Orthogonal polynomials We start with Definition 1. A sequence of polynomials {pn (x)}∞ n=0 with degree[pn (x)] = n for each n is called orthogonal with respect to the weight function w(x) on the interval (a, b) with a < b if ( Z b 0, m 6= n w(x)pm (x)pn (x) dx = hn δmn with δmn := 1, m = n. a The weight function w(x) should be continuous and positive on (a, b) such that the moments Z b µn := w(x) xn dx, n = 0, 1, 2, . . . a exist. Then the integral Z b hf, gi := w(x)f (x)g(x) dx a denotes an inner product of the polynomials f and g. The interval (a, b) is called the interval of orthogonality. This interval needs not to be finite. If hn = 1 for each n ∈ {0, 1, 2, . . .} the sequence of polynomials is called orthonormal, and if pn (x) = kn xn + lower order terms with kn = 1 for each n ∈ {0, 1, 2, . . .} the polynomials are called monic. Example. As an example we take w(x) = 1 and (a, b) = (0, 1). Using the Gram-Schmidt process the orthogonal polynomials can be constructed as follows. Start with the sequence {1, x, x2 , . . .}. Choose p0 (x) = 1. Then we have hx, 1i 1 hx, p0 (x)i p0 (x) = x − =x− , hp0 (x), p0 (x)i h1, 1i 2 p1 (x) = x − since Z Z 1 h1, 1i = dx = 1 1 and hx, 1i = 0 0 1 x dx = . 2 Further we have hx2 , p0 (x)i hx2 , p1 (x)i p0 (x) − p1 (x) hp0 (x), p0 (x)i hp1 (x), p1 (x)i µ ¶ µ ¶ hx2 , x − 12 i hx2 , 1i 1 1 1 1 2 2 = x − − x− =x − − x− = x2 − x + , 1 1 h1, 1i 2 3 2 6 hx − 2 , x − 2 i p2 (x) = x2 − since Z 2 1 hx , 1i = 0 1 x dx = , 3 and 2 Z hx − 1 2, x− 1 2i = 0 1µ 1 x− 2 Z 2 hx , x − ¶2 1 2i = 0 1µ Z dx = 0 1 2 1 µ ¶ 1 1 1 1 x x− dx = − = 2 4 6 12 2 1 x −x+ 4 2 ¶ dx = 1 1 1 1 − + = . 3 2 4 12 The polynomials p0 (x) = 1, p1 (x) = x − and p2 (x) = x2 − x + 16 are the first three monic orthogonal polynomials on the interval (0, 1) with respect to the weight function w(x) = 1. 1 Repeating this process we obtain 3 3 1 9 2 1 p3 (x) = x3 − x2 + x − , p4 (x) = x4 − 2x3 + x2 − x + , 2 5 20 7 7 70 5 20 5 5 1 p5 (x) = x5 − x4 + x3 − x2 + x − , 2 9 6 42 252 and so on. √ The orthonormal polynomials would be q0 (x) = p0 (x)/ h0 = 1, ¶ µ p √ √ p2 (x) 1 2 =6 5 x −x+ , q1 (x) = p1 (x)/ h1 = 2 3(x − 1/2), q2 (x) = √ 6 h2 µ ¶ √ p3 (x) 3 2 3 1 3 q3 (x) = √ = 20 7 x − x + x − , 2 5 20 h3 etcetera. All sequences of orthogonal polynomials satisfy a three term recurrence relation: Theorem 1. A sequence of orthogonal polynomials {pn (x)}∞ n=0 satisfies pn+1 (x) = (An x + Bn )pn (x) + Cn pn−1 (x), n = 1, 2, 3, . . . , where An = kn+1 , kn n = 0, 1, 2, . . . and Cn = − An hn · , An−1 hn−1 n = 1, 2, 3, . . . . Proof. Since degree [pn (x)] = n for each n ∈ {0, 1, 2, . . .} the sequence of orthogonal polynomials {pn (x)}∞ n=0 is linearly independent. Let An = kn+1 /kn . Then pn+1 (x) − An xpn (x) is a polynomial of degree ≤ n. Hence pn+1 (x) − An xpn (x) = n X ck pk (x). k=0 The orthogonality property now gives hpn+1 (x) − An xpn (x), pk (x)i = n X cm hpm (x), pk (x)i = ck hpk (x), pk (x)i = ck hk . m=0 This implies hk ck = hpn+1 (x) − An xpn (x), pk (x)i = hpn+1 (x), pk (x)i − An hxpn (x), pk (x)i = −An hpn (x), xpk (x)i. For k < n − 1 we have degree [xpk (x)] < n which implies that hpn (x), xpk (x)i = 0. Hence: ck = 0 for k < n − 1. This proves that the polynomials satisfy the three term recurrence relation pn+1 (x) − An xpn (x) = cn pn (x) + cn−1 pn−1 (x), n = 1, 2, 3, . . . . Further we have hn−1 cn−1 = −An hpn (x), xpn−1 (x)i = −An This proves the theorem. 2 kn−1 hn kn =⇒ cn−1 = − An hn · . An−1 hn−1 Note that the three term recurrence relation for a sequence of monic (kn = 1) orthogonal polynomials {pn (x)}∞ n=0 has the form pn+1 (x) = xpn (x) + Bn pn (x) + Cn pn−1 (x) with Cn = − hn , hn−1 n = 1, 2, 3, . . . . A consequence of the three term recurrence relation is Theorem 2. A sequence of orthogonal polynomials {pn (x)}∞ n=0 satisfies n X pk (x)pk (y) hk k=0 and n X {pk (x)}2 k=0 hk = kn pn+1 (x)pn (y) − pn+1 (y)pn (x) · , hn kn+1 x−y n = 0, 1, 2, . . . (1) ¡ ¢ kn · p0n+1 (x)pn (x) − pn+1 (x)p0n (x) , hn kn+1 n = 0, 1, 2, . . . . (2) = Formula (1) is called the Christoffel-Darboux formula and (2) its confluent form. Proof. The three term recurrence relation implies that pn+1 (x)pn (y) = (An x + Bn )pn (x)pn (y) + Cn pn−1 (x)pn (y) and pn+1 (y)pn (x) = (An y + Bn )pn (y)pn (x) + Cn pn−1 (y)pn (x). Subtraction gives pn+1 (x)pn (y) − pn+1 (y)pn (x) = An (x − y)pn (x)pn (y) + Cn [pn−1 (x)pn (y) − pn−1 (y)pn (x)] . This leads to the telescoping sum n n X X pk (x)pk (y) pk+1 (x)pk (y) − pk+1 (y)pk (x) (x − y) = hk Ak hk k=1 k=1 − n X pk (x)pk−1 (y) − pk (y)pk−1 (x) k=1 = Ak−1 hk−1 pn+1 (x)pn (y) − pn+1 (y)pn (x) k02 (x − y) − . An hn h0 This implies that (x − y) n X pk (x)pk (y) k=0 hk = pn+1 (x)pn (y) − pn+1 (y)pn (x) An hn = kn (pn+1 (x)pn (y) − pn+1 (y)pn (x)) , hn kn+1 which proves (1). The confluent form (2) then follows by taking the limit y → x: pn+1 (x)pn (y) − pn+1 (y)pn (x) y→x x−y pn (x) (pn+1 (x) − pn+1 (y)) − pn+1 (x) (pn (x) − pn (y)) = lim y→x x−y 0 0 = pn (x)pn+1 (x) − pn+1 (x)pn (x). lim 3 Zeros Theorem 3. If {pn (x)}∞ n=0 is a sequence of orthogonal polynomials on the interval (a, b) with respect to the weight function w(x), then the polynomial pn (x) has exactly n real simple zeros in the interval (a, b). Proof. Since degree[pn (x)] = n the polynomial has at most n real zeros. Suppose that pn (x) has m ≤ n distinct real zeros x1 , x2 , . . . , xm in (a, b) of odd order (or multiplicity). Then the polynomial pn (x)(x − x1 )(x − x2 ) · · · (x − xm ) does not change sign on the interval (a, b). This implies that Z b w(x)pn (x)(x − x1 )(x − x2 ) · · · (x − xm ) dx 6= 0. a By orthogonality this integral equals zero if m < n. Hence: m = n, which implies that pn (x) has n distinct real zeros of odd order in (a, b). This proves that all n zeros are distinct and simple (have order or multiplicity equal to one). Theorem 4. If {pn (x)}∞ n=0 is a sequence of orthogonal polynomials on the interval (a, b) with respect to the weight function w(x), then the zeros of pn (x) and pn+1 (x) separate each other. Proof. This follows from the confluent form (2) of the Christoffel-Darboux formula. Note that Z b hn = w(x) {pn (x)}2 dx > 0, n = 0, 1, 2, . . . . a This implies that n ¡ 0 ¢ X kn {pk (x)}2 0 · pn+1 (x)pn (x) − pn+1 (x)pn (x) = > 0. hn kn+1 hk k=0 Hence ¢ kn ¡ 0 · pn+1 (x)pn (x) − pn+1 (x)p0n (x) > 0. kn+1 Now suppose that xn,k and xn,k+1 are two consecutive zeros of pn (x) with xn,k < xn,k+1 . Since all n zeros of pn (x) are real and simple p0n (xn,k ) and p0n (xn,k+1 ) should have opposite signs. Hence we have pn (xn,k ) = 0 = pn (xn,k+1 ) and p0n (xn,k ) · p0n (xn,k+1 ) < 0. This implies that pn+1 (xn,k ) · pn+1 (xn,k+1 ) should be negative too. Then the continuity of pn+1 (x) implies that there should be at least one zero of pn+1 (x) between xn,k and xn,k+1 . However, this holds for each two consecutive zeros of pn (x). This proves the theorem. Moreover, if {xn,k }nk=1 and {xn+1,k }n+1 k=1 denote the consecutive zeros of pn (x) and pn+1 (x) respectively, then we have a < xn+1,1 < xn,1 < xn+1,2 < xn,2 < · · · < xn+1,n < xn,n < xn+1,n+1 < b. 4 Gauss quadrature If f is a continuous function on (a, b) and x1 < x2 < · · · < xn are n distinct points in (a, b), then there exists exactly one polynomial P with degree ≤ n − 1 such that P (xi ) = f (xi ) for all i = 1, 2, . . . , n. This polynomial P can easily be found by using Lagrange interpolation as follows. Define p(x) = (x − x1 )(x − x2 ) · · · (x − xn ) and consider P (x) = n X i=1 n X p(x) (x − x1 ) · · · (x − xi−1 )(x − xi+1 ) · · · (x − xn ) f (xi ) = f (xi ) . 0 (x − xi )p (xi ) (xi − x1 ) · · · (xi − xi−1 )(xi − xi+1 ) · · · (xi − xn ) i=1 Let {pn (x)}∞ n=0 be a sequence of orthogonal polynomials on the interval (a, b) with respect to the weight function w(x). Then for x1 < x2 < · · · < xn we take the n distinct real zeros of the polynomial pn (x). If f is a polynomial of degree ≤ 2n − 1, then f (x) − P (x) is a polynomial of degree ≤ 2n − 1 with at least the zeros x1 < x2 < · · · < xn . Now we define f (x) = P (x) + r(x)pn (x), where r(x) is a polynomial of degree ≤ n − 1. This can also be written as f (x) = n X f (xi ) i=1 This implies that Z b Z n X w(x)f (x) dx = f (xi ) a i=1 a b pn (x) + r(x)pn (x). (x − xi )p0n (xi ) w(x)pn (x) dx + (x − xi )p0n (xi ) Z a b w(x)r(x)pn (x) dx. Since degree[r(x)] ≤ n − 1 the orthogonality property implies that the latter integral equals zero. This implies that Z b Z b n X w(x)pn (x) dx, i = 1, 2, . . . , n. w(x)f (x) dx = λn,i f (xi ) with λn,i := 0 a a (x − xi )pn (xi ) i=1 This is the Gauss quadrature formula. This gives the value of the integral in the case that f is a polynomial of degree ≤ 2n−1 if the value of f (xi ) is known for the n zeros x1 < x2 < · · · < xn of the polynomial pn (x). If f is not a polynomial of degree ≤ 2n − 1 this leads to an approximation of the integral: Z b Z b n X w(x)pn (x) λn,i f (xi ) with λn,i := w(x)f (x) dx ≈ dx, i = 1, 2, . . . , n. (x − xi )p0n (xi ) a a i=1 The coefficients {λn,i }ni=1 are called Christoffel numbers. Note that these numbers do not depend on the function f . These Christoffel numbers are all positive. This can be shown as follows. We have Z b pn (x) λn,i = w(x)`n,i (x) dx with `n,i (x) := , i = 1, 2, . . . , n. (x − xi )p0n (xi ) a 5 Then `2n,i (x) − `n,i (x) is a polynomial of degree ≤ 2n − 2 which vanishes at the zeros {xn,k }nk=1 of pn (x). Hence `2n,i (x) − `n,i (x) = pn (x)q(x) for some polynomial q of degree ≤ n − 2. This implies that Z a b ¡ ¢ w(x) `2n,i (x) − `n,i (x) dx = Z a b w(x)pn (x)q(x) dx = 0 by orthogonality. Hence we have Z b Z b λn,i = w(x)`n,i (x) dx = w(x) {`n,i (x)}2 dx > 0. a a Now we can also prove Theorem 5. Let {pn (x)}∞ n=0 be a sequence of orthogonal polynomials on the interval (a, b) with respect to the weight function w(x) and let m < n. Then we have: between any two zeros of pm (x) there is at least one zero of pn (x). Proof. Suppose that xm,k and xm,k+1 are two consecutive zeros of pm (x) and that there is no zero of pn (x) in (xm,k , xm,k+1 ). Then consider the polynomial g(x) = pm (x) . (x − xm,k )(x − xm,k+1 ) Then we have g(x)pm (x) ≥ 0 for x ∈ / (xm,k , xm,k+1 ). Now the Gauss quadrature formula gives Z b n X w(x)g(x)pm (x) dx = λn,i g(xn,i )pm (xn,i ), a i=1 {xn,i }ni=1 where are the zeros of pn (x). Since there are no zeros of pn (x) in (xm,k , xm,k+1 ) we conclude that g(xn,i )pm (xn,i ) ≥ 0 for all i = 1, 2, . . . , n. Further we have λn,i > 0 for all i = 1, 2, . . . , n which implies that the sum at the right-hand side cannot vanish. However, the integral at the left-hand side is zero by orthogonality. This contradiction proves that there should be at least one zero of pn (x) between the two consecutive zeros of pm (x). 2 y 1 0 0.2 0.4 0.6 0.8 1 x –1 –2 The polynomials q2 (x), q3 (x) and q4 (x) 6 Classical orthogonal polynomials The classical orthogonal polynomials are named after Hermite, Laguerre and Jacobi: name pn (x) w(x) Hermite Hn (x) e−x (−∞, ∞) Laguerre Ln (x) e−x xα (0, ∞) (1 − x)α (1 + x)β (−1, 1) 1 (−1, 1) Jacobi Legendre 2 (α) (α,β) Pn (x) Pn (x) (a, b) The Hermite polynomials are orthogonal on the interval (−∞, ∞) with respect to the normal 2 distribution w(x) = e−x , the Laguerre polynomials are orthogonal on the interval (0, ∞) with respect to the gamma distribution w(x) = e−x xα and the Jacobi polynomials are orthogonal on the interval (−1, 1) with respect to the beta distribution w(x) = (1 − x)α (1 + x)β . The Legendre polynomials form a special case (α = β = 0) of the Jacobi polynomials. These classical orthogonal polynomials satisfy an orthogonality relation, a three term recurrence relation, a second order linear differential equation and a so-called Rodrigues formula. Moreover, for each family of classical orthogonal polynomials we have a generating function. In the sequel we will often use the Kronecker delta which is defined by ( 0, m 6= n δmn := 1, m = n for m, n ∈ {0, 1, 2, . . .} and the notation D= d dx (3) for the differentiation operator. Then we have Leibniz’ rule n µ ¶ X n D [f (x)g(x)] = Dk f (x)Dn−k g(x), k n n = 0, 1, 2, . . . (4) k=0 which is a generalization of the product rule. The proof is by mathematical induction and by use of Pascal’s triangle identity µ ¶ µ ¶ µ ¶ n n n+1 + = , k = 1, 2, . . . , n. k k−1 k 7 Hermite The Hermite polynomials are orthogonal on the interval (−∞, ∞) with respect to the normal 2 distribution w(x) = e−x . They can be defined by means of their Rodrigues formula: Hn (x) = (−1)n n 2 2 D w(x) = (−1)n ex Dn e−x , w(x) n = 0, 1, 2, . . . , (5) where the differentiation operator D is defined by (3). Since Dn+1 = D Dn , we obtain £ ¤ Dn+1 w(x) = D [Dn w(x)] = (−1)n D [w(x)Hn (x)] = (−1)n w0 (x)Hn (x) + w(x)Hn0 (x) £ ¤ = (−1)n+1 w(x) 2xHn (x) − Hn0 (x) , n = 0, 1, 2, . . . , which implies that Hn+1 (x) = 2xHn (x) − Hn0 (x), n = 0, 1, 2, . . . . (6) The definition (5) implies that H0 (x) = 1. Then (6) implies by induction that Hn (x) is a polynomial of degree n. Further we have that H2n (x) is even and H2n+1 (x) is odd and that the leading coefficient of the polynomial Hn (x) equals kn = 2n . The Hermite polynomials satisfy the orthogonality relation Z ∞ 1 2 √ e−x Hm (x)Hn (x) dx = 2n n! δmn , m, n ∈ {0, 1, 2, . . .}. π −∞ To prove this we use the definition (5) to obtain Z ∞ Z −x2 n e Hm (x)Hn (x) dx = (−1) −∞ ∞ −∞ (7) 2 Hm (x)Dn e−x dx. Now we use integration by parts n times to conclude that the integral vanishes for m < n. For m = n we have using integration by parts Z ∞ Z ∞ Z ∞ 2 2 2 e−x Hn (x)Hn (x) dx = (−1)n Hn (x)Dn e−x dx = Dn Hn (x) · e−x dx −∞ −∞ Z −∞ ∞ √ 2 = kn n! e−x dx = 2n n! π. −∞ This proves the orthogonality relation (7). In order to find the three term recurrence relation we start with 2 w(x) = e−x w0 (x) = −2xw(x). =⇒ Then we have by using Leibniz’ rule (4) Dn+1 w(x) = Dn w0 (x) = Dn [−2xw(x)] = −2xDn w(x) − 2nDn−1 w(x), which implies that Hn+1 (x) = 2xHn (x) − 2nHn−1 (x), 8 n = 1, 2, 3, . . . . (8) Combining (6) and (8) we find that Hn0 (x) = 2nHn−1 (x), n = 1, 2, 3, . . . . (9) Differentiation of (6) gives 0 Hn+1 (x) = 2xHn0 (x) + 2Hn (x) − Hn00 (x), n = 0, 1, 2, . . . . Now we use (9) to conclude that 2(n + 1)Hn (x) = 2xHn0 (x) + 2Hn (x) − Hn00 (x), n = 0, 1, 2, . . . , which implies that Hn (x) satisfies the second order linear differential equation y 00 (x) − 2xy 0 (x) + 2ny(x) = 0, n ∈ {0, 1, 2, . . .}. Finally we will prove the generating function e 2xt−t2 = ∞ X Hn (x) n=0 We start with n! 2 tn . 2 (10) 2 f (t) = e−(x−t) = e−x · e2xt−t . The Taylor series for f (t) is f (t) = ∞ X f (n) (0) n=0 n! tn with, by using the substitution x − t = u, · n ¸ · n ¸ d −(x−t)2 d 2 2 (n) n −u2 f (0) = e = (−1) e = (−1)n Dn e−x = e−x Hn (x) n n dt du t=0 u=x for n = 0, 1, 2, . . .. Hence we have 2 2 2 e−x · e2xt−t = e−(x−t) = f (t) = ∞ X f (n) (0) n=0 This proves the generating function (10). 9 n! tn = e−x 2 ∞ X Hn (x) n=0 n! tn . Laguerre The Laguerre polynomials are orthogonal on the interval (0, ∞) with respect to the gamma distribution w(x) = e−x xα . They can be defined by means of their Rodrigues formula: L(α) n (x) = 1 1 1 x −α n £ −x n+α ¤ Dn [w(x) xn ] = e x D e x , n! w(x) n! n = 0, 1, 2, . . . . (11) By using Leibniz’ rule (4) we have n µ ¶ X £ ¤ n n −x n+α D e x = Dk e−x Dn−k xn+α k k=0 n µ ¶ X n = (−1)k e−x (n + α)(n + α − 1) · · · (α + k + 1)xα+k k k=0 µ ¶ n X −x α k n Γ(n + α + 1) k x . = e x (−1) k Γ(k + α + 1) k=0 Hence we have L(α) n (x) µ ¶ k n X k n+α x = (−1) , n − k k! n = 0, 1, 2, . . . , (12) k=0 where µ ¶ n+α Γ(n + α + 1) (k + α + 1)n−k = = , n−k (n − k)! Γ(k + α + 1) (n − k)! k = 0, 1, 2, . . . , n. (α) This proves that Ln (x) is a polynomial of degree n. Since µ ¶ (−1)k (α + 1)n (α + 1)n (−n)k k n+α (−1) = = , n−k (n − k)! (α + 1)k n! (α + 1)k k = 0, 1, 2, . . . , n we also have for n = 0, 1, 2, . . . L(α) n (x) µ ¶ µ ¶ n (α + 1)n X (−n)k xk n+α −n = = ;x . 1 F1 n! (α + 1)k k! n α+1 (13) k=0 Note that we have L(α) n (0) µ ¶ n+α (α + 1)n = = , n n! n = 0, 1, 2, . . . (α) and that the leading coefficient of the polynomial Ln (x) equals kn = Further we have (−1)n , n! n = 0, 1, 2, . . . . µ ¶ k µ ¶ k−1 n n X d X x k n+α x k n+α (−1) = (−1) dx n − k k! n − k (k − 1)! k=0 k=1 µ ¶ k n−1 X n+α x (α+1) = −Ln−1 (x), n = 1, 2, 3, . . . . = (−1)k−1 n − k − 1 k! d (α) L (x) = dx n k=0 10 (14) Now we can prove the orthogonality relation Z ∞ Γ(n + α + 1) (α) e−x xα L(α) δmn , m (x)Ln (x) dx = n! 0 α > −1 (15) for m, n ∈ {0, 1, 2, . . .}. First of all, the integral converges if the moments Z ∞ µn = e−x xn+α dx 0 exists for all n ∈ {0, 1, 2, . . .}. This leads to the condition α > −1. Note that µn = Γ(n+α+1). Now we use (11) to obtain Z ∞ Z ∞ £ ¤ 1 −x α (α) (α) n −x n+α e x Lm (x)Ln (x) dx = L(α) e x dx. m (x)D n! 0 0 We apply integration by parts to obtain Z ∞ Z £ ¤ (α) n −x n+α n Lm (x)D e x dx = (−1) 0 ∞ 0 −x n+α Dn L(α) dx m (x)e x which equals zero for m < n. For m = n we find Z ∞ Z ∞ −x n+α Dn L(α) (x)e x dx = k n! e−x xn+α dx = (−1)n Γ(n + α + 1). n n 0 0 This proves the orthogonality relation (15). The Laguerre polynomials can also be defined by their generating function µ ¶ X ∞ xt −α−1 n (1 − t) exp − = L(α) n (x)t . 1−t (16) n=0 The proof is straightforward. Start with (13) and change the order of summation to obtain ∞ X n L(α) = n (x)t n=0 ∞ X (α + 1)n n=0 ∞ X ∞ X n! tn ∞ n ∞ X X X (−n)k xk (α + 1)n (−1)k xk tn = (α + 1)k k! (α + 1)k k! (n − k)! k=0 k=0 n=k ∞ X ∞ (−xt)k X (α + k + 1)n n t k! n! k! n! n=0 k=0 n=0 k=0 µ ¶ ∞ ∞ k X X (−xt) 1 xt k −α−k−1 −α−1 = (1 − t) = (1 − t) − . k! k! 1−t = (α + 1)n+k (α + 1)k (−1)k xk tn+k = k=0 k=0 This proves (16). If we define −α−1 F (x, t) := (1 − t) ¶ µ xt , exp − 1−t then we easily obtain µ ¶ ∂ xt −α−2 F (x, t) = −t(1 − t) exp − ∂x 1−t 11 =⇒ (1 − t) ∂ F (x, t) + tF (x, t) = 0 ∂x and ½ ¾ µ ¶ −x(1 − t) − xt xt (α + 1)(1 − t)−α−2 + (1 − t)−α−1 · exp − (1 − t)2 1−t ½ ¾ µ ¶ x xt = α+1− (1 − t)−α−2 exp − , 1−t 1−t ∂ F (x, t) = ∂t which implies that (1 − t)2 ∂ F (x, t) + [x − (α + 1)(1 − t)] F (x, t) = 0. ∂t The first relation leads to ∞ ∞ X X d (α) n n (1 − t) L (x)t + t L(α) n (x)t = 0 dx n n=0 n=0 or equivalently ∞ ∞ ∞ X X X d (α) d (α) n+1 Ln (x)tn − Ln (x)tn+1 + L(α) = 0. n (x)t dx dx n=0 n=0 n=0 Hence we have d (α) d (α) Ln+1 (x) − L (x) + L(α) n (x) = 0, dx dx n n = 0, 1, 2, . . . (17) or equivalently, by using (14), d (α) (α+1) L (x) = L(α) (x), n (x) − Ln dx n n = 0, 1, 2, . . . . The second relation leads to 2 (1 − t) ∞ X n−1 nL(α) n (x)t + [x − (α + 1)(1 − t)] n=1 ∞ X n L(α) n (x)t = 0 n=0 or equivalently ∞ X n−1 nL(α) −2 n (x)t n=1 ∞ X n nL(α) n (x)t + n=1 +x ∞ X n=0 n L(α) n (x)t − (α + 1) ∞ X n+1 nL(α) n (x)t n=1 ∞ X n L(α) n (x)t + (α + 1) n=0 ∞ X n+1 L(α) = 0. n (x)t n=0 Equating the coefficients of equal powers of t we obtain the three term recurrence relation (α) (α) (n + 1)Ln+1 (x) + (x − 2n − α − 1)L(α) n (x) + (n + α)Ln−1 (x) = 0, n = 1, 2, 3, . . . . Note that this can also be written as h i h i (α) (α) (α) (α) xL(α) (x) + (n + 1) L (x) − L (x) − (n + α) L (x) − L (x) = 0, n n n n+1 n−1 12 (18) n = 1, 2, 3, . . . . Now we differentiate and use (17) to obtain x d (α) (α) (α) Ln (x) + L(α) n (x) − (n + 1)Ln (x) + (n + α)Ln−1 (x) = 0, dx n = 1, 2, 3, . . . . This implies that x d (α) (α) Ln (x) = nL(α) n (x) − (n + α)Ln−1 (x), dx n = 1, 2, 3, . . . . (19) Now we differentiate (19) and use (17) and (19) to find x d (α) d d (α) d2 (α) Ln (x) + Ln (x) = n L(α) L (x) n (x) − (n + α) 2 dx dx dx dx n−1 ¸ · d (α) d (α) d = (n + α) L (x) − L (x) − α L(α) (x) dx n dx n−1 dx n d (α) (x) = −(n + α)Ln−1 (x) − α L(α) dx n d d (α) = x L(α) (x) − nL(α) L (x). n (x) − α dx n dx n (α) This proves that the polynomial Ln (x) satisfies the second order linear differential equation xy 00 (x) + (α + 1 − x)y 0 (x) + ny(x) = 0, n ∈ {0, 1, 2, . . .}. Finally, we use the generating function (16) to prove the addition formula L(α+β+1) (x n + y) = n X (α) β) Lk (x)Ln−k (y), n = 0, 1, 2, . . . . (20) k=0 The generating function (16) implies that ∞ X Ln(α+β+1) (x n=0 + y)t n µ ¶ (x + y)t = (1 − t) exp − 1−t ¶ µ ¶ µ yt xt −β−1 −α−1 · (1 − t) exp − = (1 − t) exp − 1−t 1−t à n ! ∞ ∞ ∞ X (α) X X X (α) (β) k (β) m = Lk (x)t · Lm (y)t = Lk (x)Ln−k (y) tn . −α−β−2 k=0 m=0 This proves (20). 13 n=0 k=0 Jacobi The Jacobi polynomials are orthogonal on the interval (−1, 1) with respect to the beta distribution w(x) = (1 − x)α (1 + x)β . They can be defined by means of their Rodrigues formula: Pn(α,β) (x) = = £ ¤ (−1)n 1 n 2 n D w(x) (1 − x ) 2n n! w(x) h i (−1)n −α −β n n+α n+β (1 − x) (1 + x) D (1 − x) (1 + x) 2n n! (21) for n = 0, 1, 2, . . .. By using Leibniz’ rule (4) we have n µ ¶ h i X n Dn (1 − x)n+α (1 + x)n+β = Dk (1 − x)n+α Dn−k (1 + x)n+β k k=0 n µ ¶ X n = (−1)k (n + α)(n + α − 1) · · · (n + α − k + 1)(1 − x)n+α−k k k=0 × (n + β)(n + β − 1) · · · (β + k + 1)(1 + x)β+k µ ¶µ ¶ n X n+β k n+α (−1) (1 − x)n+α−k (1 + x)β+k , n = 0, 1, 2, . . . . = n! k n−k k=0 This implies that Pn(α,β) (x) µ ¶µ ¶ n (−1)n X n+β k n+α = (−1) (1 − x)n−k (1 + x)k , 2n k n−k n = 0, 1, 2, . . . . (22) k=0 (α,β) This shows that Pn (x) is a polynomial of degree n. Note that we have the symmetry Pn(α,β) (−x) = (−1)n Pn(β,α) (x), and Pn(α,β) (1) n = 0, 1, 2, . . . µ ¶ µ ¶ n+α (α,β) n n+β = and Pn (−1) = (−1) , n n (23) n = 0, 1, 2, . . . . In order to find a hypergeometric representation we write for x 6= 1 µ ¶ n µ ¶µ ¶µ ¶ x+1 k x−1 nX n+α n+β (α,β) Pn (x) = , n = 0, 1, 2, . . . . n n−k 2 x−1 k=0 Now we have for x 6= 1 µ ¶ µ ¶k X ¶i k µ ¶µ x+1 k k 2 2 = 1+ = , i x−1 x−1 x−1 k = 0, 1, 2, . . . . i=0 Now we obtain by changing the order of summations for x 6= 1 and n = 0, 1, 2, . . . µ ¶µ ¶µ ¶ µ ¶ n n µ ¶i x − 1 nXX n + α n + β k 2 Pn(α,β) (x) = k n−k i 2 x−1 i=0 k=i ¶ n n−i µ ¶µ ¶µ ¶µ ¶i µ n+β i+k 2 x − 1 nXX n + α . = 2 i+k n−i−k i x−1 i=0 k=0 14 Now we reverse the order in the first sum to find for x 6= 1 and n = 0, 1, 2, . . . µ ¶ n n µ ¶µ ¶µ ¶µ ¶n−i x − 1 nXX n+α n+β n−i+k 2 (α,β) Pn (x) = 2 n−i+k i−k n−i x−1 i=0 k=0 µ ¶µ ¶µ ¶ µ ¶ n X n X n+α n+β n−i+k x−1 i = n−i+k i−k n−i 2 i=0 k=0 = n X n X i=0 k=0 = Γ(n + α + 1) (n − i + k)! Γ(i − k + α + 1) ¶ x−1 i 2 µ ¶ 1−x i Γ(i + α + 1) i! Γ(n − i + β + 1) 2 Γ(n + β + 1) (n − i + k)! × (i − k)! Γ(n − i + k + β + 1) (n − i)! k! n (−n)i Γ(n + α + 1)Γ(n + β + 1) X n! × i=0 n X (−i)k (−i − α − 1)k k=0 (n − i + β + 1)k k! µ . Since i ∈ {0, 1, 2, . . . , n} we have by using the Chu-Vandermonde summation formula µ ¶ n X (−i)k (−i − α − 1)k −i, −i − α − 1 (n + α + β + 1)i = 2 F1 ;1 = . (n − i + β + 1)k k! n−i+β+1 (n − i + β + 1)i k=0 Hence we have by using Γ(n − i + β + 1) (n − i + β + 1)i = Γ(n + β + 1) µ ¶ n Γ(n + α + 1) X (−n)i (n + α + β + 1)i 1 − x i (α,β) Pn (x) = n! Γ(i + α + 1) i! 2 i=0 µ ¶ n Γ(n + α + 1) X (−n)i (n + α + β + 1)i 1 − x i = , n = 0, 1, 2, . . . . n! Γ(α + 1) (α + 1)i i! 2 i=0 This proves the hypergeometric representation µ ¶ µ ¶ n+α −n, n + α + β + 1 1 − x Pn(α,β) (x) = F ; , 2 1 n α+1 2 n = 0, 1, 2, . . . . Note that this result also holds for x = 1. In view of the symmetry (23) we also have µ ¶ µ ¶ −n, n + α + β + 1 1 + x (α,β) n n+β Pn (x) = (−1) ; , n = 0, 1, 2, . . . . 2 F1 n β+1 2 Note that the hypergeometric representation implies that µ ¶ µ ¶ d (α,β) n + α (−n)(n + α + β + 1) 1 Pn (x) = · − dx n α+1 2 µ ¶ −n + 1, n + α + β + 2 1 − x × 2 F1 ; α+2 2 µ ¶ µ ¶ n+α+β+1 n+α −n + 1, n + α + β + 2 1 − x = F ; 2 1 2 n−1 α+2 2 n + α + β + 1 (α+1,β+1) = Pn−1 (x), n = 1, 2, 3, . . . . 2 15 (24) Another consequence of the hypergeometric representation is that the leading coefficient of (α,β) the polynomial Pn (x) equals µ ¶ n + α (−n)n (n + α + β + 1)n (−1)n (n + α + β + 1)n kn = = , n = 0, 1, 2, . . . . n (α + 1)n n! 2n 2n n! Now it can be shown that the Jacobi polynomials satisfy the orthogonality relation Z 1 2α+β+1 Γ(n + α + 1) Γ(n + β + 1) (α,β) (1 − x)α (1 + x)β Pm (x)Pn(α,β) (x) dx = δmn (2n + α + β + 1) Γ(n + α + β + 1) n! −1 for α > −1, β > −1 and m, n ∈ {0, 1, 2, . . .}. This can be shown by using the definition (21) and integration by parts. The value of the integral in the case m = n can be computed by using the leading coefficient and then writing the integral in terms of a beta integral: Z 1 n o2 (1 − x)α (1 + x)β Pn(α,β) (x) dx −1 = = = = Z h i (−1)n 1 (α,β) n n+α n+β P (x)D (1 − x) (1 + x) dx n 2n n! −1 Z 1 1 Dn Pn(α,β) (x)(1 − x)n+α (1 + x)n+β dx 2n n! −1 Z (n + α + β + 1)n 1 (1 − x)n+α (1 + x)n+β dx 22n n! −1 Z 1 Γ(2n + α + β + 1) (1 − x)n+α (1 + x)n+β dx, Γ(n + α + β + 1) 22n n! −1 n = 0, 1, 2, . . . and by using the substitution 1 − x = 2t Z 1 Z 1 n+α n+β (1 − x) (1 + x) dx = (2t)n+α (2 − 2t)n+β 2 dt −1 0 Z 1 = 22n+α+β+1 tn+α (1 − t)n+β dt = 22n+α+β+1 B(n + α + 1, n + β + 1) 0 = 22n+α+β+1 = 22n+α+β+1 Γ(n + α + 1)Γ(n + β + 1) Γ(2n + α + β + 2) Γ(n + α + 1)Γ(n + β + 1) , (2n + α + β + 1) Γ(2n + α + β + 1) (α,β) The Jacobi polynomials Pn n = 0, 1, 2, . . . . (x) satisfy the second order linear differential equation (1 − x2 )y 00 (x) + [β − α − (α + β + 2)x] y 0 (x) + n(n + α + β + 1)y(x) = 0, n = 0, 1, 2, . . . . A generating function for the Jacobi polynomials is given by ∞ X 2α+β = Pn(α,β) (x)tn , α β R(1 + R − t) (1 + R + t) n=0 16 R= p 1 − 2xt + t2 . Legendre The Legendre polynomials are orthogonal on the interval (−1, 1) with respect to the weight function w(x) = 1. They can be defined by means of their Rodrigues formula: Pn (x) = £ ¤ (−1)n n £ ¤ (−1)n 1 n 2 n D w(x) (1 − x ) = n D (1 − x2 )n , n 2 n! w(x) 2 n! This is the special case α = β = 0 of the Jacobi polynomials: ¶ µ −n, n + 1 1 − x (0,0) Pn (x) = Pn (x) = 2 F1 , ; 2 1 n = 0, 1, 2, . . . . n = 0, 1, 2, . . . . (25) (26) Further we have Pn (−x) = (−1)n Pn (x), Pn (1) = 1 and Pn (−1) = (−1)n , n = 0, 1, 2, . . . . The leading coefficient of the polynomial Pn (x) equals kn = (−n)n (n + 1)n (−1)n (2n)! = n , n (1)n n! 2 2 (n!)2 The orthogonality relation is Z 1 Pm (x)Pn (x) dx = −1 2 δmn , 2n + 1 n = 0, 1, 2, . . . . m, n ∈ {0, 1, 2, . . .}. (27) This can be shown by using the Rodrigues formula (25) and integration by parts Z 1 (−1)n Pm (x)Pn (x) dx = n 2 n! −1 Z 1 −1 Pm (x)D n £ 2 n (1 − x ) ¤ 1 dx = n 2 n! Z 1 −1 Dn Pm (x) (1 − x2 )n dx, which vanishes for m < n. For m = n we have Z 1 Z 1 Z (2n)! 1 n 2 n 2 n D Pn (x) (1 − x ) dx = kn n! (1 − x ) dx = n (1 − x2 )n dx. 2 n! −1 −1 −1 Finally we have by using the substitution 1 − x = 2t for n = 0, 1, 2, . . . Z 1 Z 1 Z 1 2 n n n (1 − x ) dx = (1 − x) (1 + x) dx = (2t)n (2 − 2t)n 2 dx −1 −1 0 = 22n+1 B(n + 1, n + 1) = 22n+1 Hence we have Z 1 −1 {Pn (x)}2 dx = Γ(n + 1) Γ(n + 1) 22n+1 (n!)2 = . Γ(2n + 2) (2n + 1)! (2n)! 22n+1 (n!)2 2 = , 22n (n!)2 (2n + 1)! 2n + 1 This proves the orthogonality relation (27). 17 n = 0, 1, 2, . . . . In order to find a generating function for the Legendre polynomials we use the hypergeometric representation (26) to find µ ¶ ∞ ∞ X n X X (−n)k (n + 1)k 1 − x k n n Pn (x)t = t (1)k k! 2 n=0 n=0 k=0 µ ¶ ∞ X ∞ X (−n)k (n + 1)k 1 − x k n = t k! k! 2 k=0 n=k µ ¶ ∞ X ∞ X (−n − k)k (n + k + 1)k 1 − x k n+k = t k! k! 2 k=0 n=0 µ ¶ ∞ ∞ X (2k)! x − 1 k k X (2k + 1)n n t t = k! k! 2 n! n=0 k=0 µ ¶k ∞ X (2k)! x − 1 tk (1 − t)−2k−1 = k! k! 2 k=0 · ¸ ∞ X (1/2)k 2(x − 1)t −1/2 k −2k−1 −1 = [2(x − 1)t] (1 − t) = (1 − t) 1− k! (1 − t)2 k=0 = £ (1 − t)2 − 2(x − 1)t ¤−1/2 1 . = (1 − 2xt + t2 )−1/2 = √ 1 − 2xt + t2 This proves the generating function ∞ X 1 √ = Pn (x)tn . 1 − 2xt + t2 n=0 (28) If we define F (x, t) = (1 − 2xt + t2 )−1/2 , then we have ∂ 1 x−t F (x, t) = − (1 − 2xt + t2 )−3/2 (−2x + 2t) = . ∂t 2 (1 − 2xt + t2 )3/2 This implies that (1 − 2xt + t2 ) ∂ F (x, t) = (x − t)F (x, t). ∂t Now we use (28) to obtain 2 (1 − 2xt + t ) ∞ X nPn (x)t n−1 = (x − t) n=1 ∞ X Pn (x)tn . n=0 This can also be written as ∞ ∞ ∞ ∞ ∞ X X X X X nPn (x)tn−1 − 2x nPn (x)tn + nPn (x)tn+1 = x Pn (x)tn − Pn (x)tn+1 n=1 n=1 n=1 n=0 n=0 or equivalently ∞ X (n + 1)Pn+1 (x)tn − x n=0 ∞ X (2n + 1)Pn (x)tn + n=0 ∞ X (n + 1)Pn (x)tn+1 = 0. n=0 This leads to P1 (x) = xP0 (x) and the three term recurrence relation (n + 1)Pn+1 (x) − (2n + 1)xPn (x) + nPn−1 (x) = 0, 18 n = 1, 2, 3, . . . . (29) Chebyshev For x ∈ [−1, 1] the Chebyshev polynomials Tn (x) of the first kind and the Chebyshev polynomials Un (x) of the second kind can be defined by sin(n + 1)θ , sin θ Tn (x) = cos(nθ) and Un (x) = The orthogonality property is given by Z 1 Z 2 −1/2 (1 − x ) Tm (x)Tn (x) dx = −1 x = cos θ, n = 0, 1, 2, . . . . (30) π cos(mθ) cos(nθ) dθ = 0, m 6= n 0 and Z Z 1 2 1/2 (1 − x ) −1 Um (x)Un (x) dx = π sin(m + 1)θ sin(n + 1)θ dθ = 0, m 6= n. 0 Both families of orthogonal polynomials satisfy the three term recurrence relation Pn+1 (x) = 2xPn (x) − Pn−1 (x), n = 1, 2, 3, . . . , since we have Tn+1 (x) + Tn−1 (x) = cos(n + 1)θ + cos(n − 1)θ = 2 cos θ cos(nθ) = 2xTn (x) and Un+1 (x) + Un−1 (x) = 2 cos θ sin(n + 1)θ sin(n + 2)θ + sin(nθ) = = 2xUn (x). sin θ sin θ Note that T0 (x) = U0 (x) = 1, T1 (x) = x and U1 (x) = 2x. We also have Tn (x) = Un (x) − xUn−1 (x), n = 1, 2, 3, . . . , since Un (x) − xUn−1 (x) = sin(n + 1)θ − cos θ sin(nθ) sin θ cos(nθ) = = cos(nθ) = Tn (x). sin θ sin θ In order to find a generating function for the Chebyshev polynomials Tn (x) of the first kind, we multiply the recurrence relation by tn+1 and take the sum to obtain ∞ X Tn+1 (x)t n+1 = 2x n=1 ∞ X Tn (x)t n+1 n=1 If we define F (x, t) = ∞ X − ∞ X Tn−1 (x)tn+1 . n=1 Tn (x)tn , |t| < 1, n=0 then we have F (x, t) − T1 (x)t − T0 (x) = 2xt [F (x, t) − T0 (x)] − t2 F (x, t). 19 This implies that (1 − 2xt + t2 )F (x, t) = T0 (x) + T1 (x)t − 2xtT0 (x) = 1 + xt − 2xt = 1 − xt. Hence we have the generating function ∞ X Tn (x)tn = F (x, t) = n=0 1 − xt , 1 − 2xt + t2 |t| < 1. In the same way we have for the Chebyshev polynomials Un (x) of the second kind: G(x, t) = ∞ X Un (x)tn , |t| < 1 n=0 where (1 − 2xt + t2 )G(x, t) = U0 (x) + U1 (x)t − 2xtU0 (x) = 1 + 2xt − 2xt = 1. Hence we have ∞ X 1 , 1 − 2xt + t2 Un (x)tn = G(x, t) = n=0 |t| < 1. This can be used, for instance, to prove that n X Tk (x)xn−k = Un (x), n = 0, 1, 2, . . . . k=0 In fact, we have for |t| < 1 ! à n ∞ ∞ X ∞ ∞ X ∞ X X X X Tk (x)xn tn+k Tk (x)xn−k tn = Tk (x)xn−k tn = n=0 k=0 = k=0 n=0 k=0 n=k ∞ X ∞ X k=0 n=0 Tk (x)tk · (xt)n = 1 − xt 1 · 2 1 − 2xt + t 1 − xt ∞ = X 1 = Un (x)tn . 2 1 − 2xt + t n=0 In a similar way we can prove that n X Pk (x)Pn−k (x) = Un (x), n = 0, 1, 2, . . . , k=0 where Pn (x) denotes the Legendre polynomial. In fact, we have for |t| < 1 à n ! ∞ X ∞ ∞ X ∞ ∞ X X X X n n Pk (x)Pn−k (x) t = Pk (x)Pn−k (x)t = Pk (x)Pn (x)tn+k n=0 k=0 k=0 n=0 k=0 n=k = ∞ X k=0 Pk (x)tk · ∞ X 1 1 Pn (x)tn = √ ·√ 2 1 − 2xt + t 1 − 2xt + t2 n=0 ∞ = X 1 = Un (x)tn . 1 − 2xt + t2 n=0 20