(September 15, 2019)

Weierstrass and Hadamard products

Paul Garrett garrett@math.umn.edu http://www.math.umn.edu/egarrett/
[This document is
http://www.math.umn.edu/egarrett/m/mfms/notes 2019-20/05c Hadamard products.pdf]
1. Weierstrass products
2. Poisson-Jensen formula
3. Hadamard products
4. Appendix: harmonic functions
Apart from factorization of polynomials, perhaps the oldest product expression is Euler’s
∞ 
sin πz Y z2 
= 1− 2
πz n=1
n

Granting this, Euler equated the power series coefficients of z 2 , evaluating ζ(2) for the first time:
∞
π2 X 1
= 2
6 n=1
n

The Γ-function factors: Z ∞

dt 1
e−t tz = Γ(z) = ∞ 
0 t Y z  −z/n
z eγz 1+ e
n=1
n

where the Euler-Mascheroni constant γ is essentially defined by this relation. The integral (Euler’s) converges
for Re (z) > 0, while the product (Weierstrass’) converges for all complex z except non-positive integers.
Granting this, the Γ-function is visibly related to sine by
∞ 
1 Y z2  z
= −z 2 1 − 2 = − · sin πz
Γ(z) · Γ(−z) n=1
n π

because the exponential factors are linear, and can cancel.

Linear exponential factors are exploited in Riemann’s explicit formula [Riemann 1859], derived from
P equality
1
of the Euler product and Hadamard product [Hadamard 1893] for the zeta function ζ(s) = n ns for
Re (s) > 1:
∞
Y 1 ea+bs Y  s  ρs Y  s  −s/2n
= ζ(s) = · 1 − e · 1 + e
1 − p−s s−1 ρ
ρ n=1
2n
p prime

where the product expansion of Γ( 2s ) is visible, corresponding to trivial zeros of ζ(s) at negative even integers,
and ρ ranges over all other, non-trivial zeros, known to be in the critical strip 0 < Re (s) < 1.
The hard part of the proof (below) of Hadamard’s theorem is adapted from [Ahlfors 1953/1966], with various
rearrangements. A somewhat different argument is in [Lang 1993]. We recall some standard (folkloric?)
proofs of supporting facts about harmonic functions, starting from scratch.

1
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

1. Weierstrass products
Given a sequence of complex numbers zj with no accumulation point in C, we will construct an entire
function with zeros exactly the zj . This is essentially elementary.

[1.1] Basic construction

Taylor-MacLaurin polynomials of − log(1 − z) will play a role: let

z2 z3 zn
pn (z) = z + + + ... +
2 3 n
We will show that there is a sequence of integers nj giving an absolutely convergent infinite product vanishing
precisely at the zj , with vanishing at z = 0 accommodated by a suitable leading factor z m :

Y z  pnj (z/zj ) Y z z z2 z3 z nj 
zm 1− e = zm 1− exp + 2 + 3 + ... + n
j
zj j
zj zj 2zj 3zj nj zj j

P Q
Absolute convergence of j log(1 + aj ) implies absolute convergence of the infinite product j (1 + aj ).
Thus, we show that X   z  z 
 log 1 − + pnj  < ∞

j
zj z j

Fix a large radius R, keep |z| < R, and ignore the finitely-many zj with |zj | < 2R, so in the following we
have |z/zj | < 12 . Using the power series expansion of log,
 z z
 1  z n+1 1  z n+2 1 |z/zj |n+1 |z/zj |n+1
 log(1 − ) − pn  ≤ ·  + ·  + ... ≤ · ≤ 2·
    
zj zj n+1 zj n+2 zj n + 1 1 − |z/zj | n+1

Thus, we want a sequence of positive integers nj such that

X |z/zj |nj +1
< ∞ (with |z| < R)
nj + 1
|zj |≥2R

Of course, the choice of nj ’s must be compatible with enlarging R, but this is easily arranged. For example,
nj = j − 1 succeeds:

X  z j X  z j X  z j X  z j X
  =   +   ≤   + 2−j
j
zj zj zj zj j
|zj |<2R |zj |≥2R |zj |<2R

Since {zj } is discrete, the sum over |zj | < 2R is finite, so we have convergence, and convergence of the
infinite product with nj = j:

Y z  pj (z/zj ) Y z z z2 z3 zj 
1− e = 1− exp + 2 + 3 + ... + j
j
zj j
zj zj 2zj 3zj jz

2
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

[1.2] Canonical products and genus

Given entire f with zeros zj 6= and a zero of order m at 0, ratios

f (z)
ϕ(z) = Y z
pnj (z/zj )
zm · 1− ·e
j
zj

with convergent infinite products are entire, and do not vanish. Non-vanishing entire ϕ has an entire
logarithm: Z z 0
ϕ (ζ) dζ
g(z) = log ϕ(z) =
0 ϕ(ζ)
Thus, non-vanishing entire ϕ is expressible as

ϕ(z) = eg(z) (with g entire)

Thus, the most general entire function with prescribed zeros is of the form
Y z
pnj (z/zj )
f (z) = eg(z) · z m · 1− ·e (with g entire)
j
zj

Naturally, with fixed f , altering the nj necessitates a corresponding alteration in g.

We are most interested in zeros {zj } allowing a uniform integer h giving convergence of the infinite product
in an expression
Y z  ph (z/zj ) Y z z z2 z3 zh 
f (z) = eg(z) · z m · 1− e = zm 1− exp + 2 + 3 + ... + h
j
zj j
zj zj 2zj 3zj hz

When f admits a product expression with a uniform h, a product expression with minimal uniform h is a
canonical product for f .

When, further, the leading factor eg(z) for f has g(z) a polynomial, the genus of f is the maximum of h and
the degree of g.

2. Poisson-Jensen formula
Jensen’s formula and the Poisson-Jensen formula are essential in the difficult half of the Hadamard theorem
(below) comparing genus of an entire function to its order of growth.

The logarithm u(z) = log |f (z)| of the absolute value |f (z)| of a non-vanishing holomorphic function f on a
∂2 ∂2
neighborhood of the unit disk is harmonic, that is, is annihilated by ∆ = ∂x 2 + ∂y 2 : expand

∂ ∂
∂ ∂
 1 1

∆ log |f (z)| = +i −i log f (z) + log f (z)
∂x ∂y ∂x ∂y 2 2

Conveniently, the two-dimensional Laplacian is the product of the Cauchy-Riemann operator and its
conjugate. Since log f is holomorphic and log f is anti-holomorphic, both are annihilated by the product of
the two linear operators. This verifies that log |f (z)| is harmonic.
Thus, log |f (z)| satisfies the mean-value property for harmonic functions:
Z 2π Z 2π
1 1
log |f (0)| = u(0) = u(eiθ ) dθ = log |f (eiθ )| dθ
2π 0 2π 0

3
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

Next, let f have zeros ρj in |z| < 1 but none on the unit circle. We manufacture a holomorphic function F
from f but without zeros in |z| < 1, and with |F | = |f | on |z| = 1, by the standard ruse
Y 1 − ρj z
F (z) = f (z) ·
j
z − ρj

Indeed, for |z| = 1, the numerator of each factor has the same absolute value as the denominator:

1 1
|z − ρj | = | − ρj | = · |1 − ρj z| = |1 − ρj z|
z |z|

For simplicity, suppose no ρj is 0. Applying the mean-value identity to log |F (z)| gives
Z 2π Z 2π
X 1 1
log |f (0)| − log |ρj | = log |F (0)| = log |F (eiθ )| dθ = log |f (eiθ )| dθ
j
2π 0 2π 0

and then the basic Jensen’s formula

Z 2π
X 1
log |f (0)| − log |ρj | = log |f (eiθ )| dθ (for |ρj | < 1)
j
2π 0

The Poisson-Jensen formula is obtained by replacing 0 by an arbitrary point z inside the unit disk. This
is obtained by replacing f by f ◦ ϕz , where ϕz is a linear fractional transformation mapping 0 → z and
stabilizing [1] the unit disk:  
1 z w+z
ϕz = : w −→
z 1 zw + 1
ρ −z
This replaces the zeros ρj by ϕ−1 j
z (ρj ) = −zρj +1 . Instead of the mean-value property expressing f (0) as an
integral over the circle, use the Poisson formula (see appendix) for f (z). This gives the basic Poisson-Jensen
formula
Z 2π
X  ρ −z 
 j 1 1 − |z|2
log |f (z)| − log   = log |f (eiθ )| · dθ (for |z| < 1, |ρj | < 1)

j
−zρj + 1 2π 0 |z − eiθ |2

More generally, for holomorphic f on a neighborhood of a disk of radius r > 0 with zeros ρj in that disk,
apply the previous to f (r · z) with zeros ρj /r in the unit disk:
Z 2π
X  ρ /r − z 
 j 1 1 − |z|2
log |f (r · z)| − log   = log |f (r · eiθ )| · dθ (for |z| < 1)

j
−zρj /r + 1 2π 0 |z − eiθ |2

Replacing z by z/r gives

Z 2π
X  ρ /r − z/r 
 j 1 iθ 1 − |z/r|2
log |f (z)| − log  = log |f (re )| · dθ (for |z| < r)

−zρj /r2 + 1 2π 0 |z/r − eiθ |2

j

[1] To verify that such maps stabilize the unit disk, expand the natural expression:

 w + z 2
     
1 −   = |zw + 1|−2 · |zw + 1|2 − |w + z|2 = |zw + 1|−2 · |zw|2 + zw + zw + 1 − |w|2 − zw − zw − |z|2
zw + 1 
 
= |zw + 1|−2 · |zw|2 + 1 − |w|2 − |z|2 = |zw + 1|−2 · (1 − |z|2 ) · (1 − |w|2 ) > 0

4
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

which rearranges slightly to the general Poisson-Jensen formula

2π
ρj − z  r2 − |z|2
Z
X  1
log |f (z)| − log   = log |f (reiθ )| · dθ (for |z| < r, |ρj | < r)

j
−zρj /r + r 2π 0 |z − reiθ |2

The case z = 0 is the general Jensen formula for arbitrary radius r:

ρ  Z 2π
X  j 1
log |f (0)| − log   = log |f (reiθ )| dθ (with |ρj | < r)
j
r 2π 0

3. Hadamard products
The order of an entire function f is the smallest positive real λ, if it exists, such that, for every ε > 0,
λ+ε
|f (z)| ≤ e|z| (for all sufficiently large |z|)

The connection to infinite products is:

[3.1] Theorem: (Hadamard) The genus h and order λ are related by h ≤ λ < h + 1. In particular, one is
finite if and only the other is finite.

Proof: First, the easier half. For f of finite genus h expressed as

Y z  ph (z/zj ) Y z z z2 z3 zh 
f (z) = eg(z) · z m 1− e = eg(z) · z m 1− exp + 2 + 3 + ... + h
j
zj j
zj zj 2zj 3zj hz

the leading exponential has polynomial g of degree at most h, so eg(z) is of order at most h. The order of
a product is at most the maximum of the orders of the factors, so it suffices to prove that the order of the
infinite product is at most h + 1.
The assumption that h is the genus of f is equivalent to
X 1
< ∞
j
|zj |h+1

We use this to directly estimate the infinite product, showing that it has order of growth λ at most h + 1.

We need an estimate on Fh (w) = (1 − w)eph (w) applicable for all w, not merely for |w| < 1. We collect some
inequalities. There is the basic

h |w|h
log |Fh (w)| = log |(1 − w)eph−1 (w) · ew /h
| ≤ log |Fh−1 (w)| + (for all w)
h
As before, for |w| < 1,

1 1 1
log |Fh (w)| ≤ · |w|h+1 + · |w|h+2 + . . . ≤ |w|h+1 · (for |w| < 1)
h+1 h+2 1 − |w|

This gives (1 − |w|) · log |Fh (w)| ≤ |w|h+1 for |w| < 1. Adding to the latter the basic relation multiplied by
|w| gives
1

log |Fh (w)| ≤ |w| · log |Fh−1 (w)| + 1 + |w|h+1 (for |w| < 1)
h
5
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

In fact, the latter inequality also holds for |w| ≥ 1 and log |Fh−1 (w)| ≥ 0, from the basic relation. For
log |Fh−1 (w)| < 0 and |w| ≥ 1, from the basic relation,

|w|h |w|h 1
h+1

log |Fh (w)| ≤ log |Fh−1 (w)| + ≤ ≤ 1+ |w| (for log |Fh−1 (w)| < 0 and |w| ≥ 1)
h h h

Now prove log |Fh (w)| h |w|h+1 , by induction on h. For h = 0, from log |x| ≤ |x| − 1,

log |1 − w| ≤ |1 − w| − 1 ≤ 1 + |w| − 1 = |w|

Assume log |Fh−1 (w)| h |w|h . For |w| < 1, we reach the desired conclusion by

1
h+1 1

log |Fh (w)| ≤ |w| · log |Fh−1 (w)| + 1 + |w| h |w| · |w|h + 1 + |w|h+1 (for |w| < 1)
h h
For |w| ≥ 1 and log |Fh−1 (w)| > 0, from the basic relation

|w|h |w|h
log |Fh (w)| ≤ log |Fh−1 (w)| + h |w|h + h |w|h+1 (for |w| ≥ 1 and log |Fh−1 (w)| > 0)
h h
For log |Fh−1 (w)| ≤ 0 and |w| ≥ 1, from the basic relation we already have

|w|h |w|h
log |Fh (w)| ≤ log |Fh−1 (w)| + ≤ h |w|h+1 (for |w| ≥ 1 and log |Fh−1 (w)| < 0)
h h

This proves log |Fh (w)| h |w|h+1 for all w.

Estimate the infinite product:
Y z  X  z  X  z h+1
log  (1 − ) · eph (z/zj )  = log (1 − ) · eph (z/zj )  h < ∞
   
zj zj zj
 
j j j

1/|zj |h+1 converges. Thus, such an infinite product has growth order λ ≤ h + 1.
P
since
Now the difficult half of the proof. Let hP≤ λ < h + 1. Jensen’s formula will show that the zeros zj are
sufficiently spread out for convergence of 1/|zj |h+1 . Without loss of generality, suppose f (0) 6= 0. From
z  Z 2π
X  j 1
log |f (0)| − log   = log |f (reiθ )| dθ (with |zj | < r)
j
r 2π 0

certainly
ρ  Z 2π
X X  j 1
log 2 ≤ − log   ε − log |f (0)| + rλ+ε dθ  rλ+ε (for every ε > 0)
r 2π 0
|zj |<r/2 |ρj |<r/2

With ν(r) the number of zeros inside the disk of radius r, this gives

ν(r)
lim = 0 (for all ε > 0)
r→+∞ rλ+ε

Order the zeros by absolute value: |z1 | ≤ |z2 | ≤ . . . and for simplicity suppose no two have the same size.
Then j = ν(|zj |) ε |z|λ+ε . Thus,
X 1 X 1 X 1
ε 1 = h+1
|zj |h+1 (j λ+ε )h+1 j λ+ε

6
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

h+1
The latter converges for λ+ε > 1, that is, for λ + ε < h + 1. When λ < h + 1, there is ε > 0 making such an
equality hold.
It remains to show that the entire function g(z) in the leading exponential factor is of degree at most h + 1,
by showing that its (h + 1)th derivative is 0.

In the Poisson-Jensen formula

Z 2π
X  z −z 
j 1 r2 − |z|2
log |f (z)| − log   = log |f (reiθ )| · dθ (for |z| < r)
 
−zzj /r + r 2π 0 |z − reiθ |2
|zj |<r

∂ ∂
application of ∂x − i ∂y annihilates the anti-holomorphic parts, returning us to an equality of holomorphic
functions, as follows. The effect on the integrand is
∂ ∂
r2 − |z|2 −z r2 − |z|2
2 −i = 2 −
∂x iθ
∂y |z − re | 2 (z − re )(z − re ) (z − reiθ )2 (z − re−iθ )
iθ −iθ

−|z|2 + zreiθ − r2 + |z|2 reiθ

= 2 iθ 2 −iθ
= 2 iθ
(z − re ) (z − re ) (re − z)2

Thus, Z 2π
f 0 (z) X 1 X zj 1 iθ 2reiθ
− + = log |f (re )| · dθ
f (z) z − zj z j z − r2 2π 0 (z − reiθ )2
|zj |<r |zj |<r

Further differentiation h times in z gives

 f 0 (z) (h) X (−1)h h! X (−1)h h! · z h+1
j (−1)h (h + 1)! 2π
Z
2reiθ
iθ
= − + log |f (re )| · dθ
f (z) (z − zj )h+1 (z j z − r2 )h+1 2π 0 (z − reiθ )h+2
|zj |<r |zj |<r

We claim that the second sum and the integral go to 0 as r → +∞.

Regarding the integral, Cauchy’s integral formula gives
Z 2π
reiθ
dθ = 0
0 (z − reiθ )h+2
Thus, letting Mr be the maximum of |f | on the circle of radius r, and taking |z| < r/2, up to sign the
integral is
Z 2π Z 2π Z 2π
 M
r
 2reiθ dθ 1  M
r
 rλ+ε
log iθ )|
· iθ )h+2
 h+1
log iθ )|
dθ ε h+1
· − log |f (reiθ )| dθ
0 |f (re (z − re r 0 |f (re r 0

Jensen’s formula gives Z 2π

1
− log |f (reiθ )| dθ ≤ − log |f (0)|
2π 0
Thus, for λ + ε < h + 1 the integral goes to 0 as r → +∞.
For the second sum, again take |z| < r/2, so for |zj | < r
 z h+1
j
 |zjh+1 | |zjh+1 | 1
≤   h+1
 
2 h+1 2 r h+1 h+1 h+1
(z j z − r ) (r − |zj | · 2 ) r (r − |zj |) r
 

We already showed that the number ν(r) of |zj | < r satisfies lim ν(r)/rh+1 = 0. Thus, this sum goes to 0 as
r → +∞. Thus, taking the limit,
 f 0 (h) X 1
= (−1)h h!
f j
(z − zj )h+1

7
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

z
Returning to f (z) = eg(z) · eph (z/zj ) , taking logarithmic derivative gives
Q
j (1 − zj )

f0 X 1 p0 (z/zj ) 
= g0 + + h
f j
z − zj zj

and taking h further derivatives gives

 f 0 (h) X (−1)h h!
= g (h+1) +
f j
(z − zj )h+1

Since the hth derivative of f 0 /f is the latter sum, g (h+1) = 0, so g is a polynomial of degree at most h.
///

4. Appendix: harmonic functions

We recall the mean-value property and Poisson’s formula for harmonic functions. The Laplacian is

∂2 ∂2
∆ = +
∂x2 ∂y 2

and continuously twice-differentiable functions u with ∆u = 0 are harmonic.

[4.1] Theorem: (Mean-value property) For harmonic u on a neighborhood of the unit disk,
Z 2π
1
u(0) = u(eiθ ) dθ
2π 0

Proof: Consider the rotation-averaged function

Z 2π
1
v(z) = u(eiθ · z) dθ (for |z| ≤ 1)
2π 0

Since the Laplacian ∆ is rotation-invariant, v is a rotation-invariant harmonic function. In polar coordinates,

for rotation-invariant functions v(z) = f (|z|), the Laplacian is
 ∂2 ∂2  p ∂ x 0
∂ y 0

∆v = 2
+ 2 f ( x2 + y 2 ) = f (|z|) + f (|z|)
∂x ∂y ∂x |z| ∂y |z|

1 0 x2 x2 1 0 y2 y2 1 0
= f − 3 f 0 + 2 f 00 + f − 3 f 0 + 2 f 00 = f 00 + f
|z| |z| |z| |z| |z| |z| |z|
The ordinary differential equation f 00 + f 0 /r = 0 on an interval (0, R) is an equation of Euler type, meaning
expressible in the form r2 f 00 + Brf 0 + Cf = 0 with constants B, C. In general, such equations are solved by
letting f (r) = rλ , substituting, dividing through by rλ , and solving the resulting indicial equation for λ:

λ(λ − 1) + Aλ + B = 0

Distinct roots λ1 , λ2 of the indicial equation produce linearly independent solutions rλ1 and rλ2 . However,
as in the case at hand, a repeated root λ produces a second solution rλ · log r.

Here, the indicial equation is λ2 = 0, so the general solution is a + b log r.

8
 Paul Garrett: Weierstrass and Hadamard products (September 15, 2019)

When b 6= 0, the solution a + b log r blows up as r → 0+ . Since f (0) = v(0) = u(0) is finite, it must be that
b = 0. Thus, a rotation-invariant harmonic function on the disk is constant. Thus, its average over a circle
is its central value. This proves the mean-value theorem for harmonic functions. ///

[4.2] Remark: The solutions a + b log r do indeed exhaust the possible solutions: given f 00 + f 0 /r = 0 on
(0, R),
∂
r · f 0 = r · f 00 + f 0 = r · − f 0 /r + f 0 = 0



∂r
Thus, r · f 0 is constant, and so on.
With some computation, from mean-value property we will obtain

[4.3] Theorem: (Poisson’s formula) For u harmonic on a neighborhood of the unit disk,
2π
1 − |z|2
Z
1
u(z) = u(eiθ ) · dθ (for |z| < 1)
2π 0 |z − eiθ |2

Proof: Composition with holomorphic

 maps
 preserves harmonic-ness. With ϕz the linear fractional
1 z
transformation given by matrix ϕz ∼ , the mean-value property for u ◦ ϕz gives
z 1
Z 2π
1
u(z) = (u ◦ ϕz )(0) = (u ◦ ϕz )(eiθ ) dθ
2π 0

Linear fractional transformations stabilizing the unit disk map the unit circle to itself. Replace eiθ by
0
eiθ = ϕ−1 iθ
z (e ) Z 2π
1
u(z) = (u ◦ ϕz )(0) = u(eiθ ) dθ0
2π 0
Computing the change of measure will yield the Poisson formula. This is computed by

0 ∂θ0 ∂ iθ0 ieiθ izeiθ (eiθ − z) ieiθ − ize2iθ + ize2iθ − ieiθ |z|2 ieiθ (1 − |z|2 )
ieiθ · = e = + = =
∂θ ∂θ 1 − zeiθ (1 − zeiθ )2 (1 − zeiθ )2 (1 − zeiθ )2

Thus,
∂θ0 1 ieiθ (1 − |z|2 ) 1 − zeiθ eiθ (1 − |z|2 ) 1 − |z|2
= iθ0 = · =
∂θ e (1 − zeiθ )2 eiθ − z (1 − zeiθ )2 |z − eiθ |2
giving the asserted integral. ///

[Ahlfors 1953/1966] L. Ahlfors, Complex Analysis, McGraw-Hill, 1953, second edition, 1966.

[Hadamard 1893] J. Hadamard, Étude sur les Propriétés des Fonctions Entières et en Particulier d’une
Fonction Considérée par Riemann, J. Math. Pures Appl. 1893, 171-215.
[Lang 1993] S. Lang, Complex Analysis, Springer-Verlag, 1993.
[Riemann 1859] B. Riemann, Über die Anzahl der Primzahlen unter ener gegebenen Grs̈se, Monats. Akad.
Berlin (1859), 671-680.
[Whittaker-Watson 1927] E.T. Whittaker, G.N. Watson, A Course of Modern Analysis, Cambridge University
Press, 1927, 4th edition, 1952.