t-Wise Independence of AES for t > 2

Statement: AI-written, not yet formalized. Proof: open – no attempt yet.

A family of permutations is \(t\)-wise independent if its behaviour on any \(t\) distinct inputs is indistinguishable from a random permutation’s. At \(t = 2\) this already rules out differential and linear cryptanalysis, and the source proves it for real AES with real S-boxes. Higher \(t\) rules out more, and the source names it as an outstanding open problem. There is a warning attached: in the degenerate width-\(1\) case, with the same patched-inverse S-box AES uses, \(4\)-wise independence provably fails at modest round counts, by an interpolation attack. So the question is not whether AES gets there but after how many rounds — and no technique currently reaches any finite answer for \(t \ge 3\).

View PDF · LaTeX source · Formal statement — not yet formalized

The setting. An SPN with word length \(b\), width \(k\) and \(r\) rounds alternates: XOR the round key; apply an S-box \(S : \mathbb{F}_{2^b} \to \mathbb{F}_{2^b}\) to each of \(k\) blocks in parallel; apply an invertible linear mixing layer. A key-alternating cipher (KAC) is the case \(k = 1\), where the mixing step can be dropped. AES is \(k = 16\), \(b = 8\), S-box the patched inverse \(x \mapsto x^{2^b - 2}\) composed with an invertible \(\mathbb{F}_2\)-affine map, mixing layer ShiftRows then MixColumns. Round keys are independent and uniform.

A family is \(\varepsilon\)-close to \(t\)-wise independent if its output distribution on any \(t\) distinct inputs is within statistical distance \(\varepsilon\) of a uniformly random permutation’s.

Conjecture. For every \(t > 2\) and every \(\varepsilon > 0\) there is a round count \(r = r(t,\varepsilon)\) such that \(r\)-round AES with independent round keys is \(\varepsilon\)-close to \(t\)-wise independent.

How the source states it. Page 7: “the two outstanding open problems that come of this work are (a) to prove \(t\)-wise independence of multi-round AES with independent round keys, for \(t > 2\); and (b) to formalize and prove security against algebraic attacks.” Item (a) is this statement; item (b) is a programme rather than a proposition, and is not.

The quantifiers are supplied here. The source names the goal without fixing \(t\), \(\varepsilon\) or a target round count, so the structure above — for every \(t\) and \(\varepsilon\), some finite \(r\) — is this statement’s choice and the weakest reading with content. A resolution should report the dependence \(r(t,\varepsilon)\); a bound polynomial in \(t\) and \(\log(1/\varepsilon)\) would be far more useful than a merely finite one.

The nearby counterexample, and why it does not settle this. The source’s Appendix B shows that for a modest number of rounds \(r \ll 2^n\), the key-alternating cipher with permutation \(x \mapsto x^{2^n - 2}\) — the same patched inverse, but at width \(1\) over the whole \(n\)-bit block — is not \(4\)-wise independent, by essentially the interpolation attack of Jakobsen and Knudsen: the output is a rational function of the input with numerator and denominator linear in \(x\).

Two consequences. It is not a refutation: AES has width \(16\) with \(8\)-bit S-boxes and a mixing layer, and the source cites Carlitz for the fact that with \(2^{\Omega(n)}\) rounds the same KAC would be indistinguishable from a random permutation — so the width-\(1\) obstruction is about round count, not impossibility. But it is a real constraint: an argument for \(t \ge 4\) cannot be insensitive to width and mixing, since the width-\(1\) specialization is false at modest \(r\). That is a sharper obstruction than one usually gets, and it is the reason to distrust any purported proof that never uses \(k\) or \(M\).

What is proved, and the line between the two halves.

Object \(t\) Result
Concrete AES, real S-box \(2\) \(6r\)-round AES is \(2^{r-1}(0.472)^r\)-close (Theorem 3.14)
Concrete AES \(> 2\) c/0091 — open
KAC with most permutations any \(t\) almost \(t\)-wise independent in \(t + o(t)\) rounds, existentially
KAC with patched inverse, width \(1\) \(4\) not \(4\)-wise independent for \(r \ll 2^n\) (Appendix B)

The existential result fixes permutations that are not AES’s S-box and says nothing about any named cipher. The gap between it and the concrete \(t = 2\) result is this statement.

Where \(t = 3\) sits. The source’s problem is stated for \(t > 2\); the counterexample is for \(t = 4\). Nothing here rules out \(t = 3\) being materially easier, and it is the natural first target — smallest case not covered by the pairwise machinery, and below the level at which the interpolation attack is known to bite. A resolution for \(t = 3\) alone is genuine progress and should be reported as partial.

Independent round keys is load-bearing. It is explicit in the source’s problem statement and is not a technicality: real AES uses a key schedule. The working expectation is that \(t\)-wise independence becomes \(t\)-wise pseudorandomness under an appropriate schedule, and the source records that “understanding the precise role of key schedules is an important open problem.”

Sources

View PDF — no proof written yet · LaTeX source — no proof written yet · Formal proof — not yet formalized

Open, with the shape of an acceptable proof unusually constrained.

Why the \(t = 2\) machinery stops. The pairwise argument works with input differences: a single difference \(\Delta\) propagates, and the S-box extraction lemma says that when the incoming difference has enough Rényi entropy the outgoing one is near-uniform. For \(t \ge 3\) there is no single difference — there are \(\binom{t}{2}\) of them, correlated — and the object to track is a joint distribution over tuples rather than a distribution over one difference. The source’s successor paper handles the \(t\)-wise case for random S-boxes via a random walk on layout graphs, an abstraction capturing equality and inequality constraints among multiple evaluations; extending that to a concrete S-box is the visible route.

Why any proof must use width and mixing. See the Appendix B counterexample above. This is the single most useful constraint available: it rules out an entire family of arguments a priori.

Checked against the literature, 2026-08-27. No result for concrete AES at \(t > 2\) found. Targeted check on this line and its successor, not an exhaustive sweep.

Why the programme is worth pursuing without pseudorandomness. There is no hard mathematical problem underlying AES to reduce to, so the alternative to cryptanalysis is to prove that specific well-studied attack classes cannot succeed. \(t\)-wise independence is the largest such class currently within reach, and each increment in \(t\) enlarges it.

Neighbouring questions in the same discussion, none of them this. A tighter tradeoff between round count and statistical distance; a direct bound on differential probability without going through statistical distance; and item (b) above, algebraic attacks. The source also notes that a \(2^{-127}\) differential bound for a \(128\)-bit block does not rule out a distinguisher, which is why closeness parameters here have to reach \(2^{-128}\).

What a reviewer should be suspicious of, in order. First, whether the result is for concrete AES or for an SPN with random or public-random S-boxes — the latter is the large-S-box model and is a different claim. Second, whether the argument uses width and mixing, since Appendix B shows one that does not must be wrong. Third, what \(r(t,\varepsilon)\) is: a finite but astronomical bound settles the statement as posed and little else.