Professional Documents
Culture Documents
Abstract: A class theorem is presented and proved: the complex Fourier transforms of a certain
class of exponential functions have all their zeros on the real line. A class of basis functions is
first considered, and the class is then extended by the method of convolutions.
In this paper we are going to consider the complex Fourier transforms of the exponential kernel
e− q (t ) , where the following conditions apply:
(i) q(t) is an even function of t which is real for all real values of t
(ii) q(it) is either a polynomial which has all its roots on the real line, or the uniform limit on
compact sets in the complex number field of such a polynomial.
It is well-known that (ii) corresponds to q(it) belonging to the Laguerre-Pólya class of functions
(denote by L-P).
Consider the case when there are non-zero roots β n . Replacing t with t/i in equation (1.1)
shows that:
q(t) = c/ i N .t N . e( µ t
2 ) +t ( γ −
∑1 / β n ) / i .∏ (1 + t 2 / β 2n) (1.2)
n
n n
If N is odd then the first part is odd and the second part is even for all real values of t. As q(t) is
an even function in t, the odd part must be zero, forcing cos( γt ) = 0 for all real t, which is
impossible, so we have obtained a contradiction. Therefore N must be even, so the first part is
even and the second odd. So sin( γt ) = 0 for all real values of t which means we must have γ = 0.
2
Denote N by 2m, and let m > 0 for simplicity because then q(0) = 0 . We can now see that c
must be real, because of (i), and that:
q(t) = k. t 2 m . eµ t .∏ (1 + t 2 / β n2)
2
(1.4)
n
In the case when there are no roots, the working above is shortened and we just obtain:
q(t) = k. t 2 m . eµ .t
2
(1.5)
Therefore, let us work with equation (1.4), and take the product to be 1 if there are no non-zero
roots β n . We have shown that if we require q(t) to be an even function of t which is real for all
real values of t, and q(it) to be either a polynomial, or the uniform limit on compact sets in the
complex number field of such a polynomial, such that q(it) has all its roots on the real line, that
this leads to q(t) being expressible in the form of equation (1.4).
Let us also require that k > 0, and that there are positive constants T, α such that for real t, q(t)
> t 2+α for t ≥ T (this condition is equivalent to there being a non-zero term in t 4 or higher in the
series expansion of q(t) about t = 0. For example, if the coefficient of t 4 in this series expansion
is a 4 > 0, then we see q(t) ≥ a 4.t 4 , and so q(t) > t 2+α for t > (1 / a 4)1/( 2−α ) , 0 < α < 2).
∞
Define G ( z ) = ∫ e − q (t ) . eizt dt = R(s,w) + i I(s,w), where z = w - is, and s,w, R, I e ℜ (2.0)
−∞
Proof of Theorem 1
The convolution hn (t ) = g n (t ) o e−t and its complex Fourier transform H n (z ) are then
4
The preceding definition of q(t) ensures that it is strictly increasing above 0 for t > 0. Hence
q(t) is a 1:1 function on (0, ∞ ), and for any natural number n, there is exactly one real root of
the equation q(t) = n in the interval (0, ∞ ). Call this root λ n (by the evenness of q, we see q(-
λ n ) = n as well ). We also note that as q(t) → ∞ as t → ∞ , it follows that λ n → ∞ as n → ∞
(each λ n is finite).
n
Define g n (t ) = χ [ −λ n ,λ n ] (t ) . y n (t ) where y n (t ) = ( ) . (1− q (t ) / n)n+1 for all real t (2.1)
n +1
3
where χ [ a ,b ] (t ) = 1 a≤t ≤b
0 otherwise
∞
Now also define a sequence of functions G n ( z ) = ∫ g n(t ). eizt .dt (2.2)
−∞
In 1927 (reference I), Professor Pólya proved that if F(t) is defined for real values of t and
( 2 +α )
equals F (−t ) , and there exist positive real constants A, α such that F (t ) < A e− t for all real
∞ ∞
values of t, then if all the zeros of ∫ F (t ). eiztdt are real, so are the zeros of ∫ φ (t ).F (t ). eizt.dt ,
−∞ −∞
provided that φ (it ) is either a polynomial with only real roots, or the uniform limit on compact
sets of such. Call this result (R1).
∞ λn
We see that ∫ χ [ − λ n , λ n ] (t ). eizt dt . = ∫ eizt .dt = [eizt /(iz )]tt ==λ−λn = ( eiz λ n − e−iz λ n )/(iz)
n
−∞ −λ n
which equals zero if and only if e = ei.2 kπ , for any integer k, in other words only when
2iz λ n
z is real and equal to k π / λ n . Now set χ [ −λ n,λ n ] (t ) = F(t) in (R1) above, we see that χ [ −λ n,λ n ] (t )
( 2 +α )
has compact support, so it easily obeys the condition F (t ) < A. e− t for some positive
constants A, α . Recalling that L-P is closed under differentiation, we observe that as q(it) lies
in L-P, then so does q ' (it ) , and so q ' (it ) is either a polynomial with only real roots, or the
uniform limit on compact sets of such, as in (R1) above.
∞ λn
Hence by (R1), we see that ∫ q ' (t ). χ [ −λ n ,λ n ] (t ). eizt dt = ∫ q' (t ). eiztdt only has real zeros
−∞ −λ n
∞ λn
Now consider ∫ (1− q (t ) / n)1. χ [ −λ n ,λ n ] (t ). eizt dt = ∫ (1− q (t ) / n)1. eizt .dt
−∞ −λ n
λn q' (t ) eizt
Integration by parts shows that this is equal to [(1− q (t ) / n)1.eizt / iz ]λ−λn - ∫ − . .dt
n
−λ n n iz
1 λn
=0+ . ∫ q' (t ).eizt dt and by the preceding result has only real zeros.
izn −λ n
∞
Now set F(t) = (1− q (t ) / n)1. χ [ −λ n,λ n ] (t ) . By (R1) we see that ∫ q ' (t ). (1− q (t ) / n)1. χ [ −λ n,λ n ] (t ). eizt dt
−∞
λn
= ∫ q' (t ). (1− q (t ) / n)1. eizt dt has only real zeros.
−λ n
∞ λn
Consider ∫ (1− q (t ) / n)2 . χ [ −λ n,λ n ](t ). eizt .dt = ∫ (1− q (t ) / n) . e .dt .
2 izt
−∞ −λ n
Again, integration by parts shows us a useful result, namely that:
λn
λ − 2 λn
∫ (1− q (t ) / n) . e .dt = [(1− q (t ) / n)2.eizt / iz ]−λ n - . ∫ q '(t ).(1− q (t ) / n)1. eizt .dt
2 izt n
−λ n izn −λ n
2 λn
∫ q ' (t ). (1− q (t ) / n) . e .dt which by the above also has only real zeros.
1 izt
=0+
izn −λ n
∞
Continuing this argument shows that ∫ (1− q (t ) / n)( n+1) . χ [ −λ n,λ n ]. eizt .dt and hence G n (z )
−∞
4
λn ( n +1)
n ∞ n
= ( ). ∫ (1− q (t ) / n)( n+1) . χ [ −λ n,λ n ]. eizt .dt = ∫ ( ).(1− q (t ) / n) . eizt .dt has only real roots.
n + 1 −∞ −λ n n +1
∞ ∞ ∞
≤ ∫ (e−q (t ) − g n (t )). eizt .dt ≤ ∫ e−q (t ) − g n (t ) e z t .dt ≤ ∫ e−q (t ) − g n (t ) . eMt .dt
−∞ −∞ −∞
∞
≤ 2 ∫ e−q (t ) − g n (t ) . eMt .dt
0
t1 ∞
therefore G ( z ) − G n ( z ) ≤ 2( ∫ e−q (t ) − g n (t ) . eMt .dt + ∫ e−q (t ) − g n (t ) . eMt .dt ) for some t1 > 0 (2.4)
0 t1
t1
Call 2 ∫ e−q (t ) − g n (t ) . eMt.dt = J n (2.5)
0
∞
and 2 ∫ e−q (t ) − g n (t ) . eMt .dt = K n (2.6)
t1
Regarding K n , we begin by noting that when x < 1, the series expansion about x = 0 of ln(1-x)
x2 x3 x4 q (t )
is -( x + + + + ..... ) ≤ − x , and replacing x with yields the relation:
2 3 4 n
Considering the case t = λ n shows that this inequality holds on the closed interval [0, λ n ]
We also observe that 0 ≤ (1-q(t)/n) ≤ 1 for t ∈ [0, λ n] and so (1− q (t ) / n)n ≥ 0 . We also note that
n
0< < 1 for n > 1.
n +1
Multiplying the left hand side of equation (2.4) by (1-q(t)/n) and n/(n+1), we see that:
n
0≤ . (1− q (t ) / n)n+1 ≤ e− q (t ) for t ∈ [0, λ n] (2.8)
n +1
5
We can now see that if t ∈ [0, λ n] then both e− q (t ) − g n (t ) ≥ 0 and g n (t ) ≥ 0 , and hence that
∞
K n ≤ 2 ∫ e−q (t ). eMt .dt (2.9)
t1
As q(t) increases more quickly than t 2+α for t ≥ T , the integral in (2.9) exists for any fixed
finite value of M, and we can make the integral less than ε / 2 by choosing t1 to be sufficiently
large. So t1 is now chosen.
n
Regarding J n , and recall that y n (t ) = e−q (t ) − ( ). (1− q (t ) / n)n+1 . We have just seen that
n +1
y n (t ) ≥ 0 for t ∈ [0, λ n] , and we note that y n ' (t ) = − q ' (t ).(e− q ( t ) − (1− q (t ) / n)n) . From equation
(1.5) we see that q ' (t ) is zero for t = 0 and strictly positive for t > 0. By equation (2.7),
e− q ( t ) − (1− q (t ) / n) ≥ 0 and so y n ' (t ) ≤ 0 for t ∈ [0, λ n] .
n
Now pick n > q( t1 ), so λ n > t1 . As y n ' (t ) ≤ 0 and y n (t ) ≥ 0 for t ∈ [0, t1] , it follows that the
maximum value of y n (t ) on the interval [0, t1 ] is y n (0) = 1/(n+1).
t1
t1 1 eMt
J n ≤ 2∫ ( ). e .dt = 2 [
Mt
] = 2 ( eM t1 -1)/(M(n+1)) < 2 eM t1 /( Mn)
0 n +1 M ( n +1) 0
Choosing any n > Max[q( t1 ), 4 eM t1 /(M ε )] ensures that the above conditions are satisfied and
that J n < ∈ / 2 and by (2.4) that G ( z ) − G n ( z ) < ε / 2 + ε / 2 = ε for all z such that z ≤ M , and
as ε is arbitrarily small, we have proved the uniform and absolute convergence of G n ( z ) to
G(z).
……………………………………………………..……………………………………………..
Now let us consider the convolutions hn (t ) = g n (t ) o e−t and h(t) = e− q (t ) o e−t defined by :
4 4
∞
hn (t ) = ∫ g n (v). e−(t −v ) .dv
4
(3.0)
−∞
∞
and h(t) = ∫ e−q (t ). e−(t −v ) .dv
4
(3.1)
−∞
∞
We see that H n ( z ) , the complex Fourier transform of hn (t ) defined by H n ( z ) = ∫ hn(t ). eizt .dt
−∞
is then equal to ∫∫ ( g n (v). e−(t −v ) ).eizt .dvdt and after substituting v = x and t = x + y
4
ℜ2
and applying Fubini’s Theorem, we see that:
6
∞ ∞ ∞
H n (z ) = ( ∫ g n ( x). eizx .dx).( ∫ e− y . eizy .dy ) = G n ( z ). F 4 ( z ) where F 4 ( z ) = ∫ e− y . eizy.dy
4 4
(3.1)
−∞ −∞ −∞
Of course equation (3.1) is just the standard result that the Fourier transform of the convolution
is the product of the individual Fourier transforms.
We saw earlier (R2) that G n ( z ) has only real zeros, and in 1923 (ref II) Pólya proved that
∞
for k ≥ 2 , all the zeros of F 2 k ( z ) = ∫ e− y . eizy .dy lie on the real line Im(z) = 0, and that there
2k
−∞
are an infinite number of them.
Hence F 4 ( z ) , and consequently H n ( z ) , have only real zeros, and further more an infinity of
them. Call this result (R3).
The order of H n ( z ) :
In the same paper previously mentioned (I), Pólya also proved that the order of the complex
Fourier transform of f(t), where f(t) is a real function of t such that f(t) = f (−t ) and there exist
2 +α
real positive constants A, α such that f (t ) < A. e− t for all real values of t, is at most
(2+ α ) /(1 + α ) < 2 . Call this result (R4).
n
Therefore, let us set f(t) = g n (t ) = χ [ −λ n,λ n ] (t ).( ). (1− q (t ) / n)n+1 . Outside the interval
n +1
[ − λ n , λ n] we observe g n (t ) = 0, and inside, by equation (2.8), 0 ≤ g n (t ) ≤ e− q (t ) .
2 +α
As q(t) > t 2+α for t ≥ T, it follows that e− q (t ) < e−t for t ≥ T, and by setting
2 +α 2 +α
A = 2 Sup[ e− q (t )+t ] on the interval 0 ≤ t ≤ T , we see e− q (t ) < A. e− t and so by (R4) above
we conclude the order of G n ( z ) is at most (2+ α )/(1+ α ) < 2
Incidentally, setting f(t) = e− q (t ) instead of g n (t ) shows us, by the same reasoning as that in the
above paragraph, that the order of G(z) is at most (2+ α )/(1+ α ) < 2 as well.
Setting f(t) = e−t shows that the order of F 4 ( z ) is at most 4/3. (Pólya mentions in (II) that the
4
2 +α
ρ 2 +α 4
≤ B e( c1+c 2 ) z ,where ρ = Max[
4/3
H n ( z ) = G n ( z ). F 4 ( z ) ≤ B ec1 z 1+α . ec 2 z , ] , for all z
1+ α 3
Note: another way of evaluating an upper limit for the order of H n ( z ) is to observe that
λn λn
hn (t ) = ∫ y n(t ). e−(t −v ) .dv and hn (t ) ≤ ∫ y n (t ) . e−( t −v ) .dv and use the result that
4 4
−λ n −λ n
7
n
y n (t ) = ( ). (1− q (t ) / n)n+1 ≤ 1 for t ∈ [− λ n , λ n] so we must have
n +1
λn t +λ n
hn (t ) ≤ ∫ e−( v−t ) .dv = − −
∫ e v .dv and if we consider t > 2λ n then hn (t ) ≤ 2 λ n .Max[ e v ]
4 4 4
−λ n t −λ n
2 +α
hn (t ) falls off more quickly than e−t , and so satisfies the conditions in (R4).
We can then set f(t) = hn (t ) and the result that the order of H n ( z ) is less than 2 follows.
As we saw in (R3) earlier, H n ( z ) has only real zeros, and an infinity of them, and we have just
seen in the preceding section that the order of H n ( z ) is less than 2.
By Hadamard’s factorisation theorem (see for example III), H n ( z ) has the product
representation:
∞ ∞
H n ( z ) = c. z N . eγz . ∏ (1 + z / wk ) e− z / wk = c. z n . e(γ −∑1/ w j ) z . ∏ (1 + z / wk ) (3.2)
k =1 k =1
We will now see that this expression can be reduced to a rather simple form, using the same
reasoning as in a previous paper by the same author (IV).
From the definitions of G n ( z ) and F 4 ( z ) in (2.2) and (3.1), we see that they are both even
functions, and therefore so is H n ( z ) = G n ( z ). F 4 ( z ) . Consequently, for every non-zero root
wk of H n ( z ) = 0, there is another root of opposite sign - wk . So we see that ∑1 / w j = 0.
j
Equation (3.2) then reduces to:
∞
H n ( z ) = c. z N . eγz . ∏ (1 − z 2 / α 2r ) (3.3)
r =1
where w1 = α 1 , w2 = − α 1 , w3 = α 2 , w4 = − α 2 etc and the α r are the positive roots of H n ( z ) = 0 .
λn n ∞
. (1− g (t ) / n)n+1.dt ).( ∫ e−t .dt ) and 1-g(t)/n > 0 when t < λ n ,
4
As H n (0) = G n (0). F 4 (0) = ( ∫
−λ n n + 1 −∞
we see H n (0) > 0 and so N = 0 and c > 0. Therefore (3.3) reduces to:
∞
H n ( z ) = c. eγz . ∏ (1 − z 2 / α 2r ) (3.4)
r =1
∞
H n ( z ) = c. ∏ (1 − z 2 / α r2) , where c > 0 (3.5)
r =1
8
Recalling that the order of F 4 ( z ) is at most 4/3, and that F 4 ( z ) has only real zeros, and
furthermore an infinity of them, the same reasoning leads to the conclusion that:
∞ ∞
F 4 ( z ) = c0 . ∏ (1 − z 2 / β k ) where c0 = ∫ e−t .dt > 0
2 4
(3.6)
k =1 −∞
∞
c. ∏ (1 − z 2 / α r2)
H n ( z) r =1 c 2
Gn ( z ) = = ∞
= .∏ (1 − z 2 / (α *j ) ) (3.7)
F 4 ( z) c0 . ∏ (1 − z 2 / β k )
2 c0 j
k =1
where the set {α *j}j≥1 is the set {α r}rr ==1∞ with the set {β k }kk ==1∞ removed ie {α r}rr ==1∞ \ {β k }kk ==1∞ .
……………………………………………………………………………………………………
In the same paper by this author previously mentioned (IV), we saw that for
∞
F 2 n ( z ) = ∫ e−t . eizt dt , where z = w - i σ ,
2n
−∞
1 σ 2m
F 2n ( z ) =
2
∑ . A2 m, 2 n ( w) (3.8a)
2 m≥0 (2m)!
∂2 m
and so that A2 m, 2 n ( w) = ( −1)m .2. [F 2n (u + w).F 2n (u − w)]u =0 (4.0a)
∂ u2m
∂2m
and we also saw that T N ,m ( w) = (−1)m .2. [P N (u + w).P N (u − w)]u =0 ≥0 (4.1a)
∂ u2m
N ∞
where P N ( z ) = c. ∏ (1 − z 2 / α 2r ) and F 2 n ( z ) = c. ∏ (1 − z 2 / α 2r ) (4.2a)
r =1 r =1
2 1 σ 2m
Gn ( z ) = .∑ . B2 m,n ( w) (3.8b)
2 m≥0 (2m)!
∂2m
and so B 2 m,n ( w) = ( −1)m .2. [Bn (u + w).Bn (u − w)]u=0 (4.0b)
∂ u2m
∂2m
with T N ,m ( w) = (−1)m .2. [P N (u + w).P N (u − w)]u =0 ≥ 0 (4.1b)
∂ u2m
c N c
and P N ( z ) = . ∏ (1 − z 2 / (α *r ) ) and G n ( z ) = . ∏ (1 − z 2 / (α *r ) )
2 2
(4.2b)
co r =1 c0 r ≥1
Whether the product representation for G n (z ) in (4.2b) above is infinite or not, we see that
B 2 m,n ( w) ≥ 0 .
We note that for any particular fixed values of n and w, it is not possible for B2 m,n ( w) to be zero
for all values of m, otherwise by equation (3.8b), there would be a continuous line of zeros of
G n (z ) along the line w = constant, and this would lead to a contradiction as the zeros of
holomorphic functions are isolated. Therefore, for any particular fixed values of n and w, at
least one of the B2 m,n ( w) must be greater than zero.
Hence by equation (3.8b) we see that if we pick a fixed value of w on the w-axis and travel in
the direction of increasing σ , then G n ( z ) must increase monotonically.
2
QED.
……………………………………………………………………………………………………..
∞
By way of example, if we set q(t) = a (cosh(t)-1), we see that e− a G(z) = e−a ∫ e−a (cosh(t )−1). eizt .dt
−∞
∞ 2
at 2 ∞ t 2 ) , which is of the
= ∫ e−a.cosh(t ). eizt .dt and we note that q(t) = 2a (sinh(t / 2))2 = . ∏ (1+
−∞ 2 m=1 4π 2 m 2
form in equation (1.4). This case was covered in Pólya’s 1927 paper (I), but it was proved
directly, and not by a class rule.
……………………………………………………………………………………………………..
We now turn our attention to the convolution of two basis functions e−q1(t ) and e−q 2(t ) defined by:
∞
c[ q1,q 2 ] (t ) = ∫ e−q1( v ). e−q 2( t −v ) .dv (4.3)
−∞
10
∞
We see c[ q1,q2 ] (−t ) = ∫ e−q1(u ) . e−q 2( −t −u ).du and as q 2 (t ) is an even function, this equals
−∞
∞ ∞
− q1(u )− q 2(t +u )
∫e .du which on replacing u with –v becomes ∫ e−q1( v )−q 2(t −v ).dv , in other words
−∞ −∞
c[ q1,q 2 ] (t ) . We have proved therefore, that the convolution is an even function of the variable t.
We note that c[ q1,q 2 ] (t ) is always strictly positive as the integrand in (4.3) is strictly positive.
Now, for convenience, consider t > 0 and differentiate the expression in (4.3) with respect to t.
We obtain:
∞
− q1( v )− q 2(t −v )
c[ q1,q 2 ] ' (t ) = ∫ − q2 ' (t − v). e .dv which after the change of variable u = t – v becomes
−∞
∞
−q1(t −u )− q 2(u )
∫ − q2' (u ). e .du and this expression may be broken up as:
−∞
∞ 0 ∞
( ∫ + ∫ )(− q2 ' (u ). e−q1(t −u )−q 2(u )).du = ∫ q 2' (u ). e−q2(u ) .(e−q1(t +u ) − e−q1(t −u )).du
0 −∞ 0
t ∞
so c[ q1,q 2 ] ' (t ) = ∫ q 2 ' (u ). e−q 2(u ) .(e−q1(t +u ) − e−q1(t −u )).du + ∫ q2 ' (u ). e−q 2(u ) .(e−q1(t +u ) − e−q1(u −t )).du (4.4)
0 t
For 0 <u <t we see t + u > t - u > 0 and so e−q1(t +u ) − e−q1(t −u ) < 0 as q1 ( x) is strictly increasing
on (0, ∞) . For u > t > 0 we see u + t > u – t > 0 and so e−q1(t +u ) − e−q1(u −t ) < 0 as well.
It follows that both integrals in (4.4) are negative, and so we conclude that:
We can see therefore that the plot of the convolution c[ q1,q 2 ] (t ) against t is a symmetric bell-
shaped curve which looks similar to that of e− q (t ) where q(t) is as defined in (1.4).
By the standard result concerning the complex Fourier transform (CFT) of convolutions, the
zeros of the CFT of c[ q1,q 2 ] (t ) are exactly those of the individual CFT’s of e−q1(t ) and e−q 2(t ) .
Therefore all the zeros of the CFT’s of the convolutions as defined in (4.3) are real.
We therefore define our class of exponential kernels whose CFT’s have only real zeros as the
set e− q (t ) where q(t) is as defined in (1.4), and their convolutions as defined by (4.3)
…………………………………………………………………………………………………….
11
References
(I) Über trigonometrische Integrale mit nur reellen Nullstellen, J.Reine Angew. Math., 158
(1927), 6 – 18
(II) On the zeros of an integral function represented by Fourier’s integral, Messenger of
Math. 52 (1923), 185 – 188.
(III) The Theory of Functions, E.C.Titchmarsh, Oxford University Press, Oxford, 1939
(IV) On the Geometric Interpretation of the Complex Fourier Transforms of a Class of
Exponential Functions, arXiv:0901.3328v1 [math.cv] (2009)