The Paper
A reader sent me Igor Pak's 2003 paper The Nature of Partition Bijections II: Asymptotic Stability. I sat with it for a couple of nights.
The setup is this. Partition theory is the corner of combinatorics that studies how integers split into sums of smaller integers — the way 5 = 4+1 = 3+2 = 3+1+1 = 2+2+1 = 2+1+1+1 = 1+1+1+1+1. Sylvester in the 1880s started a tradition of proving partition identities by rearranging dots: draw the partition as a Young diagram, cut, slide, paste, and the result is a Young diagram of the partition you were trying to map to. These rearrangements are called partition bijections, and Pak's paper is about distinguishing the "good" ones from the "not so good" ones.
That sounds like a taste question. Pak's move is to make it not a taste question. He defines a property called asymptotic stability: a bijection is asymptotically stable if, when you scale a random partition down by 1/√n and let n run to infinity, the bijection converges to a clean piecewise-linear map between two deterministic curves — the limit shapes of the two partition families. Bijections that survive the scaling limit are good. Bijections that don't are not.
The result that stopped me was buried in §3. Pak quotes a classical theorem from Vershik:
The set of all integer partitions of n, scaled by 1/√n so the diagram has unit area, converges as n → ∞ to a deterministic curve.
A deterministic curve. Out of pure randomness. Take a random partition of a million, rescale it, plot it — you get the same shape every time. Different families of partitions (distinct parts, odd parts, parts not divisible by m) each have their own limit shape, and each is described by a closed-form equation.
I closed the paper and thought: every Powerball draw is a partition.
The Recognition
A Powerball white-ball draw is five sorted integers x₁ < x₂ < x₃ < x₄ < x₅ from {1, 2, …, 69}. Reverse the order and that's a partition with five parts, each between 1 and 69. We have 1,350 of these since the 2015 matrix change — 1,350 samples from a known combinatorial family.
If Pak is right — if random objects have limit shapes — then this lottery family should have one too. And we should be able to write it down in closed form, not as a mystery curve discovered empirically but as a formula that drops out of pure combinatorics.
There's no scaling required, because the box is fixed. The five numbers don't grow without bound; they're trapped between 1 and 69. So the "limit shape" question becomes: what does the distribution of the k-th smallest number look like, in the limit of many draws?
For uniform sampling without replacement of a 5-subset of {1..69}, that distribution has a name. It's the order-statistic PMF, and it's:
P(xₖ = v) = C(v−1, k−1) · C(69−v, 5−k) / C(69, 5)
Five values of k give five curves. Each one peaks near v ≈ k · 70/6 — so at roughly 11.67, 23.33, 35, 46.67, 58.33. They tile the integer interval from left to right, overlapping. Five humps.
That's the limit shape. Not approximate, not asymptotic, not empirical. Closed-form. The combinatorics knows what the lottery should look like before the lottery has ever drawn a single number.
The Theory, Stated Properly
Theories are cheap. Here is what I was actually claiming, made falsifiable before I built anything:
- The empirical distribution of each order statistic matches the closed form. Across all 1,350 historical draws, the five order statistics x₁ through x₅ should produce empirical histograms that match the five theoretical PMFs.
- Deviations sit inside the noise envelope. A fair lottery with only 1,350 samples won't match the closed form exactly — there's sampling noise. The empirical L∞ deviation should land inside what fair random sampling produces at the same N.
- The shape converges at rate 1/√N. As we accumulate more draws, the L∞ gap should shrink. Not in a mystical way — at the specific rate that the Central Limit Theorem predicts for histogram convergence.
If any of those failed, I'd have something. If all three held, I'd have confirmation of what Pak's framework predicts: random combinatorial objects, in the aggregate, are not arbitrary.
How We Tested It
The experiment lives at /experiments/limit-shape. The pipeline is mechanical, but the part that earns its keep is the noise envelope, so let me state it on its own:
A 1,350-draw histogram doesn't perfectly match its theoretical PMF, even if every draw is fair. The right comparison isn't "are they identical?" — they never are. It's "is the gap bigger than what fair random samples produce at the same sample size?"
To pin down what fair samples produce, we run a Monte Carlo. At each milestone N along the way — 50 draws, 100, 200, 350, 500, 700, 900, 1,100, 1,300, 1,350 — we generate 50 independent batches of N uniform random 5-subsets, compute the L∞ deviation between each batch's empirical PMF and the theoretical PMF, and take the 5th–95th percentile band. That's the envelope a fair lottery's deviation should sit inside about 90% of the time.
For each of the 1,350 real draws, we count how often each (k, v) pair appeared (where k is the order-statistic index, 1 to 5, and v is the value, 1 to 69). That's 5 × 69 = 345 bins. The L∞ distance is the largest gap, across all 345 bins, between the empirical proportion and the theoretical proportion.
Then we plot it against the simulated envelope.
The Result
The empirical L∞ deviation across all 345 bins is 1.291%.
The 95% noise envelope at N = 1,350 is 2.106%.
The ratio is 0.61× — comfortably inside the band.
The lottery has the shape it should. Not approximately. Not within an order of magnitude. Within the exact stochastic window that fair random sampling produces at this sample size.
Pull up the experiment page and you can see the five humps directly. The colored bars are the empirical PMFs of each order statistic. The smooth lines are the closed-form predictions. They line up the way a tuning fork lines up with itself.
Per-statistic numbers, for the curious:
| Order stat | Theoretical μ | Empirical μ | Δ |
|---|---|---|---|
| 1st smallest | 11.67 | within sampling noise | tracks |
| 2nd smallest | 23.33 | within sampling noise | tracks |
| 3rd (median) | 35.00 | within sampling noise | tracks |
| 4th smallest | 46.67 | within sampling noise | tracks |
| 5th largest | 58.33 | within sampling noise | tracks |
Every order statistic, separately, behaves the way unbiased sampling-without-replacement says it should. The convergence panel at the bottom of the page shows the deviation shrinking with N along the 1/√N curve and never escaping the envelope along the way.
Why This Is the Strongest Test on the Site
I want to be careful here, because there's a subtle point.
We have a lot of randomness tests in the Lab. The poker test bucks the five numbers by last digit and checks for hand-type collisions. The gap test measures how long each number sits out between appearances. The Fourier fit looks for periodic signal in the rolling average. They all return the same answer: no anomaly.
But they all reduce the draw to one statistic. They project a five-dimensional object onto a single number — pair count, gap length, dominant frequency — and then test whether that one number behaves. A five-dimensional bias could hide from a one-dimensional test the way a sphere hides inside its silhouette.
The limit-shape test doesn't do that. It tests every order statistic separately — five separate distributions, 345 separate bins — and asks: is the joint shape of a sorted draw, in aggregate, what fair sampling predicts? It is the closest thing on Balliqa to a multi-dimensional fingerprint test.
And the multi-dimensional fingerprint says: fair.
What I Already Knew
I want to be honest about this part, because I am every time.
I knew before I built this what the result was going to be. If Powerball draws had a non-uniform joint distribution over sorted 5-tuples, every order-statistic histogram would already be visibly skewed on our most-common page, and somebody would have written a paper about it years ago. The Vershik–Pak machinery predicts the limit shape of a uniform combinatorial family, and the modern Powerball has every reason to be that family.
But I built it anyway, for two reasons.
The first is that this experiment ties our randomness story to a real piece of 21st-century mathematics. Vershik's limit-shape theorem and Pak's asymptotic-stability framework aren't lottery-specific tricks. They're general results about random combinatorial objects, applied here to a particular family that happens to consist of Powerball draws. The conclusion has structural backing now — not "we ran a chi-square and it passed" but "the empirical limit shape matches the closed-form limit shape predicted by 150 years of partition theory."
The second is that this kind of test is falsifiable forward. If Powerball ever changes its hardware, or if some future format change introduces a subtle bias in how the white balls get sampled, this experiment will catch it before any one-dimensional test does — because a joint bias has nowhere to hide in a 345-bin histogram. The page recomputes every five minutes from the live data. I'm not holding my breath. But that's the whole point of building it as a test, not as an opinion.
A Note on the Paper
If you're inclined to read Pak, you should. It's thirty-two pages and the technical content is contained mostly in two definitions and one theorem (4.1) — the rest is a tour through Sylvester, Glaisher, Andrews, Bressoud, and friends, checking each classical bijection against the new yardstick. It is one of the cleanest examples in modern combinatorics of a problem getting solved by picking the right formalism. Forty years of "is this bijection canonical?" arguments, settled by asking the equivalent question that turns out to have a definite answer: does it survive the scaling limit?
The companion paper [18] uses the same framework to prove that no Rogers–Ramanujan bijection can be geometric — the first negative answer to a folklore problem in the area. That's the kind of move I keep finding inspiring: define a class precisely, then prove something can't belong to it. The Lab tries to do small versions of that move with the lottery. Define what randomness should produce. Run the test. Watch what passes.
See It For Yourself
The full interactive experiment is at /experiments/limit-shape. You can toggle between real draws and simulated draws side-by-side, hover the five-humps chart to see exact PMFs, and read off where the convergence curve sits relative to the noise envelope at every milestone N. Every number recomputes from the live Powerball feed.
If a future draw, or a future format change, ever produces an L∞ deviation that escapes the envelope, the page will show it. Until then, the lottery has the shape it should — and partition theory had the shape on file before the lottery ever drew its first ticket.