Domino Draw Odds, Done Properly: The Hypergeometric Math
It's tempting to treat drawing a domino hand like flipping a weighted coin several times in a row: figure out the "chance of a double" as a single percentage, then apply it independently to every tile in your hand. That approach is called a binomial model, and it's the right tool for plenty of real problems — but a domino hand isn't one of them, and using it here produces numbers that are confidently, measurably wrong. This guide covers the model that actually fits: the hypergeometric distribution, the same math behind our Draw Probability Calculator.
Why "with replacement" math doesn't apply here
A binomial model assumes every trial is independent — like flipping a coin, where the odds of heads stay exactly 50% no matter how many times you've flipped before, because nothing is removed from play between flips. Drawing a domino hand is nothing like that: once a tile leaves the boneyard and lands in your hand, it's gone. It can't also be dealt to an opponent, and it can't come up again later in your own draw. Every tile you take changes the odds for the next one, because the population you're drawing from is one tile smaller and, if that tile happened to be a "success" (a double, say), one success poorer too.
That's exactly the situation the hypergeometric distribution was built for: sampling a fixed number of items from a finite population, without putting anything back, where some items count as successes and the rest don't. It's the same underlying math used for the odds of drawing a specific number of aces from a shuffled card deck — a domino hand is just a differently-shaped version of the same problem.
The formula
For a population of N items containing K successes, the probability of drawing exactly x successes in a sample of size k is:
P(x) = [C(K, x) × C(N−K, k−x)] / C(N, k)
where C(a, b) is "a choose b," the standard combinations count. Applied to dominoes: N is the set size, K is however many tiles count as your "success" group (doubles, or any specific number — both counts are n+1 for a double-n set), k is your hand size, and x is however many successes you're asking about.
What each piece of the formula actually counts
The formula looks dense, but each of its three combination terms answers a distinct, concrete question about how many hands (not tiles) are possible:
- C(K, x) counts how many different ways you could pick exactly x doubles out of the K doubles available in the set. For x=2 doubles out of 7 available, that's C(7,2)=21 different possible pairs of doubles.
- C(N−K, k−x) counts how many ways to fill the rest of your hand — the remaining k−x tiles — from the non-double tiles. For a 7-tile hand with 2 doubles already chosen, that's the remaining 5 tiles picked from the 21 non-doubles: C(21,5)=20,349 ways.
- C(N, k) is the total number of possible 7-tile hands from the full 28-tile set, doubles and non-doubles combined, with no restriction at all: C(28,7)=1,184,040.
Multiplying the first two together and dividing by the third is just "how many hands match what I'm asking about, divided by how many hands are possible at all" — the most basic definition of probability there is. The hypergeometric formula isn't a special trick; it's that basic definition, applied with the right counting method for a without-replacement draw.
A full worked table: double-six, 7-tile hand
For a double-six set (N=28) with doubles as the success group (K=7) and a standard 7-tile hand (k=7), running every possible value of x through the formula gives the complete distribution:
| Doubles in hand (x) | Probability |
|---|---|
| 0 | 9.82% |
| 1 | 32.08% |
| 2 | 36.09% |
| 3 | 17.69% |
| 4 | 3.93% |
| 5 | 0.37% |
| 6 | 0.01% |
| 7 | ~0% |
Those eight probabilities add up to 100%, as they must — every possible hand falls into exactly one of these eight buckets. Two doubles is the single most likely outcome (36.1%), and "at least one double" — everything except the x=0 row — sums to 90.2%, matching the calculator's headline figure for this exact set-and-hand combination.
The wrong way, for contrast
Here's what a naive binomial approximation gets, using the same 7 doubles out of 28 tiles (a 25% "success rate") applied as if each of the 7 draws were independent:
| Doubles in hand (x) | Binomial (wrong) | Hypergeometric (correct) |
|---|---|---|
| 0 | 13.35% | 9.82% |
| 1 | 31.15% | 32.08% |
| 2 | 31.15% | 36.09% |
| 3 | 17.30% | 17.69% |
| 4 | 5.77% | 3.93% |
The two models agree loosely in the middle but diverge meaningfully at the edges: the binomial shortcut overstates the chance of a doubles-free hand (13.35% vs the true 9.82%) and understates "at least one double" as a result (86.65% vs the true 90.18%). That's not rounding noise — it's a systematic bias baked into treating dependent draws as if they were independent. The binomial model doesn't know that pulling a double out of the population makes the remaining pool relatively poorer in doubles for your next tile; the hypergeometric model does, because it's built around a shrinking, finite population from the start.
Why the gap exists, intuitively
Removing tiles without replacement pulls outcomes toward the population's actual composition more tightly than independent draws would. With 7 doubles genuinely fixed in a 28-tile set, drawing 7 tiles without replacement can't help but interact with that fixed count as you go — if your first six tiles happened to include zero doubles, the seventh tile is now being drawn from a pool where doubles make up a bigger share of what's left (7 out of the remaining 22, not 7 out of 28 anymore). A binomial model has no mechanism for that adjustment; every single one of its 7 trials uses the same flat 25% regardless of what came before, which is exactly why it smooths the distribution out too much at the extremes (too many "zero doubles" hands, not quite enough tightly-clustered middle outcomes).
The same math at other set sizes
The formula doesn't change shape as the set grows — only N and K do. A double-nine set (N=55, K=10 doubles) with a 9-tile hand gives P(0 doubles)=13.94%, P(1)=33.90%, P(2)=32.12%, P(3)=15.37%, with the remaining probability trailing off across x=4 and up. A double-twelve set (N=91, K=13 doubles) with a 15-tile hand — a plausible Mexican Train deal for two players — gives P(0)=7.97%, P(1)=24.27%, P(2)=31.36%, P(3)=22.65%, P(4)=10.14%. Every one of these numbers comes directly from the same formula above; nothing here is estimated or rounded from a smaller worked case. The full reference table for every common set size, using a fixed 7-tile hand for direct comparison, is on our Domino Set Reference page.
Expected value: a useful shortcut that's still exact
One genuinely simple number falls out of the same math without needing the full distribution: the expected (average, over many hands) number of doubles in a hand is just k × K ÷ N — hand size times the success proportion. For a double-six 7-tile hand, that's 7 × 7 ÷ 28 = 1.75 doubles on average. That figure is exact, not an approximation, even though any single hand can only ever contain a whole number of doubles — it's a long-run average, the same way "2.5 children per household" is a meaningful statistic despite no household literally having half a child.
A common misreading of "90% at least one"
It's worth heading off a natural misunderstanding: a 90.2% chance of holding at least one double in a double-six 7-tile hand does not mean "9 times out of 10, you'll definitely see a double, so treat it as a sure thing." It means exactly what it says — across many, many independently dealt hands of this exact shape, roughly 90.2% of them will contain at least one double, and the other 9.8% won't. Any single hand you're dealt is one sample from that distribution, not a guaranteed outcome pulled toward the average. A doubles-free hand at 9.8% is uncommon, not impossible, and treating it as impossible is exactly the kind of overconfident misread that turns a probability into a false certainty.
What this changes about how you read a hand
The practical upshot isn't that binomial math is "close enough to ignore" — a nearly four-point swing in the odds of a doubles-free hand is large enough to change how surprised you should be by one. If your opening hand comes up with zero doubles, the honest baseline to compare it against is 9.8%, not 13.4%: a genuinely less common event than the naive shortcut would have you believe, worth treating as more informative about what might be sitting in the boneyard or an opponent's hand, exactly the kind of read our guide to reading the boneyard builds on. None of this changes any rule of the game itself — it's purely about calculating the odds correctly once you've decided to think about them at all. And to be explicit about what this is and isn't for: everything above is about reading a game position — understanding a hand you've already been dealt, or estimating what a future draw is likely to bring — not about betting on an outcome. The same math would apply equally to a purely recreational, unscored game; the probabilities describe the tiles, not any stake placed on them.