized by the estimate rather than by 85, then answers the bed question. The probability of more than 100 arrivals is 1 minus the cumulative probability of 100 or fewer. Both the estimated rate and the true rate put that chance near 5 percent.
How to use it
If you truly believe the daily counts are independent Poisson draws, average them and stop. Do not invent a more elaborate estimator for lambda until the Poisson story is the one you mean. Then put that average into the cumulative function and subtract from 1 to get the chance of exceeding capacity. Capacity in the example is 100 beds. Change the threshold and the probability changes. The lecture does not compute other thresholds.
What usually goes wrong
Treating 5 percent as exact ignores both the lecture’s “approximately” and the gap between 84.8 and 85. Another sample of 3,650 days would print another average and another tail. The pair of tails being close is the success criterion.
Questions people ask
What is the maximum likelihood estimate of a Poisson rate?
The sample mean of the counts. That is what remains after the log likelihood is differentiated and set to zero under the lecture’s independence assumption.
How close was the hospital simulation?
Counts were drawn at a true rate of 85 for 3,650 days. One reported average was about 84.8. The probability of a day above 100 beds was roughly 5 percent either with that estimate or with 85, off by a couple of basis points.
Why is the 5 percent not a cautious planning number?
The model empties every bed overnight and treats days as independent. The lecture says both choices understate the chance of running out of beds relative to a ward where stays overlap and illness clusters.