MATH2640 Introduction to Optimisation 3.

Quadratic forms and Eigenvalues, Economic applications, Constrained Optimisation, Lagrange Multipliers, Bordered Hessians : summary 3(A) Quadratic forms and eigenvalues (i) Eigenvalues are the solutions, , of the matrix equation Ax = x. (3.1)

In general, an nn matrix has n eigenvalues, though there may be less than n in special cases. The vectors x satisfying this equation are called the eigenvectors. For each eigenvalue there is a corresponding eigenvector x. (ii) 2 2 case. The quadratic form associated with the symmetric matrix A= a b b c is Q(x, y) = (x y) a b b c x y = ax2 + 2bxy + y 2 .

To nd the eigenvalues, we write equation (3.1) as (A I)x = 0, where I is the unit (identity) matrix. Then for this to have nontrivial solutions, a b b c = 0, so 2 (a + c) + ac b2 = 0

is the quadratic equation whose solutions are the two eigenvalues of the matrix A. These eigenvalues are always real for a symmetric matrix. (iii) The same methods work for an n n symmetric matrix A. Quadratic forms are always associated with symmetric matrices, where Aij = Aji , and symmetric matrices always have real eigenvalues. If there are n distinct eigenvalues then the eigenvectors of a symmetric matrix are always orthogonal to each other, xi xj = 0 for any i = j, with i and j lying between 1 and n. (iv) Rules for classifying quadratic forms using the eigenvalue method. If all eigenvalues are > 0, Q is positive denite (PD) If all eigenvalues are < 0, Q is negative denite (ND) If some eigenvalues are 0 and some are 0, Q is indenite (I). If all eigenvalues are 0, Q is positive semi-denite (PSD) If all eigenvalues are 0, Q is negative semi-denite (NSD)

(v) Normalising eigenvectors To nd the eigenvector corresponding to a particular eigenvalue, solve (A I)x = 0 for that value of . You can only solve these equations up to an arbitrary constant k. Example: A= 3 1 1 3

has eigenvalues = 2 and = 4. To nd eigenvector corresponding to = 2 , solve x1 + x 2 = 0 x1 + x 2 = 0 giving (k, k)

(vi) Normal forms

as the eigenvector. To normalise this, choose k so that (k, k) has unit length, 2k 2 = 1, so the normalised eigenvector is (1/ 2, 1/ 2). The normalised eigenvector corresponding to = 4 is (1/ 2, 1/ 2).
(1) (1) (1)

When the normalised eigenvectors corresponding to the n eigenvalues (1) , (2) , (n) , are found, (y1 , y2 , yn ); (2) (2) (2) (y1 , y2 , , yn ), etc. then the quadratic form Q can be written as a sum of squares, Q = (1) (x1 y1 + x2 y2 + )2 + (2) (x1 y1 + x2 y2 + )2 + which is said to be the normal form of the quadratic form Q. For the above example, x2 x1 x2 x1 Q = 3x2 + 2x1 x2 + 3x2 = 2( )2 + 4( + )2 . 1 2 2 2 2 2 Check this by multiplying it out. Its now obvious that positive eigenvalues make Q positive denite.
(1) (1) (2) (2)

3(B) Unconstrained Optimisation in Economics. (i) Production functions A rm produces Q items per year, which sell at a price p per item. The Revenue, R = pQ. Q depends on quantities x1 , x2 , etc. which are quantities such as the number of employees, amount spent on equipment, and so on. The vector (x1 , x2 , , xn ) is called the input bundle vector. The Cost function measures how much a rm spends on production, and this will also be a function of x1 , x2 and so on. The prot, = R C = pQ(x1 , x2 , , xn ) C(x1 , x2 , , xn ). To maximise the prot, we nd the maximum of as a function of the inputs x1 , , xn , leading to p Q C = , xi xi i = 1, n.

Problems where a rm produces several dierent products can also lead to optimisation problems. (ii) The discriminating monopolist. A monopolist produces two products at a rate Q1 and Q2 . The price obtained is not constant, but reduces if the monopolist oods the market, so p1 = a1 b1 Q1 , and p2 = a2 b2 Q2 , are the prices for the two products. The Cost function is taken as C = c1 Q1 + c2 Q2 , being simply proportional to the quantity produced. The prot = R C = p1 Q1 + p2 Q2 c1 Q1 c2 Q2 = (a1 c1 )Q1 + (a2 c2 )Q2 b1 Q2 b2 Q2 1 2 so there is a best choice of Q1 and Q2 to maximise prot, when Q1 = (a1 c1 ) 2b1 Q1 = 0, and Q2 = (a2 c2 ) 2b2 Q2 = 0.

Inspection of the Hessian shows that has a maximum here provided b1 and b2 are positive. We then also require a1 c1 and a2 c2 . The monopolist then has a prot maximising strategy allowing discrimination between the amounts of the two products to be produced. The same approach can be used when the cost function or the pricing is a more complicated function of the Qi . (iii) The Cobb-Douglas production function. A commonly used model in microeconomics is the Cobb-Douglas production function Q(x1 , x2 ) = xa xb , 1 2 where a > 0, b > 0.

The input bundle here is then just a two-dimensional vector, and we assume here the price p is independent of Q, unlike in the monopolist problem. If the Cost is linear in the input bundle, C = w1 x1 + w2 x2 , then the Prot = pxa xb w1 x1 w2 x2 1 2 and the conditions for a stationary point are apxa1 xb = w1 , 2 1 bpxa xb1 = w2 1 2

giving positive critical values x = apQ /w1 and x = bpQ /w2 , where Q is the production function at the 1 2 critical value. We want to know when this stationary value is a maximum, so we must examine the Hessian H= a(a 1)pxa2 xb 2 1 abpxa1 xb1 1 2 abpxa1 xb1 1 2 b(b 1)pxa xb2 1 2

which is negative denite provided LP M1 = a(a 1)pxa2 xb < 0 and 2 1 2a2 2b2 LP M2 = ab(1 a b)p2 x1 x2 > 0, which requires 0 < a < 1 and 0 < b < 1 and a + b < 1. If these conditions are satised, then the Cobb-Douglas production function gives a prot maximising strategy. If they are not satised, the prot increases indenitely as x1 and x2 get very large. This is possible economic behaviour: if there is no prot maximising strategy, all it means is that the bigger the rm the more prot it makes, which is of course entirely possible.

3(C) Constrained Optimisation: Lagrange multipliers Normally, economic activity is constrained, e.g. expenditure is constrained by available income. The constraints are applied to the input bundle, and may take many dierent forms. Often constraints are in the form of inequalities, e.g. we can spend up to so much on equipment but no more. However, it is often simpler to solve problems with equality constraints, e.g. we are allowed to spend 1000 pounds on equipment, and we will spend it all, so the equipment expenditure x1 = 1000 is an equality constraint. In general, we are trying to maximise the objective function f (x1 , x2 , , xn ) subject to the k inequality constraints gi (x1 , x2 , , xn ) bi , and subject to the M equality constraints hi (x1 , x2 , , xn ) ci , 1 i m. 1ik

(i) Single Equality constraint Maximise f (x) subject to h(x) = c. At a stationary point of this problem, the constraint surface must touch the level surfaces of f . To see this, note that if dx is a small displacement from the critical point x , then if the small displacement lies in the constraint surface dx h = 0. But if this displacement is at a critical point, f must not change as we move from x to x + dx, so we must have dx f = 0 also. Since this has to true for all small displacements on the constraint surface, h is parallel to f at x .When two vectors are parallel, there is a constant such that f = h, or (f h) = 0.

The constant is called the Lagrange multiplier. L = f h is called the Lagrangian. To nd the stationary (critical) points of f (x) subject to h(x) c = 0, we solve the n + 1 equations f h = 0, xi xi h(x) = c, for the n + 1 unknowns x1 , x2 xn , . As usual, we must be very careful to get all the solutions of these n + 1 equations, and not to miss some (or include bogus solutions). Example: Maximise f (x, y) = xy subject to h(x, y) = x + 4y 16 = 0. Solution: The Lagrangian L = f h = xy (x + 4y 16). The equations we need are Lx = 0, Ly = 0, or y = 0 or x 4 = 0 (A) (B) (C) i = 1, n

x + 4y = 16.

The only critical point is x = 8, y = 2 and = 2. At this point, f = 16, and this is the maximum value of f subject to this constraint. Strictly speaking, we dont yet know its a maximum, because we have not examined the Hessian (see the section on Bordered Hessians below). However, if we look at other values of x and y that satisfy the constraint, such as x = 0, y = 4, we see they give lower values of f , which strongly suggests that the critical point is a maximum. (ii) The non-degenerate constraint qualication, NDCQ The above algorithm for nding stationary points normally works well, but the argument breaks down if h = 0 at the critical point. In this case, the constraint surface doesnt a uniquely dened normal, so we cant say that f must be parallel to h. We should therefore check that at each critical point we nd the constraint qualication, usually called the non-degenerate constraint qualication NDCQ holds, i.e. h = 0 at each critical point. What happens if it doesnt hold, i.e. h = 0 at the critical point? Usually the Lagrange equations cant be solved. Example: Maximise f = x subject to h = x3 + y 2 = 0. Solution: L = x (x3 + y 2 ), so Lx = 0 and Ly = 0 give 1 = 3x2 , (A); y = 0, (B); x3 = y 2 (C).

From (A), = 0, so from (B), y = 0. Then from (C), x = 0. x = y = 0 is actually the maximising point, as we can see from the constraint (C), as y 2 must be 0, so x 0, so f = 0 is the largest value of f . However, x = y = 0 is not a well-dened solution of (A), (B) and (C) because if x = 0 equation (A) reads 1 = 0! The problem is that h = (3x2 , 2y) happens to be zero at the critical point (0, 0) so the NDCQ is not satised at this critical point. If the NDCQ constraint is not satised at the critical point, we have to use a dierent method to nd the maximum, other than Lagranges method. In this simple problem, we could just eliminate x using the constraint, so x = (y 2 )1/3 , and sketch f = (y 2 )1/3 as a function of y to nd the minimum. In general, however, the problem is much more dicult if the NDCQ is not satised. (iii) Several equality constraints If we have m equality constraints, hi (x) ci = 0, 1 i m, then we need m Lagrange multipliers, and the Lagrangian L is L = f 1 h1 2 h2 m hm . Then the n + m equations to be solved are L = 0, xi hi (x) = ci , i = 1, n i = 1, m,

for the n + m unknowns x1 , , xn and 1 , , m . If any constraint is simple, it may well be easier to use that constraint to eliminate a variable before nding the stationary points. Example: Maximise f = x + y + z subject to the constraints x2 + y 2 = 1 and z = 2. Solution: Just using the general Lagrange method mechanically, L = x + y + z 1 (x2 + y 2 ) 2 (z 2) and the ve equations we need to solve are 1 21 y = 0, 1 2 = 0, x2 + y 2 = 1, z = 2. It is easily seen that the two solutions are x = 1/ 2, y = 1/ 2, z = 2, 1 = 1/ 2, 2 = 1 and x = 1/ 2, y = 1/ 2, z = 2, 1 = 1/ 2, 2 = 1. However, it is much simpler to use z = 2 to eliminate z and just maximise f = x+y +2 subject to x 2 +y 2 = 1, which only requires three equations for three unknowns, and gives the same answer more quickly. If there are several inequality constraints, the non-degenerate constraint qualication, NDCQ, is that the m rows of the Jacobian derivative matrix, h1 h1 xn x1 , hm hm x1 xn 1 21 x = 0, are linearly independent, that is that none of the rows can be expressed as a linear combination of the other rows. If this is the case, we say the Jacobian derivative matrix is of rank m. 3(D) Constrained Optimisation: Second order conditions The Lagrange multiplier method gives the rst order conditions that determine the stationary points, and we can sometimes tell whther the stationary point is a maximum or a minimum without examining the Hessian Matrix. In more complicated examples, we need to a method to determine whether a stationary point is a local maximum or a local minimum, or neither. Note that having the constraint completely changes the nature of the quadratic expression. For example Q(x, y) = x2 y 2 is indenite, but if we apply the constraint y = 0 then Q(x, y) = Q(x, 0) = x2 which is positive denite and has a local minimum at x = y = 0. We must therefore replace our Leading Principal Minor test with something else. The Lagrangian is L = f 1 h1 2 h2 , , m hm , with h1 = c1 , , hm = cm being the constraints. Suppose we have a stationary point at x with Lagrange multipliers , Using the Taylor series, i L(x + dx) = L(x ) + dx 1 L + dxT Hdx + 2 higher order terms.

The term dx L = 0 at a stationary point, because at the stationary points L = 0. So the stationary point is a minimum if dxT Hdx > 0 and a maximum if dxT Hdx < 0. Here H is the Hessian based on the second derivatives of the Lagrangian, so the i, j component is Lxi xj = 2 2 L= (f h1 h2 + ) 1 2 xi xj xi xj h i = 0 for every constraint

Now the allowed displacements dx are restricted by the constraints to satisfy dx i = 1, m. Replacing dx by x, and form

hi by its value ui at the stationary point, the problem becomes classify the quadratic Q = xT Hx, subject to ui x = 0 i = 1, m.

(i) 2 2 with one linear constraint Note that Q(x, y) = x2 y 2 is indenite, but if we apply the constraint y = 0, i.e. u = (0, 1), then Q = x2 which is positive denite. So a constraint can change the classication of Q from indenite to denite. In general, Q(x, y) = ax2 + 2bxy + cy 2 , with ux + vy = 0. Eliminate y by setting y = ux/v to get Q(x) = x2 (av 2 2buv + cu2 ) v2

which is positive denite (minimum) if (av 2 2buv + cu2 ) > 0, and negative denite (maximum) if (av 2 2buv + cu2 ) < 0. (ii) Bordered Hessians If we have an n variable Lagrangian with m constraints, then the Bordered Hessian is an (n + m) (n + m) matrix constructed as follows: 0 B HB = BT H where the B part has m rows and n columns, with the rows consisting of the constraint vectors u i , i = 1, m, and the H part is the Hessian based on the Lagrangian. In the case where m = 2, n = 3 and the two constraints are u1 x + u2 y + u3 z = 0, v1 x + v2 y + v3 z = 0, (u1 , u2 , u3 ) being Hessian matrix is h1 and (v1 , v2 , v3 ) being HB = h2 , both evaluated at stationary point, the 5 5 Bordered 0 0 u1 u2 u3 0 0 v1 v2 v3 u1 v1 Lxx Lxy Lxz u2 v2 Lxy Lyy Lyz u3 v3 Lxz Lyz Lzz

where Lxx = fxx h1xx h2xx , 1 2 and the other components of the Hessian are dened similarly. (iii) Rules for classication using Bordered Hessians We need to classify Q = xT Hx, subject to ui x = 0,

i = 1, m.

Construct the bordered Hessian as dened above, and evaluate the leading principal minors. We only need the last (n m) LPMs, starting with LP M2m+1 and going up to (and including) LP Mn+m . LP Mn+m is just the determinant of the whole HB matrix. (A) Q is positive denite (PD) if sgn(LP Mn+m ) = sgn(det HB ) = (1)m and the successive LPMs have the same sign. Here sgn means the sign of, so e.g. sgn(2) = 1. This case corresponds to a local minimum. (B) Q is negative denite (ND) if sgn(LP Mn+m ) = sgn(det HB ) = (1)n and the successive LPMs from LP M2m+1 to LP Mn+m alternate in sign. This case corresponds to a local maximum. (C) If both (A) and (B) are violated by non-zero LPMs, then Q is indenite. The semi-denite conditions are very involved, and we dont go into them here.

Example: If n = 2 and m = 1, then nm = 1 and we only need examine LP M3 . Suppose this is LP M3 = 1. Then this has the same sign as (1)m , and so the constrained Q is positive denite. Check: if Q = x2 y 2 and the constraint is y = 0, so u = (0, 1), the bordered Hessian is 0 0 1 HB = 0 1 0 1 0 1 which has LP M3 = detHB = 1, and Q(x, 0) = x2 which is indeed positive denite. Example: Q = x2 y 2 z 2 with constraint x = 0, so now u = (1, 0, 0). Now n = 3 and m = 1 we only need examine LP M4 and LP M3 . 0 1 0 0 1 1 0 0 HB = 0 0 1 0 0 0 0 1

(iv) Evaluating determinants Evaluating 22 and 33 determinants is straightforward, but evaluating larger determinants can be dicult. The denition is det A = (1)h A1r1 A2r2 Anrn

LP M4 = 1, and LP M3 = 1. These alternate in sign there is a possibility the constrained quadratic is negative denite. To test this against criterion (B) note LP M4 = 1 which has the same sign as (1)n , and so the constrained Q is negative denite. Check: if Q = x2 y 2 z 2 and the constraint is x = 0, then Q = y 2 z 2 which is clearly negative denite.

where the summation is made over all permutations r1 r2 r3 , , rn . A permutation is just the numbers 1 to n written in a dierent order, e.g . [1 3 2] is a permutation on the rst 3 integers and corresponds to r 1 = 1, r2 = 3 and r3 = 2. Permutations are either odd or even. If it takes an odd number of interchanges to restore the permutation to its natural order, the permutation is odd. If it takes an even number, the permutation is even. In the formula for the determinant, h = 1 if the permutation is odd, h = 2 if it is even. So [2 3 4 1] is an odd permutation, as we must rst swap 2 and 1, then 2 and 3, and then 3 and 4, three swaps altogether. Since three is an odd number of swaps, the permutation is odd. Other useful facts for evaluating determinants are (i) Swapping any two rows or any two columns reverses the sign of the determinant. (ii) Adding any multiple of another row to a row leaves the determinant unchanged. Similarly adding any multiple of a column to a dierent column leaves it unchanged. using this trick, it is usually simple to generate zero elements which makes evaluating the determinant easier. APPENDIX How were the tests for the constrained quadratic forms derived? This appendix can be omitted on rst reading, as the most important issue is what the tests are and how to apply them. However, there is a way to understand where these tests come from in terms of elementary row and column operations (ERCOs). These also explain where the leading principal minor (LPM) tests for the unconstrained quadratic forms come from. An elementary row and column operation, ERCO, acts on a symmetric matrix and removes a pair of elements aij and aji from the matrix by subtracting appropriate multiples of a row and a column. To illustrate, consider the 4 4 matrix a11 a12 a13 a14 a a22 a23 a24 A = 12 a13 a23 a33 a34 . a14 a24 a34 a44 We want to remove the elements a24 from row 2, column 4 and row 4, column 2 using the elements a23 and a32 . Because the matrix is symmetric it is the same element in both places.To remove a 24 from row 2, column 4, multiply column 3 by a24 /a23 and subtract it from column 4. This gives a a14 a242313 a11 a12 a13 a a12 a22 a23 0 A = . a24 a33 a34 a23 a13 a23 a33 a24 a34 a44 a23 a14 a24 a34

This process can a11 a12 a13 a14

So the whole ERCO process is equivalent to writing A = U T AU . The important points are that U is a unit T upper triangular matrix, so U is a unit lower triangular matrix, and A has zeros where the a23 elements where in A. There are two key properties of ERCO operations. (i) The product of two unit upper triangular matrices is another unit upper triangular matrix. This means if we do two successive ERCO operations, we end up with A = U T U T AU1 U2 = V T AV
2 1

The matrix U on the right is a unit upper triangular matrix, because it has zeros below the leading diagonal and ones along the leading diagonal. Multiply this out yourself to check it really works! To remove the a 24 from row 4, column 2, multiply row 3 by a24 /a23 and subtract it from row 4. This is equivalent to the matrix multiplication U T A = A or a a14 a242313 a11 a12 a13 a a a11 a12 a13 a14 a242313 1 0 0 0 a a a12 a22 a23 0 0 1 0 0 0 12 a22 a23 = a a 0 0 1 0 a34 a242333 a13 a23 a33 a34 a242333 a13 a23 a33 a a a24 a24 a34 0 0 a23 1 a14 a24 a34 a44 a23 a a 24 a a14 a242313 0 a34 a242333 a44 2aa23 34 a a

be written as a matrix multiplication A U = A or a a14 a242313 a12 a13 a 1 0 0 0 a11 a22 a23 0 0 1 0 0 a12 = a a34 a242333 0 0 1 a24 a13 a23 a33 a23 a a14 0 0 0 1 a a44 a242334 a24 a34 a

a12 a22 a23 a24

a13 a23 a33 a34

a14 a24 . a34 a44

for some unit upper triangular matrix V . In fact, we can going on removing pairs of elements as often as we want, and still end up with A = V T AV . (ii) Removing a multiple of a row from another row, or a multiple of a column from another column, leaves the determinant of the matrix unchanged. This a standard result from the theory of determinants. Indeed, it also leaves all the leading principal minors, as dened in 2D(ii), unchanged. This result can easily be veried for a 3 3 determinant, and the extension to the n n case is straightforward. It follows that an ERCO operation leaves the leading principal minors unchanged, i.e. the LPMs of A and A are all the same. Now we apply this ERCO process to our bordered Hessian. We take the case m = 2, n = 4 so there are two constraints and four variables, u1 x1 + u2 x2 + u3 x3 + u4 x4 = 0, so that our bordered Hessian is A= 0 0 u1 u2 u3 u4 0 0 v1 v2 v3 v4 u1 v1 a11 a12 a13 a14 u2 v2 a12 a22 xa23 a24 u3 v3 a13 a23 a33 a34 u4 v4 a14 a24 a34 a44 . v1 x1 + v2 x2 + v3x3 + v4x4 = 0,

If u1 happened to be zero, we would need to reorder the variables so that a non-zero element is brought into the u1 position. Since the constraint has to be non-trivial this must be possible. Note that the U matrix applied to the vector (0, 0, x1 , x2 , x3 , x4 )T gives 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 u2 /u1 u3 /u1 u4 /u1 x1 + u2 x2 /u1 + u3 x3 /u1 + u4 x4 /u1 //x2 = x . Ux = 0 0 0 1 0 0 x3 0 0 0 0 1 0 x4 0 0 0 1 0 1

First we remove u2 , u3 and u4 by using the ERCO 0 0 u1 1 0 0 0 0 0 0 1 0 0 0 0 0 0 v1 0 0 1 0 0 0 u1 v1 a11 A= 0 0 u2 /u1 1 0 0 u2 v2 a12 0 0 u3 /u1 0 1 0 u3 v3 a13 u4 v4 a14 0 0 u4 /u1 1 0 a1

operations based on u1 to write A = U T AU giving 1 0 0 0 0 0 0 0 0 0 0 0 v2 v3 v4 0 1 0 a12 a13 a14 0 0 1 u2 /u1 u3 /u1 u4 /u1 1 0 0 a22 a23 a24 0 0 0 0 1 0 a23 a33 a34 0 0 0 0 0 0 1 0 1 a24 a34 a44

 When we apply the rst row of A to x we get zero because of the constraint, and the same is true of x Now we use the v2 element to remove v1 , v3 and v4 . If m > 2 we would applied to the rst column of A. continue until every row and column of the border just has one non-zero element in it. If v2 happened to be zero, we would again reorder the variables from x2 to xn to bring a non-zero element into the v2 position. If the only non-zero element in row 2 of A is v1 , then the second constraint is linearly dependent on the rst constraint, so its not really a second constraint. In this case the NDCQ wont hold (see 3C(iii)). The matrix A now has the form U T BU = A with B of the form 0 0 u1 0 0 0 0 0 0 v2 0 0 u1 0 a11 a12 a13 a14 B= 0 v2 a12 a22 a23 a24 . 0 0 a13 a23 a33 a34 0 0 a14 a24 a34 a44 At this stage we can check that just as applying the rst row of A to x gave zero because of the constraint, so applying the second row of B to U x also gives zero. The next step is to remove all the o-diagonal elements of the type aij elements using aij /aii multiples of the ith row and column. This reduces the matrix to the form U T CU = A with C of the form 0 0 u1 0 0 0 0 0 0 v2 0 0 u1 0 a11 0 0 0 . C= 0 0 v2 0 a22 0 0 0 0 0 a33 0 0 0 0 0 0 a44 Now we get the matrix into its nal form by removing the a11 and a22 elements using u1 and v2 respectively. In general, this means that a11 to amm are removed. Then we swap columns 1 and 3, and columns 2 and 4. Every time we swap a row we change the sign of the determinant and the principal minors, so this process multiplies the determinants and minors by a factor (1)m . After this swap the matrix now has the form U T DU = A with u1 0 0 0 0 0 0 v2 0 0 0 0 0 0 u1 0 0 0 . D= 0 0 0 0 v2 0 0 0 0 0 a33 0 0 0 0 0 0 a44
2 2 The determinant of D is simply LP M6 = u2 v2 a33 a44 . The next lowest principal minor is LP M5 = u2 v2 a33 . 1 1 T Now the U matrix applied to the vector x = (0, 0, x1 , x2 , x3 , x4 ) gives y = U x. Because of the swap of rows 1 and 3, and 2 and 4, the third and fourth elements of y are zero, and the rst element is x1 + x2 u2 /u1 + x3 u3 /u1 + x4 u4 /u1 which when multiplied by D11 gives zero from the constraint equation. Similarly, D22 times the second element y2 is zero, so that 2 2 yT Dy = xT Ax = a33 y3 + a44 y4 .

The elements y3 and y4 are linear combinations of x1 x4 . The quadratic form xT Ax is therefore positive denite if a33 and a44 are positive. This is the case if LP M5 and LP M6 are both positive. However, LP M6 and LP M5 are the LPMs of D, which are (1)m times the LPMs of A. So the test on the original matrix A to be positive dente is that LP M5 and LP M6 , the LPMs of A, both have the same sign as (1)m . Similarly, if a33 and a44 are both negative in D, then the quadratic form is negative denite. This requires the LPMs of D to have alternating signs, with LP M1 < 0. This means that LP Mn+m , the determinant of D, must have sign (1)n+m , and so LP Mn+m , the determinant of A, must have sign (1)n+m (1)m = (1)n . So the condition for negative deniteness is that the LPMs of A must alternate in sign, and the last one, the determinant of A, must have the same sign as (1)n . These are just the rules in 3D(iii). The quadratic form is semi-denite but not denite if one of a 33 or a44 is zero. The same argument holds if there are no constraints, i.e. m = 0. Then we skip the elimination of the ui and vi , and obtain the results in 2D(ii) directly. CAJ 2/11/2005