Integral Representations For Jacobi Polynomials and Some Applications
Integral Representations For Jacobi Polynomials and Some Applications
Integral Representations For Jacobi Polynomials and Some Applications
Submitted by S. Karlin
where oi > ,¶ > - 4, - 1 < x < 1, and dp(y) is a positive measure which
depends on x but not n. For p = - $ this is a result of Seidel and Szasz.
Similar results are obtained for Jacobi polynomials and the positivity of
certain sums of ultraspherical and Jacobi polynomials is obtained.
1. INTRODUCTION
(1.1)
which is as explicit an expression as one can hope for and the functions xk that
are used in the expansion are the most elementary of all functions. Unfortu-
* This work was sponsored by the Mathematics Research Center, United States
Army, Madison, Wisconsin, under Contract No.: DA-31-124-ARO-D-462.
411
409/26/2-12
412 ASKEY AND FITCH
nately (1 .I) is not a useful representation for many problems. For instance,
Pn(x) satisfies
1P,(X) i :< 1, -1.<X<l (14
and (1.2) is far from obvious from the representation (1.1).
There is another series expansion for Pn(x) from which (1.2) follows
immediately.
where
(1.4)
This expansion is a very useful one and it shows why it is useful to have
expansions with positive coefficients; or more generally, in integral expansions
it is useful to have a positive kernel. Often we are even willing to pay the price
of expanding in terms of a more complicated set of functions, if at the same
time we have a positive kernel.
In this paper we will be concerned with Jacobi polynomials and our
expansions will be series and integral expansions of Jacobi polynomials in
terms of other Jacobi polynomials. The common feature of all of the expan-
sions will be the positive kernel that occurs.
P(a~B)(~),
n the Jacobi polynomial of degree n, order (01,/3), is defined by
f$“*-qcos6)
pp;,(l) - = cos
ne U-6)
I$ *ycos
e) sin(n + 1) e (1.7)
ly l-y 1) = (n + 1) sin e
~~~~~~~~~~~
e) sin(n + 4) e
p-t)(l) = (2n + 1) sin(ej2) (l-8)
(1.11)
k=O
or equivalently
1 p(-lrB)(y)
a$+ (1 - x2)+= I^, p+J(l) (1-Y”>-* dy. (1.14)
Ll
Here we have P, on both sides of the equation and because of this we can
find a very general result for Jacobi polynomials which contains (1.17). This
result was found by Bateman [7] and is fairly well known.
Pca++qx) _ qa + p + 1) l a p:-@‘(Y)
(1 -LX)n+u p~+LB-u)(l)
Iz - & + 1) Q) J^3c(1 - y) pc$“‘(l) (y - e-l 4%
(1.18)
We could continue giving more results for trigonometric functions and then
state the corresponding formulas for Jacobi polynomials. We will content
ourselves with one more. It is easy to check that
sinnp, r
-= d9. (1.19)
n sm
(1 + xY+p pbd3+P)(X)
(1 - qz+fJ+1 n
_ 2wQfi + B + P + 1) = (1.23)
qn + p + 1) II(/L) s -1
n Ppyx) > o
--I<x<l, n = 0, l,... (1.24)
k;. P;,“‘(l) ’ ’
for 01= /3 3 0 and remarked that it also holds for some Jacobi polynomials.
Feldheim knew Bateman’s integral and so he must have known (1.24) for
012 IBI- He probably also knew (1.24) for CL> - /3 - 4 when
- 1 < /3 < - 8 . We will show that (1.24) also holds for 01> 0 for
- g < jI < 0 and for a slightly larger region than 01> - /3 - 4 for
-1.:/I<-&.
(1.19) can be used to give an easy proof of the following theorem of
Turan [31]. See [4].
416 ASKEY ASD FITCH
cN sin- nx > 0,
n=l n
O<x<a (1.25)
compares favorably with Landau’s inductive proof [23], which has been the
easiest proof up to this time.
Finally as one more application let us mention the following generalization
of a theorem of Seidel and S&z [27].
for ‘yi > max& , - 1 - &) Seidel and SzPsz prove (1.26) for 01= /3 and
k = 1.
2. HYPERGEOMETRIC FUNCTIONS
We have been unable to find a proof of (1.23) which uses just Jacobi
polynomials. For this reason we must say something about hypergeometric
functions.
(2.1)
See [5], [lo] or [28] for a general discussion of aFr’s and of more general
series of this type. The Jacobi polynomial P:*@‘(x) can be written as a 2Fl .
~=~l(-n,n+cr+~+l;cx+l;~j. (2.2)
This is an integral which was found by Bateman [7] and has since been
rediscovered by a number of people. The proof is trivial. Just expand
F(a, b; c; y) in the series (2.1) and use Euler’s formula for the beta function
1
t=-l(l - t)b-1 & = W r(b)
(2.5)
I0 r(a + b) ’
F(a, b; c; x) = (1 - x)-“F (c - a, b; c; 5) ,
had very little contact with hypergeomctric functions. Of the three books wc
referred to above only [lo] even gives Bateman’s integral and the derivation
of (2.X) and (2.9) given there, which is due to Jacobi, is much more compli-
cated than Kummer’s earlier proof.
We will need two other integrals connecting hypergeometric functions
with different parameters.
(2.10)
This can be proved in the same way that Theorem 2.1 was proved, i.e.
expand and integrate term by term.
The other is a more complicated formula.
=T(csy
+PL)
WUCL) 0
(y - x)-l (1 - X)-~-Q xc9+z, b; c; x) dx. (2.11)
= m + PI (2.12)
jly"-'(1 - y)“-l (1 - xy)-“F (a, c - 6; c; --&j dy.
m) w 0
gave Bateman’s original paper, which seems to have been the first. One
reason these results have been forgotten and then rediscovered is that most
of the people who proved them didn’t give any applications of their results.
This seems to be the reason that these results have not been included in books
on hypergeometric functions. We give one application now of Theorem 2.3
to show that these theorems can be used to obtain interesting results. If we
let p == 1 we have
(2.13)
(2.14) was found by Bailey [6]. Th e ab ove proof, including the proof of
Theorem 2.3, is easier than his proof.
3. JACOBI POLYNOMIALS
P$qx) l-x
=F -rz,n+ol+/3+1;~~+1;--- . (3.1)
PJyy 1) ( 2 i
PcM++qx) _ r(a + p + 1) l
(1 - x)oL+LL
p~a+u,"-p'(l) - Qa + 1) &) jx (1 -Y>" g$$Y - XY dY*
12
(3.3)
420 ASKEY AND FITCH
In (3.3) and in what follows we will have p > 0 and - 1 < x, y < 1. If we
use (3.2) we have
pka+qx)
(1 + x)4+* p+3+u)(- 1)
(1 + X)n+a+8p~-“.6’(x)
r(fi+a+B+l)
=r(n+ a+B+ 1- P)r(P)sz-1(1 + y)n+=+4--p z+@(y) (x y)p-1 dy.
(3.5)
Again using (3.2) we have
(1 _ X)n+a+8 p~‘8-“yx)
(1 - x)a+u p(a+u,B)(x)
(1 + X)n+a+l @+q)
and
(1 + X)BfU +a+qx)
(1 - q+i3+1 p@++ 1)
(3.5) and (3.7) include important results for ultraspherical polynomials in the
cases p = 5 4 . Ultraspherical polynomials are Jacobi polynomials with
equal parameters (y.and /3. However, for historical and some practical reasons,
they are normalized differently. We define C,yx) by
P2d(x) p($-*@3 - 1)
---= =- c2 f(X) (3.10)
P&y 1) p;‘-i’(‘) c;;+(l)
(3.11)
If we choose ,13= f 4 in (3.5) and (3.7) and use (3.10) and (3.11) and then do
some tedious calculations we see that the following results hold.
WV)
= r(h)r(V- A)slot”+2A-lCn”(Xt) (1 -
c7an(x) ta)Y-A-1dt, v > A. (3.12)
Also (3.7) holds for all values of the variable X, - 1 < x < 1, while (3.13)
only holds for 0 < 0 < n/2 (or 0 < x < I), and a different formula holds for
r/2 < 0 < V. The reason for this difference seems to be tied up with the
monotone behavior of the zeros of Jacobi polynomials Pz+,B)(x) asp increases
and the fact that the zeros of E’~+s~+~) (x) tend monotonically to zero and not
one of the end points. Both Feldheim and Vilenkin obtain the following
result which is equivalent to (3.13). It h o Id s f or all values of the variable
but now the argument of Cnn is so complicated that one is tempted to look for
another formula. (3.7) is this other formula.
C,“(COS e) e/2
+ ~) +
= r(h 2qv r(v 3 _ h) o sinzA v cos2~-2~-1(1 - sin2 ~9cos2 0)“‘2
C,“(l) I
CnA(cos 0(1 - sin2 19cos2 v)-+) ds,
7 v >A, o<eGn. (3.14)
GA(l)
In addition to (1.19) and (1.20) we should mention a few other special cases of
our results. These can be checked directly and some of them are even useful.
d sin& cos(n - 1) e
(3.16)
de n(cos ey = lcos ep+l
d
~09 e cosne = - n COS~+~
8 sin(n + 1) 8. (3.17)
28
Once you see the pattern of results of this type they are easy to write down.
These expressions are probably not new but they are far from well known
and they can be extremely useful. As we remarked earlier, (1.19) can be
used to prove an interesting theorem of P. TurPn. Later, when we give
some generalizations of the following inequality of Lyness and Moler [24]
we will show how (3.16) gives a very simple proof of their inequality.
We now return to (3.8) and give a different proof of it. There is one thing
wrong aestetically with (3.8) and that is the occurence of TV in the term
(1 - y)-*-a-u-l on the right hand side. Ideally what the formula should be is a
given function gB(y) on the right which depends on fi and on the left we
should have gB+U(x). If we can achieve this then we should be able to prove
(3.8) by taking Laplace transforms. We set
r-l s-l
y=r+l’ x=ss-1-
JACOBI POLYNOMIALS 423
s-l
s6+y1
+ s>”Pfy+@)
(s+l ) 1
s P(l + Y)nP?*s) &-$ (s - ry dr
qn+p+P++) =- %4 f 0 w+p+ 1)
(3.19)
s > 0. If we set
s-l
&3(s) =
sfl(l + S)nP$S) s+l
( 1 (3.20)
J-b + p + 1)
(3.21)
A(4
-- 1 44 (3.23)
.$3+u - $ * -yT ’
which is trivially true. Then the uniqueness theorem for Laplace transforms
would give (3.19), which is equivalent to (3.8). But we know exactly what
is. Since
which is what we wanted to show, since the only way /I enters into this formula
is x-B. However we can simplify (3.25) and obtain an integral which connects
Jacobi and Laguerre polynomials. If we let j = n - k and sum on j we see
(3.25) is
(-n-c&-n), n (-n)jXj
(3.26)
xfi+B+l( l)n ( l)n zo (1 + a)jj! *
ds = LW (3.29)
i 0 F(n+B+l) Xn+B+1’
(3.29) is not new. A similar formula was obtained by J. Chaudhuri [9] and it is
undoubtedly contained in one of the formulas in ([12], Section 4.21). One
interesting fact is that (3.29) is a consequence of (3.8). There are some other
integrals involving Laguerre polynomials and Jacobi polynomials which we
will now list. Then we will show how they can all be obtained from the
integrals we have for Jacobi polynomials.
The only well known integral is
xm+uL;+yx) 1 z (x - Y>“-‘YOLL?3Y) dY
=- p >o. (3.30)
r(n+a+p++) I
h-4 o qn+a+l) ’
(3.30) and (3.31) are ordinary fractional integrals and (3.29) and (3.32) are
Laplace transforms, or fractional integrals of infinite order if you want to
think of them in this way. It is a fruitful way, as it leads one to see where gaps
are in the literature and sometimes even leads to proofs.
These four formulas all follow easily from
and the formulas we derived at the beginning of this chapter. We omit the
details.
Similarly it is possible to derive the Sonine integrals for Bessel functions
from the above results. Here we use
1
~“+Y+1Ju+“+l(4 = pqp + 1) s o’ P’+~]&) (z” - y2)“ dr. (3.36)
Another integral found by Sonine which is extremely useful and is not given
in most of the standard references is
p-u+1 m
~-“I&) = qv _ CL) z y-“+lJu(y) (y2 - z~)“-~-~ dy, (3.37)
I
-l<vv<<2v+*. (3.37)
There are two formulas for the classical orthogonal polynomials which
contain (3.38), and they are both formulas for Laguerre polynomials.
426 ASKEY AX;D FITCH
One is
o+p+ 1)
p’*6’(x) = F(y - a) T(n + j3 + y + 1)
n r(n-k+y--)r(n+K+B+r+l)
xc k=o r(n - k + 1) I+ + k + 01+ B + 2)
(3.40)
m r(n+OL+l)r(n+k+B+y+l)r(k--++---or)
=2 r(n + 1) I+ + k + a+ B + 2) r(k - n + 1)
x m + 1) P + Y + B + 1) pjc”qX) (1 _ x) , (3.41)
r(k+y+1)Q--)2Y-”
JACOBI POLYNOMIALS 427
where we need a: > y/2 - 2 to have convergence of the series. This is given
in [3].
For p = & + , (3.40) and (3.41) contain two very useful formulas for ultra-
spherical polynomials. We use (3.10) and (3.11) in (3.40) and (3.41) and
obtain
r(v)
Cnn(x)= r(h) r(A - v)
k=O
(3.42)
(3.42) was found by Gegenbauer [20]. It contains the following well known
formula.
n @)k @>n-k
CnA(Cos e, = ,c, k!Q _ k)!
cos(n - 2k) 0. (3.44)
f(*)k sin(n + 2k + 1) 0,
(3.46)
409126/2-13
428 ASKEY AND FITCH
h(1 - Y2)
= go(u + A) C,*(X) ~-7%
> 0 for h > 0. (3.49)
(1 - 2XY + Y2)Afl
where
4. APPLICATIONS
(4.1)
Since
p$qx) = 1,
Py’(x) = ( O1 + ; + 2 ) x + E-g , (4.2)
n sin@ + 8) 19
c
k=O sin e/2
= to sin(;;;;) e’2 = l -;;;:Bf2 l) e > 0. (4.4)
n ~~~~~~~~~
e)
>, 0, o<e+, a > 0. (4.6)
k;o fy*“W
430 ASKEY AND FITCH
Since
we have
(4.8)
Applying the Bateman integral (3.3) to (4.6) and (4.8) we have (4.1) for 013 j3,
or>-/3iifp>-$andcz>-/z2-&if-l<<<-$.SinceFeldheim
knew the Bateman integral he undoubtedly knew this result. Actually for
p = - 4 we have a stronger inequality than (4.1). For 01= 4, p = - 4
recall (4.4), which was stronger than (4.1). (4.4) can be written as
11pp-u.-t+v)(x)
& ppL-Y.-*+Y)(l) (1 + XP-”
is decreasing in k we can sum the right hand side of (4.11) by parts and show
that the left hand side is nonnegative. A simple calculation show this is
equivalent to
(4.12)
k=O
h#O, (4.13)
which was proven by FejCr [14] for 0 < h < $ and by Szego [29] for
- 4 < X < 0. Another proof of (4.12) and (4.13) for 0 < X < + is to apply
(3.12) to the upper end point result, each of which we know. (4.12) can not be
extended to larger values of ;\. For COA(x)= 1 and CsA(x) = 2X(1 + X) x2 - h
so CoA(x) + C,A(x) = 1 - h + 2h(l + h) x2. For x = 0 this is negative if
x > 1.
A similar argument to that given above shows that (4.1) holds for
- 1 < /3 < - $ if 01is to the right of the hyperbola
For /3 = - 1, this hyperbola has 01= (- 1 + 1/3)/2, which is less than the
4 we had previously.
We now consider the following result of Seidel and Szhz [27] and various
generalizations of it.
Let h > 0. Then
Cn"(‘4
-= ff cosne d&B), h > 0, n = 0, l,... . (4.15)
GA(l) *i 0
Here &L(B) is a positive measure of total mass one which depends on x but
is independent of n.
Observe that (4.15) is still another integral representation and that it is
not contained in the previous section. Actually, while it isn’t contained in our
previous work and we can not find dp(8) explicitly, it is possible to reprove
(4.14) using results in Section 2 and even to find some extensions of it.
GW al CnU(Y) (4.18)
c,do = J-l Q(l) MY)
where Jtl d&) = 1 and &(y), while it depends on x, is independent of n.
PROOF. (4.16) is equivalent to
JACOBI POLYNOMIALS 433
(4.19)
Then using cos n0 cos n+ = Q [cos $0 + 4) + cos n(e - +)] we see that
k
3+ f r* n cos ?$j cos ne > 0. (4.21)
n-1 j=l
~0s
no> 0. (4.22)
axne> 0 (4.23)
sin(K--I)p, 27
,f” (- 1)‘:[ kS1 1 1; 0 , I = 1, 2,... . (4.24)
Forweletolj=~j=1,Bj=p;,kbe2Z,~:nandletr~l.Theserieswill
converge for I = 1, 2,..., so we may let Y ---f 1. Actually (4.16) can be extended
in the same way for p = 1 and we obtain
sin(K + 1) y 21 > o
&f.(- 1)”[ K+l 1 ”
1 = 1, 2,..., N = 0, l,... (4.26)
from (4.1) for 01= /3 = 3 and the convolution structure for ultraspherical
series. This may have interesting consequences for quadrature problems, for
Lyness and Moler came across (4.24) in some work on quadrature problems.
A completely trivial proof of (4.24) follows from the argument given above
when we observe that the special case of (3.13) which we can use to prove
(4.22) is just the integrated form of (3.16).
THEOREM 3.
For a! = /3 = - fr this is
a
e-cccos ntl dl9 = * [l + (- l)n+l eP”] > 0 (4.27)
s 0
THEOREM 4.
We can extend the range of (CL,8) at the expense of weakening the theorem
for the above values of (a, p).
THEOREM 5.
where
2@ + k)! T(k + *) > o
e
vJ&d ds, 2 0, a >, 0. (4.28)
s 0
REFERENCES