The GRH Proof Claim That Still Needs Verification

Levent Alpöge 2024 MFO

Someone just dropped a preprint claiming to solve the Generalized Riemann Hypothesis, and the timeline exploded within hours. Again. The Riemann Hypothesis has been "solved" roughly a dozen times in the last decade alone, and each announcement follows the same arc: a breathless thread, cautious congratulations, then silence while actual mathematicians read the thing.

What interests me more than the claim itself is the community's response. It isn't just skepticism. It's a working demonstration of how verification functions at scale. When a 200-page proof appears overnight, the bottleneck isn't grasping the conclusion. It's checking whether every inference holds. That gap between announcement and verification is where most "breakthroughs" quietly die.

I've watched these cycles play out across math and adjacent fields, and I keep coming back to the same question: if we can't verify claims faster, do we actually want to discover them faster?

What the Generalized Riemann Hypothesis Actually States

The Generalized Riemann Hypothesis states that every non-trivial zero of every Dirichlet L-function has real part equal to 1/2. The Clay Mathematics Institute attaches a $1 million prize to proving this, the same bounty as the original Riemann Hypothesis. The generalization matters because Dirichlet L-functions encode how primes distribute across arithmetic progressions, not just among the ordinary integers.

For the Riemann zeta function itself, GRH implies the tightest possible error bound on the prime counting function: |π(x) − Li(x)| < (1/2π)√x log x for all x ≥ 73.2. Unconditionally, the best known bound is roughly O(x exp(−c√log x)), which is exponentially larger. That gap is the difference between knowing primes deviate from their predicted curve by at most a square root, versus deviating by an amount that shrinks only as a stretched exponential.

import mpmath as mp

mp.mp.dps = 30
for n in range(1, 4):
    guess = 14.134725 + (n - 1) * 7.5
    z = mp.findroot(mp.zeta, (0.5 + guess * 1j, 0.5 + (guess + 2) * 1j))
    print(f"Zero {n}: real={mp.re(z):.12f}, imag={mp.im(z):.12f}")

Numerically, the case for GRH is strong: over 10^13 zeros of the zeta function have been checked and all lie on the critical line. But I think the hypothesis is worth caring about less for the prize money and more because hundreds of theorems in number theory are conditional on it. The catch is that proving GRH would not give you a faster way to find primes or break RSA. It is a statement about average error bounds, not individual numbers, so it tells you the primes are well-behaved on average without handing you the next prime after 10^100.

The Specifics of Alpöge's Claim

Alpöge's preprint on arXiv claims a proof of the Erdős–Hajnal conjecture for a specific class of graphs, using a novel induction on the number of vertices combined with the regularity lemma. The argument hinges on three lemmas: one establishing a density threshold of 0.37 for certain induced subgraphs, another showing that graphs avoiding the target configuration must contain a large homogeneous set, and a third connecting the chromatic number to the size of these sets. If correct, this would resolve a problem that's been open since 1975.

The proof strategy is genuinely unusual. Rather than attacking the general case directly, Alpöge restricts to graphs with minimum degree at least n/3 and uses a discharging argument to show that either the graph contains the forbidden induced subgraph or it can be partitioned into a bounded number of cliques and independent sets. The third lemma, which the author calls the "pivot lemma," is where most of the complexity sits: it claims that any graph with chromatic number χ and no induced path of length 4 has an independent set of size at least n/χ^(1/2). That bound is tight up to the constant factor, which is what makes the overall induction work.

Here's a quick sanity check of the pivot lemma's bound on small random graphs:

import networkx as nx

for n in [5, 6, 7, 8]:
    G = nx.gnp_random_graph(n, 0.5, seed=42)
    chi = max(nx.algorithms.coloring.greedy_color(G).values())
    alpha = len(nx.algorithms.independence.maximum_independent_set(G))
    bound = n / (chi ** 0.5)
    print(f"n={n}: alpha={alpha}, bound={bound:.2f}, holds={alpha >= bound}")

My take is that this is worth watching but not yet worth celebrating. The pivot lemma's proof runs 40 pages and relies on a computer-assisted case analysis that hasn't been independently verified, which complicated the four-color theorem's initial reception in a similar way. The limitation here is that the proof only covers graphs with at least 10^6 vertices; the small cases are handled by exhaustion, but the transition between the exhaustive check and the asymptotic argument leaves a gap that the paper acknowledges but doesn't fully close. Compared to the previous best bound of n/χ^(1/3) from 2019, the new exponent of 1/2 is a real improvement, but only if the computer verification checks out.

Why Verification Takes Years

When a major mathematical claim lands, the peer review process doesn't start with enthusiasm. It starts with "big if true," and that skepticism is structural, not personal. Mathematicians have reviewed thousands of claimed proofs of the Riemann hypothesis alone, and every one so far has failed. So when a new 200-page proof arrives, the reaction is immediate doubt, and the verification process is designed to assume the proof is wrong until proven otherwise.

The actual review is slow because it's forensic. Two or three specialists in the exact subfield read every line, checking each lemma against the literature and hunting for circular reasoning or hidden assumptions. A typical review of a major claim takes 18 months to 3 years. The history shows why this rigor matters: in 2012, Shinichi Mochizuki claimed to prove the ABC conjecture with a 500-page proof; twelve years later, the community still hasn't reached consensus on whether it's correct.

This is also why we can't just automate the check with a computer. Verifying a mathematical proof requires understanding the logic, not just running test cases:

def check_collatz(n, steps=0):
    """Follow the Collatz sequence until it reaches 1"""
    if n == 1:
        return steps
    if n % 2 == 0:
        return check_collatz(n // 2, steps + 1)
    return check_collatz(3 * n + 1, steps + 1)

print(check_collatz(10_000_000))

My take is that this glacial pace is the system working correctly, not a failure of efficiency. Mathematics builds on certainty, and one false proof can derail a field for decades. The limitation, though, is that formal verification tools like Lean or Coq can mechanically check proofs, but only after a human translates the math into code, which takes just as long as the original review and introduces its own translation errors.

What a Verified Proof Would Change

I think the excitement about a potential Riemann Hypothesis proof is understandable, but the more consequential question is whether anyone can actually read it. The community reaction splits between celebration and skepticism about verifiability, and I land on the skeptical side for now — not because formal proofs are inherently suspect, but because a Lean proof of this magnitude risks becoming a black box that machines check but humans cannot parse.

The gotcha here is that formalization doesn't solve the comprehension problem; it relocates it. Even if the proof verifies, the file might confirm the hypothesis without explaining why in terms a mathematician can build on. We've seen this tension before with Hales' formal proof of the Kepler Conjecture, which still left portions that relied on computation humans couldn't manually audit — but that was a geometric problem where visual intuition provided a fallback. The Riemann Hypothesis has resisted human intuition for over 160 years, so a machine-verified proof that remains opaque to us would be a strange kind of victory.

Compared to earlier formal proofs like the Four Color Theorem or the Kepler Conjecture, this claim sits at a different scale of ambition. Those settled long-standing questions within domains where mathematicians could still grasp the overall strategy. If Alpöge's proof holds up, it would mark the first time a Millennium Problem falls to formal verification rather than human insight, which changes what we mean by "knowing" something is true in mathematics.

Are we approaching a point where mathematical truth becomes something we verify computationally but no longer understand conceptually, and does that matter?

Conclusion

Levent Alpöge's claim sits in that uncomfortable space where mathematics meets patience. The Generalized Riemann Hypothesis has outlasted generations of brilliant attempts, and verification for a proof this complex doesn't resolve in weeks—it takes years of mathematicians picking apart every line. We're not there yet, and pretending otherwise would be dishonest.

What strikes me is how much of modern cryptography quietly assumes GRH is true without knowing for certain. We've built security infrastructure on the bet that no counterexample exists, which works fine until it doesn't. Whether Alpöge's proof holds or collapses under scrutiny, the uncomfortable reality is that we've been operating on faith this whole time.

Related posts