Vandermonde Determinants
Vandermonde Determinants
Vandermonde Determinants
Vandermonde determinants
17.1 17.2
1. Vandermonde determinants
A rigorous systematic evaluation of Vandermonde determinants (below) of the following identity uses the fact that a polynomial ring over a UFD is again a UFD. A Vandermonde matrix is a square matrix of the form in the theorem.
[1.0.1] Theorem:
1 x1 2 x1 det x3 .1 . .
1 xn 1
1 x2 x2 2 x3 2 . . .
1 xn 2
1 xn x2 n x3 n . . .
1 xn n
= (1)n(n1)/2 ( xi xj ) i<j
[1.0.2] Remark: The most universal version of the assertion uses indeterminates xi , and proves an
identity in
Z[x , . . . , x
1
n]
Proof: First, the idea of the proof. Whatever the determinant may be, it is a polynomial in x1 , . . ., xn . The
most universal choice of interpretation of the coecients is as in . If two columns of a matrix are the same, then the determinant is 0. From this we would want to conclude that for i = j the determinant is divisible by [1] xi xj in the polynomial ring [x1 , . . . , xn ]. If we can conclude that, then, since these polynomials
[1] If one treats the x merely as complex numbers, for example, then one cannot conclude that the product of the i
expressions xi xj with i < j divides the determinant. Attempting to evade this problem by declaring the xi as somehow variable complex numbers is an impulse in the right direction, but is made legitimate only by treating genuine indeterminates.
223
224
Vandermonde determinants
are pairwise relatively prime, we can conclude that the determinant is divisible by (xi xj )
i<j
Considerations of degree will show that there is no room for further factors, so, up to a constant, this is the determinant. To make sense of this line of argument, rst observe that a determinant is a polynomial function of its entries. Indeed, the formula is det M = (p) M1p(1) M2p(2) . . . Mnp(n)
p
where p runs over permutations of n things and (p) is the sign or parity of p, that is, (p) is +1 if p is a product of an even number of 2-cycles and is 1 if p is the product of an odd number of 2-cycles. Thus, for any -algebra homomorphism f to a commutative ring R with identity,
Z[ x , . . . , x
where by f (V ) we mean application of f entry-wise to the matrix V . Thus, if we can prove an identity in 1 n ], then we have a corresponding identity in any ring.
Rather than talking about setting xj equal to xi , it is safest to try to see divisibility property as directly as possible. Therefore, we do not attempt to use the property that the determinant of a matrix with two equal columns is 0. Rather, we use the property [2] that if an element r of a ring R divides every element of a column (or row) of a square matrix, then it divides the determinant. And we are allowed to add any multiple of one column to another without changing the value of the determinant. Subtracting the j th column from the ith column of our Vandermonde matrix (with i < j ), we have ... ... ... det V = det . . . ... From the identity xm y m = (x y )(xm1 + xm2 y + . . . + y m1 ) it is clear that xi xj divides all entries of the new ith column. Thus, xi xj divides the determinant. This holds for all i < j . Since these polynomials are linear, they are irreducible in [x1 , . . . , xn ]. Generally, the units in a polynomial ring R[x1 , . . . , xn ] are the units R in R, so the units in [x1 , . . . , xn ] are just 1. Visibly, the various irreducible xi xj are not associate, that is, do not merely dier by units. Therefore, their least common multiple is their product. Since [x1 , . . . , xn ] is a UFD, this product divides the determinant of the Vandermonde matrix. 11 x i xj 2 x2 i xj 3 xi x3 j . . . 1 1 xn xn i j ... ... ... ... ... 1 xj x2 j x3 j . . . n1 xj ... ... ... ... ...
To nish the computation, we want to argue that the determinant can have no further polynomial factors than the ones weve already determined, so up to a constant (which well determine) is equal to the latter
[2] This follows directly from the just-quoted formula for determinants, and also from other descriptions of
determinants, but from any viewpoint is still valid for matrices with entries in any commutative ring with identity.
225
mn 1 product. [3] To prove this, we need the notion of total degree: the total degree of a monomial xm 1 . . . xn is m1 + . . . + mn , and the total degree of a polynomial is the maximum of the total degrees of the monomials occurring in it. We grant for the moment the result of the proposition below, that the total degree of a product is the sum of the total degrees of the factors. The total degree of the product is
1=
1i<j n 1i<n
ni=
1 n(n 1) 2
To determine the total degree of the determinant, invoke the usual formula for the determinant of a matrix M with entries Mij , namely det M = ( ) Mi,(i)
i
where is summed over permutations of n things, and where ( ) is the sign of the permutation . In a Vandermonde matrix all the top row entries have total degree 0, all the second row entries have total degree 1, and so on. Thus, in this permutation-wise sum for a Vandermonde determinant, each summand has total degree 1 0 + 1 + 2 + . . . + (n 1) = n(n 1) 2 so the total degree of the determinant is the total degree of the product 1=
1i<j n 1i<n
ni=
1 n(n 1) 2
Thus, 1 x1 2 x1 det x3 .1 . .
1 xn 1
1 x2 x2 2 x3 2 . . .
1 xn 2
1 xn x2 n x3 n . . .
1 xn n
Granting this, to determine the constant it suces to compare a single monomial in both expressions. For example, compare the coecients of 1 n2 n3 0 xn x2 x3 . . . x1 n1 x n 1
1 In the product, the only way xn appears is by choosing the x1 s in the linear factors x1 xj with 1 < j . 1 n2 After this, the only way to get x2 is by choosing all the x2 s in the linear factors x2 xj with 2 < j . Thus, this monomial has coecient +1 in the product.
In the determinant, the only way to obtain this monomial is as the product of entries from lower left to upper right. The indices of these entries are (n, 1), (n 1, 2), . . . , (2, n 1), (1, n). Thus, the coecient of this monomial is (1) where is the number of 2-cycles necessary to obtain the permutation p such that p(i) = n + 1 i Thus, for n even there are n/2 two-cycles, and for n odd (n 1)/2 two-cycles. For a closed form, as these expressions will appear only as exponents of 1, we only care about values modulo 2. Because of the division by 2, we only care about n modulo 4. Thus, we have values n/2 = 0 mod 2 (for n = 0 mod 4) (n 1)/2 = 0 mod 2 (for n = 1 mod 4) = 1 mod 2 (for n = 3 mod 4) n/2 (n 1)/2 = 1 mod 2 (for n = 1 mod 4)
[3] This is more straightforward than setting up the right viewpoint for the rst part of the argument.
226
Vandermonde determinants
After some experimentation, we nd a closed expression n(n 1)/2 mod 2 Thus, the leading constant is (1)n(n1)/2 in the expression for the Vandermonde determinant. Verify the property of total degree: ///
[1.0.3] Lemma: Let f (x1 , . . . , xn ) and g(x1 , . . . , xn ) be polynomials in k[x1 , . . . , xn ] where k is a eld.
Then the total degree of the product is the sum of the total degrees.
Proof: It is clear that the total degree of the product is less than or equal the sum of the total degrees.
f1 fn en 1 Let xe 1 . . . xn and x1 . . . xn be two monomials of highest total degrees s = e1 + . . . + en and t = f1 + . . . + fn occurring with non-zero coecients in f and g , respectively. Assume without loss of generality that the exponents e1 and f1 of x1 in the two expressions are the largest among all monomials of total degrees s, t in f and g , respectively. Similarly, assume without loss of generality that the exponents e2 and f2 of x2 in the two expressions are the largest among all monomials of total degrees s, t in f and g , respectively, of degrees e1 and f1 in x1 . Continuing similarly, we claim that the coecient of the monomial en +fn M = xe1 +f1 . . . xn f1 u1 fn un en 1 is simply the product of the coecients of xe 1 . . . xn and x1 . . . xn , so non-zero. Let x1 . . . xn and v1 vn x1 . . . xn be two other monomials occurring in f and g such that for all indices i we have ui + vi = ei + fi . By the maximality assumption on e1 and f1 , we have e1 u1 and f1 v1 , so the only way that the necessary power of x1 can be achieved is that e1 = u1 and f1 = v1 . Among exponents with these maximal exponents of x1 , e2 and f2 are maximal, so e2 u2 and f2 v2 , and again it must be that e2 = u2 and f2 = v2 to obtain the exponent of x2 . Inductively, ui = ei and vi = fi for all indices. That is, the only terms in f and f1 fn en 1 g contributing to the coecient of the monomial M in f g are monomials xe 1 . . . xn and x1 . . . xn . Thus, the coecient of M is non-zero, and the total degree is as claimed. ///
2. Worked examples
[17.1] Show that a nite integral domain is necessarily a eld.
Let R be the integral domain. The integral domain property can be immediately paraphrased as that for 0 = x R the map y xy has trivial kernel (as R-module map of R to itself, for example). Thus, it is injective. Since R is a nite set, an injective map of it to itself is a bijection. Thus, there is y R such that xy = 1, proving that x is invertible. ///
[17.2]
Let P (x) = x3 + ax + b k [x]. Suppose that P (x) factors into linear polynomials P (x) = (x 1 )(x 2 )(x 3 ). Give a polynomial condition on a, b for the i to be distinct.
(One might try to do this as a symmetric function computation, but its a bit tedious.) If P (x) = x3 + ax + b has a repeated factor, then it has a common factor with its derivative P (x) = 3x2 + a. If the characteristic of the eld is 3, then the derivative is the constant a. Thus, if a = 0, gcd(P, P ) = a k is never 0. If a = 0, then the derivative is 0, and all the i are the same. Now suppose the characteristic is not 3. In eect applying the Euclidean algorithm to P and P , x3 + ax + b x x 2 3x2 + a = ax + b a = ax + b 3 3 3
227
If a = 0 then the Euclidean algorithm has already terminated, and the condition for distinct roots or factors is b = 0. Also, possibly surprisingly, at this point we need to consider the possibility that the characteristic is 2. If so, then the remainder is b, so if b = 0 the roots are always distinct, and if b = 0 Now suppose that a = 0, and that the characteristic is not 2. Then we can divide by 2a. Continue the algorithm 2 27b2 9x ax + b = a + 3x2 + a 2a 3 4a2 Since 4a2 = 0, the condition that P have no repeated factor is 4a3 + 27b2 = 0
xi
2 = 2 (x1 , . . . , xn ) =
i<j
xi xj xi xj x
3 = 3 ( x1 , . . . , x n ) =
i<j< 3 3 Express x3 1 + x2 + . . . + xn in terms of 1 , 2 , 3 .
Execute the algorithm given in the proof of the theorem. Thus, since the degree is 3, if we can derive the right formula for just 3 indeterminates, the same expression in terms of elementary symmetric polynomials will hold generally. Thus, consider x3 + y 3 + z 3 . To approach this we rst take y = 0 and z = 0, and consider x3 . This is s1 (x)3 = x3 . Thus, we next consider x3 + y 3 s1 (x, y )3 = 3x2 y + 3xy 2 As the algorithm assures, this is divisible by s2 (x, y ) = xy . Indeed, x3 + y 3 s1 (x, y )3 = (3x + 3y )s2 (x, y ) = 3s1 (x, y ) s2 (x, y ) Then consider x3 + y 3 + z 3 s1 (x, y, z )3 3 s2 (x, y, z ) s1 (x, y, z ) = 3xyz = 3s3 (x, y, z ) Thus, again, since the degree is 3, this formula for 3 variables gives the general one:
3 3 x3 1 + . . . + xn = s1 3s1 s2 + 3s3
where si = si (x1 , . . . , xn ).
[17.4] Express
i=j
We could (as in the previous problem) execute the algorithm that proves the theorem asserting that every symmetric (that is, Sn -invariant) polynomial in x1 , . . . , xn is a polynomial in the elementary symmetric functions. But, also, sometimes ad hoc manipulations can yield shortcuts, depending on the context. Here, x2 i xj =
i=j i,j
x2 i xj
i=j
x2 i xj =
i
x2 i
j
xj
i
x3 i
228
Vandermonde determinants
Thus,
2 3 3 3 x2 i xj = (s1 2s2 ) s1 s1 3s1 s2 + 3s3 = s1 2s1 s2 s1 + 3s1 s2 3s3 = s1 s2 3s3 i=j
[17.5] Suppose the characteristic of the eld k does not divide n. Let > 2. Show that
n P (x1 , . . . , xn ) = xn 1 + ... + x
is irreducible in k [x1 , . . . , x ]. First, treating the case = 2, we claim that xn + y n is not a unit and has no repeated factors in k (y )[x]. (We take the eld of rational functions in y so that the resulting polynomial ring in a single variable is Euclidean, and, thus, so that we understand the behavior of its irreducibles.) Indeed, if we start executing the Euclidean algorithm on xn + y n and its derivative nxn1 in x, we have (xn + y n ) x (nxn1 ) = y n n
Note that n is invertible in k by the characteristic hypothesis. Since y is invertible (being non-zero) in k (y ), this says that the gcd of the polynomial in x and its derivative is 1, so there is no repeated factor. And the degree in x is positive, so xn + y n has some irreducible factor (due to the unique factorization in k (y )[x], or, really, due indirectly to its Noetherian-ness).
n n Thus, our induction (on n) hypothesis is that xn 2 + x3 + . . . + x is a non-unit in k [x2 , x3 , . . . , xn ] and has no repeated factors. That is, it is divisible by some irreducible p in k [x2 , x3 , . . . , xn ]. Then in
k [x2 , x3 , . . . , xn ][x1 ] k [x1 , x2 , x3 , . . . , xn ] Eisensteins criterion applied to xn 1 + . . . as a polynomial in x1 with coecients in k [x2 , x3 , . . . , xn ] and using the irreducible p yields the irreducibility.
x3
...
(Hint: Let be an nth root of 1. If xi+1 = xi for all indices i < n, then the (j + 1)th row is times the j th , and the determinant is 0. ) Let Cij be the ij th entry of the circulant matrix C . The expression for the determinant det C =
pSn
229
where (p) is the sign of p shows that the determinant is a polynomial in the entries Cij with integer coecients. This is the most universal viewpoint that could be taken. However, with some hindsight, some intermediate manipulations suggest or require enlarging the constants to include nth roots of unity . Since we do not know that [ ] is a UFD (and, indeed, it is not, in general), we must adapt. A reasonable adaptation is to work over ( ). Thus, we will prove an identity in ( )[x1 , . . . , xn ].
Z Q
Add i1 times the ith row to the rst row, for i 2. The new rst row has entries, from left to right, x1 + x2 + 2 x3 + . . . + n1 xn x2 + x3 + 2 x4 + . . . + n1 xn1 x3 + x4 + 2 x5 + . . . + n1 xn2 x4 + x5 + 2 x6 + . . . + n1 xn3 ... x2 + x3 + 2 x4 + . . . + n1 x1 The tth of these is since n = 1. Thus, in the ring t (x1 + x2 + 2 x3 + . . . + n1 xn )
1 n ],
Q()[x , . . . , x
x1 + x2 + 2 x3 + . . . + n1 xn ) divides this new top row. Therefore, from the explicit formula, for example, this quantity divides the determinant. Since the characteristic is 0, the n roots of xn 1 = 0 are distinct (for example, by the usual computation of gcd of xn 1 with its derivative). Thus, there are n supercially-dierent linear expressions which divide det C . Since the expressions are linear, they are irreducible elements. If we prove that they are non-associate (do not dier merely by units), then their product must divide det C . Indeed, viewing these linear expressions in the larger ring ( )(x2 , . . . , xn )[x1 ]
we see that they are distinct linear monic polynomials in x1 , so are non-associate. Thus, for some c
Q(),
det C = c
1 n
x1 + x2 + 2 x3 + 3 x4 + . . . + (n1) xn
Looking at the coecient of xn 1 on both sides, we see that c = 1. (One might also observe that the product, when expanded, will have coecients in
Z.)
Exercises
17.[2.0.1] A k-linear derivation D on a commutative k-algebra A, where k is a eld, is a k-linear map D : A A satisfying Leibniz identity
D(ab) = (Da) b + a (Db) Given a polynomial P (x), show that there is a unique k -linear derivation D on the polynomial ring k [x] sending x to P (x).
230
Vandermonde determinants
17.[2.0.2] Let A be a commutative k-algebra which is an integral domain, with eld of fractions K . Let
D be a k -linear derivation on A. Show that there is a unique extension of D to a k -linear derivation on K , and that this extension necessarily satises the quotient rule.
17.[2.0.3] Let f (x1 , . . . , xn ) be a homogeneous polynomial of total degree n, with coecients in a eld k.
Let /xi be partial dierentiation with respect to xi . Prove Eulers identity, that
n
xi
i=1
f = nf xi
17.[2.0.4] Let be algebraic over a eld k. Show that any k-linear derivation D on k() necessarily gives
D = 0.