I. Introduction

The four prior papers in this suite, together with Paper 5 on the Langlands framework, have treated the Riemann hypothesis from several angles: historically, structurally, strategically, and prospectively. What none of these papers has done is treat in detail the technical theory of moments of L-functions — the body of conjectures and partial results concerning integrals of the form

∫_0^T |ζ(1/2 + it)|^{2k} dt

and their generalizations to other L-functions. This omission is, by the standards of modern analytic number theory, a substantial one. Moment theory is one of the most active research areas of the past twenty-five years, has produced some of the deepest conditional results in the field, and supplies the technical foundation on which much else depends — including the conjecture proposed in Paper 4 of this suite.

The thesis of this paper is that moment theory deserves treatment in its own right. The thesis has three components. First, moments of L-functions are technically rich: they admit multiple conjectural frameworks (random matrix theory, hybrid Euler–Hadamard products, recursive moment conjectures, autocorrelation conjectures), and each framework illuminates the others. Second, moment theory has been remarkably productive at the level of bounds: the conjectured order of magnitude of moments has been established, conditionally and in some cases unconditionally, even though the precise constants remain open. Third, moment theory is connected densely to other parts of L-function theory — to zero statistics, to extreme value theory, to the Lindelöf hypothesis, to families of L-functions, and to the broader Langlands picture — in ways that make moment progress productive across the wider field.

The structure of this paper proceeds historically and conceptually. After establishing the classical results (Hardy–Littlewood for the second moment, Ingham for the fourth), the paper traces the path through the Conrey–Ghosh and Conrey–Gonek conjectures, the Keating–Snaith random matrix framework that set the modern direction, the hybrid Euler–Hadamard product approach, the substantial body of work on lower and upper bounds (Heath-Brown, Soundararajan, Radziwiłł, Harper), the function field analogs where rigorous results match the conjectures, the moments-in-families program in the Katz–Sarnak tradition, the connections to extreme values, and the implications for the Riemann hypothesis itself. The paper closes with open problems and a concluding assessment of moment theory’s place in the broader picture.

The treatment is substantive but not encyclopedic. Moment theory has, by now, a vast literature, and any paper of reasonable length must select. The selection here emphasizes what is structurally important and what bears most directly on the broader picture of L-function theory and the Riemann hypothesis. References to the primary literature are made at points where the reader would benefit from following up; the paper does not attempt to substitute for that literature.

II. Classical Moment Results

The Second Moment: Hardy–Littlewood

The earliest substantial moment result is the Hardy–Littlewood theorem of 1918, which gives the asymptotic for the second moment of |ζ| on the critical line:

∫_0^T |ζ(1/2 + it)|² dt ~ T log T as T → ∞.

The proof uses the approximate functional equation for ζ — an asymptotic formula expressing ζ(1/2 + it) as a sum of two Dirichlet polynomial pieces, each of length approximately √(t/2π), plus a small error. Squaring this approximation and integrating produces, after careful analysis of the diagonal and off-diagonal contributions, the asymptotic above.

The structural content of the Hardy–Littlewood result is that the second moment grows like T log T, with leading coefficient 1. The leading coefficient is a specific number, computable explicitly. Higher-order terms are also accessible: the full asymptotic expansion has the form

∫_0^T |ζ(1/2 + it)|² dt = T log T + (2γ − 1 − log 2π) T + O(T^{1/2 + ε}),

where γ is the Euler–Mascheroni constant. The error term has been improved by various authors over the decades; the best known unconditional error term is currently of size O(T^{1/2} (log T)^c) for an explicit constant c.

The Fourth Moment: Ingham

The fourth moment was established by Ingham in 1926:

∫_0^T |ζ(1/2 + it)|⁴ dt ~ (1/(2π²)) T (log T)⁴.

The proof is more intricate than the second moment but proceeds along similar lines: an approximate functional equation, a careful expansion of the resulting double sum, and analysis of the cross-terms.

The leading coefficient 1/(2π²) is again specific and computable. Higher-order terms in the asymptotic expansion are also known. The best unconditional error term is of size O(T (log T)^{3 + ε}), substantially weaker relative to the main term than the corresponding error for the second moment.

The fourth moment was already a substantial achievement in 1926. The methods Ingham developed have been refined and extended over the subsequent century, but the structural difficulty of the fourth moment foreshadows the much greater difficulty of higher moments.

Why the Easy Cases Are Easy

The second and fourth moments are the “easy” cases of moment theory, in a precise sense. The reason is that for k = 1 and k = 2, the moment of |ζ|^{2k} can be computed by a method that works only for these specific values.

The method, in outline, is as follows. The approximate functional equation expresses ζ(1/2 + it) as a sum of two pieces:

ζ(1/2 + it) ≈ ∑{n ≤ X} 1/n^{1/2 + it} + (functional equation factor) · ∑{n ≤ Y} 1/n^{1/2 − it},

with X and Y both of order √(t/2π). For the 2k-th moment, one squares this approximation and integrates. The result is a sum over 2k tuples of integers, with each tuple contributing an integral that depends on the size relations among the integers.

For k = 1 and k = 2, the integrals can be evaluated explicitly. The diagonal terms (where the integers in the tuple match in pairs) give the main contribution, and the off-diagonal terms can be controlled.

For k ≥ 3, the off-diagonal contributions become genuinely difficult. The combinatorial complexity of the tuples grows rapidly, and no method has been developed that gives an explicit asymptotic for the higher moments. This is the essential obstacle that has kept the higher moments conjectural for nearly a century.

The Conrey–Ghosh Conjecture for the Sixth Moment

In 1992, Brian Conrey and Amit Ghosh proposed the form of the sixth moment:

∫_0^T |ζ(1/2 + it)|⁶ dt ~ a_3 g_3 T (log T)⁹,

with a_3 the arithmetic factor (an Euler product over primes) and g_3 the geometric factor. Their conjecture was that g_3 = 42/9!. The arithmetic factor a_3 was computable explicitly:

a_3 = ∏_p (1 − 1/p)⁴ (1 + 4/p + 1/p²).

The Conrey–Ghosh conjecture was based on a heuristic involving the asymptotic of the divisor function d_3(n) (the number of ways to write n as a product of three factors) and the analysis of the diagonal contributions in the formal expansion of |ζ|⁶.

The conjecture was significant because it gave a specific prediction for a moment that was not accessible by the methods that worked for k = 1 and k = 2. The prediction has since been refined and extended, but the basic form — arithmetic factor times geometric factor times T (log T)^{k²} — has held up.

The Conrey–Gonek Conjecture for the Eighth Moment

Conrey and Steven Gonek extended the conjecture to the eighth moment in 1998, predicting

∫_0^T |ζ(1/2 + it)|⁸ dt ~ a_4 g_4 T (log T)^{16},

with a_4 an explicit Euler product and g_4 = 24024/16!. The Conrey–Gonek prediction used a more elaborate heuristic involving the symmetry between the diagonal and off-diagonal contributions and a connection to formal recursive moment conjectures.

The Conrey–Gonek conjecture was important because it suggested that the moments for general k might fit into a coherent pattern, with the geometric constants g_k computable through some general framework. The framework, when it was eventually identified, came from random matrix theory.

III. The Keating–Snaith Framework

The Random Matrix Analogy

The connection between ζ and random matrix theory was first observed in the context of zero spacings: the Montgomery–Odlyzko law, treated in Paper 3, predicts that the imaginary parts of ζ-zeros follow GUE statistics. By the late 1990s, this connection had been verified numerically with great accuracy and had become a central organizing principle in the study of ζ.

The natural question was whether the random matrix analogy extends from zeros to values. That is: if the zeros of ζ behave statistically like eigenvalues of random unitary matrices, do the values of ζ on the critical line behave statistically like the values of the characteristic polynomial of a random unitary matrix?

The answer, proposed by Jonathan Keating and Nina Snaith in a series of papers beginning in 2000, is yes. Specifically, they conjectured that the moments of |ζ(1/2 + it)|^{2k} should be governed by the moments of the characteristic polynomial of a random matrix drawn from the Circular Unitary Ensemble (CUE), with the random matrix moment supplying the geometric constant g_k.

The Keating–Snaith Conjecture

The Keating–Snaith conjecture, in its full form, asserts:

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

with a_k the arithmetic factor

a_k = ∏p (1 − 1/p)^{k²} · ∑{m=0}^∞ (Γ(k + m)/(Γ(k) m!))² p^{−m},

and g_k the geometric factor

g_k = ∏_{j=0}^{k-1} (j!/(k+j)!) · k².

The geometric factor g_k can be expressed in closed form using the Barnes G-function, a generalization of the gamma function:

g_k = G(k+1)² / G(2k+1),

where G is the Barnes G-function, defined by the functional equation G(z+1) = Γ(z) G(z) with G(1) = 1.

For specific small k, the Keating–Snaith prediction gives:

• k = 1: g_1 = 1, recovering the Hardy–Littlewood result.
• k = 2: g_2 = 1/12, equivalent to Ingham’s coefficient 1/(2π²) after the appropriate normalization.
• k = 3: g_3 = 42/9!, agreeing with Conrey–Ghosh.
• k = 4: g_4 = 24024/16!, agreeing with Conrey–Gonek.

The agreement of the Keating–Snaith formula with the prior k = 1, 2 results (which are theorems) and with the Conrey–Ghosh and Conrey–Gonek conjectures (which were derived independently) is striking. It indicates that the random matrix framework captures, at least at the level of leading asymptotics, the correct structure of moments of ζ.

The Random Matrix Calculation

The geometric constant g_k arises, in the Keating–Snaith framework, from a calculation that can be carried out rigorously in the random matrix setting. Specifically, for U a random N × N unitary matrix drawn from CUE with Haar measure, the characteristic polynomial Z_U(θ) = det(I − U e^{−iθ}) has moments

E[|Z_U(0)|^{2k}] = G(k+1)² / G(2k+1) · N^{k²} (1 + O(1/N))

for large N. The leading-order asymptotic gives the geometric constant g_k = G(k+1)² / G(2k+1), with the factor N^{k²} matching the conjectured (log T)^{k²} when N is identified with the appropriate analog of the height T.

The rigorous nature of this calculation in the random matrix setting is one of the framework’s strengths. The constants g_k are not free parameters fitted to data; they are computed from the random matrix model and then matched against the L-function moments. The agreement is structural, not merely numerical.

The Identification of Parameters

The identification of T with N in the Keating–Snaith framework is determined by the requirement that the random matrix model reproduce the correct mean density of zeros. For ζ-zeros at height T, the mean density is (1/2π) log(T/2π), and for CUE eigenvalues of an N × N matrix on the unit circle, the mean density is N/(2π). The identification

N = log(T/2π)

makes the densities match. With this identification, the random matrix prediction for the 2k-th moment, scaled to T, gives the Keating–Snaith formula.

The identification is not just dimensional. It encodes a substantial structural assumption: that the local statistics of ζ on the critical line, at height T, are well modeled by random matrix statistics with N = log(T/2π). This assumption has been tested numerically and has substantial support, but it is, formally, part of the conjecture rather than a derived consequence.

Why This Calculation Is Convincing

The Keating–Snaith framework is convincing for several reasons.

First, it reproduces the known cases k = 1 and k = 2 from a single coherent calculation, rather than treating them as separate cases. The reproduction is exact, not approximate.

Second, it predicts the Conrey–Ghosh and Conrey–Gonek values for k = 3 and k = 4, which had been derived by independent heuristics. The agreement across different approaches suggests that the Keating–Snaith formula captures the correct structure.

Third, it provides a unified expression for all k > 0 (not only positive integers, but all positive reals), with the constant g_k smoothly interpolated through the Barnes G-function. This generalization beyond positive integers is unexpected from the L-function side but natural from the random matrix side.

Fourth, the structural reasons for the random matrix analogy — the GUE statistics of zeros, which are well established numerically — extend naturally to predict statistics of values. The same underlying picture explains both, with consistent constants.

Fifth, function field analogs of the Keating–Snaith conjectures have been proved rigorously, supplying further structural support. These analogs are treated in Section VIII below.

IV. The Hybrid Euler–Hadamard Product Approach

The Decomposition

In 2007, Steven Gonek, Christopher Hughes, and Jonathan Keating proposed a different approach to moments of ζ: the hybrid Euler–Hadamard product. The approach factors ζ formally into two pieces:

ζ(s) = P_X(s) · Z_X(s),

where P_X(s) is a finite Euler product over primes up to X, and Z_X(s) is a finite Hadamard product over zeros of ζ at heights up to (roughly) X. The decomposition is approximate; it becomes exact only in the limit X → ∞, but for finite X it provides a useful tool for analyzing |ζ|.

The idea behind the hybrid model is to split the contribution to ζ into a “primes part” and a “zeros part,” each of which can be analyzed separately. The primes part P_X(s) is a Dirichlet polynomial with an explicit expression in terms of prime contributions; the zeros part Z_X(s) captures the contribution from low-lying zeros to the overall fluctuations of ζ.

Computing Moments via the Hybrid Model

The hybrid model gives a way to compute moments by computing the moments of P_X and Z_X separately and then combining them. Specifically, the 2k-th moment of |ζ| factorizes (heuristically) as

E[|ζ|^{2k}] ~ E[|P_X|^{2k}] · E[|Z_X|^{2k}].

Each factor is computable. The arithmetic factor a_k arises from the moments of the primes part:

E[|P_X|^{2k}] ~ a_k (log X)^{k²}.

The geometric factor g_k arises from the moments of the zeros part:

E[|Z_X|^{2k}] ~ g_k.

The product is the Keating–Snaith prediction.

Why the Hybrid Approach Adds Something

The hybrid approach is significant for several reasons beyond reproducing the Keating–Snaith conjecture.

First, it makes the contribution of primes and zeros explicit. Each factor has a clear interpretation: a_k captures the arithmetic structure of the primes, g_k captures the universality from random matrix theory. The factorization makes the interaction between these two contributions transparent.

Second, it provides a tool for proving partial results. Lower and upper bounds on moments can be obtained by separately bounding each factor, with techniques tailored to each. Many of the rigorous bounds described in subsequent sections use the hybrid framework explicitly or implicitly.

Third, it suggests refinements. The leading-order Keating–Snaith prediction is the product of leading orders of P_X and Z_X. Subleading corrections to either factor produce subleading corrections to the moment, and the hybrid framework makes this expansion systematic.

Fourth, it has analogs for other L-functions. The hybrid model can be set up for any L-function in the Selberg class, with the corresponding primes and zeros contributions. The structural unity across the Selberg class is preserved.

The Connection to the Conjecture in Paper 4

The hybrid Euler–Hadamard model has direct relevance to the Stratified Zero–Prime Resonance Conjecture proposed in Paper 4. That conjecture predicts that pair correlation of ζ-zeros, weighted by character-restricted prime sums, deviates from the unconditional prediction in a specific way governed by lowest L-function zeros. In the hybrid framework, this corresponds to weighting the primes part by the character and analyzing the cross-terms with the zeros part.

The hybrid framework supplies, in this sense, the technical machinery in which the Paper 4 conjecture is most naturally stated and tested. The character-weighted prime sum S_χ(T; f) in Paper 4 is, modulo normalization, the character-twisted version of the primes factor in the hybrid model. The deviation predicted in Paper 4 arises from the cross-terms between the character-twisted primes and the zeros, which the hybrid model makes accessible to systematic analysis.

This connection is part of why moment theory deserves its own treatment: the conjecture in Paper 4, while presented there in self-contained form, depends conceptually on moment-theoretic ideas that are only fully developed in the present paper.

V. Lower Bounds for Moments

The Conditional Lower Bounds of Heath-Brown

The first substantial lower bounds for general moments were obtained by Roger Heath-Brown in the late 1970s and early 1980s. Heath-Brown proved that, conditional on the Riemann hypothesis, for every rational k > 0,

∫_0^T |ζ(1/2 + it)|^{2k} dt ≫ T (log T)^{k²}.

The bound is the conjectured order of magnitude. It establishes that moments grow at least as fast as conjectured, leaving the constants as the remaining problem.

Heath-Brown’s method uses a Dirichlet polynomial approximation to ζ^k on the critical line, combined with mean value estimates and careful analysis of the resulting cross-terms. The restriction to rational k arose from the technical structure of the proof, not from any structural reason.

Soundararajan’s Extension

In 2009, Kannan Soundararajan extended Heath-Brown’s lower bound to all real k > 0:

∫_0^T |ζ(1/2 + it)|^{2k} dt ≫ T (log T)^{k²},

conditional on RH. The extension to all real k is significant because the Keating–Snaith conjecture is most naturally stated for all real k > 0, with the random matrix analog smoothly interpolated through the Barnes G-function. Soundararajan’s proof handles the full range of k coherently, matching the conjectured form across the full range.

Soundararajan’s method introduces what is now called the resonator method. The method constructs an auxiliary Dirichlet polynomial designed to “resonate” with |ζ(1/2 + it)|^k, producing cross-terms whose mean value can be controlled. The resonator is chosen to extract the dominant contribution to the moment, yielding lower bounds of the conjectured order.

The resonator method has subsequently been refined and extended by many authors (Soundararajan, Heath-Brown, Bondarenko, Seip, Saksman, and others) and has become a standard tool in moment theory. It produces lower bounds for L-functions in many settings, including Dirichlet L-functions, modular form L-functions, and L-functions of higher rank.

The Unconditional Lower Bound of Radziwiłł and Soundararajan

In 2013, Maksym Radziwiłł and Soundararajan proved an unconditional version of the lower bound:

∫_0^T |ζ(1/2 + it)|^{2k} dt ≫ T (log T)^{k²},

without assuming RH, for all real k ≥ 1.

The unconditional result is a substantial achievement. It establishes that the conjectured order of magnitude is correct as a matter of fact, not just as a consequence of RH. The constants implicit in the bound are not the conjectured ones, but the order of magnitude is.

The proof of the unconditional bound uses the resonator method combined with techniques for handling possible zeros off the critical line. The handling is delicate: the proof must be robust enough to give the correct order of magnitude even if RH fails (in which case, off-line zeros could in principle disturb the moment growth). The proof shows that the disturbance, even in the worst case allowed by the unconditional density estimates, cannot reduce the moment below the conjectured order.

The Conceptual Content of the Lower Bound Results

The lower bound results, taken together, establish a substantial structural fact: the conjectured order of magnitude of moments is correct. The Keating–Snaith conjecture predicts moments of order T (log T)^{k²}, and the lower bounds confirm that moments are at least this large.

What remains open is the precise constant. The Keating–Snaith conjecture predicts that the constant is a_k g_k. The lower bound results establish only that the constant is at least some positive number, with the explicit lower bound substantially smaller than the conjectured a_k g_k. Closing this gap — establishing the precise constant — is the remaining problem in moment theory.

The conceptual significance of the lower bound results is that they reduce the moment problem to a question about constants. If one accepts that moments grow at the conjectured order, the question is just what the leading coefficient is. This is a substantial reduction from the original problem, where even the order of magnitude was not established.

VI. Upper Bounds for Moments

Soundararajan’s 2009 Upper Bound

The companion to the lower bound work is the upper bound work. Establishing upper bounds of the conjectured order is, in some respects, harder than establishing lower bounds, because upper bounds must rule out larger-than-expected fluctuations.

In 2009, Soundararajan proved that, conditional on RH, for every fixed k > 0,

∫_0^T |ζ(1/2 + it)|^{2k} dt ≪ T (log T)^{k² + ε}

for every ε > 0.

The bound matches the conjectured order up to an arbitrarily small loss in the exponent of the logarithm. The loss of ε is, in the analytic number theory convention, a relatively mild defect.

Soundararajan’s method uses an iterative bootstrapping argument. The idea is to obtain a moment bound at one level of generality and then use it to obtain a slightly sharper bound at the next level, iterating until the conjectured order is reached. The iteration converges to the conjectured exponent k² but cannot quite reach it within the framework of the proof — hence the residual ε.

Harper’s Refinement

In 2013, Adam Harper improved Soundararajan’s result by removing the ε:

∫_0^T |ζ(1/2 + it)|^{2k} dt ≪ T (log T)^{k²},

conditional on RH, for every fixed k > 0.

The bound matches the conjectured order exactly. Combined with the Heath-Brown–Soundararajan lower bound, this establishes that, under RH, the moment grows at exactly the conjectured order — with the constant pinned down to within a bounded ratio.

Harper’s method refines Soundararajan’s bootstrapping argument. The key innovation is a more careful tracking of the dependence of the iterated bound on k, allowing the iteration to be carried to a precise conclusion rather than stopping ε-short. The proof is technical but the structural insight is that the iterative argument, properly analyzed, converges all the way to the conjectured exponent.

What Remains Open

After Harper’s result, the conditional moment problem has the form: under RH, the moment is ~ C_k T (log T)^{k²} for some positive constant C_k. The Keating–Snaith conjecture predicts C_k = a_k g_k. Establishing this precise prediction — the constant, not just the order of magnitude — is the remaining problem.

Recent progress on the precise constant has come from several directions.

For k a positive integer, the constant a_k is the arithmetic factor and is known explicitly. The question is whether the geometric factor g_k = G(k+1)²/G(2k+1) is correct. For k = 1 and k = 2, this is verified by the Hardy–Littlewood and Ingham theorems. For k = 3 and k = 4, the Conrey–Ghosh and Conrey–Gonek conjectures predict the same value as Keating–Snaith.

For k = 3, partial progress has been made: Conrey, Farmer, Keating, Rubinstein, and Snaith have given a refined recipe (the “CFKRS recipe”) that predicts not only the leading constant but the full asymptotic expansion, including subleading terms. The leading constant agrees with Keating–Snaith, and subleading terms have been verified numerically against high-precision computation.

For higher k, the precise constant remains open. There is no known method that produces it without assuming additional structure (random matrix predictions, the CFKRS recipe, or equivalent).

The Heap–Soundararajan Direction

A recent direction of progress is due to Heap and Soundararajan and their collaborators (Conrey, Iwaniec, Soundararajan, and others), who have developed techniques for establishing precise moments in restricted settings. For families of L-functions with appropriate symmetry — quadratic Dirichlet L-functions, for instance — the moments at the central point s = 1/2 (rather than averaged over the critical line) have been computed exactly in some cases.

These results are conditional and apply only to specific families, but they confirm the random matrix predictions in those settings with full constants. The successes provide structural support for the Keating–Snaith framework as the correct prediction across the broader L-function landscape.

VII. Function Field Analogs

Why Function Fields Are Tractable

The function field setting is a recurring theme in this suite, and moment theory is no exception. As noted in Paper 2, function fields admit a fully geometric framework in which many conjectures of arithmetic number theory become theorems.

For moments specifically, the function field setting allows random matrix predictions to be made rigorous. The reason is that, in the function field setting, the Frobenius eigenvalues of L-functions form a finite set (the set of eigenvalues of a finite-dimensional operator on cohomology), and the random matrix model becomes a model for a finite ensemble rather than for an infinite one. Taking the limit as the genus or conductor grows, one obtains rigorous limiting distributions that match random matrix predictions exactly.

Keating–Roditty-Gershon–Rudnick

Jon Keating, Edva Roditty-Gershon, and Zeév Rudnick, together with various collaborators, have established function field analogs of the Keating–Snaith moment conjectures in several settings.

For families of quadratic Dirichlet L-functions over function fields, the moment predictions match those derived from the symplectic random matrix ensemble (USp). For families associated to elliptic curves, the predictions match the orthogonal ensemble. For Dirichlet L-functions with characters of large prime conductor, the predictions match the unitary ensemble.

In each case, the random matrix prediction is established as a theorem in the function field setting, with explicit rates of convergence as the relevant parameter (conductor, genus) tends to infinity. The agreement between the function field theorems and the conjectures for the corresponding number field cases provides strong structural evidence for the random matrix framework.

The Structural Lesson

The structural lesson of the function field analog work is the same as the broader function field-versus-number field disparity treated in Papers 2 and 3: the function field setting supplies a setting in which random matrix predictions are theorems, while the number field setting leaves them as conjectures.

The reason is structural. In the function field setting, the Frobenius operator on cohomology is a finite-dimensional linear operator with explicit eigenvalues, and the family of such operators (as the variety varies in a family) is controlled by the geometry of the family — typically, by the geometric monodromy group of the family in a precise sense (developed by Katz and others). The Deligne equidistribution theorem then says that the Frobenius eigenvalues equidistribute according to the Haar measure on the monodromy group, which is exactly the random matrix prediction.

In the number field setting, no analog of the geometric monodromy group is currently available. The random matrix prediction is supported by analogy and by extensive numerical evidence, but it is not derived from a geometric structure of the kind that supports the function field theorems.

The function field successes thus serve a dual role: they confirm that the Keating–Snaith framework is structurally correct in settings where rigorous proof is available, and they highlight the missing structure (a geometric monodromy for arithmetic L-function families) that would be required for analogous proofs in the number field setting.

VIII. Moments of L-Functions in Families

The Katz–Sarnak Philosophy for Moments

The Katz–Sarnak philosophy, treated in detail in Paper 7 of this suite (forthcoming), predicts that low-lying zeros of L-functions in natural families follow statistical distributions determined by the symmetry type of the family. The same philosophy extends to moments.

For an L-function in a family of symmetry type S (unitary, symplectic, or orthogonal), the moments at the central point s = 1/2 should follow the moments of the characteristic polynomial at s = 0 of a random matrix in the corresponding ensemble. The moments are computable explicitly in terms of Barnes G-function values, with formulas analogous to but different from the Keating–Snaith formula for moments along the critical line.

Symmetry Types and Their Moment Predictions

For unitary families — for instance, the family of all primitive Dirichlet L-functions of conductor q as q varies, or families of automorphic L-functions of GL(n) varying in level — the moment predictions are governed by the unitary ensemble (CUE or the related GUE). The geometric constants follow the Keating–Snaith formula.

For symplectic families — for instance, the family of quadratic Dirichlet L-functions L(s, χ_d) as d varies, or families of L-functions of self-dual representations of certain types — the moment predictions are governed by the symplectic ensemble (USp). The geometric constants are given by formulas involving Barnes G-function values, but with a different combinatorial structure than the unitary case.

For orthogonal families — for instance, the family of L-functions L(s, E_d) of quadratic twists of a fixed elliptic curve E, or families of L-functions of self-dual representations of orthogonal type — the moment predictions are governed by the orthogonal ensemble. The structure is further refined into even orthogonal and odd orthogonal subtypes, with the parity determining specific features of the predictions.

Verifications in Restricted Ranges

The moment predictions for families have been verified in restricted ranges by various authors. Notable contributions include:

Quadratic Dirichlet L-functions: The first moment was established by Jutila and others; the second moment by Soundararajan; higher moments by Soundararajan and his collaborators.

Modular form L-functions: Moments of L(1/2, f) as f varies over newforms of weight k and level N have been studied by Iwaniec, Sarnak, and others, with predictions matching the Katz–Sarnak philosophy.

L-functions of elliptic curve quadratic twists: Moments of L(1, E_d) as d varies have been studied; the predictions match the orthogonal symmetry type.

In each case, the results are partial — typically establishing the predicted constant for low moments and the predicted order of magnitude for higher moments — but they support the framework.

The Vanishing Moment

A particular focus of the family moment program is the zeroth moment: the question of how often L(1/2) = 0 across a family. For families of orthogonal symmetry type, this connects directly to the Goldfeld conjecture for elliptic curve quadratic twists and to the broader question of how often L-functions vanish at the central point.

The random matrix prediction is that, for orthogonal families, the proportion of L-functions in the family with L(1/2) = 0 is positive (one-half, in the relevant cases), and the parity of the order of vanishing matches the parity dictated by the functional equation. For unitary and symplectic families, the random matrix prediction is that L(1/2) ≠ 0 with probability 1 in the limit (vanishing is a measure-zero event).

The verifications in restricted ranges, combined with the function field analogs, provide substantial support for these predictions. The full results for natural families remain open in many cases, but the framework is well established.

IX. Connections to Extreme Values

The Maximum of |ζ| on the Critical Line

A natural question complementary to moments is: how large can |ζ(1/2 + it)| be on intervals of length T? This is the question of extreme values.

The simplest bound is the convexity bound: |ζ(1/2 + it)| ≪ t^{1/4 + ε}. The Lindelöf hypothesis predicts |ζ(1/2 + it)| ≪ t^ε. RH implies |ζ(1/2 + it)| ≪ exp(c log t / log log t). None of these is the question asked: what is the maximum of |ζ(1/2 + it)| as t varies over [0, T]?

The Predictions

Random matrix theory predicts that the maximum of |ζ(1/2 + it)| over [0, T] is, at leading order,

max_{t ∈ [0, T]} |ζ(1/2 + it)| ~ exp((1/2 + o(1)) √(log T · log log T)).

The exponent √(log T · log log T) is sharper than what RH alone supplies (which gives the smaller exponent log T / log log T) but is consistent with RH. The prediction comes from random matrix calculations of the maximum of the characteristic polynomial of a random unitary matrix.

The prediction was established by Arguin, Bourgade, Belius, Soundararajan, and others in a series of papers from the mid-2010s onward. The structural picture is that the maximum is dominated by the “freezing” of fluctuations of log |ζ| in a way that parallels the behavior of branching random walks and log-correlated random fields.

The Connection to Moments

Extreme values and moments are connected by the following heuristic. The 2k-th moment ∫_0^T |ζ|^{2k} dt is dominated, for large k, by the contribution from the largest values of |ζ|. Specifically, if the maximum is M and is attained on a set of measure roughly M^{−2k} log M (which is the prediction from the freezing picture), then the moment scales as M^{2k} · M^{−2k} log M ~ log M as k → ∞ properly normalized.

This heuristic relates the high moments to the extreme values: knowing how the moment grows with k tells one about the tail of the distribution of |ζ|, and vice versa. The Keating–Snaith moment conjecture and the maximum value prediction are, in this sense, two faces of a single underlying distributional structure.

Implications for the Riemann Hypothesis

The extreme value results have indirect implications for RH. The maximum of |ζ| on [0, T] is bounded above, conditionally on RH, by an explicit function of T. If the random matrix prediction for the maximum is correct, the actual maximum is much smaller than the worst case allowed by RH. This consistency between the predictions and the RH-conditional bounds supports both: it suggests that RH is true, and that ζ behaves “typically” in a way that matches random matrix predictions.

If the random matrix prediction for the maximum were violated — if the actual maximum grew faster than predicted — this would not directly disprove RH but would indicate that ζ has structural features not captured by the random matrix model. Conversely, confirmation of the maximum prediction (which has been increasingly substantiated by numerical computation) provides further structural support for the broader random matrix framework.

X. Implications for the Riemann Hypothesis

What Moment Results Buy Concretely

Moment theory, as it has developed, supplies several concrete inputs to the broader study of ζ and the Riemann hypothesis.

Refined PNT error terms: The Hardy–Littlewood second moment, in its sharper forms, supplies refined error terms in the prime number theorem under various assumptions. RH gives the strongest such error term, but the moment results give intermediate forms accessible without RH.

Zero-density estimates: Moment bounds for ζ on the critical line, combined with the Riemann–von Mangoldt zero counting formula, give bounds on the number of zeros in regions of the critical strip. These zero-density estimates are central tools in analytic number theory and have been treated in Paper 3.

Lindelöf hypothesis: The Lindelöf hypothesis is the assertion that |ζ(1/2 + it)| = O(t^ε) for every ε > 0. This is a moment-like statement: it follows from the assertion that the 2k-th moment grows like T (log T)^{k²} for every fixed k, plus an averaging argument. Lindelöf is conjecturally true (it follows from RH and from Keating–Snaith), but it is open in general.

Moment of L-functions in families: Moments of L-functions averaged over families give information about the average behavior of L-functions, which has direct arithmetic content (e.g., for class numbers, ranks of elliptic curves, and similar invariants).

The Indirect Path to RH

Moments do not directly prove RH. The conjectured Keating–Snaith asymptotic is consistent with RH (and indeed, the conjecture is stated assuming RH), but it does not imply RH. A proof of the Keating–Snaith conjecture would not, by itself, settle RH.

However, moments and zeros are sufficiently entangled that progress on one constrains the other. The hybrid Euler–Hadamard model makes this explicit: moments factor into a primes contribution and a zeros contribution, and progress on either factor constrains the joint behavior. A proof of full Keating–Snaith would, in particular, require a detailed understanding of zero statistics that goes well beyond what is currently known.

The path from moment theory to RH is, in this sense, indirect. It goes through structural understanding: each moment result clarifies the shape of ζ on the critical line, each zero result clarifies the location of zeros, and each constrains the other. The eventual proof of RH, when it comes, is likely to require substantial moment-theoretic input — though the proof itself need not be a moment-theoretic proof.

What Moment Theory Has Established

The state of moment theory after twenty-five years of intensive work can be summarized:

• The conjectured order of magnitude T (log T)^{k²} for the 2k-th moment is established, conditional on RH for k > 0 (Heath-Brown, Soundararajan, Harper) and unconditional for k ≥ 1 (Radziwiłł–Soundararajan).
• The conjectured leading constant a_k g_k is established for k = 1, 2 (Hardy–Littlewood, Ingham) and is supported for k = 3, 4 by independent heuristics matching Keating–Snaith.
• The full Keating–Snaith conjecture for general k is verified numerically to high precision but remains conditional on the broader random matrix framework.
• Function field analogs of Keating–Snaith are theorems, providing structural support.
• Family moment predictions in the Katz–Sarnak framework are verified in restricted ranges, with full proof in function field settings.
• Extreme value predictions have been substantiated, connecting moment theory to the broader distributional theory of ζ.

This is a substantial body of established knowledge. Moment theory has, over twenty-five years, produced some of the deepest conditional and unconditional results in analytic number theory.

XI. Open Problems in Moment Theory

The Constants

The most prominent open problem is the determination of the precise constants in the moment asymptotics. Specifically:

The constant for k = 3: ∫_0^T |ζ(1/2 + it)|^6 dt ~ a_3 g_3 T (log T)^9. The Conrey–Ghosh prediction gives g_3 = 42/9!. The conjecture is supported by the Keating–Snaith framework and by numerical computation, but is not proved.

The constant for k = 4: ∫_0^T |ζ(1/2 + it)|^8 dt ~ a_4 g_4 T (log T)^{16}. The Conrey–Gonek prediction gives g_4 = 24024/16!. Again supported but not proved.

The constants for general k: The Keating–Snaith formula predicts g_k = G(k+1)²/G(2k+1) for all k > 0. The prediction is supported across the range but proved only at k = 1, 2.

Establishing any of these constants would be a substantial advance. For k = 3, in particular, the gap between the established lower and upper bounds and the conjectured exact value is the most prominent open question.

Family Moments at the Central Point

For families of L-functions, the moments at s = 1/2 (the central point) are predicted by Katz–Sarnak. The predictions have been verified for low moments in many families, but higher moments and the precise constants remain open.

A particular focus is the family of quadratic Dirichlet L-functions L(s, χ_d): the conjectured moments at s = 1/2, predicted by the symplectic random matrix ensemble, are open for k ≥ 3.

For families of L-functions of modular forms, the moments at s = 1/2 are predicted by the orthogonal ensemble (with appropriate parity considerations). Open in many cases.

Joint Moments of Distinct L-Functions

Moments of products of distinct L-functions — for instance,

∫_0^T |ζ(1/2 + it)|^{2k} |L(1/2 + it, χ)|^{2k’} dt

for distinct characters χ — are predicted by the random matrix framework but are largely unstudied. The predictions involve cross-correlations between different L-functions and the corresponding random matrix ensembles.

These joint moments connect directly to the conjecture in Paper 4 of this suite: the Stratified Zero–Prime Resonance Conjecture predicts deviations in the joint statistics of ζ-zeros and L(s, χ)-zeros, and joint moments are the natural test of those predictions in the moment-theoretic setting.

Moments at the Edge of the Critical Strip

Moments at the edge of the critical strip — that is, of |ζ(1 + it)| or of L(1, χ) — have arithmetic content (through connections to class numbers, Mertens-type bounds, and similar). Predictions for these moments come from the same random matrix framework, with appropriate modifications for the edge.

The moments at the edge are partially understood. For ζ(1 + it), the moment of |ζ(1 + it)|^{2k} can be related to the corresponding moment on the critical line through the functional equation, but the analysis is subtle. For L(1, χ) in families, the moments connect to class number statistics and have been studied by Granville, Soundararajan, and others.

XII. Conclusion

Moment theory occupies a distinctive position in the analytic number theory of the past twenty-five years. It has been the area of most active and productive research, with the deepest conditional and unconditional results emerging in succession. It has provided technical infrastructure on which much else depends — including the conjecture proposed in Paper 4 of this suite. It has connected analytic number theory to random matrix theory, to mathematical physics, and to the broader study of L-functions across the Selberg class and the Langlands program.

The Keating–Snaith framework, introduced in 2000, set the modern direction by importing random matrix predictions into moment theory. The framework reproduces the classical results (Hardy–Littlewood, Ingham), agrees with the Conrey–Ghosh and Conrey–Gonek conjectures derived independently, and predicts moments for all k > 0 through a single coherent formula involving the Barnes G-function. The agreement across different approaches is structural: the random matrix framework captures the correct underlying distribution.

The hybrid Euler–Hadamard product model, introduced in 2007, provides a complementary framework that decomposes ζ into a primes contribution and a zeros contribution. The decomposition makes the interaction between arithmetic structure and random matrix universality transparent, and it has supported substantial subsequent work — including, implicitly, the conjecture in Paper 4.

The bounds work — Heath-Brown’s conditional lower bounds, Soundararajan’s extension and unconditional refinement with Radziwiłł, Soundararajan’s conditional upper bound and Harper’s refinement — has established the conjectured order of magnitude of moments to within a constant. The remaining problem is the precise constant, which is open for k ≥ 3 (with k = 1, 2 known classically and k = 3, 4 supported by independent heuristics).

The function field analogs are theorems. The Katz–Sarnak philosophy for families of L-functions is verified in restricted ranges and proved in function field settings. The connections to extreme values, to the Lindelöf hypothesis, and to zero statistics are dense and mutually constraining.

Why does moment theory deserve its own treatment? Because the technical depth and the conditional results are substantial; because the random matrix framework supplies a structural explanation that other parts of L-function theory rely on; because the conjecture in Paper 4 of this suite depends on moment-theoretic ideas that are most fully developed here; because the open problems in moment theory are sharply defined and likely to be addressed in coming years; and because moment theory is, on present evidence, the most productive single area of analytic number theory, with the most established conditional results and the clearest path toward continued progress.

The Riemann hypothesis itself remains where Riemann left it: probable, supported, central, and unproved. Moment theory does not, by itself, prove RH. But moment theory has substantially advanced the surrounding picture, and the cumulative effect of moment-theoretic results, family results, function field results, and zero-statistic results has been to constrain the L-function landscape with increasing precision. A proof of RH, when it comes, is likely to draw on moment-theoretic input. In the meantime, moment theory continues to produce new conditional and unconditional results, refining the picture and supplying tools that the broader theory uses.

The next paper in this suite, on zeros of L-functions in families, treats the parallel statistical theory that complements pair correlation. Family statistics and moment statistics are closely related — both involve random matrix predictions, both have function field analogs, both constrain the L-function landscape. The two together constitute the substantial body of conditional and partial results that, on present evidence, represents the most productive frontier of analytic number theory.

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