Backporch

Pillars · deep dive

The Eigenbook

One word, climbed all the way up. $A v = λ v$ — a direction a matrix only stretches — is the first rung of a ladder that does not stop at linear algebra. Replace the matrix with the Hecke operators acting on modular forms and the eigenvector becomes an eigenform; vary it p-adically and the eigenforms sweep out the eigencurve; attach arithmetic and the eigenvalues become traces of Frobenius; interpolate and you reach p-adic L-functions. Five floors, one idea.

Following Joel Bellaïche's Eigenbook. This page is a scaffold — the whole climb laid out honestly, each rung filled in over time. The base is the eigenvectors card; the q-expansions are generating functions; the p-adic size that makes the eigencurve possible is in what is a number; the analytic runway is what is a modular form.

0Eigenvectors $A v = λ v$ A direction an operator only stretches. The whole climb is this idea, in bigger and bigger spaces.
1Hecke eigenforms $T_{n} f = a_{n} f$ A modular form that is a simultaneous eigenvector of every Hecke operator. Its eigenvalues are its Fourier coefficients.
2The eigencurve $E \to W$ Assemble all finite-slope eigenforms, p-adically, into one rigid-analytic curve. Eigenvalues become functions on a geometry.
3Galois representations $tr ρ (Frob_{p}) = a_{p}$ Each eigenform carries a 2-dimensional Galois representation. The eigenvalues reappear as traces of Frobenius — analysis turned arithmetic.
4p-adic L-functions $L_{p} (f, s)$ Interpolate the critical L-values p-adically along the family. The open frontier — Iwasawa theory, Bloch–Kato, BSD.

0 · The eigenvector, one floor down

Start where the eigen card ends. A linear operator sends almost every direction somewhere new, but a few special directions it only scales:

A v = λ v .

Two moves make the rest of the climb. First, stop thinking “2×2 matrix” and think operator on a vector space — the space can be functions, and infinite-dimensional. Second, stop thinking about one operator and think about a whole commuting family at once: a single vector that is an eigenvector of every operator in the family. That simultaneous eigenvector is the object the next floor is built from.

1 · Hecke eigenforms

The space of modular forms $M_{k} (Γ)$ of weight $k$ is finite-dimensional but it carries an infinite commuting family of Hecke operators ${T_{n}}$ . A Hecke eigenform is a nonzero $f$ that is a simultaneous eigenvector of all of them:

T_{n} f = a_{n} f (\forall n), f (q) = n \geq 1 \sum a_{n} q^{n}, q = e^{2 π i τ} .

The remarkable part: once you normalize $a_{1} = 1$ , the eigenvalues $a_{n}$ are the Fourier coefficients of the form. The list of eigenvalues and the generating function are the same object read two ways. Because the operators commute, the eigenvalues are multiplicative,

a_{m} a_{n} = a_{mn} (g cd (m, n) = 1),

which is why a single arithmetic function can be built from its values at primes. The discriminant form is the textbook example: $Δ = q \prod_{n \geq 1} (1 - q^{n})^{24} = \sum_{n \geq 1} τ (n) q^{n}$ is a Hecke eigenform, and Ramanujan's $τ (n)$ is multiplicative (Mordell, 1917) for exactly this reason.

And here is why eigenforms are the secret spine of arithmetic. Take an elliptic curve, count its points mod $p$ , and the numbers $a_{p}$ you get are the Fourier coefficients of a weight-2 eigenform — two computations, one sequence. That is modularity:

conductor

N = 11

geometry — count points on

E : y^{2} + y = x^{3} - x^{2} - 10 x - 20

a_{p} = p + 1 - # E (F_{p})

a_{2}

-2

a_{3}

-1

a_{5}

a_{7}

-2

a_{11}

a_{13}

=

analysis — expand the modular form

f = η (τ)^{2} η (11 τ)^{2}

f = \sum a_{n} q^{n}, a_{p} = [q^{p}] f

a_{2}

-2

a_{3}

-1

a_{5}

a_{7}

-2

a_{11}

a_{13}

✓ identical — every

a_{p}

from counting points equals the form's coefficient. $E_{11 a}$ is modular.

the form as a generating function:

f = q - 2 q^{2} - q^{3} + 2 q^{4} + q^{5} + 2 q^{6} - 2 q^{7} - 2 q^{9} + \dots

Two computations that share nothing. On the left, pure geometry: solve $y^{2} + y = x^{3} - x^{2} - 10 x - 20$ over each finite field and count. On the right, pure analysis: multiply out an infinite product, the modular form $η (τ)^{2} η (11 τ)^{2}$ , and read its q-series coefficients. They produce the same integers, prime after prime — so $L (E, s) = L (f, s)$ . That equality is the Modularity Theorem: every elliptic curve over $Q$ is a modular form in disguise. It is also where these $a_{p}$ become traces of Frobenius, and the bedrock under Birch–Swinnerton-Dyer. Honest status: the theorem is proved — Wiles & Taylor–Wiles (semistable, with Fermat's Last Theorem), then Breuil–Conrad– Diamond–Taylor (2001) in full — by a hundreds-of-pages argument (Galois deformations, $R = T$ ). Here we only witness it, on curves whose newform is a clean eta product.

But why is $a_{p}$ an eigen-anything? Because it is literally the eigenvalue of an operator. The Hecke operator $T_{p}$ reshuffles a form's coefficients; an eigenform is the rare $f$ it returns merely scaled — apply it and watch:

apply

(T_{p} f)_{n} = a_{p n} + p^{k - 1} a_{n / p} (a_{n / p} = 0 unless p ∣ n)

f =

a_{1}

a_{2}

-2

a_{3}

-1

a_{4}

a_{5}

a_{6}

a_{7}

-2+ ⋯

T_{2} f =

b_{1}

-2

b_{2}

b_{3}

b_{4}

-4

b_{5}

-2

b_{6}

-4

b_{7}

4+ ⋯

$T_{2} f = a_{2} f$ — every coefficient just scaled by $a_{2} = - 2$ . So

f

is an eigenvector of

T_{2}

and the eigenvalue is the Fourier coefficient

a_{2}

commuting operators ⇒ the eigenvalues multiply (gcd = 1):

a_{2} \cdot a_{3} = (- 2) (- 1) = 2 = a_{6}

The Hecke operator $T_{p}$ reshuffles a form's coefficients by the rule above — it acts on the whole space $M_{k} (Γ_{0} (N))$ . Almost no form survives unchanged; the special ones that do, the eigenforms, come back scaled: $T_{p} f = a_{p} f$ . That is the whole point of the word — the coefficients $a_{p}$ are eigenvalues of a commuting family of operators, which is why they are multiplicative (the law you watched hold in the modularity lab is forced by $T_{m} T_{n} = T_{mn}$ ). One floor up, letting the weight vary p-adically threads these eigensystems into the eigencurve; attach Galois and each $a_{p}$ becomes a trace of Frobenius. $Δ$ is the textbook case — its eigenvalues are Ramanujan's $τ (n)$ , multiplicative for exactly this reason.

And the form does more than match the curve — it draws it. Over $C$ the curve is a torus, and the modular curve $X_{0} (N)$ covers it; the group law you dragged in §5 is just addition that wraps:

j = 1728

CM Z [i], square lattice

P + Q = (P+Q) mod Λ · half-points = 2-torsion (roots of the cubic)

P

0.32

P^{'}

0.18

Q

0.54

Q^{'}

0.61

Over $C$ , the curve is a torus: $E (C) ≅ C /Λ$ , glued from this cell by identifying opposite edges, with $(x, y) = (℘ (z), ℘^{'} (z))$ mapping it onto the cubic. And the group law is simply addition mod $Λ$ — slide P and Q and watch the sum wrap around; that wrapping is exactly the chord-and-tangent law of §5. The four half-points are the 2-torsion $E [2]$ — the origin $O$ and the three roots where $y = 0$ . The deep part — modularity — is that $Λ$ is the period lattice of the weight-2 form $f$ , and the modular curve covers the torus: $X_{0} (N) = H / Γ_{0} (N) τ \mapsto \int^{τ} f C /Λ ≅ E .$ The form doesn't just match the curve (the a_p agreeing) — it parametrizes it. (CM curves give the cleanest lattices: $j = 1728$ square, $j = 0$ hexagonal.)

1½ · The workbench — prove it for all $n$

Every lab above witnesses an identity on finitely many coefficients. Here is the move that turns a finite check into a theorem: $M_{k} (SL_{2} (Z))$ is finite-dimensional, so two weight- $k$ forms agreeing for $n \leq k /12$ — the Sturm bound — are equal at every coefficient, forever. Build something and prove it:

LHS×

E_{4}^{2}

E_{6}

E_{8}

E_{10}

E_{14}

Δ

E_{4}^{2}

wt 8

+×

E_{4}

E_{6}

E_{8}

E_{10}

E_{14}

Δ

0

RHS×

E_{4}

E_{6}

E_{8}

E_{10}

E_{14}

Δ

E_{8}

wt 8

q^011=q^1480480=q^261,92061,920=q^31,050,2401,050,240=

weight

k = 8

· Sturm bound: agreement for

n \leq 0

forces equality · faded chips are already free

This is the move no calculator gives you. The spaces $M_{k} (SL_{2} (Z))$ are finite-dimensional ( $dim M_{8} = dim M_{10} = dim M_{14} = 1$ !), so two weight- $k$ forms that agree for $n \leq k /12$ are equal — at every coefficient, forever: their difference would be a cusp form vanishing to higher order than the valence formula allows. One finite check, an infinite family of identities: $E_{4}^{2} = E_{8}$ is a theorem the moment one coefficient matches, and $E_{4} E_{10} = E_{6} E_{8}$ says the divisor sums $σ_{3}, σ_{5}, σ_{7}, σ_{9}, σ_{13}$ obey identities nobody would guess from arithmetic alone. (Try $E_{4}^{3}$ vs $E_{6}^{2}$ — the failure is $1728 Δ$ , Ramanujan's $τ$ hiding in the gap.) The coefficients here are exact big integers, not floats — a certificate, not a plot. Missing on purpose: $E_{12}$ , whose q-expansion needs the denominator 691 — the prime of Ramanujan's $τ (n) \equiv σ_{11} (n) (mod 691)$ , waiting on the p-adic rung. Same machinery as the Hecke lab; levels $Γ_{0} (N)$ and eta quotients join the bench next.

2 · The eigencurve

Now let the weight move. A single eigenform is a point; the Coleman–Mazur eigencurve $E$ (1998) glues all finite-slope p-adic eigenforms into one rigid-analytic curve, fibred over weight space by the weight map

κ : E ⟶ W, points of E = finite-slope systems of Hecke eigenvalues .

The eigenvalues $a_{p}$ stop being numbers attached to one form and become analytic functions on a geometry — the spectrum of the Hecke algebra, made into a space. What lets the weight vary continuously at all is p-adic size: nearby weights are weights that agree to a high power of $p$ , the p-adic world. The higher-rank generalization — eigenvarieties (Buzzard, Emerton, Chenevier) — is the same construction for other reductive groups.

You can watch this interpolation happen on the one family that's computable by hand — the Eisenstein component, where $a_{ℓ} (E_{k}) = 1 + ℓ^{k - 1}$ :

Eisenstein family,

ℓ = 2

a_{ℓ} (E_{k}) = 1 + 2^{k - 1}

classical weight

k = 2

→

a_{2} (E_{2}) = 1 + 2 = 3

= 00003(5)

k = 6

⇝

00113(5)

\equiv a_{2} (E_{2}) (mod 5^{1})

k = 22

⇝

02103(5)

\equiv a_{2} (E_{2}) (mod 5^{2})

k = 102

⇝

21003(5)

\equiv a_{2} (E_{2}) (mod 5^{3})

k = 502

⇝

10003(5)

\equiv a_{2} (E_{2}) (mod 5^{4})

k \to 2 in W (p -adically) ⟹ a_{ℓ} (E_{k}) \to a_{ℓ} (E_{2}) (p -adically) .

Watch the low base- $5$ digits lock in: as the weight $k = 2 + (p - 1) p^{n - 1}$ creeps p-adically toward $2$ , the eigenvalue $a_{ℓ}$ creeps p-adically toward $a_{ℓ} (E_{2})$ — agreeing to one more power of $5$ each step. So a Hecke eigenvalue is not one number per classical weight; it is a p-adic-analytic function of the weight, and the scattered classical eigenforms interpolate into a single rigid-analytic curve over weight space, $U \to W$ — the Coleman–Mazur eigencurve (built from overconvergent $p$ -adic modular forms and the compact $U_{p}$ operator; the points are the finite-slope eigensystems). This is its Eisenstein component, the one you can compute by hand. One floor up, attaching Galois turns these eigenvalue- functions into families of Frobenius traces, and their p-adic L-functions live on this same curve.

The rest is deep p-adic geometry: overconvergence, and why “finite slope” is the price of admission (the compact $U_{p}$ operator has a discrete spectrum only on overconvergent forms). That construction is a map here, not a full build — but the felt core, eigenvalues as p-adic-analytic functions of the weight, is exactly what the Eisenstein family above shows.

3 · Galois representations

Here the climb crosses from analysis into arithmetic. To a Hecke eigenform $f$ of weight $k \geq 2$ one attaches a continuous 2-dimensional Galois representation

ρ_{f} : Gal (\overline{Q} / Q) ⟶ GL_{2} (\overline{Q}_{ℓ}),

unramified outside $N ℓ$ , characterized by its action on Frobenius:

tr ρ_{f} (Frob_{p}) = a_{p}, det ρ_{f} (Frob_{p}) = p^{k - 1} .

The Hecke eigenvalue $a_{p}$ — a coefficient of a holomorphic function — turns out to be the trace of Frobenius, an arithmetic invariant of how primes split. This is the GL₂ case of the Langlands correspondence: Eichler–Shimura for weight 2, Deligne (1968–71) in general, Deligne–Serre (1974) for weight 1.

The cleanest way in is the $a_{p}$ you already counted: they are the traces, and Frobenius' eigenvalues sit on the circle $∣ z ∣ = p$ (Hasse). Watch them — and watch a CM curve pile up at the poles:

E : y^{2} + y = x^{3} - x

normalized Frobenius eigenvalues

e^{\pm i θ_{p}}

p \leq 160

· supersingular

(a_{p} = 0)

: 2/37

a_{5}

-2

a_{7}

-1

a_{11}

-5

a_{13}

-2

a_{17}

a_{19}

x^{2} - a_{p} x + p

Every $a_{p}$ on this circle is a number you can get by counting points ( $a_{p} = p + 1 - # E (F_{p})$ ) — and the deep fact of this rung is that it is also a trace of Frobenius: to $E$ is attached a 2-dimensional Galois representation $ρ$ with $tr ρ (Frob_{p}) = a_{p}$ and $det ρ (Frob_{p}) = p$ (Eichler–Shimura, Deligne). So Frobenius has characteristic polynomial $x^{2} - a_{p} x + p$ , its eigenvalues are $p e^{\pm i θ_{p}}$ — forced onto the circle $∣ z ∣ = p$ , which is exactly Hasse's bound $∣ a_{p} ∣ \leq 2 p$ . Switch to 32a (CM) and the angles pile up at the poles ( $a_{p} = 0$ , half the primes); the non-CM curves spread by the Sato–Tate law (a theorem now). The point count (geometry), the modular coefficient (analysis), and this Frobenius trace (arithmetic) are one number with three faces — and that $a_{p}$ is what the p-adic L-function packages next.

4 · p-adic L-functions

The top floor, and the open one. The complex L-function $L (f, s) = \sum_{n} a_{n} n^{- s}$ has special values at the critical integers; dividing by the right period makes them algebraic, and those algebraic numbers can be interpolated p-adically into a p-adic L-function

L_{p} (f, \cdot) interpolating \frac{L ( f , j )}{Ω _{f}} at critical j .

Along the eigencurve these assemble into families, and the questions become the live ones — Iwasawa main conjectures, the Bloch–Kato conjectures, and (for weight 2, via modular elliptic curves) Birch–Swinnerton-Dyer. This is the frontier the site treats honestly bounded, not faked.

To fill: the interpolation property stated cleanly, the link to the ζ machinery, and a clear marker of what is theorem versus conjecture.

One word, five floors

Read the ladder top to bottom and it is a single sentence: a direction an operator only scales (eigenvector) becomes a form fixed by every Hecke operator (eigenform), which sweeps out a p-adic geometry (eigencurve), whose points name how primes split (Frobenius eigenvalues), and whose L-values interpolate p-adically (p-adic L-functions). The same prefix carries the whole way. That is why it is worth its own deep dive — and why the elementary eigen card is not a footnote but the foundation.