Birthday Paradox: Why 23 People Give a 50% Shared Birthday

The birthday paradox simulator needs one number: how many people are in the room. Set a range — say 2 to 60 — and the tool plots the probability of at least one shared birthday at every group size in that range. You can also adjust the number of possible birthdays (default 365) and toggle a Monte Carlo simulation that generates thousands of random trials alongside the exact formula.

Leave the defaults alone for a first run. The curve that appears is the whole point: it climbs far faster than anyone expects, crossing 50% at just 23 people and 99% by 57. Once you see it, change the “days in year” to 100 or 1,000 and watch the threshold shift — that single tweak reveals why the paradox generalizes to hash collisions and cryptographic security.

Most people guess you need about 183 people for a 50% chance of a shared birthday — halfway through 365. That guess feels right but answers the wrong question. It answers “how many people until someone probably shares my birthday?” The actual question is broader: does any pair in the group match?

The difference is enormous. With 23 people there are 23 × 22 / 2 = 253 unique pairs. Each pair has a small chance of matching, but 253 small chances pile up fast. This is complementary counting at work: instead of tracking every possible match, you calculate the probability that nobody shares a birthday and subtract from 1. The first person can land on any of 365 days. The second needs one of the remaining 364. The third needs 363. Multiply all those fractions together and by person 23, the “no match” probability dips just below 0.50 — meaning a match is more likely than not.

The sample space of possible birthday arrangements grows factorially while the number of collision-free arrangements shrinks. That imbalance is why 23 is enough, and why the result shocks even professional mathematicians the first time they see it.

Set the range to 2–60, step size 1, days = 365, and enable the Monte Carlo simulation with 5,000 trials. Hit calculate. Here's what the output looks like for a few key group sizes:

The simulated column wobbles a point or two each run — that's normal sampling variation. Run 50,000 trials instead of 5,000 and the wobble shrinks. The theoretical column never changes because it's the exact product of fractions, not a random draw.

Now change days to 100 (imagine a planet with a 100-day year). The 50% threshold drops to about 12 people and the 99% threshold to about 31. The critical group size scales roughly as the square root of the number of possible birthdays — √365 ≈ 19, √100 ≈ 10 — which is the same scaling law behind birthday attacks on cryptographic hashes.

The exact formula uses complementary counting. Let D = days in the year and n = people in the room:

Monte Carlo simulation takes a different route to the same answer. Generate n random integers between 0 and D − 1, check for duplicates, repeat thousands of times, and count what fraction of trials produced a collision. With enough trials the simulated proportion converges to the theoretical value — a live demonstration of the law of large numbers.

There's also a handy approximation: P(no match) ≈ e^−n²/(2D). It overestimates the chance of no match slightly, but it's fast for back-of-envelope work. Setting that equal to 0.5 and solving gives n ≈ 1.18√D, which is where the “square root of the number of days” rule of thumb comes from.

Asking the wrong question. “Will someone match my birthday?” is not the birthday paradox. That question needs about 253 people for a 50% chance (roughly D × ln 2). The paradox asks about any pair, which involves far more comparisons for the same group size.

Treating the probability as linear. Doubling the group doesn't double the probability. Pairs grow quadratically — 23 people make 253 pairs, 46 people make 1,035 — so the probability curve is an S-shape, not a straight line.

Assuming uniform birthdays matter for the real world. Real birth rates aren't uniform. More babies are born in September than in January in the U.S. Non-uniform distributions actually make collisions more likely, not less. The uniform model is a conservative lower bound, which is useful precisely because it still gives 50% at just 23.

Ignoring leap years. February 29 birthdays are roughly four times rarer than any other date. Including leap day (D = 366 with non-uniform weighting) nudges the 50% threshold down slightly because the non-uniformity increases collision probability. The effect is tiny — less than one person — but it comes up on exams.

Why is it called a paradox if it's mathematically correct? It's a veridical paradox — a statement that sounds wrong but is provably true. The math is straightforward; it's human intuition that breaks down because we instinctively compare to our own birthday rather than scanning every possible pair.

How does this relate to hash collisions? A hash function maps inputs to a fixed set of outputs, analogous to mapping people to birthdays. The birthday bound tells you that with a hash of length b bits, you expect a collision after roughly 2^b/2 inputs — not 2^b. That's why 128-bit hashes were retired in favor of 256-bit ones: the effective security is half the bit length.

What happens at 366 people? With 365 possible birthdays and 366 people, the pigeonhole principle guarantees at least one match — probability hits exactly 100%. The simulator confirms this: every single trial finds a duplicate.

Can I use the simulator for a class presentation? Yes. The theoretical values match the standard formula taught in every introductory probability course. Run the simulation live to show Monte Carlo convergence — students remember the wobbling line settling onto the theoretical curve far better than they remember a static equation.