Current task: —

Updated 08-19-2026

Equation 50: Selberg's Central Limit Theorem for the Riemann Zeta Function

1. Statement of the Theorem

Selberg's Central Limit Theorem (CLT) states that as $T \to \infty$, if $t$ is chosen uniformly at random from $[T, 2T]$, then

$$ \frac{\log\left|\zeta\left(\tfrac{1}{2} + it\right)\right|}{\sqrt{\tfrac{1}{2}\log\log T}} \xrightarrow{d} \mathcal{N}(0, 1) $$

In words: the logarithm of the modulus of the Riemann zeta function on the critical line, appropriately normalized, converges in distribution to a standard Gaussian.

More precisely, for any fixed real numbers $a < b$:

$$ \lim_{T\to\infty} \frac{1}{T}\,\mathrm{meas}\left\{ t \in [T, 2T] : \frac{\log|\zeta(\tfrac{1}{2}+it)|}{\sqrt{\tfrac{1}{2}\log\log T}} \in [a,b] \right\} = \frac{1}{\sqrt{2\pi}}\int_a^b e^{-x^2/2}\,dx $$

Selberg actually proved a two-dimensional (complex) version: the pair

$$ \frac{\log \zeta(\tfrac12 + it)}{\sqrt{\tfrac12 \log\log T}} = \frac{\log|\zeta(\tfrac12+it)| + i\,\arg\zeta(\tfrac12+it)}{\sqrt{\tfrac12\log\log T}} $$

converges to a standard complex Gaussian — the real and imaginary parts become independent standard normals. The imaginary part is directly connected to the famous function $S(t) = \frac{1}{\pi}\arg\zeta(\tfrac12+it)$, which controls fluctuations in the zero-counting function $N(T)$ via the Riemann–von Mangoldt formula:

$$ N(T) = \frac{T}{2\pi}\log\frac{T}{2\pi e} + \frac{7}{8} + S(T) + O(1/T) $$

So Selberg's CLT for $S(t)$ says that the fluctuations of the number of zeros around its expected value are Gaussian with standard deviation $\frac{1}{\pi}\sqrt{\tfrac{1}{2}\log\log T}$ — remarkably small.

2. History and Attribution

Atle Selberg (1917–2007)

The theorem is due to Atle Selberg, the Norwegian mathematician (Fields Medal 1950, Abel Prize honorary recognition, Wolf Prize 1986). The key papers:

Historical notes:

Why it matters for the Riemann Hypothesis

Selberg's CLT says $\zeta$ on the critical line is "log-normally distributed": $|\zeta(\tfrac12+it)|$ is typically of size $e^{\pm\sqrt{\frac12\log\log T}} = (\log T)^{o(1)}$ — i.e., the zeta function is typically neither large nor small on the critical line, even though the Lindelöf Hypothesis allows it to be as large as $T^{\epsilon}$. This "typical vs. extreme value" dichotomy is the origin of the Fyodorov–Hiary–Keating program below.

3. Architecture of the Proof

The modern streamlined proof (Radziwiłł–Soundararajan, 2015) proceeds in four steps, exactly as sketched in your equation:

(i) Dirichlet polynomial approximation. Off a set of measure $o(T)$, $$ \log\left|\zeta\left(\tfrac12+it\right)\right| \approx \mathrm{Re}\sum_{p \le T^{\epsilon}} \frac{1}{p^{1/2+it}} = \sum_{p\le T^\epsilon} \frac{\cos(t\log p)}{\sqrt{p}}. $$ Justifying this requires controlling the contribution of zeros near the line — this is where zero-density estimates enter (or, in Radziwiłł–Soundararajan's proof, a clever comparison of $\zeta$ with $\zeta$ evaluated slightly off the line at $\sigma = \tfrac12 + \frac{W}{\log T}$, combined with a mollifier).

(ii) Quasi-independence. The functions $t \mapsto \cos(t\log p)$ for distinct primes $p$ behave like independent random variables uniform on the circle, because $\{\log p\}$ are linearly independent over $\mathbb{Q}$ (unique factorization!). Each term $\cos(t\log p)/\sqrt{p}$ has mean $0$ and variance $\frac{1}{2p}$.

(iii) Variance computation. By Mertens' theorem, $$ \sum_{p \le X} \frac{1}{2p} \sim \frac{1}{2}\log\log X \quad\Longrightarrow\quad \mathrm{Var} \sim \frac{1}{2}\log\log T. $$ The divergence of $\sum 1/p$ is essential: it is what makes a CLT possible (Lindeberg-type condition holds since individual terms $1/\sqrt{p}$ are tiny relative to the total variance).

(iv) Method of moments. Computing $\frac1T\int_T^{2T} \left(\sum_p \cos(t\log p)/\sqrt p\right)^{2k}dt$ via orthogonality of $p^{it}$ shows the moments match Gaussian moments $\frac{(2k)!}{k!2^k}\sigma^{2k}$, hence convergence in distribution.

4. The Modern Refinement: Fyodorov–Hiary–Keating

Selberg's CLT describes the typical value at a random point. The maximum over a short interval is governed by extreme-value theory of log-correlated Gaussian fields (the same universality class as branching random walk and the 2D Gaussian free field):

FHK Conjecture (2012): For $t$ random in $[T, 2T]$, $$ \max_{|h| \le 1} \log\left|\zeta\left(\tfrac12 + i(t+h)\right)\right| = \log\log T - \frac{3}{4}\log\log\log T + M_T, $$ where $M_T$ is a tight sequence of random variables with explicitly conjectured (randomly-shifted Gumbel) limiting law. The peculiar coefficient $\frac{3}{4}$ (instead of $\frac{1}{4}$ for i.i.d. Gaussians) is the signature of log-correlation.

Status: The leading order was proved by Najnudel (on RH) and Arguin–Belius–Bourgade–Radziwiłł–Soundararajan (unconditionally); the second-order term and tightness of $M_T$ were proven by Arguin–Bourgade–Radziwiłł (2020–2023). The conjecture is now essentially a theorem up to the exact limiting law of $M_T$.

5. Key Research Papers (with links)

Paper Link
Radziwiłł & Soundararajan, "Selberg's central limit theorem for log |ζ(1/2+it)|" (Enseign. Math., 2017) — a beautiful 6-page self-contained proof https://arxiv.org/abs/1509.06827
Fyodorov, Hiary & Keating, "Freezing transition, characteristic polynomials of random matrices, and the Riemann zeta function" (PRL 2012) https://arxiv.org/abs/1202.4713
Fyodorov & Keating, "Freezing transitions and extreme values: random matrix theory, ζ(1/2+it), and disordered landscapes" (Phil. Trans. R. Soc. A, 2014) https://arxiv.org/abs/1211.6063
Najnudel, "On the extreme values of the Riemann zeta function on random intervals of the critical line" (2016, assumes RH) https://arxiv.org/abs/1611.05562
Arguin, Belius, Bourgade, Radziwiłł & Soundararajan, "Maximum of the Riemann zeta function on a short interval of the critical line" (CPAM 2019, unconditional leading order) https://arxiv.org/abs/1612.08575
Arguin, Bourgade & Radziwiłł, "The Fyodorov–Hiary–Keating Conjecture. I" (upper bound, second order) https://arxiv.org/abs/2007.00988
Arguin, Bourgade & Radziwiłł, "The Fyodorov–Hiary–Keating Conjecture. II" (lower bound / tightness) https://arxiv.org/abs/2307.00982
Keating & Snaith, "Random matrix theory and ζ(1/2+it)" (CMP 2000) — CUE analogue of Selberg's CLT https://link.springer.com/article/10.1007/s002200000261
Harper, "Sharp conditional bounds for moments of the Riemann zeta function" (moment upper bounds via the same multiplicative-chaos ideas) https://arxiv.org/abs/1305.4618
Selberg's original 1946 paper: Arch. Math. Naturvid. 48 (1946), 89–155 — reprinted in Selberg's Collected Papers, Vol. I (Springer) (no free online copy; see Collected Papers, Springer 1989)

6. Computational Examples in Python

Example 1: Empirical verification of Selberg's CLT

We sample $\log|\zeta(\tfrac12+it)|$ at random heights $t \in [T, 2T]$ and compare the normalized histogram with the standard Gaussian.

import numpy as np
import mpmath as mp
import matplotlib.pyplot as plt

mp.mp.dps = 15  # working precision

# ---------------------------------------------------------------
# Example 1: Selberg's CLT -- histogram of normalized log|zeta|
# ---------------------------------------------------------------
T = 1.0e6          # base height (loglog converges VERY slowly!)
N = 3000           # number of random sample points
rng = np.random.default_rng(42)

sigma = np.sqrt(0.5 * np.log(np.log(T)))   # Selberg standard deviation
print(f"T = {T:.0e},  predicted sigma = sqrt(0.5*loglog T) = {sigma:.4f}")

samples = np.empty(N)
for k in range(N):
    t = rng.uniform(T, 2*T)
    z = mp.zeta(mp.mpc(0.5, t))
    samples[k] = float(mp.log(abs(z)))

normalized = samples / sigma

print(f"empirical mean = {normalized.mean():+.4f}  (theory: 0)")
print(f"empirical std  = {normalized.std():.4f}   (theory: 1)")

# Plot histogram vs standard normal density
x = np.linspace(-4, 4, 400)
plt.figure(figsize=(8, 5))
plt.hist(normalized, bins=60, density=True, alpha=0.6,
         color='steelblue', label=r'normalized $\log|\zeta(1/2+it)|$')
plt.plot(x, np.exp(-x**2/2)/np.sqrt(2*np.pi), 'r-', lw=2,
         label=r'$\mathcal{N}(0,1)$')
plt.title(f"Selberg's CLT,  T = {T:.0e},  N = {N} samples")
plt.xlabel(r'$\log|\zeta(1/2+it)| \,/\, \sqrt{\frac{1}{2}\log\log T}$')
plt.ylabel('density')
plt.legend(); plt.tight_layout()
plt.savefig('selberg_clt.png', dpi=150)
plt.show()

Expected output (approximately):

T = 1e+06,  predicted sigma = sqrt(0.5*loglog T) = 1.1462
empirical mean = -0.05...
empirical std  =  0.97...

Caveat: convergence is at rate $\sim 1/\sqrt{\log\log T}$ — triple logarithmically slow. Even at $T = 10^6$, $\log\log T \approx 2.63$, so finite-size corrections (a slight negative skew, since $\log|\zeta| = -\infty$ at zeros) are visible. This slowness is itself an important, well-documented phenomenon.

Example 2: The Dirichlet polynomial approximation (Step (i)–(ii))

import numpy as np
import mpmath as mp
from sympy import primerange

mp.mp.dps = 15
rng = np.random.default_rng(7)

T = 1.0e6
X = 1000                      # prime cutoff ~ T^epsilon
primes = np.array(list(primerange(2, X)), dtype=float)
logp   = np.log(primes)
invsqp = 1.0 / np.sqrt(primes)

# Theoretical variance of the prime sum:  sum 1/(2p) ~ (1/2) loglog X
var_prime_sum = np.sum(1.0 / (2.0 * primes))
print(f"sum_(p<{X}) 1/(2p)      = {var_prime_sum:.4f}")
print(f"(1/2) log log {X}       = {0.5*np.log(np.log(X)):.4f}")

print(f"\n{'t':>14} | {'log|zeta|':>10} | {'prime sum':>10}")
print("-" * 44)
for _ in range(10):
    t = rng.uniform(T, 2*T)
    lz = float(mp.log(abs(mp.zeta(mp.mpc(0.5, t)))))
    ps = float(np.sum(np.cos(t * logp) * invsqp))   # Re sum p^{-1/2-it}
    print(f"{t:14.2f} | {lz:10.4f} | {ps:10.4f}")

# Correlation over many samples
M = 800
lz = np.empty(M); ps = np.empty(M)
for k in range(M):
    t = rng.uniform(T, 2*T)
    lz[k] = float(mp.log(abs(mp.zeta(mp.mpc(0.5, t)))))
    ps[k] = float(np.sum(np.cos(t * logp) * invsqp))
print(f"\ncorrelation(log|zeta|, prime sum) = {np.corrcoef(lz, ps)[0,1]:.3f}")

You should see a strong positive correlation ($\approx 0.8$–$0.9$), confirming that the short prime sum captures most of the fluctuation of $\log|\zeta|$ — the heart of Selberg's proof.

Example 3: A "random model" verification of the variance and Gaussianity (Steps (ii)–(iv))

Replace $p^{-it}$ by genuinely independent uniform phases $e^{i\theta_p}$ — the random Euler product model:

import numpy as np
from sympy import primerange

rng = np.random.default_rng(0)
X = 100000
primes = np.array(list(primerange(2, X)), dtype=float)
invsqp = 1.0 / np.sqrt(primes)

M = 200000
# random model: sum_p cos(theta_p)/sqrt(p) with theta_p iid Uniform[0,2pi)
theta = rng.uniform(0, 2*np.pi, size=(M, 1))          # memory-friendly loop:
S = np.zeros(M)
block = 2000
for i in range(0, len(primes), block):
    th = rng.uniform(0, 2*np.pi, size=(M, min(block, len(primes)-i)))
    S += (np.cos(th) * invsqp[i:i+block]).sum(axis=1)

print(f"model variance      = {S.var():.4f}")
print(f"sum 1/(2p), p<{X}  = {np.sum(1/(2*primes)):.4f}")
print(f"(1/2) loglog X      = {0.5*np.log(np.log(X)):.4f}")
# 4th moment test of Gaussianity: kurtosis should be 3
print(f"kurtosis            = {np.mean(S**4)/S.var()**2:.4f}  (Gaussian: 3)")

Example 4: Fyodorov–Hiary–Keating maxima over short intervals

import numpy as np
import mpmath as mp

mp.mp.dps = 15
rng = np.random.default_rng(1)

T = 1.0e6
n_intervals = 60
grid = np.linspace(0, 2*np.pi, 200)   # sample the interval [t, t+2pi]

maxima = []
for _ in range(n_intervals):
    t0 = rng.uniform(T, 2*T)
    vals = [float(mp.log(abs(mp.zeta(mp.mpc(0.5, t0 + h))))) for h in grid]
    maxima.append(max(vals))
maxima = np.array(maxima)

LLT  = np.log(np.log(T))
pred = LLT - 0.75 * np.log(LLT)     # FHK: loglog T - (3/4)logloglog T
print(f"mean observed max over [t, t+2pi] = {maxima.mean():.3f}")
print(f"FHK prediction                    = {pred:.3f}")
print(f"typical value (Selberg scale)     = 0 +/- {np.sqrt(0.5*LLT):.3f}")

This illustrates the key contrast: the typical value of $\log|\zeta|$ is $O(\sqrt{\log\log T}) \approx 1.1$, while the maximum over even a tiny window is already near $\log\log T \approx 2.6$ — and at astronomically large $T$ this gap becomes dramatic.

7. Summary for Your Paper