Take a fresh deck of cards out of the box. It’s in perfect order — spades, hearts, clubs, diamonds, each suit ace through king. Shuffle it once, the ordinary way, riffling the two halves together. It’s still obviously ordered if you look closely: you can see the ghost of the original sequence, broken into interleaved chunks. Shuffle it a second time, and a third. Something strange happens around the seventh shuffle: the deck reaches the conventional practical benchmark for randomness. An eighth shuffle substantially improves it further. Seven remains the number traditionally cited.
This isn’t a superstition or a card-table myth. It’s a theorem, proved in 1992 by mathematicians Dave Bayer and Persi Diaconis, and it’s one of the most delightful pieces of applied probability I know — because it turns a question that sounds impossibly vague (“when is something random enough?”) into a number you can compute.
What a Riffle Shuffle Actually Does
Before we can ask how many shuffles it takes to randomize a deck, we need a mathematical model of a shuffle. The standard model, called the Gilbert-Shannon-Reeds shuffle, works like this. Cut the deck into two piles, where the size of each pile is chosen according to a specific probability rule — cutting close to 26/26 is much more likely than cutting 5/47, matching how real riffles behave. Then interleave the two piles back together card by card, where at each step the probability that the next card drops from a given pile is proportional to how many cards remain in that pile.
That last rule is the crucial one. It means a real shuffle isn’t a clean, perfect interleave like a machine would produce. Cards drop in little clumps — two or three from the left pile, then one from the right, then four more from the left — because your thumb doesn’t release cards with perfect one-at-a-time regularity. Bayer and Diaconis showed that this sloppy, human, clump-dropping process has a closely related inverse description. Assign each card a random bit, 0 or 1, sort the cards by that bit while preserving their relative order within each bit-group, and you’ve simulated one inverse riffle shuffle exactly.
Counting Rising Sequences
To measure how “unshuffled” a deck still is, Bayer and Diaconis used a quantity called the number of rising sequences. A rising sequence is a maximal sequence of consecutive card labels that appear in their original order, even if they’re no longer physically adjacent. A perfectly ordered deck has exactly one rising sequence — the whole thing. A single riffle shuffle of a 52-card deck can produce at most two rising sequences, because you’ve merely interleaved two already-sorted piles.
Here’s the key fact their theorem establishes: after k riffle shuffles, the deck can have at most 2^k rising sequences. But that ceiling alone does not establish mixing. A fully random arrangement of 52 cards will typically have around 26 rising sequences, while probabilities under the GSR model depend on the rising-sequence counts of individual arrangements. Establishing randomness requires comparing the full probability distribution with the uniform one.
Although 2^6 already exceeds 26, that does not establish mixing. The exact total-variation calculation — not merely clearing the rising-sequence count — produces the conventional seven-shuffle benchmark.
The Total Variation Distance
Rising sequences tell you the mechanism; what actually settles the question is a precise measure of “how far from random” the deck’s probability distribution is. Mathematicians use something called total variation distance, which compares the actual probability of every possible arrangement of the deck to the probability it would have under a perfectly uniform shuffle, and adds up the worst-case gap.
Bayer and Diaconis computed this distance exactly for each number of shuffles. The results are striking:
- After 1 shuffle: distance is essentially 1 (completely non-random — obviously so)
- After 4 shuffles: distance is essentially 1 (still almost entirely non-random)
- After 5 shuffles: distance is about 0.924
- After 6 shuffles: distance is about 0.614
- After 7 shuffles: distance is about 0.334
- After 8 shuffles: distance is about 0.167
- After 10 shuffles: distance is about 0.043
Notice the shape: the distance stays near one through four shuffles, then drops sharply from five through eight. This rapid transition is called a cutoff phenomenon, and it’s the real mathematical reason “seven” became the answer — not because seven achieves some perfect zero, but because it lies within the narrow transition. An eighth shuffle cuts the remaining distance roughly in half.
Why the Curve Has a Knee at All
Cutoff phenomena show up all over probability theory, not just in card shuffling — they appear in models of how gossip spreads through a network, how a gas approaches equilibrium, and how random walks on complex mathematical structures settle down. The common thread is a system with an enormous number of possible states (52 cards can be arranged in 52 factorial ways, a number with 68 digits) but a mixing process that, for a long while, can’t reach most of those states at all. Then abruptly it can reach almost all of them, and shortly after that, it becomes statistically indistinguishable from having reached all of them uniformly.
Think of it like filling a swimming pool through a narrow hose: for a long time the water level barely rises because there’s so much volume to fill, then at some point it climbs fast, then it slows again as it tops off near the rim. Six shuffles is still climbing fast. Seven is within that steep transition. An eighth substantially improves the result again, while later shuffles continue reducing the remaining distance.
What Casinos Actually Do About It
This isn’t purely academic. Casinos care enormously about deck randomness, because a predictable deck is a deck that can be exploited by card counters and cheats. Standard casino procedure for blackjack typically calls for a specific shuffle count precisely because of results like Bayer and Diaconis’s, though many casinos now sidestep the whole question by using automatic shuffling machines that implement a different randomization process entirely, one that can achieve true randomness in far fewer physical passes, or by using continuous shuffling machines that never let the deck rest in a fixed order at all.
There’s a sharper practical lesson buried in all this for anyone who plays cards at home, though. Four shuffles — a very typical amount for casual play — leaves a total variation distance essentially equal to 1, which is to say the deck is still enormously predictable, still carrying detectable fingerprints of its original order. If you’ve ever noticed that home poker games seem to produce suspiciously repetitive patterns, straights and flushes clustering strangely, this is at least part of the mathematical reason why: most people simply don’t shuffle enough.
The One Number Worth Remembering
You don’t need to compute a total variation distance at the kitchen table to use this. The one clean, transferable fact is the number itself: five or six ordinary riffle shuffles leave a deck meaningfully non-random, while seven is the conventional practical benchmark for a 52-card deck and an eighth still substantially improves the result. If you deal cards for money, or just care about your Friday night poker game being fair, seven is the traditional number to count to — not because seven is magic, but because it lies within the sharp mathematical transition from “still obviously ordered” toward “genuinely anyone’s game.”
Leave a Reply