The Jordan Form Given a square matrix A, we want to choose M so that M~!AM is as nearly diagonal as possible. In the simplest case, A has a complete set of eigenvectors and they become the columns of M—otherwise known as S. The Jordan form is J = M~'AM = A; it is constructed entirely from 1 by 1 blocks J; = 4;, and the goal of a diagonal matrix is completely achieved. In the more general and more difficult case, some eigenvectors are missing and a diagonal form is impossible. That case is now our main concern. We repeat the theorem that is to be proved: Ifamatrix A has s linearly independent eigenvectors, then itis similar to a matrix J that is in Jordan form, with s square blocks on the diagonal: a J=M1AM= Js Each block has one eigenvector, one eigenvalue, and Is just above the diagonal: Ai od 1 Ai ‘An example of such a Jordan matrix is 8 1000 81 I 08000 0 8 J=|0 001 0J= 01 = A 00000 00 00000 {0] 4b The double eigenvalue 2. = 8 has only asingle eigenvector, in the first coordinate direction e1 = (1, 0, 0, 0, 0); as a result, 4 =8 appears only in a single block J}. The triple eigen- value 2 =0 has two eigenvectors, e3 and es, which correspond to the two Jordan blocks Jy and Jy. If A had 5 eigenvectors, all blocks would be 1 by 1 and J would be diagonal. The key question is this: /f A is some other 5 by 5 matrix, under what conditions will its Jordan form be this same J? When will there exist an M such that M1AM = J? As a first requirement, any similar matrix A must share the same eigenvalues 8, 8, 0, 0, 0. But the diagonal matrix with these eigenvalues is not similar to J—and our question really concerns the eigenvectors.AppendixB The Jordan Form 423 ‘To answer it, we rewrite M~!AM = J in the simpler form AM = MJ: 8) d 08 Al x, x2 x3 Xq Xs] =] x1 x2 x3 xq XS ol 00 0 Carrying out the multiplications a column at a time, Ax, =8x, and Ax = 8x +x (10) Ax; = 0x; and Axy=Oxy+x3 and Axs =Oxs. ay ‘Now we can recognize the conditions on A. It must have three genuine eigenvectors, just as J has, The one with 4 = 8 will go into the first column of M, exactly as it would have gone into the first column of S: Ax; = 8xy. The other two, which will be named x3 and x5, go into the third and fifth columns of M: Ax; = Axs = 0. Finally there must be two other special vectors, the generalized eigenvectors xz and xs. We think of x2 as belonging to a string of vectors, headed by x; and described by equation (10). In fact, Xz is the only other vector in the string, and the corresponding block J; is of order 2. Equation (11) describes two different strings, one in which x4 follows x3, and another in which xs is alone; the blocks Jz and Js are 2 by 2 and 1 by 1. The search for the Jordan form of A becomes a search for these strings of vectors, each one headed by an eigenvector: For every i, either © AX; = Aix, or AX SAX +X (12) The vectors x; go into the columns of M, and each string produces a single block in J. Es- sentially, we have to show how these strings can be constructed for every matrix A. Then if the strings match the particular equations (10) and (11), our J will be the Jordan form of A. I think that Filippov’s idea makes the construction as clear and simple as possible.* It proceeds’by mathematical induction, starting from the fact that every 1 by 1 matrix is already in its Jordan form. We may assume that the construction is achieved for all matrices of order less than n—this is the “induction hypothesis”—and then explain the steps for a matrix of order n. There are three steps, and after a general description we apply them to a specific example. Step 1 If we assume A is singular, then its column space has dimension r < n. Looking only within this smaller space, the induction hypothesis guarantees that a Jordan form is possible—there must be r independent vectors w; in the column space such that either = Aw;=A;w; or = Aw; = Ayw; + Wi-1. (13) Step 2 Suppose the nullspace and the column space of A have an intersection of dimension p. Of course, every vector in the nullspace is an eigenvector corresponding to 2 = 0. Therefore, there must have been p strings in step 1 that started from this eigenvalue, and we are interested in the vectors w; that come at the * A.E Filippov, A short proof of the reduction to Jordan form, Moscow Univ. Math. Bull., volume 26 (1971) pp. 70-71.dixB The Jordan Form end of these strings. Each of these p vectors is in the column space, so each one is acombination of the columns of A: w; = Ay; for some yj. Step 3 The nullspace always has dimension n — r, Therefore, independent from its p-dimensional intersection with the column space, it must contain n —r — p additional basis vectors z; lying outside that intersection. Now we put these steps together to give Jordan’s theorem: The r vectors w;, the p vectors y;, and the n — r — p vectors 2; form Jordan strings for the matrix A, and these vectors are linearly independent. They go into the columns of M, and J = M7! AM is in Jordan form. If we want to renumber these vectors as x1,..., Xn, and match them to equation (12), then each y; should be inserted immediately after the w; it came from; it completes a string in which 4; = 0. The z’s come at the very end, each one alone in its own string; again the eigenvalue is zero, since the z’s lie in the nullspace. The blocks with nonzero eigenvalues are already finished at step 1, the blocks with zero eigenvalues grow by one row and column at step 2, and step 3 contributes any I by 1 blocks J; = [0]. Now we try an example, and to stay close to the previous pages we take the eigen- values to be 8, 8, 0, 0, 0: Step 1. The column space has dimension r = 3, and is spanned by the coordinate vectors ¢), €2, és. To look within this space we ignore the third and fourth rows and columns of A; what is left has eigenvalues 8, 8, 0, and its Jordan form comes from the vectors 8 0 0 0 1 8 w= 0], w= 0], ws = Oo}. 0 0 0 0 1 0 ‘The w; are in the column space, they complete the string for 4 = 8, and they start the string for 4 = 0: Aw, = 8w1, Au = 8u.+ wi, Aw; = 0ws. (4) Step 2 The nullspace of A contains e, and ¢3, so its intersection with the column space is spanned by >. Therefore p = 1 and, as expected, there is one string in equation (14) corresponding to 4 = 0. The vector ws comes at the end (as well as the beginning) of that string, and w; = A(e4 —e1). Therefore y = e4 — €). Step3 The example has n—r ~ p = 5—3—1 = 1, and z = es is in the mullspace but outside the column space. It will be this z that produces a 1 by 1 block in J. If we assemble all five vectors, the full strings are Aw, =8u, Aw, =8u.+w1, Aw;=0w;, Ay =O0y+ws, Az = 0z.Appendix B The Jordan Form 4925 Comparing with equations (10) and (11), we have a perfect match—the Jordan form of our example will be exactly the J we wrote earlier. Putting the five vectors into the columns of M must give AM = MJ, or MAM = J: 8 00 -1 0 018 OO M=/]0 00 0 1}. ooo 10 o10 00 We are sufficiently trustful of mathematics (or sufficiently lazy) not to multiply out M-'AM. In Filippov’s construction, the only technical point is to verify the independence of the whole collection w;, y;, and z;. Therefore, we assume that some combination is zero: Law + Va+ dais =0. (is) Multiplying by A, and using equations (13) for the w; as well as Az; = 0, Aw: Ya or + ay; Aiw; + Wi-r (16) The Ay; are the special w; at the ends of strings corresponding to 2; = 0, so they cannot appear in the first sum. (They are multiplied by zero in A;1;.) Since equation (16) is some combination of the w;, which were independent by the induction hypothesis—they supplied the Jordan form within the column space—we conclude that each d; must be zero. Returning to equation (15), this leaves > ¢;wy = — J gizj, and the left-hand side is in the column space. Since the z’s were independent of that space, each g; must be zero. Finally, > ¢;1; = 0, and the independence of the w; produces c; = 0. If the original A had not been singular, the three steps would have been applied instead to A’ = A — cI. (The constant c is chosen to make A’ singular, and it can be any one of the eigenvalues of A.) The algorithm puts ’ into its Jordan form M~!A’M = J’ by producing the strings x; from the w;, y; and z;. Then the Jordan form for A uses the same strings and the same M: MAM =M'A'M+M"\cM =J'+cl =J. This completes the proof that every A is similar to some Jordan matrix J. Except for a reordering of the blocks, it is similar to only one such J; there is a unique Jordan form for A. Thus, the set of all matrices is split into a number of families, with the following property: All the matrices in the same family have the same Jordan form, and they are all similar to each other (and to J), but no matrices in different families are similar. Inevery family, J is the most beautiful—if you like matrices to be nearly diagonal. With this classification into families, we stop. 0 1\2 oli} witha =0,0,0. 0 00 ehdixB The Jordan Form ‘This matrix has rank r = 2 and only one eigenvector. Within the column space, there is a single string w), w2, which happens to coincide with the last two columns: ‘je = a Aw; =0 and Aw = Ow, + wi. or ‘The nullspace lies entirely within the column space, and it is spanned by w. Therefore p = lin step 2, and the vector y comes from the equation 2 0 w,= {1}, whose solutions y = |0]. 0 1 Finally, the string w1, wa, y goes into the matrix M: 120 010 M={j|0 1 O}, and M'AM=|0 0 ll =J. 000 Ay ool Application to du/dt = Au ‘As always, we simplify the problem by uncoupling the unknowns. This uncoupling is complete only when there is a full set of eigenvectors, and u = Sv; the best change of variables in the present case is u = Mv. This produces the new equation M du/dt = AMp, or dv/dt = Jv, which is as simple as the circumstances allow. It is coupled only by the off-diagonal 1s within each Jordan block. In the preceding example, which has a single block, du/dt = Au becomes eaaste iOierlee da/dt =b a= ay + bot + cot?/2 Galo 0 fv or db/dt=c or bo + cot t lo 00 dc/dt =0 Co. The system is solved by working upward from the last equation, and a new power of ¢ enters at every step. (An £ by £ block has powers as high as r'~',) The exponentials of J, in this case and in the earlier 5 by 5 example, are e& te 0 0 0 sse¢serta/2: 0 & 000 ef=/O 1 ¢ and 0 oO lf Of. oo 1 0 0 010 o 0 001 ‘You can see how the coefficients of a, b, and c appear in the first exponential. And in the second example, you can identify all five of the “special solutions” to du /dt = Aw. Three of them are the pure exponentials u; = ex, us = e%xs, and us = e%x5, formed as usual from the three eigenvectors of A. The other two involve the generalized eigenvectors x2 and x4: uy =e" (tx; tx) and tg = ™ (x3 +4). a7Appendix B The Jordan Form 427 The most general solution to du/dt = Au is a combination cyu, + «++ + esis, and the combination that matches uo at time ¢ = 0 is again Ug = CX +o + C5X5, or = uo = Mc, or c= Muy. This only means that u = Me! M~‘up, and that the S and A inthe old formula Se’ S~!ug, have been replaced by M and J. Problem Set B 1. Find the Jordan forms (in three steps!) of tae (Onsitist2, [i | and =B=|0 0 Oj. 000 2. Show that the special solution 2 in equation (17) does satisfy du/dt = Au, exactly because of the string Ax, = 8x1, Axo = 8x2 +21. A 3. For the matrix B in Problem 1, use Me!‘ M~! to compute the exponential e®*, and compare it with the power series J + Br + (Bt)?/21 +++». 4, Show that each Jordan block J; is similar to its transpose, J’ = P~!J; P, using the Permutation matrix P with 1s along the cross-diagonal (lower left to upper right). Deduce that every matrix is similar to its transpose. 5. Find “by inspection” the Jordan forms of i oO 4S and. B [ 6. Find the Jordan form J and thé“matrix M for A and B (B has eigenvalues 1, 1, 1, -1). What is the solution to du/dt = Au, and what is e“"? 00100 00010 tg: A=|]0 0001 Braedon ele sce eessstaesentacetacteee 00000 obec eeeerte 00000 7. Suppose that A? = A. Show that its Jordan form J = M~!AM satisfies J? = J. Since the diagonal blocks stay separate, this means J? = J; for each block; show by direct computation that J; can only be a 1 by I block, J; = [0] or J; = [1]. Thus, Ais similar to a diagonal matrix of Os and 1s. Note — This is a typical case of our closing theorem: The matrix A can be diagonal- ized if and only if the product (A — dy1)(A — AgI) +++ (A — Apl), without including any repetitions of the 2’s, is zero. One extreme case is a matrix with distinct eigen- values; the Cayley-Hamilton theorem says that with n factors A ~ AZ we always get zero. The other extreme is the identity matrix, also diagonalizable (p = 1 and ‘A-I = 0). The nondiagonalizable matrix A = [3 t satisfies not (A—1) = Obut only (A = 1)? = 0—an equation with a repeated root.