I. Introduction: What a Proof of RH Would Have to Look Like

After more than a century and a half of effort, the Riemann hypothesis has accumulated a substantial dossier of attempted proofs, partial results, and structural frameworks. None of these has succeeded in establishing the hypothesis. But each has illuminated, in its own way, what a successful proof would require. Taken collectively, the accumulated body of work imposes several structural constraints on any prospective proof — constraints that any new strategy must satisfy if it is to have any chance of succeeding.

The first constraint is that the proof must distinguish the critical line from neighboring lines. The Riemann hypothesis is a one-codimension assertion: it claims that the nontrivial zeros lie on a particular line within the two-dimensional critical strip. Any method that produces, as its output, only that the zeros lie within some open region — even a very narrow region around the critical line — is structurally insufficient. The method must somehow detect the line itself, not merely a neighborhood of it. This rules out, on present understanding, approaches that proceed by progressive narrowing of zero-free regions: the asymptotic shrinkage of such regions toward the critical line is a project of infinite depth that cannot be completed in finitely many steps.

The second constraint is that the proof must use information specific to ζ that is absent from arbitrary L-functions in broader classes lacking the necessary structural features. There exist L-functions in extended classes — for instance, Davenport–Heilbronn-style functions, or certain combinations of L-functions in a wider Selberg class — that satisfy partial analogs of the structural conditions defining the Selberg class but for which the Riemann hypothesis is known to fail. The Davenport–Heilbronn function, in particular, has all the analytic features one might naively associate with an “L-function” except for the Euler product, and it has zeros off the critical line. Any proof of RH that did not use the Euler product, or some equivalent multiplicative structure, would prove something false. The proof must be sensitive to features of ζ that are not shared by all functions satisfying its analytic shape.

The third constraint is that the proof must explain the function field success or differ from it deliberately. The Weil and Deligne proofs of the function field Riemann hypothesis succeeded by means of finite-dimensional cohomology, geometric structure, and positivity. Any proof of RH over Q must either supply analogs of these features for the integers (which is what the F_1, Arakelov, and Connes programs attempt), or must succeed by genuinely different means whose absence in the function field case can be accounted for. A proof that proceeded by analogy with Weil but did not produce the requisite geometry would be incoherent. A proof that proceeded by methods unrelated to the function field case would owe an explanation of why those methods do not also yield the function field result, or alternatively, why the function field proof’s methods are not the only path to a Riemann hypothesis-style theorem.

These three constraints, taken together, narrow the space of plausible proof strategies considerably. The remainder of this paper surveys the strategies that have been seriously developed, examines what each has achieved and where each has stalled, and considers the meta-question of whether the accumulated obstructions point toward any particular path forward.

II. The Hilbert–Pólya Program

Origins of the Spectral Interpretation

In the early twentieth century, David Hilbert and George Pólya independently suggested — neither in print, but in letters and conversations later reported — that the imaginary parts of the nontrivial zeros of ζ might be the eigenvalues of some self-adjoint operator on a Hilbert space. The suggestion was speculative; neither Hilbert nor Pólya offered a candidate operator. But the suggestion has been enormously generative. The Hilbert–Pólya program, as it is now called, is the program of finding such an operator and using its self-adjointness to establish the Riemann hypothesis.

The program rests on a single observation that, if it could be realized, would close the proof immediately. If H is a self-adjoint operator on a Hilbert space with real eigenvalues λ_n, and if there is a natural way to associate to ζ a function whose nontrivial zeros are precisely 1/2 + iλ_n for these eigenvalues, then the eigenvalues being real is equivalent to the zeros lying on the critical line. The Riemann hypothesis would follow from the spectral theorem.

The challenge of the program is that the existence of such an operator is far from automatic. Constructing a self-adjoint operator whose eigenvalues match a given sequence is, in general, easy: one can simply define a diagonal operator with those eigenvalues. But this trivial construction does not connect the operator to ζ in any meaningful way. What the program requires is an operator that arises naturally from the arithmetic of the integers, in such a way that its spectrum is related to ζ by something more substantial than fiat.

Early Candidates and Their Limitations

For most of the twentieth century, no convincing candidate for the Hilbert–Pólya operator was identified. Various spectral interpretations of zeta values were noted — for instance, the trace formula on hyperbolic surfaces, which produces a Selberg zeta function whose zeros are known to lie on the critical line — but these did not yield ζ itself.

The case of the Selberg zeta function is instructive. For a quotient X = Γ\H of the upper half-plane by a Fuchsian group Γ, the Laplace–Beltrami operator on X is self-adjoint, with spectrum consisting of nonnegative real eigenvalues. The Selberg zeta function Z_Γ(s) has zeros at points related to these eigenvalues, and these zeros lie on the critical line. The Selberg zeta function thus satisfies its own Riemann hypothesis as a consequence of the self-adjointness of the Laplacian — a clean realization of the Hilbert–Pólya idea in a setting where the relevant operator exists naturally.

The trouble is that the Selberg zeta function is not the Riemann zeta function. The two share certain analytic features — both have functional equations, both have Euler-product-like representations — but they are different functions, attached to different geometric and arithmetic objects. The success of the Hilbert–Pólya idea for Z_Γ does not transfer to ζ, and the search for the operator that does the job for ζ has remained open.

The Berry–Keating Conjecture

In 1999, Michael Berry and Jonathan Keating proposed a candidate for the Hilbert–Pólya operator, motivated by considerations from quantum chaos. The proposed operator is essentially

H = (xp + px)/2,

where x is the position operator, p is the momentum operator, and the symmetrized product accounts for the noncommutativity of x and p. This operator has been studied in mathematical physics under the name of the Berry–Keating Hamiltonian.

The Berry–Keating proposal is suggestive but incomplete. The operator as written is not self-adjoint on the natural Hilbert space L²(R); it requires careful boundary conditions, and the choice of boundary conditions is precisely what determines the spectrum. The conjecture, in its full form, is that there exists a choice of self-adjoint extension whose spectrum yields the imaginary parts of the zeros of ζ. Establishing this — finding the right boundary conditions, proving that the resulting spectrum has the predicted form — has not been accomplished.

Subsequent work, including by Connes (treated separately below), has extended the Berry–Keating idea in various directions. The general intuition is that the operator should be a quantization of a classical dynamical system whose periodic orbits correspond to prime numbers. The classical system in question — a particle moving on a configuration space related to the integers, with a Hamiltonian generating dilation — is well defined heuristically. Its rigorous quantization, in a way that produces ζ-zeros as eigenvalues, has not been established.

The Status of the Program

The Hilbert–Pólya program has produced substantial mathematics. It has motivated the development of random matrix theory in the context of L-functions (treated in the next section). It has connected number theory to mathematical physics. It has produced candidate operators that, even if not the eventual answer, have illuminated the structural features that any successful operator must possess.

The program has not produced a proof. The principal obstacle is that constructing a self-adjoint operator whose spectrum captures ζ is, on present evidence, of the same order of difficulty as proving RH directly. Each candidate operator either fails to be self-adjoint, or fails to have the predicted spectrum, or is defined in a way that requires the Riemann hypothesis as input rather than producing it as output. The program is genuinely promising as an organizing framework, but its realization in concrete form has remained elusive.

III. Random Matrix Theory and the Montgomery–Odlyzko Law

Montgomery’s Pair Correlation Conjecture

In 1973, Hugh Montgomery, then a graduate student, was studying the statistical distribution of zeros of ζ. Assuming RH, he wrote each nontrivial zero as 1/2 + iγ_n with γ_n real, and considered the statistics of the spacings between consecutive γ_n. Montgomery conjectured that the pair correlation of these zeros — the average density of pairs (γ_m, γ_n) with the difference γ_m − γ_n falling in a given interval — has a specific form:

F(α) = 1 − (sin(πα)/(πα))²

for the appropriately normalized correlation function F.

Montgomery proved his conjecture in a restricted range of test functions, contingent on RH, and conjectured the full form. He took the conjecture to Hugh Odlyzko, who had access to substantial computational resources, and Odlyzko verified the prediction numerically with remarkable accuracy at very large heights.

The Encounter with Dyson

The story of Montgomery’s discovery has become canonical. Montgomery, while visiting the Institute for Advanced Study in Princeton in the early 1970s, was introduced to Freeman Dyson at tea. When Montgomery described his conjectured pair correlation function, Dyson recognized it immediately as the pair correlation function for eigenvalues of large random Hermitian matrices drawn from the Gaussian Unitary Ensemble (GUE). The connection between the zeros of ζ and the eigenvalues of random matrices was thus established by chance encounter, and the resulting framework — that the local statistics of ζ-zeros match those of GUE eigenvalues — has come to be called the Montgomery–Odlyzko law.

The connection is, in retrospect, structurally suggestive. Random matrix theory had been developed in the 1950s and 1960s by Wigner, Dyson, and others to model the energy levels of large complex quantum systems (originally heavy atomic nuclei). The empirical observation was that the spacings of energy levels in such systems followed universal statistics determined by the symmetries of the Hamiltonian: GUE for systems with no special symmetry, GOE (Gaussian Orthogonal Ensemble) for systems with time-reversal symmetry, GSE (Gaussian Symplectic Ensemble) for certain other symmetry classes.

The fact that the zeros of ζ follow GUE statistics is, on the Hilbert–Pólya picture, exactly what one would expect: if the zeros are eigenvalues of some self-adjoint operator, and if that operator has no special symmetries, then GUE is the predicted statistics. The Montgomery–Odlyzko law thus provides indirect evidence for the Hilbert–Pólya program — evidence that whatever operator might eventually be found, its symmetry class is GUE.

Odlyzko’s Verifications

Andrew Odlyzko’s numerical work over several decades has verified the Montgomery–Odlyzko law to extraordinary precision. Computing zeros of ζ at heights around the 10^{20}-th zero, Odlyzko established that the pair correlation function, as well as higher correlation functions and the nearest-neighbor spacing distribution, agree with GUE predictions to within statistical error.

The numerical agreement is among the most striking pieces of evidence in mathematics. It is not merely that ζ-zeros lie on the critical line up to high heights; it is that, conditional on lying there, they distribute themselves with statistics that match a precise theoretical prediction from a quite different domain (random matrix theory) to multiple decimal places. The agreement is too detailed to be coincidence and too quantitative to be vague. It encodes some deep structural fact about ζ that has not yet been articulated in proof form.

The Keating–Snaith Conjectures

Building on the Montgomery–Odlyzko framework, Jonathan Keating and Nina Snaith in 2000 proposed conjectures for the moments of |ζ(1/2 + it)|. Specifically, they conjectured that

(1/T) ∫_0^T |ζ(1/2 + it)|^{2k} dt ~ a_k g_k (log T)^{k²}

where a_k is an arithmetic constant computed from an Euler product over primes, and g_k is a “geometric” constant computed from random matrix theory — specifically, from moments of characteristic polynomials of matrices in the Circular Unitary Ensemble (CUE).

The Keating–Snaith conjectures generalize earlier results of Hardy–Littlewood (k = 1) and Ingham (k = 2), where the constants were determined by direct calculation. For k > 2, the constants had been mysterious, and Keating–Snaith provided a coherent prediction by importing random matrix moments into the picture. The conjectures have been checked numerically with high accuracy, and they have generated substantial subsequent work on moments of L-functions in families.

What Random Matrix Theory Constrains, and What It Does Not

The random matrix framework constrains the predictions of RH in a specific way: if the hypothesis is true, then the zeros must distribute themselves according to GUE statistics, and any proof of RH should, in principle, be compatible with this distribution. The framework also generates conditional predictions — about moments, about spacings, about extreme values of ζ on the critical line — that go beyond what RH alone implies.

What random matrix theory does not do is prove RH. The framework assumes RH as input (one cannot speak of statistics of imaginary parts of zeros if one does not know they are real) and produces predictions about the statistical distribution of those imaginary parts. The framework is descriptive of the conjectured world, not constructive of a proof.

The deeper question of whether random matrix theory could be used to prove RH — by, say, identifying ζ with a specific random matrix whose eigenvalues are then constrained to be real — is structurally analogous to the Hilbert–Pólya program. Both programs require constructing a specific operator (or matrix) whose spectrum is the zeros, and both face the same fundamental obstacle: such constructions, when they can be carried out at all, tend to require RH as input rather than producing it as output.

IV. Connes’ Noncommutative Geometric Approach

The Adèle Class Space

Alain Connes has developed, over several decades, an approach to RH through noncommutative geometry. The framework’s central object is the adèle class space of Q. The adèles A_Q are the restricted product of all completions of Q (the real numbers and the p-adic numbers for each prime p), with appropriate restrictions on which factors can be unbounded. The multiplicative group Q* acts on A_Q, and the adèle class space is the quotient X_Q = A_Q/Q*.

This quotient is a noncommutative space in the sense of Connes’s noncommutative geometry: the action of Q* on A_Q is not free, and the quotient does not exist as a reasonable point-set space. But noncommutative geometry provides tools — operator algebras, cyclic cohomology, spectral triples — that allow one to do geometry on such quotients, treating the noncommutative structure of the algebra of functions as the substitute for the missing point-set structure.

The Spectral Realization

The Connes framework constructs, on the adèle class space, a flow whose “periodic orbits” correspond to prime numbers. More precisely, the action of the idele class group, which contains Q* and acts on A_Q with the quotient action descending to X_Q, has a natural decomposition whose components are indexed by primes.

Connes’s central conjecture in this framework is that there exists a spectral realization of the zeros of ζ as eigenvalues of an operator constructed from this flow. The operator is a regularized version of the dilation generator on the adèle class space. The regularization is necessary because the naive operator is unbounded and not directly analyzable; producing a self-adjoint extension whose spectrum is the imaginary parts of zeros is the central technical task.

The Trace Formula and Its Positivity

A trace formula of Selberg type, applied to the dilation operator on the adèle class space, would give a relation between primes (the periodic orbits) and the spectrum (the zeros). Connes has formulated a precise version of this trace formula and has shown that, if a certain positivity statement holds, the spectrum lies on the real line — which would be equivalent to RH.

The positivity statement in question is, structurally, the analog of the Weil positivity in the function field setting. In Weil’s proof for curves, the positivity of certain intersection numbers on the surface C × C, supplied by the Hodge index theorem, constrains the Frobenius eigenvalues to lie on the critical line. In Connes’s framework, the analogous positivity would constrain the spectrum of the dilation operator. The question of whether this positivity holds is, on Connes’s framing, the question of whether RH is true.

The Status of the Program

The Connes program has produced a substantial body of mathematics. It has connected number theory to noncommutative geometry, dynamical systems, and mathematical physics. It has formulated the Riemann hypothesis as a positivity question in a precise framework, parallel in form to the function field case. It has identified specific structures whose existence would yield a proof.

The program has not produced a proof. The positivity statement that would close the proof has not been established, and there is no current strategy for establishing it that does not amount to assuming RH directly. The framework is, in this sense, a translation of RH into a different language rather than a route to its proof. The translation is illuminating — it shows what kind of object RH is — but the underlying problem has not yielded.

The Connes program also faces specific technical objections. The adèle class space is mathematically delicate, and certain steps in the construction have been the subject of debate. The relation between Connes’s spectral framework and the established Hilbert–Pólya program is not entirely settled. Whether the program represents the right framework for an eventual proof, or whether it captures an essential structure that will be incorporated into a different proof, is open.

V. The Function Field Analog as a Template

What Weil and Deligne Achieved

The function field Riemann hypothesis, treated in detail in Paper 2, was proved by Weil for curves in 1948 and extended by Deligne to general smooth projective varieties over finite fields in 1974. The proofs use, in essential ways, three structural ingredients:

1. A geometric setting: the variety X over F_q is a concrete geometric object, with subvarieties, products, fibrations, and Lefschetz pencils available as tools.
2. A finite-dimensional cohomology: étale cohomology H^i_{ét}(X, Q_l) is a finite-dimensional Q_l-vector space, and the Frobenius operator acts on this finite-dimensional space.
3. A positivity statement: in Weil’s original proof for curves, the Hodge index theorem on the surface C × C; in Deligne’s general proof, the weight filtration and monodromy arguments that give the analogous constraint on Frobenius eigenvalues.

The combination of these three features produces the Riemann hypothesis as a consequence: the eigenvalues of Frobenius on the i-th cohomology group have absolute value q^{i/2}, which translates directly to the zeros of the zeta function lying on the critical line.

Why Direct Translation Fails

The natural strategy for proving RH over Q would be to find analogs of these three ingredients for the integers. The strategy faces a structural obstacle at each step.

For the geometric setting, Spec(Z) does not present itself naturally as a smooth projective variety. It is a one-dimensional scheme, but it is not complete: the Archimedean place is missing. Arakelov theory supplies a partial replacement (treated in Paper 2), but the resulting object is not a variety in the ordinary sense, and the geometric tools available on it are limited.

For the cohomology, no finite-dimensional cohomology of Spec(Z) is known whose Frobenius eigenvalues are the imaginary parts of zeros of ζ. The classical cohomology theories (Betti, étale, de Rham, crystalline) all give finite-dimensional groups for varieties over fields, but Spec(Z) is not a variety over a field in the relevant sense, and the analogous groups for it are either trivial or not directly connected to ζ.

For the positivity, no analog of the Hodge index theorem for arithmetic schemes is known that would constrain ζ-zeros. Arakelov theory has its own intersection theory and its own Hodge-index-style theorems (developed by Faltings, Gillet–Soulé, and others), but these have not been connected to ζ in a way that would produce RH.

The structural reason these obstacles exist is that the function field setting works because F_q is a field — a self-contained algebraic object — and varieties over F_q form a category in which the standard tools of algebraic geometry apply directly. Z is not a field; it is a ring with a unique nontrivial structure (the integers) that does not fit the variety-over-a-field template. The “field with one element” program, treated in Paper 2 and revisited briefly below, attempts to construct a base field over which Z sits as a curve. The Connes program attempts to construct a noncommutative replacement. Neither has succeeded.

The Lesson of the Template

The function field analog provides the clearest extant template for what a proof of RH should look like. It also provides the clearest extant evidence that the proof, when it comes, will require structures over Z that are not yet known. The disparity between the proved function field case and the unproved number field case is, at present, the most informative single fact about the difficulty of RH.

This has a practical consequence for evaluating proof strategies. Any proposed proof of RH can be tested, in part, by asking whether its methods would also yield the function field case. If they would, and the function field case has independent proofs, the consistency is a positive sign. If they would not, the proposer owes an explanation of why the method is specific to Q. If they would yield the function field case but by a route that contradicts the existing Weil/Deligne proofs, something is wrong. This kind of triangulation has, over the years, helped to identify errors in announced proofs.

VI. Selberg Trace Formula and Spectral Approaches

The Selberg Zeta Function

For a Fuchsian group Γ acting on the upper half-plane H, with quotient X = Γ\H of finite hyperbolic volume, the Selberg zeta function is defined as

Z_Γ(s) = ∏γ ∏{k=0}^∞ (1 − e^{−(s+k) ℓ(γ)}),

where the outer product is over primitive closed geodesics γ on X and ℓ(γ) is the length of γ. The function Z_Γ(s) is an entire function with zeros related to the spectrum of the Laplace–Beltrami operator on X.

Specifically, Z_Γ has trivial zeros at certain negative integers and “nontrivial” zeros at points 1/2 ± i r_n, where r_n are determined by the eigenvalues λ_n = 1/4 + r_n² of the Laplacian on X. Since the Laplacian is a self-adjoint nonnegative operator, the eigenvalues λ_n are real and nonnegative, which forces r_n to be real or pure imaginary. The pure imaginary case corresponds to “small eigenvalues” λ_n < 1/4 and gives zeros off the critical line. Generically, however — for groups Γ without small eigenvalues — all r_n are real, and all nontrivial zeros of Z_Γ lie on the critical line Re(s) = 1/2.

The Selberg zeta function thus satisfies its own Riemann hypothesis as a direct consequence of the self-adjointness of the Laplacian. This is the cleanest realization of the Hilbert–Pólya idea.

The Selberg Trace Formula

The relation between geodesics (the primes of this setting) and eigenvalues (the zeros of Z_Γ) is given by the Selberg trace formula, an identity relating spectral data to geometric data on X. The trace formula is structurally analogous to the explicit formula relating ζ-zeros to primes, and it can be regarded as a vast generalization of the Poisson summation formula.

In its general form, the trace formula equates a sum over the spectrum of the Laplacian (or a related operator) to a sum over conjugacy classes in Γ, with explicit weights on each side. The spectral side encodes eigenvalues; the geometric side encodes lengths of closed geodesics. The two sides are equal as distributions, and equating them in various ways produces identities that have been deeply exploited in number theory.

Why This Doesn’t Transfer to ζ

The Selberg setting works because the Laplacian on a hyperbolic quotient is a canonical, self-adjoint operator. The space X is a natural geometric object; the Laplacian is its natural differential operator; self-adjointness is automatic from the Riemannian structure. The Riemann hypothesis for Z_Γ follows from this canonical structure without further input.

For ζ, no canonical operator is known. Constructing one — finding the operator whose eigenvalues are the imaginary parts of ζ-zeros — is precisely the unsolved Hilbert–Pólya problem. The Selberg setting shows that if such an operator could be found, the proof would be straightforward; it does not provide a method for finding the operator.

There are nonetheless suggestive analogies. The Selberg trace formula is structurally parallel to the explicit formula. The geodesics on X play the role of primes. The eigenvalues of the Laplacian play the role of imaginary parts of zeros. These analogies have motivated substantial work — by Sarnak, Iwaniec, and others — on connections between automorphic forms and L-functions, and on whether some automorphic setting might supply the canonical operator for ζ.

The current state of this work is that automorphic L-functions have their own conjectured Riemann hypotheses (the Grand Riemann Hypothesis for the Selberg class), and that proving these hypotheses appears, on present evidence, to be at least as difficult as proving RH for ζ itself. The Selberg setting illuminates the structure of the problem but has not opened a route to its solution.

VII. de Branges’ Attempted Proofs

Hilbert Spaces of Entire Functions

Louis de Branges, a mathematician at Purdue University who in the 1980s gave the proof of the Bieberbach conjecture in complex analysis, has developed over several decades a substantial theory of Hilbert spaces of entire functions. The theory generalizes classical Hardy space theory and provides a framework in which entire functions of restricted growth can be analyzed through reproducing kernel methods.

The basic objects are spaces H(E) defined by a Hermite–Biehler entire function E(z), consisting of entire functions f such that f/E and f*/E are both bounded in a half-plane in an appropriate L² sense. These spaces have natural inner products, reproducing kernels, and orthonormal bases, and they admit a structure theorem due to de Branges that relates them to canonical systems of differential equations.

de Branges has produced multiple announced proofs of RH using this theory, in various forms over the years. The general strategy is to exhibit a Hilbert space of entire functions in which ζ (or a function closely related to ζ) appears, with structural properties that force its zeros onto the critical line.

The Strategy and the 1998 Objection

The most-discussed version of de Branges’s strategy proceeds by associating to ζ a specific space of entire functions and showing that a certain positivity condition — analogous in spirit to the Weil positivity and the Connes positivity — holds in that space. If the positivity holds, the zeros of ζ are constrained to the critical line.

In 1998, J. Brian Conrey and Xian-Jin Li examined a specific form of de Branges’s strategy in detail. They showed that the positivity condition required by that specific strategy is in fact violated. The argument was technical but conclusive for the form of the strategy at issue: the positivity that de Branges was assuming did not hold, and so that particular path to RH was closed.

de Branges has continued to refine and modify his approach since then, posting updated manuscripts on his university webpage. The wider mathematical community has not accepted any version as a proof. Reviewing successive versions of the manuscripts, identifying specific points where they fail, and engaging with de Branges’s responses to objections has been a substantial task that experts in analytic number theory have undertaken intermittently over the past two decades.

What Survives

The de Branges theory of Hilbert spaces of entire functions, considered apart from its application to RH, is a genuine and useful body of mathematics. It has applications in spectral theory of differential operators, in moment problems, in the theory of de Branges modules over Schrödinger operators, and elsewhere. Whether or not the Riemann hypothesis ever yields to a method along these lines, the theoretical framework has standing as mathematics.

The case of de Branges’s RH attempts illustrates several features common to long-running attempts on the hypothesis. It illustrates that even mathematicians of substantial accomplishment can persist for decades on a strategy that has been shown not to work in identifiable forms. It illustrates the difficulty of definitively closing a proof attempt: each new manuscript can be modified, and each new modification requires expert engagement. It illustrates the toll on attention: each iteration consumes the time of those qualified to evaluate it. And it illustrates the gravitational pull that the hypothesis exerts on careers: the prestige of a successful proof is large enough that mathematicians who have devoted years to a strategy may continue beyond the point where outside observers would conclude the strategy is unworkable.

These observations are sociological rather than mathematical. They do not establish that no proof along de Branges’s lines can ever be found. They do, however, suggest that any new announcement should be evaluated on its specific technical merits rather than on the reputation of its author or the elegance of its general framework.

VIII. Density Theorems and Zero-Free Region Strategies

Vinogradov–Korobov Bounds

Unconditionally, the strongest known zero-free region for ζ is due to I. M. Vinogradov and N. M. Korobov, working independently in the late 1950s. Their result is that ζ(s) ≠ 0 in the region

Re(s) > 1 − c/((log |t|)^{2/3} (log log |t|)^{1/3})

for sufficiently large |t|, where c is a positive constant. This zero-free region is wider than the de la Vallée Poussin region (which had log |t| in the denominator rather than the (2/3)-power expression), and it yields a corresponding improvement in the error term of the prime number theorem:

π(x) = Li(x) + O(x exp(−c (log x)^{3/5} (log log x)^{−1/5})).

The Vinogradov–Korobov method is based on a sophisticated analysis of exponential sums, using mean value estimates due to Vinogradov. The bounds it produces have been refined incrementally over the decades, but the basic 2/3-power has not been improved unconditionally. Improving it to a 1/2-power, or equivalently producing the bound that would follow from RH, would represent a substantial structural advance.

Density Estimates

A complementary line of attack proceeds not by ruling out zeros in a region, but by bounding the number of zeros that can lie in a region. Define N(σ, T) to be the number of nontrivial zeros ρ = β + iγ of ζ with β ≥ σ and 0 < γ ≤ T. Under RH, N(σ, T) = 0 for σ > 1/2; unconditionally, one wants upper bounds.

The density hypothesis asserts that N(σ, T) = O(T^{2(1−σ) + ε}) for every ε > 0 and every σ ≥ 1/2. This is a weaker statement than RH but stronger than what is currently known unconditionally. Various conditional and unconditional density estimates have been proved, with improvements due to Ingham, Huxley, Heath-Brown, Bourgain, and others.

One striking application is the case σ = 1: the assertion that N(1 − δ, T) is small implies, by partial summation arguments, sharp results about primes in short intervals. Specifically, Huxley’s density estimate implies that the number of primes between x and x + x^{7/12 + ε} is asymptotic to the expected count, an unconditional result that would follow from RH with a substantially better exponent.

The “Proof by Attrition” Critique

A natural question is whether the cumulative effect of density estimates, combined with progressive narrowing of zero-free regions, could eventually amount to a proof of RH. The answer, on present understanding, is no.

The structural reason is that RH is an assertion about a line of measure zero in the critical strip. Any method that produces, at each stage, a bound on N(σ, T) for σ bounded away from 1/2 — or a zero-free region that does not reach the critical line itself — does not, in any number of iterations, prove RH. The asymptotic shrinkage of zero-free regions toward the critical line is a project of infinite depth, and finitely many improvements never close it.

This is a structural rather than a contingent constraint. A proof of RH must, at some essential step, detect the line itself — must use information that distinguishes the line Re(s) = 1/2 from neighboring lines. Methods that proceed by uniform improvements over open regions cannot do this. They can prove statements arbitrarily close to RH, but they cannot prove RH.

What Density Theorems Can Do

Density theorems, despite this limitation, have substantial unconditional consequences. They yield results about primes in short intervals, distribution of primes in arithmetic progressions, and other questions that would follow from RH but are weaker. They provide rigorous benchmarks against which RH-based predictions can be compared.

They also provide a kind of negative evidence for RH, in the following sense. If RH were false, with some zeros having β > 1/2, then the density of such zeros would have to be consistent with the unconditional density estimates. As those estimates have been improved, the room for off-line zeros has shrunk. The density of any potential off-line zeros is now constrained to be very small. This does not prove RH, but it does mean that any failure of RH would have to take a specific and limited form — there could be at most a sparse set of off-line zeros, satisfying explicit upper bounds on their count.

The density theorem program is, in this sense, a parallel approach to the question. It does not aim to prove RH directly; it aims to determine, as completely as possible, what the structure of zeros could be without RH. The results of this program inform what RH would have to add and constrain alternative scenarios in which RH might fail.

IX. Probabilistic and Heuristic Considerations

The Cramér Random Model

Harald Cramér in 1936 proposed a probabilistic model of the primes that has come to be called the Cramér model. The model treats the primes as if they were a random sequence: the integer n is “prime” with probability 1/log n, independently for each n. Under this model, one can compute expected values and variances of various functions of the primes, and these expectations often match the conjectured behavior of the actual primes with reasonable accuracy.

The Cramér model predicts, for instance, that the largest gap between consecutive primes near x is asymptotically (log x)², and that primes in short intervals [x, x + h] for h = (log x)² should be Poisson distributed. These predictions are not quite accurate — the actual primes have additional structure that the Cramér model does not capture — but they are accurate to leading order in many cases.

For the Riemann hypothesis specifically, the Cramér model predicts that the deviation of π(x) from Li(x) should be of order √x log log log x almost surely (in the Cramér probability space). RH itself implies a deviation of at most O(√x log² x), a slightly weaker bound. The two predictions are consistent: RH is roughly what one would expect if the primes were a Cramér-random sequence with appropriate corrections.

The Cramér model thus provides a heuristic argument for RH: under a reasonable probabilistic model of how primes might be expected to distribute, RH-type bounds emerge naturally. This is evidence for the truthfulness of RH, but it is not a proof. The Cramér model does not capture all of the structure of the primes — for instance, it does not capture the parity of integers (every prime greater than two is odd, a deterministic constraint that the model does not respect) — and refining the model to capture more structure has been an active area of research (Granville, Soundararajan, and others).

Heuristic Evidence for RH

Beyond the Cramér model, there are various heuristic arguments that support RH. The Montgomery–Odlyzko law, treated above, is one. The agreement of Keating–Snaith moment predictions with numerical computation is another. The function field analog, where the corresponding hypothesis is proved, provides a structural argument. The wide network of conditional consequences of RH, including many that have been established unconditionally and would follow more sharply under RH, provides indirect evidence — RH would, if true, predict outcomes that are consistent with what is observed.

This accumulation of heuristic and structural evidence is, on any reasonable assessment, strong. It is not a proof, but it explains why most experts in the field treat RH as essentially certain to be true. The question is not whether RH is true (the working assumption is that it is); the question is what proof method will eventually establish it.

The Status of Heuristic Arguments

Heuristic arguments have a complicated standing in mathematics. They are not proofs and cannot substitute for proofs. They can be wrong: there are many examples of conjectures supported by extensive heuristic and numerical evidence that turned out to be false (Skewes’s bound on the first sign change of π(x) − Li(x), discussed in Paper 1, is an example of a phenomenon that defied long-standing heuristic expectations).

But heuristics can also be substantially right. They can identify the correct structural framework, predict the correct quantitative form of the answer, and guide proof attempts in productive directions. The Montgomery–Odlyzko law is, on present evidence, substantially right: the local statistics of ζ-zeros really are GUE statistics, to whatever precision computation has been able to confirm. A proof of this fact, going beyond the partial pair correlation result Montgomery established under RH, would be a substantial advance, and it would likely come hand-in-hand with progress on RH itself.

The role of heuristic arguments in the overall picture is, then, to constrain plausible proof strategies and to provide evidence about what the eventual theorem must say. They do not provide proof methods; they provide guidance about what proof methods should yield.

X. Negative Results and Obstructions

What Is Forbidden

The accumulated body of work on RH has established not only what is conjectured but also what is forbidden. Various plausible-looking statements that would imply false versions of RH (or implausible refinements of it) have been proved false, and these negative results constrain proof strategies.

The Davenport–Heilbronn function, mentioned in the introduction, is a key example. This function is a linear combination of two Dirichlet L-functions chosen to satisfy a functional equation but not an Euler product. It has zeros off the critical line, despite satisfying the analytic shape that one might naively associate with RH. Any proof of RH that does not use the Euler product, or some equivalent structural feature, would prove a false statement when applied to Davenport–Heilbronn. This rules out a wide class of “purely analytic” arguments that try to derive RH from the functional equation alone.

A further negative constraint comes from the fact that the Selberg class includes many L-functions, and the Grand Riemann Hypothesis asserts RH for all of them. Any proof method that establishes RH only for ζ but not for, say, Dirichlet L-functions of small modulus, would be suspicious: there is no obvious structural reason for ζ to be uniquely special. Conversely, a method that establishes GRH for all primitive L-functions, including those for which the corresponding hypothesis is currently open, would be either a major breakthrough or evidence of an error.

Lehmer Pairs and Near-Failures

Derrick Lehmer’s discovery in 1956 of pairs of ζ-zeros that come very close together — pairs where the function Z(t) almost fails to have a sign change between them — provides another kind of constraint. Lehmer pairs show that ζ does not behave with the kind of rigid regularity that would make a violation of RH at some sufficiently large height inconceivable. The function does come close to the kind of failure that would require a zero off the line.

This has been formalized in the de Bruijn–Newman constant Λ, defined so that Λ ≤ 0 is equivalent to RH. Newman conjectured in 1976 that Λ ≥ 0, with Λ = 0 expressing the idea that RH is “barely true” in a precise sense. In 2018, Brad Rodgers and Terence Tao proved Newman’s conjecture: Λ ≥ 0. Combined with computational upper bounds on Λ (currently around 10^{-12}), this means that if RH is true, it is true with no room to spare — the function ζ is, in this sense, extremely close to having zeros off the line, but does not.

The Rodgers–Tao result is striking. It says that any proof of RH must establish a precise inequality Λ ≤ 0 with no slack, rather than a robust inequality of the form Λ ≤ −ε for some positive ε. The hypothesis is true on the boundary of failure. This is consistent with the Hilbert–Pólya picture (eigenvalues of a self-adjoint operator are real, but small perturbations can push them off the line), and it constrains proof strategies: any method that produces a robust inequality with slack would be inconsistent with the Rodgers–Tao result and so cannot be correct.

The Question of a Natural Proofs Obstruction

In computational complexity theory, the “natural proofs barrier” of Razborov and Rudich identifies a structural obstruction to certain kinds of lower-bound proofs in circuit complexity. The barrier shows that a wide class of proof methods (those whose construction is “natural” in a precise sense) cannot prove the lower bounds they are aimed at, conditional on certain cryptographic assumptions.

It is reasonable to ask whether an analogous obstruction exists for RH. Is there a class of proof methods that can be shown, for structural reasons, to be incapable of proving RH? On present evidence, no such formal barrier has been identified. But the accumulated negative results — the Davenport–Heilbronn obstruction, the Lehmer near-failures, the Rodgers–Tao result, and the structural disparity between function field and number field cases — function as informal barriers. They constrain proof strategies even in the absence of a formal impossibility result.

Whether these informal barriers will eventually be formalized into a “natural proofs”-style theorem for RH, or whether some combination of methods will overcome them, is an open meta-question of the subject.

XI. The Question of Multiple Necessary Ingredients

Why a Single Method May Not Suffice

The accumulated state of knowledge on RH suggests, to many practitioners, that no single existing technique is likely to suffice for a proof. The reasoning is structural. The function field proof required the conjunction of geometric setting, cohomological framework, and positivity statement. Each of these is a substantial body of mathematics in its own right; their combination produced the proof. By analogy, a proof of RH over Q is likely to require a similar conjunction: a spectral or cohomological framework (Hilbert–Pólya, Connes, or some new construction), a positivity input (a statement constraining the spectrum or the eigenvalues), and a structural connection of these to the arithmetic of Z.

No existing program supplies all three ingredients in compatible form. The Hilbert–Pólya program supplies the spectral framework but lacks the positivity. The Connes program supplies a noncommutative framework with a candidate positivity statement, but the positivity has not been established. Random matrix theory supplies statistical predictions but no operator. Density theorems and zero-free regions supply quantitative information but no structural framework.

The hypothesis that a proof will require synthesis of multiple programs, rather than completion of any single one, is not provable but is suggestive. It would explain why each individual program has stalled at a different obstacle. It would explain why the function field proof, which had all three ingredients in compatible form, succeeded relatively quickly once the framework was established (within a few decades of the conjecture being formulated for varieties in general). And it would explain why a proof, when it comes, may be the product of substantial cross-disciplinary effort rather than the achievement of a single mathematician working in a single tradition.

What a Synthesis Might Look Like

Speculation about the form of an eventual proof is, of necessity, speculative. But certain features can be identified as likely. The proof will probably use a cohomology theory or spectral framework that does not yet fully exist — possibly an extension of étale cohomology to arithmetic schemes, possibly a noncommutative cohomology of the type Connes has been developing, possibly something else. The proof will probably use a positivity statement whose discovery is itself a major mathematical event, comparable in difficulty to the Hodge index theorem in the function field setting. The proof will probably draw on connections among number theory, mathematical physics, geometry, and possibly other fields, in ways that integrate methods that are currently developed in relative isolation.

The proof will probably not be a clever combination of existing analytic techniques applied to ζ directly. The space of such combinations has been thoroughly explored, and the structural constraints identified above suggest that no such combination will suffice. The proof will probably not be elementary in the sense of avoiding deep machinery: the function field proof is not elementary, and the obstacle to its translation to Q is precisely the absence of comparably deep machinery for Z.

These predictions could be wrong. Mathematics has surprised observers many times, and a proof of RH by means no one currently anticipates remains possible. But the accumulated evidence suggests that a proof, when it comes, will be substantial — not a short paper that resolves the hypothesis through a clever trick, but a major construction that integrates large bodies of existing mathematics with new structures yet to be developed.

XII. Conclusion

The Riemann hypothesis has resisted proof for one hundred sixty-six years. The resistance is not a failure of effort: substantial mathematics has been developed in the attempt, and the developed mathematics is, in itself, of lasting value. Random matrix theory, the Selberg class framework, the Connes noncommutative geometric program, the de Branges theory of Hilbert spaces of entire functions, the density theorem program, the Vinogradov–Korobov methods — all of these are substantial bodies of mathematics motivated, at least in part, by RH. They have applications throughout number theory and beyond, regardless of whether RH is ever proved.

The resistance has, however, identified the structural shape of the problem with increasing clarity. A proof of RH must distinguish the critical line from neighboring lines, must use information specific to ζ (or to L-functions with appropriate structure), and must explain the function field success or differ from it deliberately. The function field analog provides a template for what a successful proof would look like: geometric setting, finite-dimensional cohomology, and positivity statement, in compatible form. The number field setting lacks all three of these in workable form.

Each of the major programs has supplied one or more of the missing ingredients in some form. None has supplied all of them in compatible form. The Hilbert–Pólya program supplies the spectral framework conceptually but lacks the operator. Connes supplies a noncommutative framework but lacks the positivity. The F_1 program (treated in Paper 2) supplies a candidate base over which Z might be a curve but lacks the geometry to make this concrete. Each program has produced substantial mathematics; none has produced the proof.

The structural constraints identified in this paper — the requirement of a fine-grained method, the disparity with the function field setting, the negative results on Davenport–Heilbronn and similar functions, the Lehmer near-failures and the Rodgers–Tao result, and the implicit “natural proofs”-style barrier — suggest that a proof, when it comes, will not be the product of incremental improvement of any single existing program. It will more likely be the product of synthesis: a combination of programs, with new structural ideas supplying the missing connections.

The next paper in this suite proposes a specific novel conjecture in this direction — a refinement of pair correlation conditional on RH that aims to be sharp, falsifiable, and connected to the structure of L-functions in a way that, if correct, would generate testable predictions and possibly new arithmetic consequences. The conjecture is offered not as a proof of RH or a path to one, but as the kind of forward-looking hypothesis whose investigation is the natural successor to the present survey. The Riemann hypothesis itself remains where Riemann left it — probable, supported, central, and unproved.

═══════════════════════════════════════════════