You are on page 1of 205

Matrices : Theory & Applications

Denis Serre

Ecole Normale Superieure

de Lyon
Contents
Topics 2
Themes of the exercises 4
Exercises 19
Index of scientists (mostly mathematicians) 201
Index of stamps 205
Notation
Unless stated otherwise, a letter / or 1 denoting a set of scalars, actually denotes a (commu-
tative) eld.
Related web pages
See the solutions to the exercises in the book on
http://www.umpa.ens-lyon.fr/ serre/exercises.pdf,
and the errata about the book on
http://www.umpa.ens-lyon.fr/ serre/errata.pdf.
http://www.umpa.ens-lyon.fr/ serre/open.pdf.
1
Topics
Calculus of variations, dierential calculus: Exercises 2, 3, 49, 55, 80, 109, 124, 184, 215,
249, 250, 292, 321, 322, 334, 371, 389
Complex analysis: Exercises 7, 74, 82, 89, 104, 180, 245, 354, 391
Commutator: 81, 102, 115, 170, 203, 232, 236, 244, 256, 289, 306, 310, 315, 321, 368, 368,
369
Combinatorics, complexity, P.I.D.s: Exercises 9, 25, 28, 29, 45, 62, 68, 85, 115, 120, 134,
135, 142, 154, 188, 194, 278, 314, 315, 316, 317, 320, 332, 333, 337, 358, 381, 393
Algebraic identities: Exercises 6, 11, 24, 50, 68, 80, 84, 93, 94, 95, 114, 115, 119, 120, 127,
129, 135, 144, 156, 159, 172, 173, 174, 203, 206, 242, 243, 244, 265, 288, 289, 335, 349,
379, 385, 388
Inequalities: Exercises 12, 27, 35, 40, 49, 58, 63, 69, 77, 82, 87, 101, 106, 110, 111, 125, 128,
139, 143, 155, 162, 168, 170, 182, 183, 198, 209, 210, 218, 219, 221, 222, 253, 254, 264,
272, 279, 285, 334, 336, 341, 361, 363, 364, 368, 371, 386
Bistochastic or non-negative matrices: Exercises 9, 22, 25, 28, 76, 96, 101, 117, 138, 139,
145, 147, 148, 150, 152, 154, 155, 164, 165, 192, 193, 212, 230, 231, 251, 322, 323, 332,
333, 341, 347, 350, 365, 380, 383, 384, 389, 392, 394
Determinants, minors and Pfaan: Exercises 8, 10,11, 24, 25, 50, 56, 84, 93, 94, 95, 112,
113, 114, 119, 120, 127, 129, 143, 144, 146, 151, 159, 163, 172, 181, 188, 190, 195, 206,
207, 208, 213, 214, 216, 221, 222, 228, 234, 246, 270, 271, 272, 273, 274, 278, 279, 285,
286, 290, 291, 292, 308, 335, 336, 349, 357, 372, 374, 383, 384, 385, 386, 387, 393
Hermitian or real symmetric matrices: Exercises 12, 13, 14, 15, 16, 27, 40, 41, 51, 52, 54,
58, 63, 70, 74, 75, 77, 86, 88, 90, 92, 102, 105, 106, 110, 113, 116, 118, 126, 128, 143, 149,
151, 157, 164, 168, 170, 180, 182, 184, 192, 198, 199, 200, 201, 209, 210, 211, 215, 216,
217, 218, 219, 223, 227, 229, 230, 231, 239, 241, 248, 258, 259, 263, 264, 271, 272, 273,
274, 279, 280, 285, 291, 292, 293, 294, 301, 307, 308, 312, 313, 319, 326, 328330, 334,
336, 347, 351, 352, 353, 355, 360, 361, 364, 371, 376, 382, 383, 386, 389, 396
Orthogonal or unitary matrices: Exercises 21, 64, 72, 73, 81, 82, 91, 101, 116, 158, 226,
236, 255, 257, 276, 277, 281, 297, 302, 314, 343, 344, 348, 357, 358, 366, 395
Norms, convexity: Exercises 7, 17, 19, 20, 21, 40, 65, 69, 70, 71, 77, 81, 82, 87, 91, 96, 97,
100, 103, 104, 105, 106, 107, 108, 109, 110, 111, 117, 125, 128, 131, 136, 137, 138, 140,
141, 153, 155, 162, 170, 183, 199, 223, 227, 242, 243, 261, 297, 299, 300, 301, 302, 303,
313, 323, 361, 366, 368, 380
Eigenvalues, characteristic or minimal polynomial: Exercises 22, 31, 37, 41, 47, 61, 69,
71, 83, 88, 92, 98, 99, 101, 113, 116, 121, 123, 126, 130, 132, 133, 145, 146, 147, 156, 158,
2
168, 175, 177, 185, 186, 187, 194, 195, 205, 252, 256, 260, 267, 269, 276, 277, 294, 295,
304, 305, 339, 340, 344, 359, 362, 367, 373, 376, 379, 384, 398
Diagonalization, trigonalization, similarity, Jordanization: Exercices 4, 5, 85, 100, 187,
189, 191, 205, 212, 240, 282, 284, 287, 324, 325, 326330, 338, 339, 340, 345, 359, 360,
378, 395
Singular values, polar decomposition, square root: Exercises 30, 46, 51, 52, 69, 82, 100,
103, 108, 109, 110, 111, 139, 147, 162, 164, 185, 198, 202, 226, 240, 254, 261, 262, 281,
282, 296, 304, 354, 363, 370, 394, 395, 396
Classical groups, exponential: Exercises 30, 32, 48, 53, 57, 66, 72, 73, 78, 79, 106, 123, 148,
153, 176, 179, 196, 201, 202, 241, 246, 268, 295, 316, 321
Numerical analysis: Exercises 33, 42, 43, 65, 67, 140, 194, 197, 344, 353, 355, 392
Matrix equations: Exercises 16, 34, 38, 59, 60, 64, 166, 167, 237, 238, 319, 325, 331, 399
Other real or complex matrices: Exercises 26, 35, 39, 46, 80, 103, 118, 131, 166, 171, 185,
220, 221, 232, 242, 243, 245, 247, 249, 250, 253, 268, 269, 284, 290, 298, 303, 304, 306,
309, 311, 324, 331, 351, 356, 360, 369, 373, 391
Dierential equations: Exercises 34, 44, 122, 123, 124, 145, 195, 196, 200, 236, 241, 263,
295, 365
General scalar elds: Exercises 1, 2, 4, 5, 6, 10, 18, 23, 36, 115, 133, 160, 161, 167, 169, 173,
174, 178, 179, 181, 186, 187, 204, 206, 207, 213, 214, 224, 225, 237, 238, 239, 244, 256,
265, 266, 283, 288, 289, 310, 316, 317, 318, 320, 339, 340, 345, 349, 359, 372, 375, 377,
390, 399
Algorithms: 11, 24, 26, 42, 43, 67, 88, 95, 112, 344, 353, 355, 392, 397
Factorizations. 41, 344, 349, 351, 370
Just linear algebra: 233, 235, 375, 382, 388
More specic topics
Numerical range/radius: 21, 65, 100, 131, 223, 253, 269, 306, 309, 356, 369
H det H, over SPD
n
or HPD
n
: 58, 75, 209, 218, 219, 271, 292, 308
Contractions: 82, 104, 220, 221, 272, 300, 380
Hadamard product: 250, 279, 285, 322, 335, 336, 342, 343, 371, 383, 389
3
Functions of matrices: 51, 52, 63, 74, 80, 82, 107, 110, 111, 128, 179, 249, 250, 280, 283,
334, 354
Commutation: 38, 115, 120, 202, 213, 237, 238, 255, 256, 277, 281, 297, 345, 346
Weyls inequalities and improvements: 12, 27, 69, 168, 275, 363
Hessenberg matrices: 26, 113, 305
Themes of the exercises
1. Similarity within various elds.
2. Rank-one perturbations.
3. Minors ; rank-one convexity.
4, 5. Diagonalizability.
6. 2 2 matrices are universaly similar to their matrices of cofactors.
7. RieszThorin by bare hands.
8. Orthogonal polynomials.
9. Birkhos Theorem ; wedding lemma.
10. Pfaan ; generalization of Corollary 7.6.1.
11. Expansion formula for the Pfaan.
12. Multiplicative Weyls inequalities.
13, 14. Semi-simplicity of the null eigenvalue for a product of Hermitian matrices ; a contrac-
tion.
15, 16. Toepliz matrices ; the matrix equation A +

A
1
= H.
17. An other proof of Proposition 3.4.2.
18. A family of symplectic matrices.
19, 20. BanachMazur distance.
21. Numerical range.
22. Jacobi matrices ; sign changes in eigenvectors ; a generalization of Perron-Frobenius for
tridiagonal matrices.
4
23. An other second canonical form of square matrices.
24. A recursive algorithm for the computation of the determinant.
25. Self-avoiding graphs and totally positive matrices.
26. Schur parametrization of upper Hessenberg matrices.
27. Weyls inequalities for weakly hyperbolic systems.
28. SL
+
2
(Z) is a free monoid.
29. Sign-stable matrices.
30. For a classical group G, G U
n
is a maximal compact subgroup of G.
31. Nearly symmetric matrices.
32. Eigenvalues of elements of O(1. :).
33. A kind of reciprocal of the analysis of the relaxation method.
34. The matrix equation

A + A = 2A 1
n
.
35. A characterization of unitary matrices through an equality case.
36. Elementary divisors and lattices.
37. Companion matrices of polynomials having a common root.
38. Matrices M
3
(/) commuting with the exterior product.
39. Kernel and range of 1
n
1

1 and 21
n
1 1

when 1 is a projector.
40. Multiplicative inequalities for unitarily invariant norms.
41. Largest eigenvalue of Hermitian band-matrices.
44. Controllability, Kalmans criterion.
45. Nash equilibrium. The game scissors, stone,. . . .
46. Polar decomposition of a companion matrix.
47. Invariant subspaces and characteristic polynomials.
48. Eigenvalues of symplectic matrices.
49. From Compensated-Compactness.
5
50. Relations (syzygies) between minors.
51. The square root map is analytic.
52. The square root map is (globally) H olderian.
53. Lorentzian invariants of electromagnetism.
54. Spectrum of blockwise Hermitian matrices.
55. Rank-one connections, again.
56. The transpose of cofactors. Identities.
57. A positive denite quadratic form in Lorentzian geometry.
58. A convex function on HPD
n
.
59. When elements of  +H
n
have a real spectrum.
60. When elements of  +H
n
are diagonalizable.
61. A sucient condition for a square matrix to have bounded powers.
62. Euclidean distance matrices.
63. A Jensens trace inequality.
64. A characterization of normal matrices.
65. Pseudo-spectrum.
66. Squares in GL
n
(R) are exponentials.
67. The Le VerrierFadeev method for computing the characteristic polynomial.
68. An explicit formula for the resolvent.
69. Eigenvalues vs singular values (New formulation.)
70. Horn & Schurs theorem.
71. A theorem of P. Lax.
72, 73. The exchange map.
74. Monotone matrix functions.
75. o log det o is concave, again.
76. An application of PerronFrobenius.
6
77. A vector-valued form of the HahnBanach theorem, for symmetric operators.
78. Compact subgroups of GL
n
(C).
79. The action of U(j. ) over the unit ball of M
qp
(C).
80. Balians formula.
81. The projection onto normal matrices.
82. Von Neumann inequality.
83. Spectral analysis of the matrix of cofactors.
84. A determinantal identity.
85. Flags.
87. Parrotts Lemma.
88. The signs of the eigenvalues of a Hermitian matrix.
89. The necessary part in Picks Theorem.
90. The borderline case for Pick matrices.
91. Normal matrices are far from triangular.
92. Tridiagonal symmetric companion matrix.
93. An extension of the formula det = (Pf)
2
.
94. The Pfaan of + r .
95. A formula for the inverse of an alternate matrix.
96. Isometries of (R
n
; /
p
) when j ,= 2.
97. Iteration of non-expansive matrices.
98. The eigenvalues of a 4 4 alternate matrix.
99. An orthogonal companion matrix.
100. The numerical range of a nilpotent matrix.
101. The trace of 1, when and 1 are normal.
102. Compensating matrix.
7
103. A characterization of singular values.
104. The assumption in von Neumann Inequality cannot be weakened.
105. An inequality for Hermitian matrices.
106. The Lipschitz norm of c
iA
over Hermitian matrices.
107. When ( H
+
n
) =(|
2
| = ||
2
).
108. Convexity and singular values.
109. Rank-one convexity and singular values.
110. The square root in operator norm.
111. And now, the cubic root.
112. A criterion for a polynomial to have only real roots.
113. Some non-negative polynomials on Sym
n
(R).
114. Invariant factors of the matrix of cofactors.
116. How far from normal is a product of Hermitian matrices.
117. The extremal elements among symmetric, bistochastic matrices.
118. Non-negative linear forms on M
n
(C).
119. Hilbert matrices.
120. When blocks commute, the determinant is recursive.
121. The characteristic polynomial of some blockwise structured matrix.
122. The Green matrix for ODEs.
123. Stable, unstable and neutral invariant subspaces.
124. The Lopatinski condition in control theory.
125. A trace inequality.
126. Can we choose a tridiagonal matrix in Theorem 3.4.2 ?
127. 4 4 matrices.
128. The Holder inequality for

.
8
129. A determinantal identity.
130. A disk  a la Gershgorin.
131. Making the diagonal constant.
132. Connected components of real matrices with simple eigenvalues.
133. Characteristic and minimal polynomials.
134. The converse of Proposition 8.1.3.
136. Unitary invariant norms and adjunction.
137, 138. Convex subsets of the unit sphere of M
n
(R).
139. An other inequality from von Neumann.
141. Bounded semigroups of matrices.
142. Commutators and palindromes.
143. Determinantal inequalities for SDP
n
Toepliz matrices.
144. Generalized DesnanotJacobi formula.
145. An entropy inequality for positive matrices.
146. Exterior power of a square matrix.
147. The spectrum of a totally positive matrix.
148. Totally non-negative semi-groups.
149. More about Euclidean distance matrices.
150. Farkas Lemma.
151. The hyperbolic polynomial A
2
0
A
2
1
A
2
r
.
152. Another proof of Birkhos Theorem.
153. SO
2
(R) and O

2
(R) are linked inside the unit sphere of M
2
(R).
154. The combinatorics of
3
as a polytope.
155. Estimates of the joint spectral radius.
9
156. The characteristic polynomial of a cyclic matrix.
157. Flat extension of a Hermitian matrix.
158. Spectrum of a normal matrix and of its principal submatrices.
159. The characteristic polynomial of a cyclic matrix (bis).
160. Schurs Lemma.
161. Isomorphic but not conjugated nilpotent sub-algebras of M
n
(/).
162. The unitarily invariant norms.
163. The mapping A
T
A on the space of symmetric matrices.
164. Spectral gap for tridiagonal symmetric stochastic matrices.
165. Primitive non-negative matrices.
166. Solvability of A1 = C.
167. Solvability of A A1 = C.
168. The spectrum of +1 when . 1 are Hermitian and 1 is small (a case of the A. Horns
problem).
169. Theorem 6.2.1 is not valid in rings that are not PID.
170. Heisenberg Inequality.
171. Lattices in C
m
.
172. Schurs Pfaan identity.
173. Hua Identity.
174. Jordan and Thedy Identities for symmetric matrices.
175. Diagonalization in the reals and projectors.
176. Exponential in a Jordan algebra.
177. A characterization of the minimal polynomial.
178. The rank of Schur complement.
179. Group-preserving functions.
180. Hyperbolic Hermitian pencils.
10
181. The reciprocal of Exercise 127.
182. Sums of eigenvalues of a Hermitian matrices.
183. The j-norms of a matrix and its absolute value.
184. Inf-convolution.
185. A generalization of the fundamental Lemma of Banach algebras.
186. The characteristic polynomial and the invariant factors of submatrices.
187. The Dunford decomposition.
188. The Smith determinant.
189. Similar permutation matrices.
190. The determinant of
T
.
191. Matrices that are conjugated to their inverse.
192. Two kinds of positivity.
193. A criterion for the spectral radius of non-negative matrices.
194. An improvement in Le Verriers method.
195. Strongly stable matrices.
196. A formula for exp(t).
197. Symmetric eigenvalue problem: The order of the Jacobi method when : = 3.
198. The geometric mean of positive semi-denite Hermitian matrices.
199. Interpolation of Hermitian spaces.
200. Homogenization of elliptic operators.
201. Hamiltonian matrices.
202. The exponential and the index.
203. A polynomial identity in M
n
(/).
204. Modular subgroups of GL
n
(Z) are torsion free (with one exception).
205. Partial similarity and characteristic polynomial.
206. An identity in M
3
(/).
11
207. Rank and kernel.
208. A determinantal identity.
209. H (det H)
1/n
is concave over HPD
n
.
210. A scalar inequality for vectors, which implies an inequality between Hermitian matrices.
211. An extension problem for Hermitian positive semi-denite matrices.
212. Conjugation classes in SL
2
(Z).
213. Tensor product and determinant (I).
214. Tensor product and determinant (II).
215. Dierentiation over Sym
n
.
216. A converse of Exercise 13 of Chapter 3.
217. Diagonalizability of pencils of real quadratic forms.
218. H (det H)
1/n
is concave over HPD
n
(bis).
219. H (det H)
1/n
is concave over HPD
n
(ter)
220. Symmetry group of the unit ball of M
pq
(C).
221. Huas Inequality.
222. Eigenvalues vs singular values (II).
223. A convex range for quadratic maps.
224. Nilpotence.
225. The dimension of nilpotent subspaces.
226. The Aluthge transform.
227. Computation of a convex envelop.
228. The topology of Alt
4
(R) GL
4
(R).
229. Schur complement for positive semi-denite matrices.
230, 231. Doubly stochastic :-tuples.
232. A matricial equation.
233. Search of an adapted orthogonal basis in R
3
.
12
234. The Pfaan of simple alternate matrices.
236. Isospectral ows.
237, 238. The commutant of a square matrix.
239. Sums of squares.
240. Unitarily similar matrices.
241. Symmetry of a resolvant.
242. Zariski closure of the unit sphere of (M
2
(R). | |
2
).
243. The algebraic nature of the unit sphere of (M
n
(R). | |
2
).
244. A polynomial identity in M
2
(/).
245. Pencils in GL
n
(C).
246. A determinantal identity for orthogonal matrices.
247. Small subgroups of GL
n
(C) are nite.
248. A positive denite Hermitian matrix.
249. Dierentials of power and exponential.
250. Dierentiating matrix functions.
251. The strong PerronFrobenius property.
252. Strange minimal polynomial !
253. The power inequality for the numerical radius.
254. A generalized CauchySchwarz inequality.
255. Fugledess theorem and its consequences.
256. Commutators and nilpotence.
257. A rational map from U
n
to U
m
.
258. The product of Hermitian matrices of which one is positive denite.
259. The spread of a Hermitian matrix.
260. The eigenvalues of
1

.
13
261. Projection onto the unit ball of M
n
(C).
262. An algorithm for the polar decomposition.
264. A Cauchy-Schwarz inequality for the geometric mean.
265. Near generalized inverses.
266. Polynomial group action over nilpotent matrices.
267. Matrices whose roots have integral entries.
268. A characterization of nite subgroups of U
n
.
269. The convex hull of the spectrum vs the numerical range.
270. Matrices whose roots have integral entries (bis).
271. A generalization of Huas Inequality.
272. Bellmans positive matrix.
273. Symplectic matrices and Siegel domain.
274. Some 3 3 matrices, and planar triangles.
275. The determinant of a sum in HPD
2
.
276. The diagonal and the spectrum of normal matrices.
277. The property L.
278. Sylvesters Lemma.
279. Oppenheims Inequality.
280. Positive unital linear maps.
281. Normality of . 1. 1 and 1.
282. Unitary and orthogonal similarity.
283. Matrix version of Lagrange interpolation.
284. Invariant planes for real matrices.
285. An improvement of the Oppenheims Inequality.
286. A snail determinant.
14
287. Matrices being diagonalizable within M
n
(F
p
).
288. A preparation for the next one.
289. The AmitsurLevitzki Theorem (after S. Rosset).
290. The determinant and trace for positive denite matrices.
291. The theorem of Craig & Sakamoto.
292. The Legendre transform of H log det H.
293. Andos supremum of Hermitian matrices.
294. Stability issues in Hamiltonian systems.
295. The exponential as a polynomial.
296. The negative second moment identity.
297. Von Neumanns proof of Fugledes theorem.
298. Embeddings from M
m
(C) into M
n
(C).
299. The operator norm of a nilpotent matrix of order two.
300. Contractions and Toepliz matrices.
301. Positive operators that sum up to the identity.
302. A characterization of unitary matrices.
303. Canonical anticommutative systems.
304. Singular values vs restriction over subspaces.
305. Unit Hessenberg matrix with prescribed Ritz values.
306. Equivalence classes under unitary conjugation in M
2
(C).
307. From a projection to an orthogonal projection.
308. Conjugation of a convex function over H
n
.
309. The resolvant and the numerical range.
310. A characterization of nilpotent matrices.
311. The real part of a nilpotent matrix.
312. The orthogonal of a positive denite Hermitian matrix.
15
313. The method of alternate projections.
314. Hadamard matrices of size 2
m
.
315. Zero trace matrix is a commutator.
316. The unipotent group.
317. Non-isomorphic groups whose elements have the same orders.
318. Elements in GL
n
(Z) (1
n
+ jM
n
(Z)).
319. Matrices 1 + 1 with . 1 HPD
n
.
320. Subsets with odd cardinals and even intersections.
321. Formula for c
A+B
when [. 1] = .
322. Generators of Markovian semi-groups.
323. Log-convexity of the PerronFrobenius eigenvalue.
324. A simple algebra.
325. The Cecioni-Frobenius theorem.
326. Symmetric matrices that are compatible, in the sense of linear elasticity.
327330. Horn & Johnsons Theorem 4.1.7: matrices similar to real ones, matrices similar to
their Hermitian adjoint, products of Hermitian matrices.
331. Sums of squares in M
n
(R).
332. Sums of permutation matrices.
333. The maximum of diagonal sums.
334. Operator monotone functions; Loewners Theorem.
335. The relative gain array (after C. R. Johnson & H. Shapiro).
336. The relative gain array. Positive denite case.
337. A bound for the permanent.
338. Similarity in GL
2
(Z).
339. Diagonalizable companion matrix.
340. Test for diagonalizability.
16
341. The inverse of a strictly diagonally dominant matrix.
342. The relative gain array. Strictly diagonally dominant case.
343. The relative gain array. The case of permutation matrices.
344. Shifted QR method (Francis algorithm.)
345. Matrix commuting with a non-derogatory matrix.
346. Matrix commuting with a non-derogatory matrix and its transpose.
347. Semipositive denite symmetric matrices that have non-negative entries.
348. Parametization of SO
3
(R) by unitary quaternions.
349. Explicit 1l factorization.
350. Tridiagonal bistochastic matrices.
351. The instability of factorization  = H
n
H
n
for  M
n
(R).
352. Maximal positive / negative subspaces of a Hermitian matrix and its inverse.
353. Irrelevance of the choice of angle in the method of Jacobi.
354. A square root cannot be dened over GL
2
(C).
355. The GaussSeidel algorithm for positive semi-denite matrices.
356. The numerical range of and
1
.
357. In orientable Riemannian geometry, the Hodge star operator is an isometry.
358. Five-fold symmetry of lattices (after M. Krzek, J.

Solc and A.

Solcov a).
359. The similarity of a matrix and its transpose (after O. Taussky and H. Zassenhaus).
360. When a real matrix is conjugate to its transpose through a symmetric positive denite
matrix.
361. The convex cone spanned by the squares of skew-symmetric matrices.
362. Computing the multiplicity of an eigenvalue from a polynomial equation j() = 0.
363. Singular values vs eigenvalues.
364. Tr c
H
c
K
vs Tr c
H+K
when H. 1 are Hermitian.
365. Linear dierential equation whose matrix has non-negative entries.
17
366. The supporting cone of SO
n
(R) at 1
n
.
367. The spectrum of (1. C) ((1A AC)A
T
. (A
T
(AC 1A)).
368. A proof of the B ottcherWenzel Inequality.
369. The numerical radius of a commutator.
370. Choleski vs polar factorization.
371. Loewners Theory of operator monotone functions.
372. Subspaces formed of singular matrices.
373. A problem about coins and weights.
374. The rank of some integer matrices.
375. Mixing two idempotent matrices.
376. Spectrum and the equation  = +

.
377. Finding 0
n
by multiplying factors : + t1, (:. t) ,= (0. 0).
378. Polynomials 1 such that 1() is diagonalizable for every .
379. A universal annihilator on M
n
(F
p
).
380. Non-negative matrices and the Hilbert distance.
381. Left- and right- annhilators of M
n
(1) where 1 is nite.
382. Linear forms over quadratic forms.
383. The cone of non-negative symmetric matrices with positive entries.
384. The determinant and the permanent as eigenvalues.
385. Frobenius determinant formula.
386. An other proof of Hadamard Inequality.
387. An integral fraction.
388. Special composition of symmetries.
389. A nonlinear eigenvalue problem.
390. Idempotent matrices that are sums of idempotent matrices.
391. A third proof of Fugledes Theorem.
18
392. Iterative method for a linear system: the case of positive matrices.
393. A matrix related to Boolean algebra.
394. Singular values of bistochastic matrices.
395. 1

1 and 11

are unitarily similar.

396. A decomposition lemma in H
+
n
.
397. Numerical range vs diagonal.
398. Achieving a prescribed characteristic polynomial.
399. The Waring problem for matrices.
Exercises
1. Let 1 be a eld and .  M
n
(1). Let / be the subeld spanned by the entries of
and . Assume that  is similar to  in M
n
(1). Show that  is similar to  in
M
n
(/). Compare with Exercise 2, page 55.
2. (a) When .  GL
n
(/), show that rk( ) = rk(
1

1
).
(b) If GL
n
(/) and r. /
n
are given, let us dene 1 = + r
T
. If 1 GL
n
(/),
show that 1
1
=
1
1
1
r
T

1
. Compute 1
1
r and deduce an explicit formula
for 1
1
, the ShermanMorrison formula.
(c) We now compute explicitly det 1.
i. We begin with the case where = 1
n
. Show that det(1
n
+ r
T
) = 1 +
T
r
(several solutions).
ii. We continue with the case where is non-singular. Deduce that det 1 =
(det )(1 +
T

1
r).
iii. What is the general algebraic expression for det 1 ? Hint: Principle of algebraic
identities. Such an identity is unique and can be deduced form the complex case
/ = C.
iv. Application: If is alternate and r /
n
, prove that det( + trr
T
) det .
(d) Let t vary in /. Check that +tr
T
is invertible either for all but one values of t, or
for all values, or for no value at all. In particular, the set GL
+
n
(R) of real matrices
with positive determinant is rank-one convex (see the next exercise), in the sense
that if . 1 GL
+
n
(R) and rk(1 ) = 1, then the interval (. 1) is contained in
GL
+
n
(R).
(e) We now specialize to / = R. Check that det(+r
T
) det(r
T
) det
2
. Show
that when the rank of 1 is larger than one, det( + 1) det( 1) can be larger
than det
2
.
19
3. Given a map ) : GL
n
(/) /, we dene )

: GL
n
(/) / by )

() := )(
1
) det .
(a) Check that )

= ).
(b) If ) is a linear combination of minors, prove that )

is another such combination.

Mind that the void minor, which is the constant function equal to 1, is allowed in
these combinations. More precisely, prove that for every sets 1 and J of row and
column indices, with [1[ = [J[, the (1. J)-minor of
1
is given by the formula
(det )
1
_
1
J
_
= c(1. J)
_
J
c
1
c
_
for an appropriate sign c(1. J).
(c) Let GL
+
n
(R) be the set of real matrices with det 0. We say that ) : GL
+
n
(R) R
is rank-one convex if its restrictions to segments (. 1) is convex whenever 1
has rank one. Show that, if ) is rank-one convex, then )

is rank-one convex (use

the previous exercise).
(d) According to J. Ball, we say that ) : GL
+
n
(R) R is polyconvex if there exists a
convex function p : R
N
R such that )() = p(:()), where :() is the list of
minors of . Show that, if ) is polyconvex, then )

is polyconvex. Hint: A convex

function is the supremum of some family of ane functions.
4. Assume that the characteristic of the eld / is not equal to 2. Given  GL
n
(/), show
that the matrix
:=
_
0
n

1
0
n
_
is diagonalisable. Compute its eigenvalues and eigenvectors. More generally, show that
every involution (
2
= 1) is diagonalisable.
5. Let 1 /[A] have simple roots in /. If M
n
(/) is such that 1() = 0
n
, show that
is diagonalisable. Exercise 4 is a special case of this property.
6. If : = 2, nd a matrix 1 GL
2
(/) such that, for every matrix M
2
(/), there holds
1
1
1 =

(

is the matrix of cofactors.)
Nota: If SL
2
(/), we thus have 1
1
1 =
T
, meaning that the natural represen-
tation of SL
2
(/) into /
2
is self-dual.
7. Prove the norm inequality (1 < j < )
||
p
||
1/p
1
||
1/p

. M
n
(C).
by a direct comptutation using only the H older inequality and the explicit formulae for
the norms ||
1
and ||

. Remark: Exercise 20 of Chapter 4 corresponds to the case

j = 2, where Holder is nothing but CauchySchwarz.
20
8. (See also Exercise 119). Let j be a probability measure on R, with a compact support.
We assume that this support is not nite. Dene its moments
:
k
:=
_
R
r
k
dj(r). / N.
and the determinants
1
n
:=

:
0
:
1
:
n
:
1
:
2
:
n+1
.
.
.
.
.
.
:
n
:
n+1
:
2n

. 1
n
(r) :=

:
0
:
1
:
n
:
1
:
2
:
n+1
.
.
.
.
.
.
:
n1
:
n
:
2n1
1 r r
n

.
Dene at last
j
n
(r) =
1

1
n1
1
n
1
n
(r).
(a) Prove that the leading order term of the polynomial j
n
is c
n
r
n
for some c
n
0.
Then prove that
_
R
j
n
(r)j
m
(r)dj(r) =
m
n
.
In other words, the family (j
n
)
nN
consists in the orthonormal polynomials relatively
to j.
(b) What happens if the support of j is nite ?
(c) Prove the formula
1
n
=
1
(: + 1)!
_
R
n+1

0i<jn
(r
i
r
j
)
2
dj(r
0
) dj(r
n
).
Hint: Compute 1
n
, a determinant of size : + 1, in (: + 1)! dierent ways. Then
use the expression of a Vandermonde determinant.
9. Find a short proof of Birkhos Theorem (Theorem 5.5.1), using the wedding Lemma,
also known as Halls Theorem: let : be a positive integer and let G (girls) and 1 (boys)
two sets of cardinality :. Let 1 be a binary relation on G 1 (p1/ means that the
girl p and the boy / appreciate each other). Assume that, for every / = 1. . . . . :, and for
every subset 1

of 1

(that is, the set of those p such

that p1/ for at least one / 1

) has cardinality larger than or equal to /. Then there

exists a bijection ) : 1 G such that )(/)1/ for every / 1. Nota: the assumption is
politically correct, in the sense that it is symmetric in 1 and G (though it is not clear at a
rst sight). The proof of the wedding Lemma is by induction. It can be stated as a result
on matrices: given an : : matrix  of zeroes and ones, assume that, for every / and
every set of / lines, the ones in these lines correspond to at least / columns. Then  is
21
larger than or equal to a permutation matrix. Hint: Here are the steps of the proof. Let
 be bistochastic. Prove that it cannot contain a null block of size / | with / +| :.
Deduce, with the wedding lemma, that there exists a permutation such that :
i(i)
0
for every i. Show that we may restrict to the case :
ii
0 for every i. Show also that
we may restrict to the case where 0 < :
ii
< 1. In this case, show that (1 c) + c1
n
is
bi-stochastic for [c[ small enough. Conclude.
10. Let / be a eld. Dene the symplectic group Sp
m
(/) as the set of matrices  in M
2m
(/)
that satisfy
T
J
m
 = J
m
, where
J
m
:=
_
0
m
1
m
1
m
0
m
_
.
Check that the word group is relevant. Using the Pfaan, prove that every symplectic
matrix (that is,  Sp
m
(/)) has determinant +1. Compare with Corollary 7.6.1.
11. Set : = 2:.
(a) Show the following formula for the Pfaan, as an element of Z[r
ij
; 1 i < , :] :
Pf(A) =

(1)

r
i
1
i
2
r
i
2m1
i
2m
.
Hereabove, the sum runs over all the possible ways the set 1. . . . . : can be parti-
tionned in pairs :
1. . . . . : = i
1
. i
2
i
2m1
i
2m
.
To avoid redundancy in the list of partitions, one normalized by
i
2k1
< i
2k
. 1 / :.
and i
1
< i
3
< < i
2m1
(in particular, i
1
= 1 and i
2m
= 2:). At last, is the
signature of the permutation (i
1
i
2
i
2m1
i
2m
).
Compute the number of monomials in the Pfaan.
(b) Deduce an expansion formula with respect to the i-th row for the Pfaan: if i is
given, then
Pf(A) =

j(=i)
(i. ,)(1)
i+j+1
r
ij
Pf(A
ij
).
where A
ij
M
n2
(/) denotes the alternate matrix obtained from A by removing
the i-th and the ,-th rows and columns, and (i. ,) is +1 if i < , and is 1 if , < i.
12. (a) Let . 1 be :: Hermitian positive denite matrices. Denote by
1
()
2
()
and
1
(1)
2
(1) their eigenvalues. Remarking that 1 is similar to

1, show that the spectrum of 1 is real, positive, and satises

i
()
1
(1)
i
(1)
i
()
n
(1).
Hint: Use Theorem 3.3.2.
22
(b) Compare with point a) of Exercise 6, Chapter 7. Show also that the conclusion still
holds if . 1 are only positive semi-denite.
(c) More generally, prove the inequalities

j
()
k
(1)
i
(1)
j
()
l
(1).
whenever , + / i + 1 and , + | i + :.
(d) Set : = 2. Let c
1
c
2
, /
1
/
2
, j
1
j
2
be non-negative numbers, satisfying
j
1
j
2
= c
1
c
2
/
1
/
2
and the inequalities
c
1
/
1
j
1
minc
1
/
2
. c
2
/
1
. maxc
1
/
2
. c
2
/
1
j
2
c
2
/
2
.
Prove that there exist 2 2 real symmetric matrices and 1, with spectra c
1
. c
2

and /
1
. /
2
, such that j
1
. j
2
is the spectrum of 1. Hint: Begin with the
case where some of the inequalities are equalities. Then use the intermediate value
Theorem.
(e) Set : 2, and assume that the Hermitian matrix 1 has eigenvalues /
1
= 0, /
2
=
= /
n
= 1. Show that

1
(1) = 0.
i1
()
i
(1)
i
().
Conversely, if
j
1
= 0 c
1
j
2
c
2
j
n1
c
n
.
show that there exists a Hermitian matrix with eigenvalues c
1
. . . . . c
n
, such that
1 has eigenvalues j
1
. . . . . j
n
.
Generalize to the case /
0
= = /
k
= 0, /
k+1
= = /
n
= 1.
13. (From J. Gro and G. Trenkler.) Given . 1 two Hermitian positive semi-denite
matrices, show that C
n
= 1(1) ker(1).
14. (a) Assume that M
np
(C) is injective. Prove that H :=

is positive denite.
Show that H
1

is the orthogonal projector onto 1().

(b) Given two injective matrices
j
M
np
j
(C), dene H
j
as above. Dene also 1 :=

2
and then  := H
1
2
1

H
1
1
1. Using the previous exercise, show that the
eigenvalues of  belong to [0. 1].
15. Let M
n
(C) and H HPD
n
be given. If R, dene () := H + c
i
+ c
i

.
Dene a matrix 
k
H
kn
by

k
=
_
_
_
_
_
_
H 0
n

.
.
.
.
.
.
.
.
.
0
n
.
.
.
.
.
.
.
.
.

H
_
_
_
_
_
_
.
23
(a) Decomposing vectors r C
kn
as / blocks r
j
C
n
, nd an expression of r


k
r in
terms of (2|,/) (| = 1. . . . . /), r
1
and r
k
.
(b) We assume that (2|,/) 0
n
(| = 1. . . . . /). Show that there exist two positive
constants c
k
, d
k
such that
r

k
r c
k
|r|
2
d
k
|r
k
| |r
1
|.
Deduce that there exists a t
k
k
1
n
in the bottom-right block,
moves 
k
to a positive denite matrix.
(c) Under the same assumption as above, prove that
1
. . . . . 
k1
are positive denite.
16. Let M
n
(C) and H HPD
n
be given, with ,= 0
n
. We are interested in the equation
A +

A
1
= H.
where A HPD
n
is the unknown.
(a) Show that a necessary condition for A to exist is (Property (P))
the Hermitian matrix () := H + c
i
+ c
i

is positive semi-denite for

every ,
the map det(()) does not identically vanish.
Hint: Factorize (), and more generally H + . + .
1

.
We shall prove later on that Property (P) is also a sucient condition. The proof
relies more or less upon innite dimensional matrices called Toepliz matrices. For a
general account of the theory, see M. Rosenblum & J. Rovniak, Hardy classes and
operator theory, Oxford U. Press, 1985).
(b) Check that (P) is fullled in the case where H = 1
n
and is real skew-symmetric.
Then show that a solution does exist, which is not unique. Hint: First, solve the
case : = 2. The nature of the solution depends on the sign of t
2
1,4, where
=
_
0 t
t 0
_
.
(c) Let be invertible. Find a transformation that reduces the equation to the case
H = 1
n
(though with a dierent ). Verify that this transformation preserves the
validity or the failure of Property (P).
(d) From now on, we assume that H = 1
n
. Show that A is a solution of A+

A
1
=
1
n
, if and only if 1
n
A is a solution of 1 + 1
1

= 1
n
.
(e) We temporarily assume that the equation admits at least one solution A HPD
n
.
Here is an algorithm for the approximation of a solution (the largest one):
A
0
= 1
n
. A
n+1
:= 1
n

A
1
n
.
24
i. First, show that A 1
n
in H
n
.
ii. Prove inductively that A
k
A for every /.
iii. Prove inductively that A
k
is non-increasing in HPD
n
. Deduce that it converges
to some limit, and that this limit is a solution.
iv. Deduce that the equation admits a largest solution.
v. Show that the equation also admits a smaller solution in HPD
n
.
(f) We now turn to the existence of a solution.
i. Dene a block-tridiagonal matrix
k
H
kn
by

k
=
_
_
_
_
_
_
1
n
0
n

.
.
.
.
.
.
.
.
.
0
n
.
.
.
.
.
.
.
.
.

1
n
_
_
_
_
_
_
.
If (0) 0
n
and (P) holds, show that
k
HPD
kn
for every / 1. Hint: use
Exercise 15.
ii. Then show that there exists a unique blockwise lower-bidiagonal matrix 1
k
, with
diagonal blocks in HPD
n
, such that 1
k
1

k
= 
k
.
iii. Then prove that there exist matrices 1
j
HPD
n
and C
j
M
n
(C), such that,
for every / 1, there holds
1
k
=
_
_
_
_
_
_
1
1
0
n
C
1
.
.
.
.
.
.
0
n
.
.
.
.
.
. 0
n
.
.
. C
k1
1
k
_
_
_
_
_
_
.
iv. Write the recursion satised by (1
j
. C
j
), and check that A
k
:= 1
2
k
satises the
algorithm above. Then, show that A
k
converges as / +, and that its limit
is a solution of our equation (therefore the greatest one).
v. Assuming (P) only, show that we may assume (0) 0
n
(consider the matrix
c
i
instead of , with suitably chosen). Conclude.
17. (a) Let c R
n
have positive entries. Recall (Exercise 20.a of Chapter 5) that the
extremal points of the convex set dened by the inequality / ~ c are obtained
from c by permutation of its entries.
Show
(1) (/ ~ c) =(

j
/
j

i
c
i
).
(b) Deduce an other proof of Proposition 3.4.2, with the help of Theorem 3.4.1. (One
may either deduce (1) from Proposition 3.4.2 and Theorem 3.4.2. These are rather
long proofs for easy results !)
25
18. Let M
n
(/) be invertible and dene  M
2n
(/) by
 :=
_
0
n

1

1
_
.
Show that  is symplectic: 
T
J = J, with
J :=
_
0
n
1
n
1
n
0
n
_
.
19. (The BanachMazur distance.)
Stefan Banach.
(a) Let  and

be two norms on /
n
(/ = R or C).
If M
n
(/), we may dene norms
||

:= sup
x=0


(r)
(r)
. |
1
|

:= sup
x=0
(
1
r)

(r)
.
Show that ||

|
1
|

achieves its upper bound.

We shall denote by (.

) the minimum value.

Verify
0 log (.

) log (.

) + log (

.

).
When  = | |
p
, we shall write /
p

= | |
q
, we write | |
p,q
for | |

.
(b) In the set A of norms on /
n
, let us consider the following equivalence relation:


if and only if there exists an GL

n
(/) such that 

=  . Show that
log induces a metric d on the quotient set Norm := A, . This metric is called
the BanachMazur distance.How many classes of Hermitian norms are there ?
(c) Compute |1
n
|
p,q
for 1 j. : (there are two cases, depending on the sign of
j). Deduce that
(/
p
. /
q
) :
[
1
p

1
q
[
.
(d) Show that (/
p
. /
q
) = (/
p

. /
q

), where j

are the conjugate exponents.

(e) i. Given H H
+
n
, nd that the average of r

Hr, as r runs over the set dened

by [r
j
[ = 1 for all ,s, is Tr H (the measure is the product of : copies of the
normalized Lebesgue measure on the unit disk). Deduce that

Tr 

||
,2
for every  M
n
(/).
ii. On the other hand, prove that
||
p,
= max
1in
|
(i)
|
p
.
where
(i)
denotes the i-th row vector of .
26
iii. Deduce that (/
2
. /

) =

:.
iv. Using the triangle inequality for log , deduce that
(/
p
. /
q
) = :
[
1
p

1
q
[
whenever j. 2, and then for every j. such that (j 2)( 2) 0. Nota:
The exact value of (/
p
. /
q
) is not known when (j 2)( 2) < 0.
v. Remark that the curves /
p
; 2 j and /
p
; 1 j 2 are geodesic,
in the sense that the restrictions of the BanachMazur distance to these curves
satisfy the triangular equality.
(f) When : = 2, prove that (/
1
. /

) = 1. On the other hand, if : 3, then (/

1
. /

)
1.
(g) A Theorem proved by F. John states that the diameter of (Norm. d) is precisely
1
2
log :. Show that this metric space is compact. Nota: One may consider the norm
whose unit ball is an :-agon in R
2
, with : even. Denote its class by 
m
. It seems
that d(/
1
.
m
) =
1
2
log 2 when 8[:.
20. (Continuation of Exercise 19.) We study here classes of norms (in Norm) that contain a
pair | |. | |

. We recall that the dual norm of | | is dened by

|r|

:= inf
x=0
1(

r)
||
.
As shown in the previous exercise, | | and | |

are in the same class if and only if there

exists an GL
n
(C) such that
(2) |r|

= |r|. r C
n
.
(a) Let , | | and | |

satisfy (2). Show that

is an isometry of (C
n
. | |), where

:= (

)
1
. Deduce that

is diagonalizable, with eigenvalues of modulus

one. Nota: The whole exercise is valid with the eld R instead of C, but the latter
result is a bit more dicult to establish.
(b) Let 1 GL
n
(C) be such that 1 := 1
1

1 is diagonal. Dene a norm  | |

by (r) := |1r|. Show that 1 is an isometry of (C
n
. ), and that

(1r) = (r). r C
n
.
where 1 := 1

1.
(c) Using Exercise 7 of Chapter 3 (page 56), prove that the class of | | contains a norm
A such that
A

(r) = A(r). r C
n
for some diagonal matrix . Show also that A(

r) A(r). Show that one

may choose unitary.
27
(d) Find more than one example of such classes of norms on C
2
.
21. (Numerical range.)
Given M
n
(C), dene :
A
(r) = (r. r) = r

r. The numerical range of is

1() = :
A
(r) ; |r|
2
= 1.
(a) We show that if : = 2, then 1() is an ellipse whose foci are the eigenvalues of .
i. First check that it suces to consider the cases of matrices
_
0 2c
0 0
_
.
_
1 2c
0 1
_
. c R
+
.
ii. Treat the rst case above.
iii. From now on, we treat the second case. First prove that 1() is the union of
circles with center j [1. 1] and radius :(j) = c
_
1 j
2
.
iv. We dene the (full) ellipse c C R
2
by the inequality
r
2
1 + c
2
+

2
c
2
1.
Show that 1() c.
v. Dene j p(j) :=
2
+ (r j)
2
:(j)
2
over [1. 1]. If (r. ) c, prove that
min p 0 ; deduce that p vanishes, and thus that (r. ) 1(). Conclude.
vi. Show that for a general 2 2 matrix , the area of / equals

4
[det[

. ][
1/2
.
(b) Deduce that for every :, 1() is convex (ToeplitzHausdor Theorem). See an
application in Exercise 131.
(c) When is normal, show that 1() is the convex hull of the spectrum of .
(d) Let us dene the numerical radius u() by
u() := sup[.[ ; . 1().
Prove that
u() ||
2
2u().
Hint: Use the polarization principle to prove the second inequality.
Deduce that u() is a norm (there are counter-examples showing that it is not
submultiplicative.)
(e) Let us assume that there exists a complex number 1() such that [[ = ||
2
,
say that = :
A
(r) for some unit vector r. Prove that is an eigenvalue of , with
r an eigenvector. Thus () = ||
2
= u(). Prove also that r is an eigenvector of

, associated with

. Hint: Show at rst that

r +

r = 2[[
2
r.
28
(f) If
2
= 0
n
, show that u() =
1
2
||
2
22. (Jacobi matrices.) Let M
n
(R) be tridiagonal, with c
i,i+1
c
i+1,i
0 when i = 1. . . . . :
1.
(a) Show that is similar to a tridiagonal symmetric matrix o with :
i,i+1
< 0.
(b) Deduce that has : real and simple eigenvalues. We denote them by
1
< <
n
.
(c) For , = 1. . . . . :, let
(j)
be the principal submatrix, obtained from by keeping
the rst , rows and columns. Without loss of generality, we may assume that the
o-diagonal entries of are non-positive, and denote /
j
:= c
j,,j+1
0. If is an
eigenvalue, show that
r := (/
1
1
/
1
j1
1
j1
())
1jn
. 1
j
(A) := det(
(j)
A1
j
)
is an eigenvector associated with .
(d) Deduce that the sequence of coordinates of the eigenvector associated with
j
has
exactly , 1 sign changes (one says that this eigenvector has , 1 nodes). Hint:
Use the fact that (1
1
. . . . . 1
n
) is a Sturm sequence.
How could this be proven rapidly when , = 1 ?
23. Let / be a eld with zero characteristic. Let M
n
(/) and | 1 be given. We form
the block-triangular matrix
:=
_
_
_
_
_
_
_
_
1
n
0
n
. . . 0
n
0
n
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0
n
.
.
.
.
.
.
.
.
. 1
n
0
n
. . . . . . 0
n

_
_
_
_
_
_
_
_
M
nl
(/).
(a) Let 1 /[A] be a polynomial. Show that 1() is block-triangular, with the
(i. i + :)-block equal to
1
r!
1
(r)
().
(b) If an irreducible polynomial divides a polynomial Q, together with its derivatives
Q
(r)
up to Q
(s)
, prove that
s+1
divides Q.
(c) Assume that the minimal polynomial of is irreducible. Compute the minimal
polynomial of . Compute the lists of its invariant polynomials.
(d) Deduce an alternate and somehow simpler second canonical form of square matrices.
24. Given  M
n
(/), denote [
i
j
[ the minor obtained by removing the i-th row and the
,-column. We dene similarly [
i,l
j,k
[.
29
(a) Prove DesnanotJacobi formula, also called Dodgson
1
condensation formula,
[
1,n
1,n
[ det  = [
1
1
[ [
n
n
[ [
1
n
[ [
n
1
[.
See Exercise 144 for a generalization.
Left: Charles L. Dodgson.
(b) Deduce a recursive algorithm for the
computation of det , in about 4:
3
,3
operations.
(c) Does this algorithm always produce a
result?
25. Consider the oriented graph whose vertices are the points of Z
2
and edges are the hor-
izontal (left to right) and vertical (downwards) segments. Given two vertices and 1,
denote by :(. 1) the number of paths from to 1. Thus :(. 1) 1 i
1
1
1
and

2
1
2
.
Given 2: points
1
. . . . .
m
and 1
1
. . . . . 1
m
, we assume that
i+1
(respectively 1
i+1
) is
strictly upper right with respect to
i
(resp. 1
i
) and that 1
m
is lower right with respect
to
1
.
(a) Consider :-tuples of paths
j
, each one joining
j
to 1
j
. Prove that the number of
such :-tuples, for which the
j
s are pairwise disjoint, equals the determinant of
 :=
_
:(
i
. 1
j
)
_
1i,jm
.
(b) Prove that the matrix  is totally positive.
26. Let l U
n
be upper Hessenberg. Up to a multiplication by a unitary diagonal matrix,
we may assume that the entries
k
:= n
k+1,k
are real non-negative. Prove that there
1
Charles Lutdwidge Dodgson, English mathematician. Educated people, as well as children, prefer to call
him Lewis Caroll.
30
exist numbers
k
C such that [
k
[
2
+
2
k
= 1 for / = 1. . . . . : 1, [
n
[ = 1 and
l = G
1
(
1
) G
n
(
n
), where
G
k
(
k
) := diag
_
1
k1
.
_

k

k

k

k
_
. 1
nk1
_
. / = 1. . . . . : 1.
and
G
n
(
n
) := diag(1. . . . . 1.
n
).
This is the Schur parametrization. Notice that the matrices G
k
(
k
) are unitary.
27. (P. D. Lax, H. F. Weinberger (1958).) Let \ be a linear subspace of M
n
(R), with the
property that the spectrum of every matrix  \ is real. We shall denote
1
()

n
() the eigenvalues of  \ , repeated with multiplicities.
(a) Prove that the functions 
k
() are continous over \ .
(b) Let . 1 \ , with
1
(1) 0 (one says that 1 is positive.) Given R, show that
the polynomial r det(1
n
r1) has : real roots (counting with multiplicities.)
Nota: This is really a dicult question, but just assume that the functions
k
are
innitely dierentiable away from the origin, a fact that is true when the eigenvalues
are simple for every non-zero  \ .
(c) With . 1 as above, prove that j
k
(+j1) is strictly increasing. Deduce that
j
k
( + j1) j
1
(1) is non-decreasing, and that j
k
( + j1) j
n
(1)
is non-increasing.
(d) Given A. 1 \ , show that

k
(A) +
1
(1 )
k
(A + 1 )
k
(A) +
n
(1 ).
Deduce that
1
and
n
are respectively a concave and a convex functions.
(e) Prove that the subset of positive matrices is a convex cone in \ . Deduce that the
relation 1, dened on \ by

k
()
k
(1). / = 1. . . . . :.
is an order relation.
(f) Give an example of such a subspace \ , of dimension :(: + 1),2. Did we already
know all the results above in this case ?
28. Denote by SL
+
2
(Z) the set of non-negative matrices with entries in Z and determinant
+1. Denote by 1. 1 the elementary matrices:
1 =
_
1 1
0 1
_
. 1 =
_
1 0
1 1
_
.
Show that SL
+
2
(Z) is the disjoint union of
1
2
. 1 SL
+
2
(Z) and 1 SL
+
2
(Z).
31
Deduce that SL
+
2
(Z) is the free monoid spanned by 1 and 1, meaning that for every
SL
+
2
(Z), there exists a unique word : in two letters, such that = :(1. 1). Notice
that 1
2
corresponds to the void word.
Show that

m
:=
_
_
1 : :
: 1 + :
2
0
: 0 1 + :
2
+ :
4
_
_
. : N
is an element of SL
+
3
(Z), which is irreducible, in the sense that
m
=  and . 
SL
+
3
(Z) imply that  or  is a permutation matrix (the only invertible elements in
SL
+
3
(Z) are the matrices of even permutations.) Deduce that SL
+
3
(Z) cannot be generated
by a nite number of elements.
29. (R. M. May, C. Jeries, D. Logofet, Ulianov.) We distinguish three signs . 0. + for real
numbers, which exclude each other. In mathematical terms, = (. 0), 0 = 0 and
+ = (0. +). The product of signs is well-dened.
Two given matrices . 1 M
n
(R) are said sign-equivalent if the entries c
ij
and /
ij
have
same sign, for every pair (i. ,). Sign-equivalence is obviously an equivalence relation.
An equivalence class is written as a matrix o, whose entries are signs. Here are three
examples of sign-classes:
o
1
= diag(. . ). o
2
=
_
0
+
_
. o
3
=
_
_
0 0
+ 0 0
+ +
_
_
.
In some applications of dynamical system theory, we are concerned with the asymptotic
stability of the origin in the system r = r. For some reason, we do know the signs
of the entries of , but the magnitude of non-zero entries is unknown. This arises for
instance in ecology or in the study of chemical reactions. Hence we ask whether the
sign structure of (its sign-class) ensures that the whole spectrum lies in the half-space
:= . C; 1. < 0, or not. If it does, then we say that this class (or this matrix) is
sign-stable.
Given a sign-class o, we denote ((o) the oriented graph whose vertices are the indices
, = 1. . . . . :, and arrows correspond to the non-zero entries :
ij
with i ,= ,.
(a) Show that the classes o
1
and o
2
above are sign-stable, but that o
3
is not.
(b) Actually (Hint), o
3
is reducible. Show that the determination of the stable sign-
classes reduces to that of the stable irreducible sign-classes.
From now on, we restrict to irreducible sign-classes.
(c) Given a class o, we denote by

o its closure, where = (. 0) is replaced by
(. 0], and similarly for +.
If a class o is sign-stable, show that

o is weakly sign-stable, meaning that its elements
have their spectra in the closed half-space

. Deduce the following facts:
32
i. :
ii
0 for every i = 1. . . . . :,
ii. ((o) does not contain cycles of length j 3.
(d) We restrict to sign-classes that satisfy the two necessary conditions found above.
i. Considering the trace of a matrix in o, show that sign-stability requires that
there exists a / such that :
kk
= .
ii. Deduce that, for a class to be stable, we must have, for every pair i ,= ,, either
:
ij
= :
ji
= 0 or :
ij
:
ji
< 0. In ecology, one says that the matrix is a predation
matrix. Hint: Use the irreducibility and the absence of cycle of length j 2.
iii. Under the two additional restrictions just found, show that every monomial (
a permutation)
c():
1(1)
:
n(n)
is either 0 or ()
n
. Deduce that the sign of the determinant is not ambiguous:
Either every element of o satises ()
n
det 0, or every element of o satises
det = 0. In the latter case, every monomial in det o must vanish.
iv. Check that sign-stability requires that the sign of the determinant be ()
n
.
(e) Check that the following class satises all the necessary conditions found above, but
that it is not sign-stable because it contains an element with eigenvalues i:
o
5
=
_
_
_
_
_
_
0 0 0 0
+ 0 0 0
0 + 0
0 0 + 0
0 0 0 + 0
_
_
_
_
_
_
.
(f) Show that the following class satises all the necessary conditions found above, but
one (det o
7
= 0):
o
7
=
_
_
_
_
_
_
_
_
_
_
0 0 0 0 0 0
+ 0 0 0 0
0 + 0 0
0 0 + 0 0 0 0
0 0 + 0 0 0
0 0 0 0 +
0 0 0 0 0 + 0
_
_
_
_
_
_
_
_
_
_
.
Actually, show that every element of o
7
satises r = 0 for some non-zero vector
r with r
1
= r
2
= r
3
= r
6
= 0.
30. Let G be a classical group of real or complex :: matrices. We only need that it satises
Proposition 7.3.1. Let G
1
be a compact subgroup of G, containing G U
n
.
(a) Let  be an element of G
1
, with polar decomposition QH. Verify that H belongs
to G
1
HPD
n
.
33
(b) Using the fact that H
m
G
1
for every : Z, prove that H = 1
n
.
(c) Deduce that G
1
= G l
n
. Hence, G U
n
is a maximal compact subgroup of G.
31. Following C. R. Johnson and C. J. Hillar (SIAM J. Matrix Anal. Appl., 23, pp 916-928),
we say that a word with an alphabet of two letters is nearly symmetric if it is the product
of two palindromes (a palindrome can be read in both senses; for instance the French city
LAVAL is a palindrome). Thus 111 = (11)1 is nearly symmetric. Check
that every word in two letters of length / 5 is nearly symmetric. Show that if a word
:(. 1) is nearly symmetric, then the matrix :(o
1
. o
2
) is diagonalizable with positive
real eigenvalues, for every symmetric, positive denite matrices o
1
. o
2
(see Exercise 258).
32. In R
1+m
we denote the generic point by (t. r)
T
, with t R and r R
m
. Let (
+
be the
cone dened by t |r|. Recall that those matrices of O(1. :) that preserve (
+
form the
subgroup G
+
. The quadratic form (t. r) |r|
2
t
2
is denoted by .
Let  belong to G
+
.
(a) Given a point r in the unit closed ball 1 of R
m
, let (t. )
T
be the image of (1. r)
T
under . Dene )(r) := ,t. Prove that ) is a continous map from 1 into itself.
Deduce that it has a xed point. Deduce that  has at least one real positive
eigenvalue, associated with an eigenvector in the closure of (
+
. Nota: If : is odd,
one can prove that this eigenvector can be taken in the light cone t = |r|.
(b) If  = and () ,= 0, show that [[ = 1.
(c) Let = (t. r) and u = (:. ) be light vectors (that is () = (u) = 0), linearly
independent. Show that

Ju ,= 0.
(d) Assume that  admits an eigenvalue of modulus dierent from 1, being an
eigenvector. Show that 1, is also an eigenvector. Denote by u a corresponding
eigenvector. Let < . u

be the orthogonal of and u with respect to . Using

the previous question, show that the restriction
1
of to < . u

is positive
denite. Show that < . u

is invariant under  and deduce that the remaining

eigenvalues have unit modulus.
(e) Show that, for every  G
+
, () is an eigenvalue of .
33. Assume that M
n
(C) is tridiagonal, with an invertible diagonal part 1. Assume that
the relaxation method converges for every parameter in the disc [ 1[ < 1.
(a) Show that, for every in the circle [ 1[ = 1, the spectrum of /

is included in
the unit circle.
(b) Deduce that the spectrum of the iteration matrix J of the Jacobi method is included
in the interval (1. 1). Compare with Theorem 9.4.1 and Exercise 7 of the book.
34. Let M
n
(C) be given, and l(t) := exp(t).
34
(a) Show that |l(t)| exp(t||) for t 0 and any matrix norm. Deduce that the
integral
_
+
0
c
2t
l(t)

l(t) dt
converges for every ||.
(b) Denote H

/ = 0 of / l(/)

l(/), by two dierent methods, deduce that H

is a solution
of
(3)

A + A = 2A 1
n
. A HPD
n
.
(c) Let be larger than the supremum of the real parts of eigenvalues of . Show
that Equation (3) admits a unique solution in HPD
n
, and that the above integral
converges.
(d)
Alexandr M. Lyapunov.
In particular, if the spectrum of  has positive
real part, and if 1 HPD
n
is given, then the
Lyapunov equation


H + H = 1. H HPD
n
Let r(t) be a solution of the dierential equation
r + r = 0, show that t r

Hr decays, and
strictly if r ,= 0.
35. Show that if  M
n
(C) and if Tr 

 :, then [ det [ 1, with equality if, and

only if,  is unitary.
36. Let / be a eld of characteristic zero, meaning that 1 spans an additive subgroup, iso-
morphic to Z. By a slight abuse of notation, this subgroup is denoted by Z. We call a
lattice of rank : if there exists a basis r
1
. . . . . r
n
of /
n
such that
= Zr
1
Zr
n
.
Such a basis of /
n
is called a basis of .
35
Let and

. Prove that there exist a ba-

sis
1
. . . . .
n
of , together with integers d
1
. . . . . d
n
, such that d
1
[d
2
, d
2
[d
3
, . . . , and
d
1

1
. . . . . d
n

n
is a basis of

. Show that d
1
. . . . . d
n
are uniquely dened by this prop-
erty, up to their signs. Finally, prove that the product d
1
d
n
equals the order [ :

]
of the quotient group ,

.
37. (From E. S. Key.) Given the companion matrix of a polynomial A
n
c
1
A
n1
c
n
,
in the form
_
_
_
_
_
0 1 0
.
.
.
.
.
.
.
.
. 0
0 0 1
c
n
c
1
_
_
_
_
_
.
and given a root r of 1, compute an eigenvector associated with r. Deduce that, if
1
1
. . . . . 1
k
have a common root ., then .
k
is an eigenvalue of the product of their com-
panion matrices.
38. Dene the wedge product in /
3
in the same way as in R
3
. Given a non-zero vector c in
/
3
, nd all matrices M
3
(/) with the property that (c) r = (c r) = c (r)
for every r /
3
. Hint: The result depends on whether c c = c
2
1
+ c
2
2
+ c
2
3
vanishes or
not.
39. (From Y. Tian.) Let 1 be a projector (that is 1
2
= 1) on a real or complex nite
dimensional space.
(a) Prove that 1
n
1

1 is positive semi-denite if and only if 1 is an orthogonal

projector, that is 1(1)ker(1).
(b) In general, prove the equalities
ker(1
n
1

1) = 1(1) 1(1

) = ker(21
n
1 1

).
and deduce that
1(1
n
1

1) = ker 1 + ker 1

= 1(21
n
1 1

).
40. Let | | be a unitary invariant norm on M
n
(C), and let HPD
n
and 1 M
n
(C)
n
, a matrix such that
exp(log ) = . For complex numbers ., we thus dene
z
:= exp(. log ).
(a) Dene 1(.) :=
z
1
1z
. Show that |1(.)| = |1(1.)|, then that 1 is bounded
on the strip 0 1. 1.
(b) If |1| 1 and |1| 1, deduce that |
1/2
1
1/2
| 1. More generally,
|
1/2
1
1/2
|
2
|1| |1|.
Compare with Exercise 21, page 79.
36
(c) Replacing 1 by 1

:= (1 +c|1|)
1
1 and by

:= +c1
n
with c 0, show that
the same result holds true if we suppose only that H
n
is positive semi-denite.
(d) More generally, if H. 1 are Hermitian positive semi-denite, prove that
|H11|
2
|H
2
1| |11
2
|.
41. Let H H
n
be positive semidenite and have bandwidth 2/ 1, meaning that [, i[ /
implies :
ij
= 0.
(a) Prove that the Choleski factorization H = 11

inherits the band-property of .

(b) Deduce that the largest eigenvalue of H is lower than or equal to the maximum
among the sums of / consecutive diagonal entries. Compare to Exercise 20, page 59
of the book.
Let r = / be a linear system whose matrix is symmetric positive denite (all entries
are real.) Recall that the convergence ratio of the Conjugate gradient method is the
number

GC
= log
_
1() 1
_
1() + 1
.
that behaves like 2,
_
1() when 1() is large, as it uses to be in the real life. The
number 1() :=
max
(),
min
() is the condition number of .
Preconditioning is a technique that reduces the condition number, hence increases the
convergence ratio, through a change of variables. Say that a new unknown is := 1
T
r,
so that the system is equivalent to

=

/, where

:= 1
1
1
T
. / = 1

/.
For a given preconditioning, we associate the matrices C := 11
T
and 1 := 1
n
C
1
.
Notice that preconditioning with C or
1
C is essentially the same trick if 0, although
1 = 1() diers signicantly. Thus we mere associate to C the whole family
1() = 1
n
C
1
; 0.
(a) Show that

is similar to C
1
.
(b) Consider the decomposition =   with  =
1
C and  =
1
C . This
yields an iterative method
C(r
k+1
r
k
) = / r
k
.
whose iteration matrix is 1(). Show that there exist values of for which the
method is convergent. Show that the optimal parameter (the one that maximizes
the convergence ratio) is

opt
=
2

min
(

) +
max
(

)
.
37
with the convergence ratio

opt
= log
1(

) 1
1(

) + 1
.
(c) When 1(

) is large, show that

GCP

opt

_
1(

).
where
GCP
stands for the preconditioned conjugate gradient, that is the conjugate

.
Conclusion ?
43. (Continuation.) We now start from a decomposition =   and wish to construct
a preconditioning.
Assume that  + 
T
, obviously a symmetric matrix, is positive denite. We already
know that |
1
|
A
< 1, where | |
A
is the Euclidean norm associated with (Lemma
9.3.1.)
(a) Dene 1 := (1
n

T
)(1
n

1
). Prove that |1|
A
< 1. Deduce that the
symmetric method
r
k+1/2
= r
k
+ /.
T
r
k+1
= 
T
r
k+1/2
+ /
is convergent (remark that =
T

T
.)
This method is called symmetric S.O.R., or S.S.O.R. when  is as in the relaxation
method.
(b) From the identity 1 = 1
n

T
(+
T
)
1
. we dene C = (+
T
)
1

T
.
Express the corresponding preconditioning C() when  and  come from the
S.O.R. method:
 =
1

1 1. (0. 2).
This is the S.S.O.R. preconditioning.
(c) Show that
max
(C()
1
) 1, with equality when = 1.
(d) Compute (1) and 1(

) when is tridiagonal with c
ii
= 2, c
i,i1
= 1 and c
ij
= 0
otherwise. Compare the S.S.O.R. method and the S.S.O.R. preconditioned conjugate
44. In control theory of linear systems, we face dierential equations
r = r + 1n.
where M
n
(R) and 1 M
nm
(R). We call r(t) the state and n(t) the control.
Controllability is the property that, given a time 1 0, an initial state r
0
and a nal
state r
f
, it is possible to nd a control t n(t) such that r(0) = r
0
and r(1) = r
f
.
38
(a) Assume rst that r(0) = 0. Express r(1) in terms of 1n and c
tA
(0 t 1) in a
closed form. Deduce that controllability is equivalent to r
f
H, where
H :=
n1
+
k=0
1(
k
1).
This is Kalmans criterion.
(b) Reversing the time, show that controllability from a general r(0) to r
f
= 0 is equiva-
lent to Kalmans criterion. Conclude that controllability for general initial and nal
states is equivalent to Kalmans criterion.
(c) Prove the following forms of Kalmans criterion:
i. ker 1
T
does not contain any eigenvector of
T
.
ii. For every complex number ., the matrix
_

T
.1
n
1
T
_
has rank :.
(d) Assume : = 1: The control is scalar. We shall denote / instead of 1, since it is a
vector. Furthermore, assume controllability. Show that there exists a vector c R
n
such that c
T
(+1
n
)
1
/ =
k
1
for / = 1. . . . . :. Deduce that the spectrum of +/c
T
reduces to 1. Hence the feedback n(t) = c(t) r(t) yields stabilization, since then
r(t) decays exponentially for every initial data.
45. Consider a zero-sum repeated game between two players A and B. Each player chooses
one object among a list C
1
. . . . . C
n
. When A chooses C
i
and B chooses C
j
, the payo
is :
ij
R, which is positive if A wins, negative if B wins. Obviously, the matrix  is
skew-symmetric.
Players play a large number of games. One may represent their strategies by vectors
r
A
, r
B
, where r
A
i
is the probability that A chooses the i-th object. Hence r
A
0 and

i
r
A
i
= 1, and the same for r
B
. Given a pair of strategies, the expectation has the form
(r
A
. r
B
) where (r. ) := r
T
. Player A tries to maximize, while B tries to minimize
the expectation.
A Nash equilibrium ( r. ) is a pair of strategies that is optimal for both players, in the
sense that, for all strategies r and :
(r. ) ( r. ) ( r. ).
(a) Prove that a Nash equilibrium always exists, and that the set of Nash equilibria is
a product C 1 of convex subsets.
(b) Deduce that, for every skew-symmetric matrix with real entries, there exists a non-
negative vector r ,= 0, such that r is non-negative, and r
j
(r)
j
= 0 for each
, = 1. . . . . :.
39
(c) Example: The list of objects consists in scissors, a stone, a hole and a sheet of
paper. The payo is 1, according to the following natural rules. Scissors win
against paper, but looses against the hole and the stone. Paper wins against the
hole and the stone. The hole wins again the stone.
Find the (unique and rather counter-intuitive) Nash equilibrium.
Remark: In many countries, this game is more symmetric, with only scissors, stone
and the sheet of paper. It was illustrated during WWII, when the leaders of Great
Britain, USSR and Germany each had their own choice. Ultimately, the cissors
and the stone defeated the sheet of paper. During the Cold War, there remained
the cissors and the stone, untill Stalines death in 1953 and Churchills loss of 1955
elections. One had to wait untill 1991 to see the cissors defeating the stone, unlike
in the usual game.
46. (P. Van den Driessche, H. K. Wimmer.)
(a) Characterize the complex numbers c. /. c and the vectors A C
n1
, such that the
following matrix is unitary
l :=
_
cA

c
1
n1
/AA

A
_
.
(b) Let C be a companion matrix, given in the form
C =
_
0

:
1
n1
\
_
. : C. \ C
n1
.
Find the polar decomposition C = QH. Hint: Q equals l, where l is as in the
previous question.
47. Let 1 be an invariant subspace of a matrix  M
n
(R).
(a) Show that 1

is invariant under 
T
.
(b) Prove the following identity between characteristic polynomials:
(4) 1
M
(A) = 1
M|E
(A)1
M
T
|E
(A).
48. (See also Yakubovich & Starzhinskii, Linear dierential equations with periodic coe-
cients. Wiley & Sons, 1975.)
Let  belong to Sp
n
(R). We recall the notations of Chapter 7: 
T
J = J and
J
2
= 1
2n
as well as J
T
= J.
(a) Show that the characteristic polynomial is reciprocal:
1
M
(A) = A
2n
1
M
_
1
A
_
.
Deduce a classication of the eigenvalues of .
40
(r) := 2r
T
Jr.
Verify that  is a -isometry.
(c) Let (c
i
. c
i
) be a pair of simple eigenvalues of  on the unit circle. Let be the
corresponding invariant subspace:
:= ker(
2
2(cos ) + 1
2n
).
i. Show that J

is invariant under .
ii. Using the formula (4) above, show that c
i
are not eigenvalues of [
J
.
iii. Deduce that R
2n
= J

.
(d) (Continued.)
i. Show that does not vanish on 0. Hence denes a Euclidian structure
on .
ii. Check that [

iii. Show that [

is a rotation with respect to the Euclidian structure dened by

, whose angle is either or .
(e) More generally, assume that a plane is invariant under a symplectic matrix ,
with corresponding eigenvalues c
i
, and that is not Lagrangian: (r. )
T
Jr is
not identically zero on . Show that [

acts as rotation of angle . In particular,

if  = J, show that = +,2.
(f) Let H be an invariant subspace of , on which the form is either positive or
negative denite. Prove that the spectrum of [
H
lies in the unit circle and that
[
H
is semisimple (the Jordan form is diagonal).
(g) Equivalently, let be an eigenvalue of  (say a simple one) with , R and [[ , = 1.
Let H be the invariant subspace associated with the eigenvalues (.

. 1,. 1,

).
Show that the restriction of the form to H is neither positive nor negative denite.
Show that the invariant subspace 1 associated with the eigenvalues and

is
-isotropic. Thus, if [
H
is non-degenerate, its signature is (2. 2).
49. (From L. Tartar.) In M
n
(R), prove the inequality
(Tr )
2
(rk ) Tr(
T
).
Use the latter to built as many as possible non-trivial quadratic forms on M
n
(R), non-
negative on the cone of singular matrices.
50. The minors of general matrices are not independent on each other. For instance, each
entry is a minor (of order one) and the determinant (an other minor) is dened in terms of
entries. An other instance is given by the row- or column-expansion of the determinant.
41
Denote 1
2m
the set of partitions 1 J of 1. . . . . 2: into two sets 1 and J of equal lengths
:. If (1. J) 1
m
, let (1. J) be the signature of the permutation (i
1
. . . . . i
m
. ,
1
. . . . . ,
m
),
where
1 = i
1
i
m
. J = ,
1
,
m
.
Prove that, for every matrix M
2mm
(/), there holds

(I,J)Pm
(1. J)(1)(J) = 0.
with
(1) :=
_
1 :
i
1
i
m
_
.
Find other algebraic relations (syzygies) between minors.
51. (a) Let belong to SPD
n
. Prove that the linear map + is an automorphism
of Sym
n
(R). Hint: Consider the spectrum of
2
. Show that it is real non-negative
on one hand, non-positive on the other hand. Then conclude.
(b) Let belong to SPD
n
. Compute the dierential of the map
2
.
(c) Deduce that the square root map o

o is analytic on SPD
n
. Remark: The
same result holds true on HPD
n
, same proof.
52. We consider real symmetric : : matrices. We use the SchurFrobenius norm | |
F
.
The result would be the same for complex Hermitian matrices.
(a) Given two matrices = diag(c
1
. . . . . c
n
) and 1 = 1diag(/
1
. . . . . /
n
)1
T
, with 1 an
orthogonal matrix, verify that
|1 |
2
F
=

i,l
j
2
il
(c
i
/
l
)
2
.
(b) Assume that both and 1 are positive denite and as above. Prove that
|

|
4
F
:|1 |
2
F
.
Hint: Use [

c[
2
[/ c[ for positive real numbers, together with Cauchy
Schwarz inequality.
(c) Deduce that the square root map, dened on SPD
n
, is H olderian with exponent
1,2. Verify that the supremum of
|

|
2
F
|1 |
1
F
.
taken either on SPD
n
or on the subset of diagonal matrices, takes the same value.
Remark: See the improvement of this result in Exercise 110.
42
53. The electromagnetic eld (1. 1) must be understood as an alternate 2-form:
= dt (1 dr) + 1
1
dr
2
dr
3
+ 1
2
dr
3
dr
1
+ 1
3
dr
1
dr
2
.
In coordinates (t. r
1
. r
2
. r
3
), it is thus represented by the alternate matrix
=
_
_
_
_
0 1
1
1
2
1
3
1
1
0 1
3
1
2
1
2
1
3
0 1
1
1
3
1
2
1
1
0
_
_
_
_
M
4
(R).
In another choice of Lorentzian coordinates, is thus represented by the new matrix

:=
T
. Matrix  O(1. 3) is that of change of variables. With

are associated
(1

. 1

). Namely, the decomposition of the electro-magnetic eld into an electric part

and a magnetic one depends on the choice of coordinates. The purpose of this exercise is
to nd Lorentzian invariants, that is quantities associated with that do not depend on
the choice of coordinates.
(a) If  = diag(1. Q) with Q O
3
(R), express 1

. 1

in terms of 1. 1 and Q.
(b) Verify that  belongs to O(1. 3) SPD
4
if and only if there exists a unit vector
and numbers c. : with c
2
:
2
= 1, such that
 =
_
c :
T
: 1
3
+ (c 1)
T
_
.
Find how  transforms 1 and 1.
(c) Verify that [1[
2
[1[
2
and [1 1[ are Lorentzian invariants.
(d) We now show that [1[
2
[1[
2
and [1 1[ are the only Lorentzian invariants. Thus
let (1
i
. 1
i
) and (1
f
. 1
f
) be two pairs of vectors in R
3
, dening two 2-forms
i
and

f
. We assume that
[1
f
[
2
[1
f
[
2
= [1
i
[
2
[1
i
[
2
=: c. [1
f
1
f
[ = [1
i
1
i
[ =: .
i. Choose S
2
that is orthogonal to both 1
i
and 1
i
. If R is given, dene
)() := [(cosh )1 + (sinh )1 [
2
.
Show that the range of ) covers exactly the interval [
1
2
(c +

c
2
+ 4). +).
ii. Deduce that there exist a pair (1
m
. 1
m
) of vectors, both belonging to the plane
spanned by 1
i
and 1
i
and with
[1
m
[
2
= [1
f
[
2
. [1
m
[
2
= [1
f
[
2
. [1
m
1
m
[ = .
and a symmetric positive denite Lorentzian matrix 
i
such that

i
(
i
r. 
i
)
m
(r. ).
43
iii. Show also that there exists an orthogonal Lorentzian matrix 1
f
such that

m
(1
f
r. 1
f
)
f
(r. ).
Conclude.
54. Let H
n
be given blockwise in the form
=
_
1 C
C

1
_
. 1 H
m
. 1 H
nm
.
Assume that the least eigenvalue of 1 is greater than the highest eigenvalue j of 1.
Prove that the spectrum of is included into (. j] [. +). Prove also that [. +)
contains exactly : eigenvalues, counting with multiplicities.
55. Let / be a eld and . 1. C be non-colinear matrices in M
n
(/), such that 1, 1 C
and C have rank one. Show that every distinct matrices 1. Q in the ane plane
spanned by . 1. C dier from a rank-one matrix.
On the contrary, nd in
2
(/) four matrices . 1. C. 1 such that 1, 1 C, C 1,
1 have rank one, but for instance C has rank two.
56. Given an abelian ring , recall that for a square matrix  M
n
transpose of the cofactors of , so that the following identity holds
(adjM) = (adj) = (det )1
n
.
(a) Prove that the rank of adj equals either :, 1 or 0. Characterize the three cases by
the rank of . Describe in more details adj when rk = : 1 and is a eld.
(b) Given : 2, show that det(adj) = (det )
n1
and adj(adj) = (det )
n2
.
In the special case : = 2, the latter means adj(adj) = . In particular, if : 3
(but not when : = 2), we have
(det  = 0) =(adj(adj) = 0
n
).
Hint: First prove the formul when is a eld and det  ,= 0. Then, considering
that the required identities are polynomial ones with integral coecients, deduce
that they hold true in Z[A
11
. . . . . A
nn
] by choosing = Q, and conclude.
57. (A. Majda, Thomann) Let := diag1. 1. . . . . 1 be the matrix of the standard
scalar product . in Lorentzian geometry. The vectors belong to R
1+n
r = (r
0
. . . . . r
n
)
T
. The forward cone 1
+
is dened by the inequalities
r. r 0. r
0
0.
The words non-degenerate and orthogonal are used with respect to the Lorentzian scalar
product.
44
(a) Let . : 1
+
be given. Show the reverse CauchySchwarz inequality
(5) . :. : . :
2
.
with equality if, and only if, and : are colinear. This is a special case of a deep
result by L. Garding.
(b) More generally, if 1
+
and : R
1+n
, prove (5).
(c) Given j R and . : 1
+
H

(r) := j. r :. r . : r. r.
i.
Hendrik Lorentz.
In the case = : = c
0
:= (1. 0. . . . . 0)
T
,
check that H

is positive denite when-

ever j 1.
Deduce that if j 1 and if . : are
colinear, then H

is positive denite.
Hint: Use a Lorentz transformation to
drive to c
0
.
ii. We now assume that and : are not colinear. Check that the plane 1 spanned
by and : is non-degenerate and that the Lorentz norm is denite on 1

.
Deduce that H

is positive denite if, and only if, its restriction to 1 has this
property.
iii. Show that H

is not positive denite for j 1, as well as for large positive js.

On the other hand, show that it is positive denite for every j in a neighbour-
hood of the interval
_
2.
2. :
2
. :
2
. :. :
_
.
58. (a) Given a matrix  M
n
(C) that have a positive real spectrum, show that
det(1
n
+ ) 2
n

det .
(b) Deduce that, for every positive denite Hermitian matrices H. 1, there holds
(det(H + 1))
2
4
n
(det H)(det 1).
Nota: This inequality can be improved into
(det(H + 1))
1/n
(det H)
1/n
+ (det 1)
1/n
.
the proof of which being slightly more involved. This can be viewed as a consequence
of an inequality of L. Garding about hyperbolic polynomials. See Exercise 218 for a
proof without polynomials. See Exercise 219 for an improvement of the inequality
above.
45
(c) Show that the map
H log det H
is strictly convex upon HPD
n
. Notice that Gardings inequality tells us the better
result that H (det H)
1/n
is concave. This is optimal since this map is homogeneous
of degree one, and therefore is linear on rays R
+
H. However, it depends on :, while
H log det H does not.
(d) Deduce that the set of positive denite Hermitian matrices such that det H 1 (or
greater or equal to some other constant) is convex.
59. Let  M
n
(C) be given, such that every matrix of the form  + H, H Hermitian, has
a real spectrum.
(a) Prove that there exists a matrix  in  +H
n
that has only simple eigenvalues.
(b) Because of simplicity, the eigenvalues  + H
j
are (

-functions for H small.

Compute their dierentials:
d
j
() 1 =
1

i
1A
i
1

i
A
i
.
where A
i
, 1
i
are eigenvectors of  and

, respectively.
(c) Show that 1
i
and A
i
are colinear. Hint: Take 1 of the form rr

with r C
n
.
(d) Deduce that  is Hermitian.
60. One wishes to prove the following statement: A matrix  M
n
(C), such that  +H is
diagonalizable for every H H
n
, has the form ic1
n
+1 where c R and 1 is Hermitian.
(a) Prove the statement for : = 2.
(b) Let  satisfy the assumption. Show that there exists a Hermitian matrix 1 such
that +1 is upper triangular with pure imaginary diagonal entries. Then conclude
by an induction.
61. (P. D. Lax) Let M
n
(/) be given, with / = R or C. Assume that for every r /
n
,
the following bound holds true:
[r. r[ |r|
2
.
Deduce that the sequence of powers (
m
)
mN
is bounded. Hint: Prove that the unitary
eigenvalues are semi-simple. Then use Exercise 10 of Chapter 4.
62. Let c R
n
denote the vector (1. . . . . 1)
T
. A square matrix  M
n
(R) is called a
Euclidean distance matrix (EDM) if there exist vectors j
1
. . . . . j
n
in a Euclidean vector
space 1, such that :
ij
= |j
i
j
j
|
2
for every pair (i. ,).
46
(a) Show that every Euclidean distance matrix,
besides being non-negative and symmetric
with a diagonal of zeroes (obvious condi-
tions), also denes a non-positive quadratic
form on the hyperplane c

.
Left: Euclid
(b) Given a symmetric matrix  with a diagonal of zeroes, assume that it denes a
non-positive quadratic form on the hyperplane c

. Check that  0 (in the sense

that the entries are nonnegative.) Prove that there exists a vector R
n
such that
(r) := ( r)(c r)
1
2
r
T
r
is non-negative. Let o be the symmetric matrix associated with , and let 1 be a
symmetric square root of o. Prove that  is an EDM, associated with the column
vectors j
1
. . . . . j
n
of 1.
(c) Prove that the minimal dimension : of the Euclidean space 1, equals the rank of
JJ, where J is the orthogonal projection onto c

:
J := 1
n

1
:
cc
T
.
63. (F. Hansen, G. Pedersen) Hereafter, we denote by (r. ) =

j
r
j

j
the usual Hermitian
product in C
n
. Given a numerical function ) : 1 R dened on an interval, and given
a Hermitian : : matrix H, with Sp(H) 1, we dene )(H) in the following natural
way: Let H = l

1l be a diagonalization of H in a unitary basis, 1 = diagd

1
. . . . . d
n
,
then )(H) := l

)(1)l, where
)(1) = diag)(d
1
). . . . . )(d
n
).
(a) Find a polynomial 1 R[A], that depends only on ) and on the spectrum of H,
so that )(H) = 1(H). Deduce that the denition above is not ambiguous, namely
that it does not depend on the choice of the unitary eigenbasis.
(b) Let : be any positive integer and H
1
. . . . . H
m
be Hermitian. We also give : matrices

1
. . . . .
m
in M
n
(C), with the property that

1
+ +

m
= 1
n
.
Finally, we dene
H :=

1
H
1

1
+ +

m
H
m

m
.
47
i. Let 1 be an interval of R that contains all the spectra of H
1
. . . . . H
m
. Show that
H is Hermitian and that 1 contains Sp(H).
ii. For each 1, we denote by 1
k
() the orthogonal projector on ker(H
k
). If
is a unit vector, we dene the (atomic) measure j

by
j

(o) =
m

k=1

S
(1
k
()
k
.
k
).
Show that j

is a probability. Also, show that if is an eigenvector of H, then

(H. ) =
_
dj

().
iii. Under the same assumptions, show that
()(H). ) = )
__
dj

()
_
.
iv. If ) is convex on 1, deduce that
Tr)(H) Tr
_
m

k=1

k
)(H
k
)
k
_
.
Hint: Use Jensens inequality, plus the fact that
Tr =
n

l=1
(
l
.
l
).
for every unitary basis
1
. . . . .
n
, for instance an eigenbasis if  is Hermitian.
64. We deal with complex : : matrices. We denote by () the spectrum of and by
() its complement, the resolvant set of . We use only the canonical Hermitian norm
on C
n
and write || for the induced norm on M
n
(C) (we wrote ||
2
in the book). We
denote dist(.; 1) the distance from a complex number . to a closed subset 1 in C.
(a) Prove that for every matrix M
n
(C) and complex number . (), there holds
(6) |(. )
1
|
1
dist(.; ())
.
(b) When is normal, prove that the equality holds in (6).
(c) Conversely, we consider a matrix such that the equality holds in (6).
i. Show that we may assume, up to a unitary conjugation, that be block-
triangular
=
_
A

0 1
_
.
with A C
n1
. Hint: Apply Theorem 3.1.3 (Schur).
48
ii. When is block-triangular as above, compute the inverse of . blockwise,
when . is close to (but distinct from) . Establish the following inequality
21
_
A

(. 1)
1

_
+[. [
2
|(. 1)
1
|
2
||
2
.
for every complex number and C
nm
. Deduce that A = 0.
iii. Conclude, by an induction, that is diagonal. Finally, show that a matrix
satises for every . the equality in (6), if and only if it is normal.
65. We use the notations of the previous exercise. In addition, if c 0 we dene the c-
pseudospectrum as

() := ()
_
. () ; |(. )
1
|
1
c
_
.
We recall (Exercise 21 in this list) that the numerical range
1() := :
A
(r) ; |r|
2
= 1
is a convex compact subset.
(a) Prove that

() =
_
B
( + 1).
(b) Prove also that

() . C; dist(.; 1()) c.
66. (From notes by M. Coste.) This exercise shows that a matrix  GL
n
(R) is the
exponential of a real matrix if, and only if, it is the square of another real matrix.
(a) Show that, in M
n
(R), every exponential is a square.
(b) Given a matrix M
n
(C), we denote / the C-algebra spanned by , that is the
set of matrices 1() as 1 runs over C[A].
i. Check that / is commutative, and that the exponential map is a homomorphism
from (/. +) to (/

. ), where /

denotes the subset of invertible matrices (a

multiplicative group.)
ii. Show that /

is an open and connected subset of /.

iii. Let 1 denote exp(/), so that 1 is a subgroup of /

. Show that 1 is a neigh-

bourhood of the identity. Hint: Use the Implicit Function Theorem.
iv. Deduce that 1 is closed in /

; Hint: See Exercise 21, page 135. Conclude that

1 = /

.
v. Finally, show that every matrix 1 GL
n
(C) reads 1 = exp(1(1)) for some
polynomial 1.
49
(c) Let 1 GL
n
(R) and 1 C[A] be as above. Show that
1
2
= exp(1(1) +

1(1)).
Conclusion ?
67. (The Le VerrierFadeev method.) Given M
n
(/), we dene inductively a sequence
(
j
. c
j
. 1
j
)
1jn
by

j
= 1
j1
(or
1
= ). c
j
=
1
,
Tr
j
. 1
j
=
j
+ c
j
1
n
.
Rowan Hamilton.
Show that the characteristic polynomial of is
A
n
+ c
1
A
n1
+ + c
n
.
Apply CayleyHamiltons theorem and compare with
Exercise 25, page 37.
68. Given M
n
(/), with its characteristic polynomial
1
A
(A) = A
n
+ c
1
A
n1
+ + c
n
.
we form a sequence of polynomials by the Horners rule:
j
0
(A) := 1. j
1
(A) := A + c
1
. j
j
(A) = Aj
j1
(A) + c
j
. . . .
Prove that
(A1
n
)
1
=
1
1
A
(A)
n1

j=0
j
j
()A
nj1
.
69. Let M
n
(C) be given, with eigenvalues
j
and singular values
j
, 1 , :. We
choose the decreasing orders:
[
1
[ [
2
[ [
n
[.
1

2

n
.
Recall that the
j
s are the square roots of the eigenvalues of

.
We wish to prove the inequality
k

j=1
[
j
[
k

j=1

j
. 1 / :.
50
(a) Prove directly the case / = 1. Show the equality in the case / = :.
(b) Working within the exterior algebra, we dene
p
End(
p
(C
n
)) by

p
(r
1
r
p
) := (r
1
) (r
p
). r
1
. . . . . r
p
C
n
.
Prove that the eigenvalues of
p
are the products of j terms
j
with pairwise
distinct indices. Deduce the value of the spectral radius.
(c) We endow
p
(C
n
) with the natural Hermitian norm in which the canonical basis
i
1
e
ip
with i
1
< < i
p
, is orthonormal. We denote by . the
scalar product in
p
(C
n
).
i. If r
1
. . . . . r
p
.
1
. . . . .
p
C
n
, prove that
r
1
r
p
.
1

p
= det (r

j
)
1i,jp
.
ii. For  M
n
(C), show that the Hermitian adjoint of 
p
is (

)
p
.
iii. If l U
n
, show that l
p
is unitary.
iv. Deduce that the norm of
p
equals
1

p
.
(d) Conclude.
70. Use Exercise 20.a of Chapter 5 to prove the theorem of R. Horn & I. Schur : The set
of diagonals (/
11
. . . . . /
nn
) of Hermitian matrices with given spectrum (
1
. . . . .
n
) is the
convex hull of the points (
(1)
. . . . .
(n)
) as runs over the permutations of 1. . . . . :.
71. (A theorem by P. D. Lax.)
Assume that a subspace \ of M
n
(R) has dimension 3, and that its non-zero elements
have a real spectrum, with pairwise distinct eigenvalues. When  \ and  ,= 0
n
,
denote

1
() < <
n
()
its eigenvalues. We equip R
n
with the standard Euclidean norm.
(a) Verify that

j
() =
nj+1
().
(b) Prove that there exists a continuous map
 B() = :
1
(). . . . . :
n
().
dened on \ 0
n
, such that B() is a unitary basis of . Hint: The domain
\ 0
n
is simply connected.
(c) Let choose 
0
a non-zero element of \ . We orient R
n
in such a way that B(
0
) be
a direct basis. Show that B() is always direct.
(d) Show also that for every ,, there exists a constant
j
= 1 such that, for every
non-zero , there holds
:
j
() =
j
:
nj+1
().
51
(e) From the former questions, show that if : 2. 3 (mod 4), then
n

j=1

j
= 1.
(f) On another hand, show that there always holds

nj+1
= 1.
Deduce that : , 2 (mod 4).
72. (The exchange, gyration or sweep operator.) Given a matrix  M
n
(/) in block form
=
_
1
C 1
_
. GL
p
(/).
we dene a matrix ( := : j)
exc() :=
_
1
p
0
pq
C 1
_

_
1
0
qp
1
q
_
1
=
_

1
1
C
1
1 C
1
1
_
.
(a) Show that the exchange map is an involution:
exc(exc()) = .
(b) If 1 is non-singular, prove that exc() is non-singular, with
exc(
1
) = exc()
1
.
(c) Let J := diag1
p
. 1
q
. Show that
exc(JJ)
T
= exc(
T
).
(d) We restrict to / = R. Recall that O(j. ) is the orthogonal group associated with J.
Show that the exchange map is well dened on O(j. ). With the previous formul,
prove that it maps O(j. ) on a subset of O
n
(R).
(e) Show that the image of the exchange map is a dense open subset of O
n
(R).
73. Using the quadratic forms of R
n
that are preserved by elements of the groups O(j. ) and
O
n
(R), nd a simpler proof of the fact that the exchange map maps the former into the
latter.
74. Given a function ) : (0. +) R, we may dene a map
 )().
SPD
n
Sym
n
(R)
in the same way as we dened the square root. The uniqueness is proved with the same
argument (see for instance Exercise 63.a). We say that ) is a monotone matrix function
if, whenever 0
n
<  <  in the sense of quadratic forms, there holds )() < )().
52
(a) Prove that )(:) := 1,: is a monotone matrix function.
(b) Verify that the set of monotone matrix functions is a convex cone. Deduce that, for
every nonnegative, non-zero measure :,
)(:) :=
_
+
0
d:(t)
: + t
is a monotone matrix function.
(c)
Henri Poincare.
Prove that, given numbers c 0, / R and a non-
negative bounded measure :, such that (c. :) ,=
(0. 0),
(7) )(:) := c: + /
_
+
0
1 :t
: + t
d:(t)
is a monotone matrix function (Loewners theo-
rem asserts that every monotone matrix function
is of this form ; such functions have a holomor-
phic extension to the domain C R

, and send
the Poincare half-space . 0 into itself. The
formula above is the Nevanlinna representation of
such functions.)
(d) Prove that, given a non-negative measure ,= 0,
:
__
+
0
d(t)
: + t
_
1
is a monotone matrix function (this function is the inverse of the Cauchy transform
of the measure .)
(e) Compute
_

0
dt
t

(: + t)
.
Deduce that )

(:) = :

is a monotone matrix function for every (0. 1).

(f) Prove that 0
n
<   implies log  log . Hint: Consider the map
1
()

1).
(g) Find two matrices . 1 SPD
2
(R) such that 1 + 1 , SPD
2
(R). Deduce
that )
2
(:) = :
2
is not a monotone matrix function.
53
75. Given o SPD
n
(R), prove the formula

n/2

det o
=
_
R
n
c
x
T
Sx
dr.
Deduce that
o log det o
is concave on SPD
n
(R). Nota: an other proof was given in Exercise 58.
76. Let  M
n
(R) be non-negative and let us choose numbers :
1
. . . . . :
n
in (1. +). For
R, dene
o

:= diag:

1
. . . . . :

n
.
Assume that  o

is non-singular for every 0. prove that () < 1.

77. (From T. Barbot.) Let o M
m
(R) and 1 M
pm
(R) be given, with o symmetric. Our
goal is to prove that there exists a symmetric M
p
(R) such that
(8)
_
_
_
_
_
o 1
T
1
__
_
_
_
2

_
_
_
_
_
o
1
__
_
_
_
2
=: .
Indeed, since the reverse inequality is always true, we shall have an equality. Nota: This
is a particular case of Parrotts lemma. See Exercise number 87.
This property may be stated as a vector-valued version of the HahnBanach theorem for
symmetric operators: Let n : 1 1, dened on a subspace of 1, a Euclidian space, with
the symmetric property that n(r). = r. n() for every r. 1, then there exists
a symmetric extension l /(1), such that |l| |n| (and actually |l| = |n|) in
operator norm.
By homogeneity, we are free to assume = 1 from now on. In the sequel, matrix
inequalities hold in the sense of quadratic forms.
(a) Show that o
2
+ 1
T
1 1
m
.
(b) For [j[ 1, show that
H

:= j1
p
1(j1
m
o)
1
1
T
is well-dened, and that the map j H

is monotone increasing on each of the

intervals (. 1) and (1. +).
(c) We begin with the case where (o) < 1. Then (1
m
o
2
)
1
is well-dened and
symmetric, positive denite. Prove that |1(1
m
o
2
)
1/2
|
2
1. Deduce that
|(1
m
o
2
)
1/2
1
T
|
2
1. Conclude that H
1
H
1
.
(d) Let be symmetric with H
1
H
1
. For instance, = H
1
is convenient.
Prove the inequality (8). Hint: Consider an eigenvector (r. )
T
for an eigenvalue
j 1. Compute
T
and reach a contradiction. Proceed similarly if j < 1.
54
(e) In the general case, we have by assumption (o) = |o|
2
1. Replace o by to
with 0 < t < 1 and apply the previous result. Then use a compactness argument as
t 1

.
78. Let 1 be a compact subgroup of GL
n
(R). We admit the existence of a Haar measure,
that is a probability j on 1, with the left-invariance property:
_
K
(p/) dj(/) =
_
K
(/) dj(/). ((1). p 1.
(a) Let [ [ denote the canonical Euclidian norm, and (. ) its scalar product. For
r. R
n
, dene
r. :=
_
K
(/r. /) dj(/). |r| := r. r
1/2
.
Show that . is a scalar product, for which every element of 1 is an isometry.
(b) Deduce that 1 is conjugated to a subgroup of O
n
(R). Similarly, prove that every
compact subgroup of GL
n
(C) is conjugated to a subgroup of U
n
.
79. (Thanks to P. de la Harpe and E. Ghys.) Let j. 1 be integers. We endow C
p
and C
q
with the canonical Hermitian scalar products . . :=
1
.
1
+ . On C
p+q
, we consider
their dierence. The corresponding Hermitian form is
Q(.) = [.
1
[
2
+ [.
p+1
[
2
.
Denote by A the set of linear subspaces of C
p+q
on which the restriction of Q is positive
denite, and which are maximal for this property.
(a) Show that 1 A if, and only if, it is the graph
(r. r) [ r C
p

of a matrix  M
qp
(C) with || < 1 ; this norm is taken with respect to the
Hermitian norms of C
p
and C
q
, in particular, ||
2
= (

) = (

).
(b) Let 2 U(j. ), written blockwise as
2 =
_
1
C 1
_
.
Verify that, if 1 A, then 21 A. Deduce that if  lies in the unit ball B of
M
qp
(C), then so does the matrix

Z
() := ( + 1)(C + 1)
1
.
(c) Show that : U(j. ) Bij(B) is a homomorphism. In other words, it is a group
action of U(j. ) over B.
55
(d) Let  B be given. Prove that there exists a unique element 2 of U(j. ) HPD
n
such that  =
Z
(0
qp
). This shows that the group action is transitive.
(e) Find the stabilizer of 0
qp
, that is the set of 2s such that
Z
(0
qp
) = 0
qp
.
(f) Deduce that B is dieormorphic to the homogeneous space
U(j. ),(U
p
U
q
).
80. Given a function : R R, we admit that there is a unique way to dene a :
Sym
n
(R) Sym
n
(R) such that (C
T
oC) = C
T
oC if C is orthogonal, and (1) =
diag(d
1
). . . . . (d
n
) in the diagonal case (see for instance Exercise 63.a). It is clear
that the spectrum of (o) is the image of that of o under .
Let 1 M
n
(R) be symmetric and M
n
(R) be diagonal, with diagonal entries
c
1
. . . . . c
n
. If is of class (
2
, prove Balians formula:
lim
t0
t
2
Tr(( + t1) + ( t1) 2()) =

i
/
2
ii

(c
i
) +

i,j (j=i)
/
2
ij

(c
j
)

(c
i
)
c
j
c
i
.
Hint: Prove the formula rst in the case where (t) = t
m
for some integer :. Then pass
to general polynomials, then to (
2
functions.
81. We denote by | |
F
the Frobenius norm: ||
2
F
= Tr(

).
(a) Show that the set A
n
of :: normal matrices is closed. Deduce that if M
n
(C),
there exists an  in A
n
for which | |
F
is minimum.
(b) Given / H
n
and t R, exp(it/) is unitary. Therefore we have
| |
F
| c
ith
c
ith
|
F
.
By letting t 0, deduce that
(9) ( )

( ) H
n
.
(c) Using a unitary conjugation, show that we may assume that  is diagonal.
In that case, write  = diagd
1
. . . . . d
n
. Then:
i. Show that d
j
= c
jj
. Hint: Compare with other diagonal matrices.
ii. Suppose that d
j
= d
k
(=: d) for some pair (,. /) (, ,= /). Verify that
|1 d1
2
|
F
|1 :|
F
. : A
2
.
where
1 :=
_
c
jj
c
jk
c
kj
c
kk
_
.
Deduce that c
jk
= c
kj
= 0 (in other words, 1 = d1
2
).
56
iii. From (9) and the previous question, deduce that one can dene a Hermitian
matrix H, such that
/
jk
(d
k
d
j
) = c
jk
. ,. / (, ,= /).
(d) In conclusion, prove that for every in M
n
(C), there exist a normal matrix  (the
one dened above) and a Hermitian one H, such that
=  + [H. ].
82. (von Neumann inequality.)
John von Neumann.
Let  M
n
(C) be a contraction, meaning that
||
2
1. In other words, there holds 

1
n
in the sense of Hermitian matrices. We re-
call that |

|
2
= ||
2
, so that we also have


1
n
. We denote o =

1
n


and
1 :=

1
n


. Such positive square roots do

exist, from unitary diagonalisation ; they turn out
to be unique, but we do not use this fact.
Given an integer / 1 and 1 = 2/ + 1, we dene a matrix \
k
M
Kn
(C) blockwise:
\
k
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
.
.
1
n
1
n
o 

1
1
n
1
n
.
.
.
1
n
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
.
where the dots represent blocks 1
n
, while missing entries are blocks 0
n
. The column and
row indices range from / to /. In particular, the central block indexed by (0. 0) is .
All the other diagonal blocks are null.
(a) We begin with the easy case, where  is normal. Prove that
|j()|
2
= max[j()[ ; Sp().
(b) We turn to the general case. Check that o = 1 and o

=

1. Deduce that
\
k
is unitary.
57
(c) Show that, whenever 2/, the central block of the -th power \
q
k
equals 
q
.
Deduce that if j C[A] has degree at most 2/, then the central block of j(\
k
)
equals j().
(d) Then, still assuming d
o
j 2/, show that |j()|
2
|j(\
k
)|
2
.
(e) Deduce von Neumann inequality:
|j()|
2
max[j()[ ; o
1
.
where o
1
is the unit circle.
83. Let / be eld and M
n
(/) be given. We denote 1 := adj =

T
the transpose
of the cofactors matrix. We recall 1 = 1 = (det )1
n
. Denote also the respective
characteristic polynomials
j
A
(A) = A
n
c
1
A
n1
+ + (1)
n
c
n
. j
B
(A) = A
n
/
1
A
n1
+ + (1)
n
/
n
.
(a) Prove the identity
/
n
j
A
(A) = (1)
n
A
n
j
B
_
c
n
A
_
.
(b) Deduce that, if det ,= 0, there holds /
j
= c
nj
c
j1
n
for , = 1. . . . . :.
(c) Extend these formulas to the general case. Hint: Apply the previous question when
/ is replaced by the ring [c
11
. . . . . c
nn
], where the indeterminates c
ij
are the entries
of a general matrix . See for instance the proof of Theorem 2.1.1.
(d) Conclude that the spectrum of 1 is given, counting with multiplicities, by

1

j1

j+1

n
. , = 1. . . . . :.
where
1
. . . .
n
are the eigenvalues of .
(e) Compare with the additional exercise 56.
84. We consider an :: matrix A whose entries r
ij
are independent indeterminates, meaning
that the set of scalars is the ring := Z[r
11
. r
12
. . . . . r
nn
]. We embed into its eld of
fractions / = Z(r
11
. r
12
. . . . . r
nn
).
(a) Prove that A is non-singular.
(b) Let 1 j : 1 be an integer. Consider the block forms
A =
_
A
p

_
. A
1
=
_

1
np
_
.
where A
p
M
p
() and 1
np
M
np
(/). Prove the identity
det 1
np
=
det A
p
det A
.
58
85. Let / be a eld and \ a nite dimensional /-vector space. A ag in \ is a sequence
1 = (\
1
. . . . . \
n
= \ ) of subspaces with the properties dim\
m
= : and \
m
\
m+1
.
In particular, : = dim\ . A basis A
1
. . . . . A
n
is adapted to the ag if for every :,
A
1
. . . . . A
m
is a basis of \
m
versely, an adapted basis determines uniquely the ag. Two adapted bases dier only by
a triangular change of basis:
1
m
= c
mm
A
m
+ c
m,m1
A
m1
+ + c
m1
A
1
. c
mm
,= 0.
Identifying \ to /
n
, we deduce that the set of ags is in one-to-one correspondence
with the set of right cosets GL
n
(/),T
sup
(/), where T
sup
(/) denotes the subgroup of
upper triangular matrices whose diagonal is non-singular. Therefore, questions about
ags reduce to questions about GL
n
(/),T
sup
(/).
(a) Consider the statement
(B): given two ags 1 and 1

in \ , there exists a basis adapted to 1, of

which a permutation is adapted to 1

.
Prove that (B) is equivalent to
(M): given GL
n
(/), there exists 1. 1

T
sup
(/) and a permutation
matrix 1 such that = 111

.
(b) In the statement (M), prove that the permutation is necessarily unique.
(c) We turn to the existence part. Thus we give ourselves in GL
n
(/) and we look for
1. 1

and a permutation , such that

c
ij
=
n

r=1
t
i(r)
t

rj
.
Show the necessary condition
(1) = maxi ; c
i1
,= 0.
(d) Show that there exists an index i ,= (1) such that (we use the notation of Section
2.1 for minors)

_
(1) i
1 2
_
,= 0.
Hint: This amounts to nding syzygies between minors taken from two given
columns, here the rst and the second.
Then prove that (2) must be the maximum of such indices i.
(e) By induction, prove the necessary condition that (,) is the largest index i with the
properties that i ,= (1). . . . . (, 1) and

_
(1) (, 1) i
1 , 1 ,
_
,= 0.
59
(f) Deduce the theorem that for every two ags 1 and 1

, there exists a basis adapted

to 1, of which a permutation is adapted to 1

.
(g) In the particular case / = R or C, prove that, for every in a dense subset of GL
n
(/),
(M) holds true, with 1 the matrix associated with the permutation (,) = :+1,.
86. Show that a complex matrix M
n
(C) is Hermitian if and only if r. r is real for
every r C
n
, where . is the standard scalar product.
87. Let / = R or C. The matrix norms that we consider here are subordinated to the /
2
-norms
of /
d
.
Given three matrices M
pq
(/), 1 M
ps
(/) and C M
rq
(/), we consider the
ane set J of matrices \ M
nm
(/) of the form
\ =
_
1
C 1
_
.
where 1 runs over M
rs
(/). Thus : = j + : and : = + :.
Denoting
1 =
_
1
0
_
. Q = ( 1 0 )
the projection matrices, we are going to prove (Parrotts Lemma) that
(10) min|\| ; \ J = max|Q\|. |\1|.
where the right hand side does not depend on 1:
\1 =
_

C
_
. Q\ = ( 1 )
(a) Check the inequality
inf|\| ; \ J max|Q\|. |\1|.
(b) Denote j(1) := |\|. Show that the inmum of j on J is attained.
(c) Show that it is sucient to prove (10) when : = 1.
(d) From now on, we assume that : = 1, and we consider a matrix 1
0
M
r1
(/) such
that j is minimal at 1
0
. We denote by \
0
the associated matrix. Let us introduce
a function 1 (1) = j(1)
2
. Recall that is the largest eigenvalue of \

\. We
denote )
0
its multiplicity when 1 = 1
0
.
i. If )
0
2, show that \

0
\
0
has an eigenvector with
m
= 0. Deduce that
j(1
0
) |\1|. Conclude in this case.
60
ii. From now on, we suppose )
0
= 1. Show that (1) is a simple eigenvalue for
every 1 in a small neighbourhood of 1
0
. Show that 1 (1) is dierentiable
at 1
0
, and that its dierential is given by

2
||
2
1[(Q\
0
)

Q] .
where is an associated eigenvector:
\

0
\
0
= (1
0
).
iii. Deduce that either Q = 0 or Q\
0
= 0.
iv. In the case where Q = 0, show that j(1
0
) |\1| and conclude.
v. In the case where Q\
0
= 0, prove that j(1
0
) |Q\| and conclude.
Nota: Parrott is the name of a mathematician. Therefore, Parrotts Lemma has nothing
to do with the best seller Le Theoreme du Perroquet, written by the mathematician Denis
Guedj.
88. For a Hermitian matrix , denote by 1
k
1
k
:= det

c
11
c
1k
.
.
.
.
.
.
.
.
.
c
k1
c
kk

.
When / = 0, we also set 1
0
= 1. Finally, we set
c
k
:= sign
1
k
1
k1
1. 0. 1. / = 1. . . . . :.
(a) We assume that 1
k
,= 0 for all /. Prove that the number of positive eigenvalues of
is precisely the number of 1s in the sequence c
1
. . . . . c
n
.
Hint: Argue by induction, with the help of the interlacing property (Theorem 3.3.3).
(b) We assume only that 1
n
= det ,= 0. Prove that the number of negative eigenvalues
of is precisely the number of sign changes in the sequence c
1
. . . . . c
n
(the zeros are
not taken in account).
89. We dene a Hilbert space H
2
of holomorphic functions on the unit disc D, endowed with
the scalar product
). p :=
1
2
_
2
0
)(c
i
)p(c
i
) d.
2
as the completion of the space
of polynomials under the norm
|)| =
_
). ).
61
or as the set of 1
2
-functions on the unit circle T that have a holomorphic extension to D.
In other words, 1
2
-functions ) : T C such that
_
2
0
)(c
i
)c
im
d = 0. : = 0. 1. 2. . . .
We dene the Szego kernel
/(. j) =
1
1 j
.
When D, we dene a holomorphic function
/

(.) := /(.. ).
(a) If n H
2
, prove
/

. n = n().
(b) For holomorphic and bounded on D, the operator

: n n is bounded on H
2
.
Prove


= () /

.
(c) Prove also that
|

| = |

| = sup[(.)[ ; . T.
(d) We assume moreover that : D D. Given  distinct numbers
1
. . . . .
N
in D,
we form the Hermitian matrix with entries
c
ij
:= (1 u
i
u
j
)/(
i
.
j
).
where u
j
= (
j
). Prove that is semi-denite positive.
Hint: Write that 

1 on the space spanned by the /

j
s.
Nota: Picks Theorem tells us that, given the
j
s and the u
j
s, 0
N
is equivalent
to the existence of a holomorphic function : D D such that (
j
) = u
j
. This is
an interpolation problem. The space H
2
is a Hardy space.
90. Let H
n
be a semi-denite positive Hermitian matrix, with c
jj
0 and c
jk
,= 0 for
every (,. /). Let us form the Hermitian matrix 1 such that /
jk
:= 1,c
jk
. Assume at last
that 1 is semi-denite positive too.
(a) Prove that [c
jk
[ =

c
jj
c
kk
, using principal minors of rank 2.
(b) Using principal minors of rank 3, show that c
jk
c
kl
c
jl
is real positive.
(c) Deduce that is rank-one: There exists a C
n
such that =

.
(d) What does it tell in the context of the previous exercise ?
91. Let M
n
(C) be a normal matrix. We decompose = 1 + 1 + l in strictly lower,
diagonal and strictly upper triangular parts. Let us denote by /
j
the Euclidean length of
the ,-th column of 1, and by n
j
that of the ,-th row of l.
62
(a) Show that
k

j=1
n
2
j

k

j=1
|
2
j
+
k

j=1
j1

m=1
n
2
mj
. / = 1. . . . . : 1.
(b) Deduce the inequality
|l|
S

: 1|1|
S
.
where | |
S
is the SchurFrobenius norm.
(c) Prove also that
|l|
S

1

: 1
|1|
S
.
(d) Verify that each of these inequalities are optimal. Hint: Consider a circulant matrix.
92. The ground eld is R.
(a) Let 1 and Q be two monic polynomials of respective degrees : and : 1 (: 2).
We assume that 1 has : real and distinct roots, strictly separated by the : 1 real
and distinct roots of Q. Show that there exists two real numbers d and c, and a
monic polynomial 1 of degree : 2, such that
1(A) = (A d)Q(A) c
2
1(A).
(b) Let 1 be a monic polynomial of degree : (: 2). We assume that 1 has : real and
distinct roots. Build sequences (d
j
. 1
j
)
1jn
and (c
j
)
1jn1
, where d
j
. c
j
are real
numbers and 1
j
is a monic polynomial of degree ,, with
1
n
= 1. 1
j
(A) = (A d
j
)1
j1
(A) c
2
j1
1
j2
(A). (2 , :).
Deduce that there exists a tridiagonal matrix , which we can obtain by algebraic
calculations (involving square roots), whose characteristic polynomial is 1.
(c) Let 1 be a monic polynomial. We assume that 1 has : real roots. Prove that
one can factorize 1 = Q
1
Q
r
, where each Q
j
has simple roots, and the factoriza-
tion requires only nitely many operations. Deduce that there is a nite algorithm,
involving no more than square roots calculations, which provides a tridiagonal sym-
metric matrix , whose characteristic polynomial is 1 (a tridiagonal symmetric
companion matrix).
93. (D. Knuth.) Let be an alternate matrix in M
n+1
(/), / a eld. We index its rows and
columns from 0 to : (instead of 1. . . . . :+1), and form the matrices
j
i
by removing from
the i-th row and the ,-th column. We denote also by PF[|
1
. . . . . |
2m
] the Pfaan of the
alternate matrix obtained by retaining only the rows and columns of indices |
1
. . . . . |
2m
.
(a) Prove the following formulas, either (if : is even):
det
0
1
= PF[0. 2. . . . . :] PF[1. 2. . . . . :].
63
or (if : is odd):
det
0
1
= PF[0. 1. 2. . . . . :] PF[2. . . . . :].
Hint: If : is odd, expand the identity det( + A1) = (Pf( + A1))
2
where
/
12
= /
21
= 1 and /
ij
= 0 otherwise. Then use Exercise 11.b) in the present list. If
: is even, expand the identity det (A. 1 ) = (Pf(A. 1 ))
2
, where
(A. 1 ) :=
_
_
_
_
_
_
_
_
_
0 A 1 0 0
A
1
0
.
.
.
0
_
_
_
_
_
_
_
_
_
.
(b) With the rst formula, show that if is an : : alternate matrix, then the
transpose matrix of cofactors adj is symmetric for : odd! Prove that in fact adj
has the form 22
T
where 2 /
m
. Compare with the additional exercise 56.
94. Let / be a eld and : an even integer. If r. /
n
, denote by r the alternate matrix
r
T
r
T
. Show the formula
Pf( + r ) = (1 +
T

1
r) Pf
for every non-singular alternate : : matrix .
Hint: We recall the formul
det(+r
T
) = (1+
T

1
r) det . (+r
T
)
1
= 
1

1
1 +
T

1
r

1
r
T

1
.
95. Let / be a eld, : be an even integer and be an : : non-singular alternate matrix.
Using the odd case of Exercise 93 above, prove the formula

1
=
1
Pf
_
(i. ,)(1)
i+j+1
Pf
ij
_
1,i,jn
.
where
ij
is obtained from be removing the i-th and ,-th rows and columns, and (i. ,)
is according to the sign of , i. Compare this formula with Exercise 11.b) above.
In particular, show that Pf divides, as a polynomial, every entry of adj.
96. (Banach.) Let j [1. +] be such that j ,= 2. We consider matrices  M
n
(R) which
are isometries, namely
|r|
p
= |r|
p
. r R
n
.
(a) Let us begin with the case 2 < j < +. We give r. R
n
such that r
i

i
= 0 for
every i :. Dene n = r and = , and let H be the set of indices , such
that n
j

j
,= 0.
64
i. Show that the function
(:) :=

jH
([:n
j
+
j
[
p
[:n
j
[
p
[
j
[
p
)
vanishes identically.
ii. Computing the second derivative of , show that H is void.
(b) (Continued, 2 < j < +.) Let :
k
be the number of non-zero entries in the /-th
column of , and 1
k
be the vector space spanned by the other columns. Using
the previous question, show that dim1
k
: :
k
. Then deduce that :
k
= 1. At
last, show that  is the product of a diagonal matrix diag(1. . . . . 1) and of a
permutation matrix.
(c) If 1 < j < 2, prove the same conclusion. Hint: Apply the previous result to 
T
.
(d) If j = 1, prove directly that if r
i

i
= 0 for every index i, then (r)
j
()
j
= 0 for
every index ,. Hint: Use (1) = (1) = 0. Conclude.
(e) If j = +, conclude by applying the case j = 1 to
T
.
97. (Lemmens & van Gaans.) We endow R
n
with some norm | |. Let  M
n
(R) be
non-expansive: |r| |r| for every r R
n
.
(a) Let 1 be the unit ball. Show that
1 :=

k1

k
1
is a compact symmetric convex set. We denote by 1 the vector space spanned by
1.
(b) Show that 1 = 1 and that the restriction of  to 1 is an isometry.
(c) Let (/
j
)
jN
be increasing sequence such that
k
j
. Prove that 1 = 1.
(d) Show that there exists an increasing sequence (/
j
)
jN
such that 
k
j
converges.
Prove that
k
j+1
k
j
converges towards a projector 1 whose range is 1, and which
is non-expansive.
98. Given an alternate 4 4 matrix , verify that its characteristic polynomial equals
A
4
+ A
2

i<j
c
2
ij
+ Pf()
2
.
We dene
1
+
() = (c
12
+ c
34
)
2
+ (c
23
+ c
14
)
2
+ (c
31
+ c
24
)
2
.
1

() = (c
12
c
34
)
2
+ (c
23
c
14
)
2
+ (c
31
c
24
)
2
.
65
Factorize 1
A
in two dierent ways and deduce the following formula for the eigenvalues
of , in characteristic dierent from 2:
i
2
_

_
1
+
()
_
1

()
_
.
where the signs are independent of each other.
99. (a) Verify that the characteristic polynomial 1
V
of a real orthogonal matrix \ can be
factorized as
1
V
(A) = (A 1)
r
(A + 1)
s
A
m
Q
_
A +
1
A
_
where Q R[A] a monic polynomial whose roots lie in (2. 2).
(b) Conversely, we give a monic polynomial Q R[1 ] of degree :, whose roots lie in
(2. 2), and we consider
1(A) = A
m
Q
_
A +
1
A
_
.
Let be a tridiagonal symmetric matrix whose characteristic polynomial is Q (see
Exercise 92.)
i. Prove that 1
m

1
4

2
is positive denite.
ii. Let us dene
1 :=
_
1
n

1
4

2
.
Prove that
\ :=
_
1
1
_
is orthogonal.
iii. Prove 1
V
= 1 (\ is a companion matrix of 1.)
100. (Inspired by M. T. Karaev.) Let 1 be a nite dimensional Hilbert space. We are interested
in nilpotent endomorphisms. Recall that n /(1) is nilpotent of order : if n
m
= 0
E
but n
k
,= 0
E
if / < :.
(a) Let 1 be the orthogonal of ker n and let G be n(1). Prove that there exists an
orthonormal basis of 1, whose image is an orthogonal basis of G. Hint: This is
essentially the Singular Value Decomposition.
(b) Deduce that, if : = 2, there exists an orthonormal basis of 1, in which the matrix
of n has the Jordan form
(11)
_
_
_
_
_
_
_
_
0 c
2
0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0
.
.
.
.
.
. c
n
0 0
_
_
_
_
_
_
_
_
.
66
(c) Arguing by induction, prove the same result for every : 2.
(d) Deduce that if  M
n
(C) is nilpotent, then  is unitarily similar to a matrix of
the form (11).
(e) If  and  are unitarily similar, check that their numerical ranges (see Exercise
21) are equal, and that ||
2
= ||
2
.
(f) Let  M
n
(C) be nilpotent of order :. Prove that 1() is a disk centered at
the origin. Show that its radius is less than or equal to (Haagerupde la Harpe
inequality)
||
2
cos

: + 1
.
Hint: It is enough to work in the case  is of the form (11) and the c
j
s are non-
zero. Then it is easy to show that 1() is rotationaly invariant. Since it is convex
(Exercise 21), it is a disk. The triangular inequality leads to the computation of the
_
_
_
_
_
_
_
_
0 1,2 0 0
1,2
.
.
.
.
.
.
.
.
.
.
.
.
0
.
.
.
.
.
. 0
.
.
.
.
.
.
.
.
. 1,2
0 1,2 0
_
_
_
_
_
_
_
_
.
101. (From de Oliveira.) Let (
1
. . . . .
n
) C
n
and (
1
. . . . .
n
) C
n
be given. Let us form
the diagonal matrices := diag
1
. . . . .
n
and 1 := diag
1
. . . . .
n
.
(a) If \ is unitary, show that
Tr(\ 1\

) =
_

1

n
_
o
_
_
_

1
.
.
.

n
_
_
_
.
for some orthostochastic matrix o.
(b) Using Birkhos Theorem, deduce that Tr(\ 1\

) belongs to the convex hull of

the set of numbers
n

j=1

(j)
. S
n
.
(c) More generally, given two normal matrices . 1 M
n
(C) whose respective spectra
are (
1
. . . . .
n
) and (
1
. . . . .
n
), prove that Tr(1) belongs to this convex hull.
Nota: See Exercise 139 for a related result.
(d) With the same notations as above, prove that
min
Sn
n

j=1
[
j

(j)
[
2
| 1|
2
F
max
Sn
n

j=1
[
j

(j)
[
2
.
67
where | |
F
is the Frobenius norm. The rst inequality constitutes Homan
Wielandt theorem. The case for Hermitian matrices was found earlier by Loewner
(L owner).
102. (Y. Shizuta, S. Kawashima.) Let . 1 be : : Hermitian matrices, with 1 0
n
. We
denote by c
1
. . . . . c
r
the distinct eigenvalues of and by 1
1
. . . . . 1
r
the corresponding
eigenprojectors, so that =

j
c
j
1
j
.
(a) Prove that 1
j
is Hermitian and that 1
j
1
k
= 0
n
when / ,= ,.
(b) Let us assume that there exists a skew-Hermitian matrix 1 such that 1 + [1. ] is
positive denite. Show that ker 1 does not contain any eigenvector of .
(c) Let us assume that the intersection of the kernels of 1, [1. ], [[1. ]. ],... (take
successive commutators with ) equals 0. Prove that ker 1 does not contain any
eigenvector of .
(d) Conversely, we assume that ker 1 does not contain any eigenvector of .
i. Dene
1 :=

i=j
1
c
i
c
j
1
i
11
j
.
Show that 1 is skew-Hermitian, and that 1 + [1. ] is positive denite.
ii. Show that the intersection of the kernels of 1, [1. ], [[1. ]. ],... equals 0.
103. Let M
nm
(/) be given, with / = R or / = C. We put the singular values
1

2

in decreasing order. We endow M
nm
(/) with the norm | |
2
. Prove that the distance
of to the set 1
l
of matrices of rank less than or equal to |, equals
l+1
(if | = min:. :,
put
l+1
= 0). Hint: Use Theorem 7.7.1 of Singular Value Decomposition.
104. (a) Let 1
0
M
n
(C) be such that |1
0
|
2
1. Prove
(12) |(1
0
u1
n
)
1
|
2

1
[u[ 1
. u C; [u[ 1
Hint: Use von Neumann Inequality (Exercise 82), while approximating the function
)(.) := (. u)
1
by polynomials, uniformly on a neighbourhood of Sp(1
0
).
(b) Let 1 M
n
(C) be of the form 1 = c1
n
+  where c C and 
2
= 0
n
.
i. Prove that
(|1|
2
1)
_
||
2
+[c[
2
1
_
.
Deduce that |1|
2
is the largest root of the quadratic equation
:
2
||
2
: [c[
2
= 0.
ii. Prove that
_
|(1 u1
n
)
1
|
2

1
[u[ 1
. u C; [u[ 1
_

_
1
2
||
2
+[c[ 1
_
.
68
iii. Deduce that the converse of (12) does not hold if : 2. In particular, one cannot
replace the assumption |1|
2
1 by (12) in the inequality of von Neumann.
iv. However, prove that for such a 1 = c1
n
+ , (12) implies |1|
2
2, and the
equality is achieved for some (c. ).
(c) Let 1 M
n
(C) satisfy Property (12). Prove that
|1|
2
c.
Hint: Use the formula
1 =
1
2i/
_
r
.
k
(.1
n
1)
k
d..
(Interestingly enough, this part of the exercise is true not only for the norm | |
2
,
but also in every Banach algebra.)
105. Let . 1 be two positive semi-denite Hermitian matrices.
(a) Prove that for every r C
n
, |1r|
2
2
|1|
2
1r. r.
(b) Deduce that 11 ||
2
|1|
2
1 in the sense of Hermitian matrices.
106. (From L. Tartar.)
(a) Let )(.) =

n0
c
n
.
n
be a series, converging for [.[ < 1. We denote 1(.) :=

n0
[c
n
[.
n
. Check that whenever / is a Banach algebra and c. / /, there holds
|)(/) )(c)|
1(|/|) 1(|c|)
|/| |c|
|/ c|.
(b) Deduce that if . 1 H
n
(C), then
|c
iB
c
iA
|
2
|1 |
2
.
Hint: Choose an integer : 1. Apply the previous result to )(.) = c
iz/m
, and
decompose c
iB
c
iA
as a sum of : products having c
iB/m
c
iA/m
as one factor and
unitary matrices otherwise. Then let : +.
107. Let / be R or C. We consider a norm on /
n
such that the induced norm has the property
|
2
| = ||
2
for every Hermitian, positive semi-denite matrix .
(a) Show that || = () for every H
+
n
.
(b) Deduce that, for every r /
n
, there holds |r|
2
2
= |r| |r|

(recall that | |

is the
dual norm of | |).
(c) Let be the unit sphere of norm | |. Given r , show that is on one side of
the plane dened by 1(

r) = |r|
2
.
69
(d) Deduce that | | is proportional to | |
2
. Hint: Given a point r
0
, show
that the smallest convex cone with vertex r
0
, which contains , is the half-plane
1(

r) |r|
2
(this uses the previous question). Deduce that is a dierentiable
manifold of codimension one. Then conclude with the previous question.
(e) Let | | be and induced norm on M
n
(/) such that the square root

is
1,2-Holderian on HPD
n
(or SPD
n
if / = R), with constant one (compare with
Exercise 52):
(13) |

| |1 |
1/2
.
Prove that the norm on /
n
is proportional to | |
2
. Comment: For the validity of
(13) for | |
2
, see Exercise 110.
108. Let :
j
() (1 , :) denote the singular values of a matrix  M
n
(R), labelled in
increasing order: :
1
() :
n
().
(a) We dene
j
:= :
j
+ + :
n
. Prove the formula

nj+1
() = supTr(1Q) ; 1 M
jn
(R). Q M
nj
(R)
s.t. 11
T
= 1
j
. Q
T
Q = 1
j
.
Deduce that
j
is a convex function, and thus is a norm. Do you recognize the norm

n
= :
n
?
(b) Let c := (c
1
. . c
n
) be a given :-uplet of non-negative reals numbers with c
1

c
n
. We dene
1(c) :=  M
n
(R) ; :
j
() = c
j
. for all 1 , :.
Verify that the set 1

(c) dened below is convex, and that it contains the convex

hull of 1:
1

(c) :=  M
n
(R) ;
j
() c
j
+ + c
n
. for all 1 , :.
(c) Show that the extremal points of 1

(c) belong to 1(c), and deduce that 1

(c) is
the convex hull of 1(c). Hint: The set ext(1

(c)) is left- and right-invariant under

multiplication by orthogonal matrices. Thus one may consider diagonal extremal
points.
(d) Deduce that the convex hull of O
n
(R) is the unit ball of | |
2
. Remark: Here,
the convex hull of a set of small dimension (:(: 1),2) has a large dimension :
2
.
Thus it must have faces of rather large dimension ; this is precisely the contents of
Corollary 5.5.1, when applied to the induced norm | |
2
. See Exercise 137 below for
a more accurate description.
70
109. (Continuation.) We say that a subset A of M
n
(R) is rank-one convex if whenever . 1
A and 1 is of rank one, then the segment (. 1) is included in A. Rank-one
convexity is preserved under intersection. If 1 is a subset of M
n
(R), its rank-one-convex
hull is the smallest rank-one convex subset that contains 1 . We recall that a function
) : A R+ is rank-one convex if it is convex on every segment of the type above.
(a) Let ) : M
n
(R) R + be rank-one convex. Show that the sets dened by
)() are rank-one convex.
(b) Verify that  [ det [ is rank-one convex.
(c) Prove the formula for products of singular values:
:
n
() :
nj+1
() = supdet(1Q) ; 1 M
jn
(R). Q M
nj
(R)
s.t. 11
T
= 1
j
. Q
T
Q = 1
j
.
(d) Deduce that
j
:= :
n
:
j
is rank-one convex.
(e) Deduce that the rank-one-convex hull of the set 1(c) (see the previous exercise) is
included in
1

(c) :=  M
n
(R) ;
j
() c
j
c
n
. for all 1 , :.
Comment: A result of B. Dacorogna tells us that 1

(c) is actually equal to the

rank-one-convex hull of 1(c).
110. Let . 1 be : : positive semi-denite Hermitian matrices.
(a) Using the formula (Taylor expansion)
1
2

2
= (1 )
2
+ (1 ) + (1 ).
prove that (1 )
2
|1
2

2
|
2
. Hint: Use an eigenvector of 1 . The
formula above might not be the useful one in some case.
(b) Deduce that (13) holds true for the operator norm | |
2
. Comment: The resulting
inequality |

|
2
|1 |
1/2
2
is much more powerful than that of Exercise
52, since it does not depend on the dimension :. In particular, it holds true for
bounded self-adjoint operators in Hilbert spaces.
111. (Continuation.) Likewise, write the Taylor expansion
1
3
=
3
+ + H
3
. H := 1 .
Then, using an eigenvector c of H, associated with (H), show that
c

(1
3

3
)c (H)
3
|c|
2
2
= 2(H)|c|
2
2
+ c

Hc + 3(H)
2
c

c.
Prove the bound [c

Hc[ (H)|c|
2
2
and deduce that
|1 |
3
2
|1
3

3
|
2
.
1/3
over HPD
n
1/4
(simpler than
what you think in a rst instance) ?
71
112.
Isaac Newton.
Given a real polynomial
1 = A
n
c
1
A
n1
+ + (1)
n
c
n
R[A].
whose roots are r
1
. . . . . r
n
(repeted with multiplic-
ities), we dene the Newton sums
:
k
:=

r
k

(:
0
= :).
We recall that :
k
is a polynomial in c
1
. . . . . c
k
, with
integer coecients.
Let us form the Hankel matrix
H :=
_
_
_
_
_
:
0
:
1
:
n1
:
1
:
2
:
n
.
.
.
.
.
.
:
n1
:
n
:
2n2
_
_
_
_
_
.
Let Q be the quadratic form associated with H. Write Q as a sum of squares. Deduce
that the rank of H equals the number of distinct complex roots of 1, while the index
of H (the number of positive squares minus the number of negative squares) equals the
number of distinct real roots of 1. Conclude that the roots of 1 are all real if, and only
if, H is positive semi-denite.
113. (Continuation. Newell (1972/73), Ilyushechkin (1985), Lax (1998), Domokos (2011) ;
special thanks to L. Tartar.) In the previous exercise, set 1 := det(A1
n
) where is
a general matrix in Sym
n
(R). Check that the :
k
s are polynomials in the entries of .
Show that H is positive semi-denite. Deduce that every principal minor of H,
H
_
i
1
. . . i
r
i
1
. . . i
r
_
.
is a polynomial in the entries of , and that it takes only non-negative values.
According to Artins Theorem (see J. Bochnak M. Coste & M.-F. Roy: Real algebraic
geometry, Springer-Verlag (1998), Theorem 6.1.1), this property ensures that these minors
are sums of squares of rational functions. Hilbert pointed out that not every non-negative
polynomial is a sum of squares of polynomials. However, it turns out that these principal
minors are sums of squares of polynomials : Let us endow Sym
n
(R) with the scalar
product . 1 := Tr(1).
(a) Check that H is a Gram matrix: /
ij
= \
i
. \
j
for some \
i
Sym
n
(R).
(b) Show that the exterior algebra of Sym
n
(R) is naturally endowed with a scalar prod-
uct.
72
(c) Show that
H
_
i
1
. . . i
r
i
1
. . . i
r
_
= |\
i
1
\
ir
|
2
and conclude. Remark: The scalar product of a Euclidean space 1 extends in a
natural way to the exterior algebra
r
1.
Nota: This formulation, due to Ilyushechkin, gives the discriminant over Sym
n
(R) as
the sum of :! squares of polynomials. This upper bound has been improved by Domokos
into
_
2: 1
: 1
_

_
2: 3
: 1
_
.
The minimal number of squares is not yet know, except for : = 2 and 3, where 2 and 5
squares suce.
114. Let be a principal ideal domain. If  M
n
() and  = 11Q with 1. Q GL
n
()
and 1 diagonal, prove the following equality about cofactors matrices:
co() = (det 1)(det Q)1
T
co(1)Q
T
.
Prove
1

(co()) = (det )
1
1
n
()
and deduce the value of d

(co()), the /-th invariant factor of co(). Compare with the

result of Exercise 56.
115. (Potter.) Let / be a eld and an element of /.
(a) Prove that there exists polynomials 1
r,j
Z[A] such that, for every integer : 1,
every element in / and every pair of matrices . 1 
n
(/) such that
(14) 1 = 1.
there holds
(15) ( + 1)
r
=
r

j=0
1
r,j
()1
j

rj
.
Matrices satisfying (14) are said to -commute. Remark that they satisfy
p
1
q
=

pq
1
q

p
, a formula that is a discrete analogue of the Stonevon Neumann formula.
(b) Dene polynomials

l
(A) =
l

s=1
(1 + A + + A
s1
).
Show the formula

rj
1
r,j
=
r
.
Hint: Proceed by induction over :.
73
(c) Assume that is a primitive root of unity, of order : (then : is not the characteristic
of /). Deduce that (14) implies (Potters Theorem)
( + 1)
r
=
r
+ 1
r
.
Remark. It is amazing that when : is prime, the identity (15) can occur in two
cases : either and 1 -commute with respect to a primitive root of unity of order
:, or : is the characteristic of / and [. 1] = 0
n
; and both cases exclude each other !
(d) Let 1 be a cyclic matrix in the sense of Section 5.4:
1 :=
_
_
_
_
_
_
_
_
0
1
0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0
0
.
.
. 
r1

r
0 0
_
_
_
_
_
_
_
_
.
We recall that the diagonal blocks are square (null) matrices, of respective sizes
:
1
. . . . . :
r
. Let us dene := diag(1
n
1
. 1
n
2
. . . . .
r1
1
nr
). We assume again that
is a primitive root of unity, of order :. Prove that (1. ) -commute and deduce
that
( + 1)
r
= 1
n
+ diag(
1
. . . . .
r
).
where 
j
:=
j

j+1

j1
. For instance,

1
=
1

2

r
. 
r
=
r

1

2

r1
.
116. Let H HPD
n
and / H
n
be given. We recall that  := /H is diagonalizable with
real eigenvalues (see Exercise 258). We recall that a matrix is normal if and only if its
eigenvectors form a unitary basis. Thus the angles between eigenvectors of  measure
the deviation of  from normality. We compute here the minimum angle

as / runs
over H
n
.
The space C
n
is endowed with its Hermitian norm | |
2
.
(a) Let and u be two eigenvectors of  associated with distinct eigenvalues. Prove
that

H
u (i.e. u

H = 0).
(b) Let C
n
be given. We dene a linear form 1(u) :=

u with domain the H-

orthogonal to . Check that
|1| inf
rR
| + :H|
2
.
Deduce that
:= sup
_
[

u[
||
2
|u|
2
; u
H

_
inf
rR
(1
n
+ :H) =
1(H) 1
1(H) + 1
.
where 1(H) is the condition number of H.
74
(c) Looking for a specic pair (. u), prove that there holds actually
=
1(H) 1
1(H) + 1
.
(d) Deduce that
sin

=
2
1(H) + 1
.
117. Let us dene 1
n
:=
n
Sym
n
(R), the set of symmetric, bistochastic matrices.
(a) Show that 1
n
is the convex hull of the set of matrices of the form
Q

:=
1
2
(1

+ 1

1). ( S
n
)
where 1

denotes the permutation matrix associated with .

(b) If = (1. 2. . . . . :) is a cycle, prove that Q

is extremal in 1
n
if and only if either
: is odd or : = 2. Hint: If : is even, show that Q

=
1
2
(Q
+
+ Q

) where Q

are
permutation matrices associated with involutions, and Q
+
,= Q

if : 4. If : is
odd, consider the graph of pairs (i. ,) for which
ij
,= 0 in Q

; an edge between
(i. ,) and (i

. ,

or , = ,

. The graph is a cycle of length

2:, and d((i. ,). (,. i)) = : is odd. If Q

=
1
2
(1 + o) with 1. o 1
n
, show that
:
ji
+ :
ij
= 1 along and conclude.
(c) Deduce that ext(1
n
), the set of extremal points of 1
n
, consists in the matrices Q

for the permutations that are products of disjoint cycles of lengths as above (either
odd or equal to two).
118. Let 1 : M
n
(C) C be a linear form with the properties that if H is Hermitian, then
1(H) is real, and if moreover H 0
n
, then 1(H) 0. Prove that 1(

) = 1() for
every M
n
(C). Then prove, for every pair of matrices,
[1(

1)[
2
1(

)1(1

1).
(a) Prove the determinantal identity (Cauchys double alternant)
_
_
_
_
1
c
i
+ /
j
_
_
_
_
1i,jn
=

i<j
(c
j
c
i
)

k<l
(/
k
/
l
)

i,k
(c
i
+ /
k
)
.
75
Hint: One may assume that
c
1
. . . . . c
n
. /
1
. . . . . /
n
are indetermi-
nate, and then work in the eld
Q(c
1
. . . . . c
n
. /
1
. . . . . /
n
). This determi-
nant is a homogeneous rational function
whose denominator is quite trivial. Some
specialisations make it vanishing ; this
gives an accurate information about the
numerator. There remains to nd a scalar
factor. That can be done by induction on :,
with an expansion with respect to the last
row and column.
Augustin Cauchy.
(b)
David Hilbert.
Dene the Hilbert matrix of order : as the Gram
matrix
/
ij
:= r
i1
. r
j1

L
2
(0,1)
. (1 i. , :)
where
). p
L
2
(0,1)
:=
_
1
0
)(r)p(r) dr.
Use the formula above to compute (H
1
)
nn
.
(c) In particular, nd that |H
1
|
2

1
32n
16
n
(asymptotically, there holds a better bound
1
8
16
n
.) Hint: The binomial (2:)!,(:!)
2
is larger than 2
2m
,(2: + 1).
(d) Deduce that the Hilbert matrix is pretty much ill-conditionned:
(H) := |H|
2
|H
1
|
2

1
32:
16
n
.
Remark: Hilberts matrix satises |H|
2

, and this estimate is the best one to
be uniform, in the sense that, denoting H
n
the Hilbert matrix of size : :, one has
sup
n
|H|
2
=

. See for instance G. P olya & G. Szeg o, Problems and theorems in
analysis, vol. I. Fourth edition (1970), Springer-Verlag. Part III, chapter 4, exercise
169.
76
120. (I. Kovacs, D. Silver & S. Williams) Let M
pq
(/) be in block form
=
_
_
_

11

12

21
.
.
.
.
.
.
_
_
_
where the blocks are j j. Let us assume that these blocks commute pairwise. Prove
that
det = det . :=

Sq
c()
q

m=1

m(m)
.
Notice that is nothing but the determinant of , considered as a matrix with
entries in the abelian ring 1 generated by the
ij
s. One may therefore write
det
pq
= det
p
det
R
.
where the subscripts indicate the meaning of each determinant. Hint: Use Schurs for-
mula for the determinant of a matrix with four blocks. Argue by induction over .
121. Let us dene the tridiagonal and block-tridiagonal matrices
J
p
:=
_
_
_
_
_
_
_
_
0 1
1
.
.
.
.
.
. 0
.
.
.
.
.
.
.
.
.
0
.
.
.
.
.
. 1
1 0
_
_
_
_
_
_
_
_
M
p
(/).
pq
:=
_
_
_
_
_
_
_
_
J
p
1
p
1
p
.
.
.
.
.
. 0
.
.
.
.
.
.
.
.
.
0
.
.
.
.
.
. 1
p
1
p
J
p
_
_
_
_
_
_
_
_
M
pq
(/).
Denote by 1
p
the polynomial
1
p
(A) := det(A1
p
+ J
p
).
Using the previous exercise, prove that the characteristic polynomial 1
pq
(1 ) of
pq
is the
resultant
Res(1
q
( 1 ).

1
p
).

1(A) := 1(A).
Nota: Since these matrices have integer coecients, their characteristic polynomials do
not depend of the scalar eld /. It is therefore enough to consider the real case.
122. Let M
n
(C) have no purely imaginary eigenvalue (one says that is hyperbolic). The
aim of this exercise is to prove the existence and uniqueness of a Green matrix. This is a
matrix-valued function G : R M
n
(C) that is bounded, dierentiable for t ,= 0, which
has left and right limits G(0), and satises
dG
dt
(t) = G(t). (t ,= 0). G(0+) G(0) = 1
n
.
77
(a) We begin with the case where the eigenvalues of have negative real part. We recall
that there exists a positive and a nite C such that | exp(t)| Cc
t
for t 0.
Prove that
G(t) :=
_
0
n
. t < 0.
exp(t). t 0
denes a Green matrix.
Prove that there are constants as above such that | exp(t)| C

c
t
for t < 0
(Hint: Use the inequality 1 || |
1
|). Deduce that the Green matrix is
unique.
(b) Treat the case where = diag(1. C), where the eigenvalues of 1 (resp. of C) have
negative (resp. positive) real parts.
(c) Treat the general case.
(d) Show that actually the Green matrix decays exponentially fast at innity.
(e) Let ) : R C
n
be bounded continuous and dene
(t) :=
_
R
G(t :))(:).
Show that is the unique bounded solution of

(t) = (t) + )(t).

123. Let M
n
(C) be given. The spectrum of is split into three parts

.
+
.
0
according
to the sign of the real part of the eigenvalues. For instance,
0
is the intersection of the
spectrum with the imaginary axis.
(a) Show that the following denitions of a subspace are equivalent:
The sum of the generalized eigenspaces associated with the eigenvalues of neg-
ative real part,
The largest invariant subspace on which the spectrum of the restriction of lies
in the open left half-space of C.
The set of data c C
n
such that the solution of the Cauchy problem
r

(t) = r(t). r(0) = c

tends to zero as t +.
This subspace is called the stable invariant subspace (more simply the stable sub-
space) of and denoted by o(). Prove that if c o(), the solution of the Cauchy
problem above actually decays exponentially fast.
(b) Let us dene the unstable subspace by l() := o(). Give characterizations of
l() similar to above.
(c) Show that the following denitions of a subspace are equivalent:
78
The sum of the generalized eigenspaces associated with the purely imaginary
eigenvalues,
A subspace C(), invariant under and such that
C
n
= o() C() l().
The set of data c C
n
such that the solution of the Cauchy problem above
is polynomially bounded for t R: there exists : : 1 and c
0
such that
|r(t)| c
0
(1 +[t[
m
).
The subspace C() is called the central subspace of .
(d) Prove that the spectra of the restrictions of to its stable, unstable and central
subspaces are respectively

.
+
and
0
.
(e) If M
n
(R), prove that the stable, unstable and central subspaces are real, in the
sense that they are the complexications of subspaces of R
n
.
(f) Express o(

). . . . in terms of o(). . . ..
Nota: The case where
0
is void corresponds to a hyperbolic matrix, in the sense
of the previous exercise.
124. In control theory, one meets a matrix H M
2n
(C) given by the formula (notations are
equivalent but not identical to the standard ones)
H :=
_
11

_
.
Hereabove, , 1, C have respective sizes : :, : : and : :. Without loss of
generality, one may assume that 1 is one-to-one and C is onto ; hence : :.
One says that the pair (. 1) is stabilizable if the smallest invariant subspace of ,
containing the range of 1, contains C() l(). One also says that the pair (. C) is
detectable if the largest invariant subspace of , contained in ker C, is contained in o().
(a) Prove that (. C) is detectable if and only if (

. C

) is stabilizable.
(b) From now on, we assume that (. 1) is stabilizable and (. C) is detectable. If
R, show that ( i1
n
. 1) is stabilizable and ( i1
n
. C) is detectable.
(c) Prove that H is non-singular. Hint: Let (r. )
T
belong to ker H. Find an a priori
estimate and deduce that
Cr = 0. 1 = 0. r = 0.

= 0.
(d) Deduce that H is hyperbolic (see the previous exercise).
(e) Let (0.
0
)
T
belong to o(H), the stable subspace. Dene (r(t). (t))
T
the solution of
the Cauchy problem .

(t) = H(t), .(0) = (0.

0
)
T
. Establish an integral estimate ;
prove that 1

0, Cr 0 and
r

(t) = r(t).

(t) =

(t).
79
Deduce that r 0. Using the fact that (. 1) is stabilizable, prove that 0.
Hint: The space spanned by the values (t) is invariant under

and annihilated
by 1

.
(f) Likewise, if (r
0
. 0) belongs to l(H), use the detectability of (. C) to prove that
r
0
= 0.
(g) Deduce the Lopatinski condition
C
2n
= (0 C
n
) o(H).
125. Let . 1. C be matrices with complex entries, such that the product 1C makes sense
and is a square matrix. Prove
[Tr(1C)[ ||
S
|1|
S
|C|
S
.
with ||
S
=

Tr

the SchurFrobenius norm. Hint: Apply repeatedly the Cauchy

Schwarz inequality.
Nota: This is the discrete analogue of the following inequality for functions dened on
the plane:

_
R
3
)(r. )p(. .)/(.. r)dr d d.

|)|
L
2
(R
2
)
|p|
L
2
(R
2
)
|/|
L
2
(R
2
)
.
126. Let H be a tridiagonal Hermitian 3 3 matrix:
H =
_
_
c
1
/
1
0

/
1
c
2
/
2
0

/
2
c
3
_
_
(c
j
R. /
j
C).
We denote the characteristic polynomial of H by 1.
(a) Prove the formula (c
3
c
1
)[/
1
[
2
= 1(c
1
). Therefore, the signs of 1(c
1
) and 1(c
3
)
are determined by that of c
3
c
1
.
(b) Construct a pair (c. ) R
3
R
3
such that c ~ , with c
1
< c
3
and

j
(
j
c
1
) < 0.
(c) Deduce that there exists a pair (c. ) R
3
R
3
such that c ~ (and therefore there
is a Hermitian matrix with diagonal c and spectrum , from Theorem 3.4.2), but no
such matrix being tridiagonal.
Nota: A theorem of M. Atiyah asserts that the Hermitian matrices with given
diagonal and given spectrum form a connected set (not true for real symmetric
matrices). A strategy could have been to show that in such a set, there exists a
tridiagonal element. False alas !
80
(d) Prove Atiyahs result for 2 2 Hermitian matrices. Prove also that it becomes false
in 2 2 real symmetric matrices.
127. If  is a square matrix, we denote

 the transpose of the matrix of its cofactors. Recall
that

 =

 = (det )1
n
.
Let . 1. C. 1 be given 2 2 matrices. We form a 4 4 matrix
 :=
_
1
C 1
_
.
(a) Prove that
det  = det(1) + det(1C) 2Tr(1

1C

).
(b) Deduce that

 =
_
(det 1)

C1

1

_
.
Compare with formula of Corollary 8.1.1.
(c) Show that rk 2 holds if and only if

C1

1 = (det 1)

, ... or equivalently
1

1C = (det 1), ... Deduce that
(rk 2) =(det(1) = det(1C)).
(d) More generally, let . 1. C. 1 be given in M
n
(/), such that the matrix  dened
blockwise as above is of rank at most : (this is a (2:) (2:) matrix). Prove that
det(1) = det(1C). Hint: Use the rank decomposition.
Nota: This ensures that, for every matrix 

equivalent to  (

= 1Q with
1. Q non-singular), there holds det(

) = det(1

)
128. Let ) : (0. +) R be a continous monotone matrix function (see Exercise 74). We
recall that ) has a representation of the form (7) with c 0, / R and : a non-negative
bounded measure, with (c. :) ,= (0. 0).
(a) Show that ) is concave increasing.
(b) From now on, we assume that ) is continous at : = 0, meaning that )(0+) .
Prove that for every :. t 0, there holds )(: +t) )(:) +)(t) )(0). Deduce that
if H
n
is non-negative, and if : 0, then )( + :1
n
) )() + ()(:) )(0))1
n
.
(c) Deduce that is . 1 H
n
are non-negative and if : 0, then 1 + :1
n
implies
)() )(1) + ()(:) )(0))1
n
.
(d) Prove at last that for every non-negative Hermitian . 1, there holds
|)(1) )()|
2
)(|1 |
2
) )(0).
Hint: Choose : cleaverly in the above inequality.
Compare with the results of Exercises 110 and 111. We recall (Exercise 74) that
: :

is a monotone matrix function for 0 < 1.

81
129. (Krattenthaler) Let A
1
. .... A
n
,
2
. ....
n
and 1
1
. .... 1
n1
be indeterminates. Prove the
identity
det
1i,jn
((A
i
+
n
) (A
i
+
j+1
)(A
i
+ 1
j1
) (A
i
+ 1
1
)) =

i<j
(A
i
A
j
)

i<j
(1
i

j
).
Hint: Prove that the right-hand side divides the left-hand side. Then compare the
degrees of these homogeneous polynomials. At last, compute the ratio by specializing the
indeterminates.
130. (R. Piche.) We recall (see Exercise 49 for the case with real entries) that the Hermitian
form /() := (: 1) Tr(

) [ Tr [
2
takes non-negative values on the cone of
singular matrices in M
n
(C). Use this result to prove that the spectrum of a general matrix
M
n
(C) is contained in the disk of center :
1
1
_
[ Tr [
2
+ :/().
Check (on : = 2 for instance) that this result is not implied by Gershgorins Theorem.
131. Given a complex : : matrix , show that there exists a unitary matrix l such that
 := l

l has a constant diagonal:

:
ii
=
1
:
Tr . i = 1. .... :.
Hint: Use the convexity of the numerical range (see Exercise 21).
In the Hermitian case, compare with Schurs Theorem 3.4.2.
132. We aim at computing the number of connected components in the subsets of M
n
(R)
or of Sym
n
(R) made of matrices with simple eigenvalues. The corresponding sets are
denoted hereafter by sM
n
(R) and sSym
n
(R). We recall that GL
n
(R) has two connected
components, each one characterized by the sign of the determinant. We denote by GL
+
n
the connected set dened by det  0.
(a) Let sM
n
(R) be given. Show that there exists a matrix 1 GL
+
n
such that
11
1
is block-diagonal, with the diagonal blocks being either scalars (distinct real
numbers), or 2 2 matrices of rotations of distinct nonzero angles.
(b) Let . 1 be given in sM
n
(R), with the same number of pairs of complex conjugate
eigenvalues.
i. Find a path from the spectrum of to that of 1, such that each intermediate
set is made of distinct real numbers and distinct pairs of complex conjugate
numbers.
ii. By using the connectedness of GL
+
n
, prove that and 1 belong to the same
connected component of sM
n
(R).
(c) Deduce that sM
n
(R) has exactly
_
n+1
2

connected components.
82
(d) Following the same procedure, prove that sSym
n
is connected. Comment: Here
is a qualitative explanation of the discrepancy of both results. The complement of
the set of matrices with simple eigenvalues is the zero set of a homogeneous polyno-
mial (), the discriminant of the characteristic polynomial. In the general
case, this polynomial takes any sign, and the complement of sM
n
(R) is an alge-
braic hypersurface. It has codimension one and splits M
n
(R) into several connected
components. In the symmetric case, takes only non-negative values, because the
spectra remain real. It therefore degenerates along its zero set, which turns out to
be of algebraic codimension two (see V. I. Arnold, Chapitres supplementaires de la
theorie des equations dierentielles ordinaires, Mir (1980)). Consequently sSym
n
is
connected.
133. In paragraph 6.3.1, one shows that the minimal polynomial of a companion matrix equals
its characteristic polynomial. On another hand, Proposition 10.1.1 tells us that eigen-
values of an irreducible Hessenberg matrix are geometrically simple. Show that these
results are variants of the more general one: The minimal polynomial of an irreducible
Hessenberg matrix equals its characteristic polynomial.
134. Given M
nm
(/) and 1 M
mn
(/), we form the matrix  M
2n+m
(/):
=
_
_
1
n
0
0 1
m
1
0 0 1
n
_
_
.
Compute the inverse 
1
. Deduce that if we need d

(d. independent of ) operations

to invert an   matrix, we can multiply two matrices such as and 1 in d(2:+:)

operations. In particular, the converse of Proposition 8.1.3 holds true.

135. Let : be an even integer (: = 2:). You check easily that given a row r and a column
of : scalars, their product satises the relation
r =
m

k=1
(r
2k1
+
2k
)(r
2k
+
2k1
)

k
r
2k
r
2k1

2k

2k1
.
Deduce a way to compute a matrix product in M
n
(/) in
n
3
2
+ C(:
2
) multiplications and
:
3
+ C(:
2
3
+ C(:
2
) of each by the naive method. Comment:
This is Winograds calculation. In the 60s, the computational cost of a multiplication
was two or three times that of an addition. Thus it was valuable to divide the number
of multiplications by some factor (here by two), keeping the number of additions roughly
the same.
Can this idea be used recursively, as for Strassens multiplication ?
136. Let / be R or C and | | be a unitary invariant norm on M
n
(/). Prove that for every
matrix M
n
(/), there holds |

| = ||.
83
137. Given a norm | | on R
n
, we denote by o the unit sphere of the corresponding induced
norm on M
n
(R).
(a) Given c. ) R
n
such that |)| = |c| , = 0, we dene
1(c. )) :=  M
n
(R) ; || 1 and c = ).
Prove that 1(c. )) is a convex subset of o.
(b) Conversely, let 1 be a convex subset of o. We assume that the norm | | in R
n
is
strictly convex, meaning that if |r +| = |r| +||, then r and are proportional
(with the same sense of course). Prove that there exists a pair (c. )) as above, such
that 1 = 1(c. )). Hint: Consider an internal point  of 1 and a unit vector c for
which |c| = 1.
(c) In this question, the norm of R
n
is | |
2
and the sphere o is denoted by o
2
.
i. Prove that every 1()

. )) is linearly isometric to 1(c. c) with c = (1. . . . . 1)

T
.
ii. Show that 1(c. c) is the set of matrices such that c = c,
T
c = c and the
restriction of  to c

(an endomorphism of course) has operator norm at most

one. Hence 1(c. c) is linearly isometric to the unit ball of M
n1
(R) equipped
with the induced norm | |
2
. Compare this result with Corollary 5.5.1.
iii. In particular, show that 1(c. c), and therefore each 1(c. )), is maximal among
the convex subsets of o
2
.
138. (Continuation.) We keep c = (1. . . . . 1)
T
but consider the norm | |
p
, where 1 j .
The corresponding set 1(c. c) is denoted by 1
p
.
(a) For j = 1, show that 1
1
reduces to the set
n
of bistochastic matrices.
(b) For j = , show that 1

is the set of stochastic matrices, dened by c = c and

 0.
(c) If : (j. ), show that 1
p
1
q
1
r
.
(d) Assume that 1 < j < . Making a Taylor expansion of |c+c|
p
and of |c+c|
p
,
as c 0, prove that 1
p
1
2
.
(e) Prove that 1

is not a subset of 1
2
. Hint: Compare the dimensions of these convex
sets.
(f) Show that j 1
p
is non-decreasing on [1. 2] and non-increasing on [2. ) (we have
seen above that the monotonicity fails on [2. ]).
(g) If 1 j < 2, prove the right-continuity
1
p
:=

p<q2
1
q
.
84
(h) If 1 < j < , show that 
T
is an isometry from 1
p
onto 1
p
, where j

is
the conjugate exponent (the calculations above show that this is false for j = 1 or
j = ). Deduce that 1
q
is left-continuous on (2. ) (of course it is not at
= since 1

is much too big).

(i) Deduce that

2q<
1
q
=
n
.
139. Let d. R
n
be given, together with unitary matrices Q. 1. We form the diagonal
matrices 1 = diag(d) and = diag().
(a) Show that Tr(1Q1) equal d
T
o, where the matrix of moduli [o[ is majorized by
a bi-stochastic matrix  (see also Exercise 101).
(b)
John von Neumann.
Deduce von Neumanns inequality in M
n
(C):
(16) [Tr(1)[

i
:
i
():
i
(1).
where :
1
(C) :
n
(C) denote the singular
values of C.
Nota: B. Dacorogna and P. Marechal have proven the more general inequality
(17) Tr(1) (sign det(1)):
1
():
1
(1) +

i2
:
i
():
i
(1).
140. Let / be R or C. Given a bounded subset 1 of M
n
(/), let us denote by 1
k
the set of
all possible products of / elements in 1. Given a matrix norm | |, we denote |1
k
| the
supremum of the norms of elements of 1
k
.
(a) Show that |1
k+l
| |1
k
| |1
l
|.
(b) Deduce that the sequence |1
k
|
1/k
converges, and that its limit is the inmum of the
sequence.
(c) Prove that this limit does not depend on the choice of the matrix norm.
This limit is called the joint spectral radius of the family 1, and denoted (1). This
notion is due to G.-C. Rota and G. Strang.
(d) Let (1) denote the inmum of |1| when | | runs over all matrix norms. Show
that (1) (1).
85
(e) Given a norm  on /
n
and a number c 0, we dene for every r /
n
|r| :=

l=0
((1) + c)
l
max(1r) ; 1 1
l
).
i. Show that the series converges, and that it denes a norm on /
n
.
ii. For the matrix norm associated with | |, show that || (1) + c for every
1.
iii. Deduce that actually (1) = (1). Compare with Householders Theorem.
141. (G.-C. Rota & G. Strang.) Let / be R or C. Given a subset 1 of M
n
(/), we consider the
semi-group T generated by 1. It is the union of sets 1
k
dened in the previous exercise,
as / runs over N. We have 1
0
= 1
n
, 1
1
= 1, 1
2
= 1 1,. . .
If T is bounded, prove that there exists a matrix norm | | such that || 1 for every
1. Hint: In the previous exercise, take a sup instead of a series.
142. Let dene the two matrices

0
=
_
1 1
0 1
_
.
1
=
_
1 0
1 1
_
.
Given a map : : 1. .... : 0. 1 (i.e. a word in two letters), we dene
(:) :=
s(1)

s(2)

s(r)
.

(:) :=
s(r)

s(r1)

s(1)
(

(:) is the palindrome of (:)). Show that

(:) (:) =
_
:(:) 0
0 :(:)
_
.
where :(:) is an integer.
143. Let 1 Sym
n
(R) be Toepliz, meaning that
1 =
_
_
_
_
_
_
_
_
c
1
c
2
c
3
c
n
c
2
c
1
c
2
.
.
.
c
3
c
2
c
1
.
.
.
.
.
.
.
.
.
.
.
.
c
n
c
1
_
_
_
_
_
_
_
_
.
We denote by
j
the principal minors, in particular
1
= c and
n
= det 1.
(a) Prove the inequality

n2

2
n1
.
Hint: Use DesnanotJacobi formula of Exercise 24.
86
(b) If
n2
,= 0 and c
1
. . . . . c
n1
are given, show that there exists a unique value c
n
such
that the above inequality becomes an equality.
(c) When 1 is positive denite, deduce the inequalities

1/n
n

1/(n1)
n1

1
.
(d) Let us assume that 1 is positive denite. Prove that for every  :, 1 may be
completed into an SDP
N
Toepliz matrix.
144. Let M
n
(/) be given, with : = j + . For 1 i. , j, let us dene the minor
d
ij
:=
_
i j + 1 :
, j + 1 :
_
.
With the entries d
ij
, we form a matrix 1 M
p
(/). Prove the DesnanotJacobi formula
(see Exercise 24 for the case p=2)
det 1 =
p1
det . :=
_
j + 1 :
j + 1 :
_
.
Hint: Develop d
ij
with the help of Schurs determinant formula. Then apply once more
Schurs formula to det .
145. (B. Perthame and S. Gaubert.) Given a non-negative matrix  M
n
(R) that is irre-
ducible, let denote  :=  ()1
n
. From PerronFrobenius Theorem, = 0 is a
simple eigenvalue of , associated with a positive eigenvector A. Let also 1 denote a
positive eigenvector of
T
, again for the zero eigenvalue.
Given an initial data r
0
R
n
, let t r(t) be the solution of the ODE r = r, such
that r(0) = r
0
.
(a) Show that
(18) 1 r(t) 1 r
0
.
(b) Let H be a (
1
-function over R. Show that
d
dt

j
1
j
A
j
H
_
r
j
A
j
_
=

j,k
A
k
1
j
_
H
_
r
j
A
j
_
H
_
r
k
A
k
_
(19)
+
_
r
k
A
k

r
j
A
j
_
H

_
r
j
A
j
__
.
(c) Show that the kernel of the quadratic form
(r. .)

j,k
:
jk
1
j
A
j
(r
j
.
k
r
k
.
j
)
2
is exactly the line spanned by ..
87
(d) Let : := r 1,A 1 . Deduce that the expression
1(t) :=

j
1
j
A
j
(r
j
:A
j
)
2
satises a dierential inequality of the form

1 + c1 0.
where c 0 depends only on . Hint: Use both (18) and (19).
(e) Verify that r(t) converges towards :A, exponentially fast.
(f) State a result for the solutions of r = r.
146. To do this exercise, you need to know about the exterior algebra 1. Recall that if 1 is a
1-vector space of dimension :, the exterior algebra 1 is the direct sum of the subspaces

k
1 of the tensor algebra, spanned by the vectors r
1
r
k
, where r := rr
whenever r. 1, and is associative. We have
dim
k
1 =
_
:
/
_
.
If c
1
. . . . . c
n
is a basis of 1, then a basis of
k
1 is given by the vectors c
j
1
c
j
k
as 1 ,
1
< < ,
k
:.
(a) Let n be an endomorphism in 1. Prove that there exists a unique endomorphism in

k
1, denoted by n
(k)
, such that
n
(k)
(r
1
r
k
) = n(r
1
) n(r
k
). r
1
. . . . . r
k
1.
(b) Let c
1
. . . . . c
n
be a basis of 1 and be the matrix of n in this base. Show that
the entries of the matrix associated with n
(k)
are the minors

_
i
1
i
k
,
1
,
k
_
with i
1
< < i
n
and ,
1
< < ,
n
.
(c) Let r
1
. . . . . r
k
be linearly independent vectors of n, associated with the eigenvalues

1
. . . . .
k
. Prove that r
1
r
k
is an eigenvector of n
(k)
. What is the corresponding
eigenvalue ?
(d) If 1 = C, show that the spectrum of n
(k)

iI

i
. [1[ = /.
where
1
. . . . .
n
is the spectrum of n. Hint: Consider rst the case where

1
. . . . .
n
are distinct. Then proceed by density.
88
(e) Prove that the above property is true for every scalar eld 1. Hint: The case
1 = C provides an algebraic identity with integral coecients.
147. (Continuation.) We now take 1 = R. Recall that a matrix  M
n
(R) is totally positive
if all the minors

_
i
1
i
k
,
1
,
k
_
with / :, i
1
< < i
n
and ,
1
< < ,
n
are positive. Total positiveness implies
positiveness.
(a) If  is positive, prove that () is a positive simple eigenvalue with the property
that () [[ for every other eigenvalue of .
(b) Let  M
n
(R) be totally positive. Prove that its eigenvalues are real, positive and
pairwise distinct (thus simple). Hint: Proceed by induction on :. Use the fact that

1

n
is the unique eigenvalue of 
(n)
.
(c) Likewise, show that the singular values of  are pairwise distinct.
148. (Loewner.)
(a) Let be a tridiagonal matrix with non-negative o-diagonal entries (i ,= ,).
i. Let us consider a minor

_
i
1
i
k
,
1
,
k
_
with / 2, i
1
< < i
k
and ,
1
< < ,
k
. Show that it is either the
product of minors of smaller sizes (reducible case), or it is a principal minor
(,
1
= i
1
. . . . . ,
k
= i
k
).
ii. Deduce that there exists a real number r such that r1
n
+ is totally non-
negative. Hint: Choose r large enough. Argue by induction on :.
(b) Let 1 M
n
(R) be tridiagonal with non-negative o-diagonal entries. Deduce from
above and from Trotters formula
exp  = lim
k+
_
1
n
+
1
/

_
k
.
that the semi-group (exp(t1))
t0
is made of totally non-negative matrices.
(c) Conversely, let M
n
(R) be given, such that the semi-group (exp(t))
t0
of totally non-negative matrices.
i. Show that the o-diagonal entries of are non-negative.
ii. If , i + 1, show that c
ij
0. Hint: consider the minor

t
_
i i + 1
i + 1 ,
_
of
t
:= exp(t).
89
iii. Deduce that is tridiagonal, with non-negative o-diagonal entries.
149. We denote by c the vector of R
n
whose every component equals one.
(a) Let be a Hermitian matrix, semi-positive denite, and denote c R
n
the vector
whose components are the diagonal entries of . Show that 1 := cc
T
+cc
T
2 is a
Euclidean distance matrix (see Exercise 62). Hint: Use a factorization = 
T
.
(b) Conversely, show that every Euclidean distance matrix is of this form for some semi-
positive denite Hermitian matrix.
(c) Deduce that the set of Euclidean distance matrices is a convex cone. Find also a
direct proof of this fact.
(d) Let : R
n
be such that c
T
: = 1. Let 1(:) be the cone of semi-positive denite
Hermitian matrices 1 such that 1: = 0. Show that the restriction to 1(:) of the
map cc
T
+ cc
T
2 is injective, and that its inverse is given by

1
2
(1
n
c:
T
)(1
n
:c
T
).
150. Let M
nm
(R) and / R
n
be given. Dene two sets
A := r R
m
; r / and r 0. 1 := R
n
;
T
0. 0 and / < 0.
where the inequalities stand for vectors, as in Chapter 5.
Prove that exactly one of both sets is void (Farkas Lemma).
151. Consider the homogeneous polynomial j
r
(A) := A
2
0
A
2
1
A
2
r
for some integer
: 1.
(a) For : = 1 and : = 2, show that there exist real symmetric matrices o
0
. . . . . o
r
such
that
j
r
(A) = det(A
0
o
0
+ + A
r
o
r
).
(b) When : 3 show that there does not exist matrices
0
. . . . .
r
M
2
(R) such that
j
r
(A) = det(A
0

0
+ + A
r

r
).
Hint: Consider a vector R
r+1
such that
0
= 0 and the rst row of
1

1
+ +

r
vanishes.
152. Here is another proof of Birkhos Theorem. Let be a bistochastic matrix of size :.
(a) Prove that there does not exist a submatrix of size /| with null entries and /+| :.
Hint: Count the sum of all entries of .
Then Exercise 10 of Chapter 2, page 32, tells you that there exists a permutation
such that c
i(i)
,= 0 for every i = 1. .... : (this result bears the name of Frobenius
K onig Theorem). In the sequel, we denote by 1 the permutation matrix with entries
j
ij
=
(i)
j
.
90
(b) Let c be the minimum of the numbers c
i(i)
,= 0, so that c (0. 1]. If c = 1, prove
that = 1. Hint: Again, consider the sum of all entries.
(c) If c < 1, let us dene
1 =
1
1 c
( c1).
Show that 1 is bistochastic. Deduce that if is extremal in
n
, then = 1.
153. This is a sequel of Exercise 26, Chapter 4 (notice that this exercise does not exist in
the French edition of the book). We recall that denotes the unit sphere of M
2
(R)
for the induced norm | |
2
. Also recall that is the union of the segments [:. :] where
: 1 := SO
2
(R) and : o, the set of orthogonal symmetries. Both 1 and o are circles.
At last, two distinct segments may intersect only at an extremity.
(a) Show that there is a unique map : o 1, such that  belongs to some
segment [(). :) with : o. For which  is the other extremity : unique ?
(b) Show that the map above is continuous, and that coincides with the identity
over 1. We say that is a retraction from o onto 1.
(c)
Luitzen Brouwer.
Let ) : 1 be a continuous function, where
1 is the unit disk of the complex plane, such that
)(exp(i)) is the rotation of angle . Show that
)(1) contains an element of o.
Hint: Otherwise, there would be a retraction of 1
onto the unit circle, which is impossible (an equiv-
alent statement to Brouwer Fixed Point Theorem).
Meaning. Likewise, one nds that if a disk 1

is
immersed in , with boundary o, then it contains
an element of 1. We say that the circles 1 and o
154. Recall that
3
denotes the set of 33 bistochastic matrices. As the convex hull of a nite
set (the permutation matrices), it is a polytope. It thus has /-faces for / = 0. 1. 2. 3. Of
course, 0-faces are vertices. Justify the following classication:
There are 6 vertices,
15 1-faces, namely all the segments [1. Q] with 1. Q permutation matrices,
18 2-faces, all of them being triangles. Each one is characterized by an inequality
:
ij
+ :
i

j
0, where i ,= i

and , ,= ,

,
9 3-faces, all of them being 3-simplex. Each one is characterized by an inequality
:
ij
0 for some pair (i. ,).
91
Leonhard Euler.
Hint: To prove that a convex subset of dimension / 3 is
a face, it is enough to characterize it by a linear inequality
within
3
. Notice that the alternate sum 6 15 + 18 9
vanishes, as the EulerPoincare characteristics of the sphere
S
3
is zero. Be cautious enough to prove that there is not any
other face.
155. (From V. Blondel & Y. Nesterov.) Let 1 =
1
. . . . .
m
be a nite subset of
n
(/)
(/ = R or / = C). We denote by (
1
. . . . .
m
) the joint spectral radius of 1 (see Exercise
140 for this notion). Prove that
1
:
(
1
+ +
m
) (
1
. . . . .
m
).
Suppose that / = R and that
1
. . . . .
m
are non-negative. Prove that
(
1
. . . . .
m
) (
1
+ +
m
).
156. Let : = |: and
1
. . . . .
m
M
l
(/) be given matrices. Let us form the matrix
:=
_
_
_
_
_
_
_
_
0
l
0
l

1

2
.
.
. 0
l
0
l
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
l
0
l

m
0
l
_
_
_
_
_
_
_
_
.
Prove that
det(1
n
) = det(1
l

m

1
).
Deduce the formula involving characteristic polynomials:
1
A
(A) = 1
A
m
A
1(A
m
).
157. Let H be a Hermitian matrix, given in block form as
H =
_
1
1

C
_
.
92
Assume that rk(H) = rk() (we say that H is a at extension of ). Prove that the
number of positive (resp. negative) eigenvalues of and H are equal. In particular:
(H 0) ( 0).
158. Let M
n
(C) be a normal matrix. We dene 1 M
n1
(C) by deleting the last row
and the last column from . Let
1
. . . . .
n
be the spectrum of , and j
1
. . . . . j
n1

be that of 1. Finally, denote c := (0. . . . . 0. 1)

T
.
(a) Show the identity
( )
1
c. c =
det(1
n1
1)
det(1
n
)
.
(b) Deduce that the rational function
1() :=
det(1
n1
1)
det(1
n
)
has simple poles with non-negative residues.
(c) Conversely, let
1
. . . . .
n
and j
1
. . . . . j
n1
be given tuples of complex numbers,
such that the rational function
1() :=

n1
j=1
( j
j
)

n
k=1
(
k
)
has simple poles with non-negative residues. Prove that there exists a normal matrix
such that
1
. . . . .
n
is the spectrum of , and j
1
. . . . . j
n1
is that of 1.
159. Let : = |: and
1
. . . . .
m
. 1
1
. . . . . 1
m
M
l
(/) be given matrices. Let us form the
matrix
:=
_
_
_
_
_
_
_
_
_
_
_
0
l
0
l
1
1

2
.
.
. 0
l
1
2
1
3
.
.
.
.
.
. 0
l
0
l
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
l
0
l
1
m

m
0
l
_
_
_
_
_
_
_
_
_
_
_
.
We dene also the product
(A) :=
_
0
l
A1
l
1
m

m
_

_
0
l
A1
l
1
1

1
_
.
93
(a) Let / and : :=
_
:
t
_
/
2l
be such that (): =
m
:. Construct an
eigenvector 1 of , for the eigenvalue , such that we have in block form
1 =
_
_
_
_
_
_
_
t
1
.
.
.
t
m2
t
m1
= :
t
m
= t
_
_
_
_
_
_
_
.
(b) Conversely, let be an eigenvalue of . Assume that is not zero. Prove that
m
is an eigenvalue of ().
(c) Let us dene the multivariate polynomial
Q(A
1
. . . . . A
m
) := det
_
A
1
A
m
1
2l

_
0
l
A
m
1
l
1
m

m
_

_
0
l
A
1
1
l
1
1

1
__
.
Show that if at least one of the A
j
s vanishes, then Q vanishes. Therefore A
1
A
m
factorizes in Q(A
1
. . . . . A
m
).
(d) In the case where has : distinct non-vanishing eigenvalues, deduce that
(20) A
n
1
A
(A) = det(A
m
1
2l
(A)).
where 1
A
is the characteristic polynomial of .
(e) Using the principle of algebraic identities, show that (20) holds for every scalar eld
/ and every matrices
1
. . . . . 1
m
.
160. Let o be a set and :. : be positive integers. Let (
s
)
sS
and (1
s
)
sS
be two families
indexed by o, with
s
M
m
(/), 1
s
M
n
(/). We assume that the only subspaces of /
m
(respectively /
n
) invariant by every
s
(respectively 1
s
), that is
s
1 1, are 0 and
/
m
(respectively /
n
) itself.
Let  M
nm
(/) be such that 1
s
= 
s
for every :. Prove (Schurs Lemma) that
either  = 0
nm
, or : = : and  is non-singular. Hint: Consider the range and the
kernel of .
161. Let : = j + with 0 < j < be given. We denote by / the subset of M
n
the matrices with block form
_
0
p
0
pq
0
q
_
.
Likewise, B is made of the matrices
_
0
q
0
qp
1 0
p
_
.
Both / and B are subalgebras of M
n
(/), with dimension j and the property that  =
0
n
for every two elements (of the same algebra). Prove that / and B are not conjugated
in M
n
(/). Show however that B is conjugated to /
T
in M
n
(/).
94
162. (von Neumann)
John von Neumann.
Recall that a norm | | on M
n
(C) is unitarily
invariant if |l\ | = || for every l. \ U
n
and M
n
(C). In this exercise, we dene
: M
n
(C) R
n
by () = (:
1
(). . . . . :
n
()),
where :
1
() :
n
() are the singular values
of .
(a) Let | | be a unitarily invariant norm. Prove that there exists a unique function
p : R
n
R, even with respect to each coordinates and invariant under permutations,
such that | | = p .
In the sequel, such an invariant norm on R
n
is called a gauge.
(b) What are the gauges associated with | |
2
and to the Frobenius norm (Tr(

))
1/2
?
(c) Conversely, let p be a gauge on R
n
. We denote by p

n
. Verify
that p

is a gauge. Then prove that

(21) p

() sup
B=0n
1Tr(

1)
p((1))
.
Hint: First consider the case where is diagonal and non-negative, and use only
the 1s that are diagonal an non-negative.
(d) Prove that (21) is actually an equality. Hint: Use von Neumanns inequality of
Exercise 139.
(e) Deduce that if G is a gauge, then G is a unitarily invariant norm on M
n
(C).
Hint: Apply the results above to p = G

.
163. Given M
n
(/), we dene a linear map 1
A
by A 1
A
A :=
T
A for A Sym
n
(/).
The goal of this exercise is to compute the determinant of 1
A
.
(a) If = Q
T
1Q with Q orthogonal, prove that 1
A
and 1
D
are conjugate to each other.
(b) Compute det 1
D
when 1 is diagonal.
(c) Verify 1
QS
= 1
S
1
Q
, whence det 1
QS
= det 1
Q
det 1
S
.
(d) Consider the case / = R. Deduce from above that if is itself symmetric, then
det 1
A
= (det )
2
.
(e) (Case / = R, continuing). Let Q be real orthogonal. Show that 1
n
is a cluster point
of the sequence (Q
m
)
mN
. Deduce that the identity of Sym
n
(R) is a cluster point of
the sequence (1
m
Q
)
mN
. Thus det 1
Q
= 1. Using the connectedness of SO
n
, show
that actually det 1
Q
= 1.
95
(f) (Case / = R, continuing). Using the polar decomposition of GL
n
(R), prove
that det 1
A
= (det )
2
. Show that this formula extends to every M
n
(R).
(g) Check that the formula det 1
A
= (det )
2
is a polynomial identity with integer
coecients, thus extends to every scalar eld.
164. (Boyd, Diaconis, Sun & Xiao.) Let 1 be a symmetric stochastic : : matrix:
j
ij
= j
ji
0.

j
j
ij
= 1 (i = 1. . . . . :).
We recall that
1
= 1 is an eigenvalue of 1, which is the largest in modulus (Perron
Frobenius). We are interested in the second largest modulus j(1) = max
2
.
n
where

1

n
is the spectrum of 1 ; j(1) is the second singular value of 1.
(a) Let R
n
be such that ||
2
= 1 and

j

j
= 0. Let u. . R
n
be such that
(j
ij
,= 0) =
_
1
2
(.
i
+ .
j
)
i

j

1
2
(u
i
+ u
j
)
_
.
Show that
2

j
.
j
and
n

j
u
j
. Hint: Use Rayleigh ratio.
(b) Taking

j
=
_
2
:
cos
(2, 1)
2:
. .
j
=
1
:
_
cos

:
+
cos
(2j1)
n
cos

n
_
.
deduce that j(1) cos

n
for every tridiagonal symmetric stochastic : : matrix.
(c) Find a tridiagonal symmetric stochastic : : matrix 1

such that
j(1

) = cos

:
.
Hint: Exploit the equality case in the analysis, with the and . given above.
(d) Prove that 1 j(1) is a convex function over symmetric stochastic :: matrices.
Comment: S.-G. Hwang and S.-S. Pyo prove conversely that, given real numbers

1
= 1
2

n
satisfying
1
:
+

2
:(: 1)
+

3
(: 1)(: 2)
+ +

n
2 1
0.
there exists a symmetric bistochastic matrix whose spectrum is
1
. . . . .
n
. This
matrix is not necessarily tridiagonal, since j may be smaller than the bound cos

n
.
It may even vanish, for the matrix
1
n
11
T
.
165. We deal with non-negative matrices in the sense of PerronFrobenius theory. We say
that a non-negative matrix M
n
(R) is primitive if it is irreducible thus the spectral
radius is a simple eigenvalue and this eigenvalue is the only one of maximal modulus.
Let us denote by r and positive eigenvectors of and
T
, according to PerronFrobenius
Theorem. We normalize them by
T
r = 1.
96
(a) Assume rst that () = 1. Remarking that (r
T
) is less than one, prove that

m
r
T
tends to zero as : +. Deduce that
m
0 for : large enough.
(b) Deduce the same result without any restriction over ().
Wielandts Theorem asserts that
m
0 for : :
2
2: + 2.
(c) Conversely, prove that if is non-negative and irreducible, and if
m
0 for some
integer :, then is primitive.
166. Let . 1. C be complex matrices of respective sizes : :, : : and : :. Prove that
the equation
A1 = C
is solvable if, and only if,

C1

1 = C.
In this case, verify that every solution is of the form

C1

+ 1

1 11

.
where 1 is an arbitrary : : matrix. We recall that 

is the MoorePenrose inverse of

.
167. (Roth) Let / be a eld and M
n
(/), 1 M
m
(/), C M
nm
(/) be given matrices.
If the equation A A1 = C is solvable, prove that the following (: + :) (: + :)
matrices are similar:
1 :=
_
0
0 1
_
. 1 :=
_
C
0 1
_
.
We now prove the converse, following Flanders & Wimmer.
(a) We dene two homomorphisms
j
of M
n+m
(/):

0
(1) := 11 11.
1
(1) := 11 11.
Prove that the kernels of
1
and
0
are isomorphic, hence of equal dimensions. Hint:
This is where we use the assumption.
(b) Let 1 be the subspace of M
m(n+m)
(/), made of matrices (1. o) such that
11 = 1. 1o = o1.
Verify that if
1 :=
_
1 Q
1 o
_
ker
j
(, = 0 or 1).
then (1. o) 1. This allows us to dene the projections j
j
(1) := (1. o), from
ker
j
to 1.
97
(c) Verify that ker j
0
= ker j
1
, and therefore 1(j
0
) and 1(j
1
) have equal dimensions.
(d) Deduce that j
1
is onto.
(e) Show that there exists a matrix in ker
1
, of the form
_
1 A
0 1
m
_
.
Conclude.
Remark. With the theory of elementary divisors, there is a nite algorithm which
computes a matrix conjugating given similar matrices. However, the knowledge of
such a conjugator between 1 and 1 does not give an explicit construction of a
solution.
168. (Wielandt.) As in Exercise 146, we use the exterior algebra 1 =
n
k=0

k
1, where now
1 = C
n
. For a given matrix  M
n
(C), we dene
(k)
End(
k
1) by

(k)
r
1
r
k
= (r
1
) r
2
r
k
+ + r
1
r
k1
(r
k
).
(a) Let
1
. . . . .
n
be the eigenvalues of . Verify that the spectrum of 
(k)
consists of
the numbers
I
, [1[ = /, where

I
:=

jI

j
.
(b) The canonical scalar product in 1 extends to
k
1 ; its denition over decomposable
vectors is
r
1
r
k
.
1

k
= det(r
i
.
j
)
1i,jk
.
If  is Hermitian, show that 
(k)
is Hermitian.
(c) From now on, we assume that and 1 are Hermitian, with respective eigenvalues
(they must be real) j
1
j
n
and
1

n
. We from C := + 1, whose
eigenvalues are
1

n
.
i. What is the largest eigenvalue of 1
(k)
?
ii. Show that there is a permutation 1 1

of subsets of 1. . . . . : of cardinal /,
such that

I
j
I
+
k

i=1

i
.
iii. We now assume that has simple eigenvalues: j
1
j
n
, and that 1 is
small, in the sense that
(22) max
i
[
i
[ < min
j
(j
j
j
j+1
).
98
Show that the permutation mentionned above is the identity. Deduce the set of
inequalities
(23)
I
j
I
+
k

i=1

i
. 1 ; [1[ = /.
(d) Conversely, we give ourselves two lists of real numbers j
1
j
n
and
1

n
, satisfying the smallness assumption (22). Let
1

n
be given, satisfying
the list of inequalities (23), together with
(24)

i
=

i
j
i
+

i
.
We dene

i
:=
I
j
i
, and is a re-ordering of

, so that
1

n
.
i. Show that . Deduce that there exists a orthostochastic matrix , such
that = . We recall that an orthostochastic matrix is of the form :
ij
=
[n
ij
[
2
, where l is unitary.
ii. Show that is the spectrum of C = + 1, where := diag(j) and 1 :=
l

diag()l, where l is as above. Therefore the inequalities (23), together with

(24), solve the A. Horns problem when 1 is small.
169. Let / be a eld, = /[1. 2] be the ring of polynomials in two indeterminates and
1 = /(1. 2) be the corresponding eld of rational fractions.
(a) We consider the matrix
(A. 1. 2) :=
_
A + 1 2
2 A 1
_
M
2
([A]).
From Theorem 6.2.1, we know that there exist 1. Q GL
2
(1[A]) such that
1 =
_
1 0
0 A
2
1
2
2
2
_
Q.
Show that one cannot nd such a pair with 1. Q GL
2
([A]), namely with poly-
nomial entries in (A. 1. 2). Hint: The top row entries would vanish at the origin.
(b) Let us now consider the matrix
(A. 1 ) :=
_
A + 1 1
1 A 1
_
M
2
(

[A]).

:= /[1 ].
Show that there exist o. 1 GL
2
(

[A]) such that

o =
_
1 0
0 A
2
1
2
1
_
1.
Hint: multiply  left and right by appropriate elementary matrices.
99
(c) One has o. 1 GL
2
(/[A. 1 ]). Explain why this does not contradict the previous
result.
170. (Heisenberg Inequality.) Let and 1 be two Hermitian matrices of size :. We employ
the canonical Hermitian product . on C
n
and the associated norm.
(a) For every r C
n
, prove
[[. 1]r. r[
2
4|r|
2
|1r|
2
.
Werner Heisenberg.
(b) If |r| = 1 and C is Hermitian (quantum physi-
cists say that C is an observable), we dene the
expectation and the variance of C by
1(C; r) := Cr. r. \ (C; r) := |Cr1(r)r|
2
.
Prove the Heisenberg Inequality: for r, and 1
as above,
[1(i[. 1]; r)[
2
4\ (; r) \ (1; r).
The Heisenberg inequality is therefore a manifesta-
tion of the non-commutativity between operators.
171. A lattice in R
m
is a discrete subgroup of maximal rank. Equivalently, it is the set of vectors
of which the coordinates over a suitable basis of R
m
are integers. It is thus natural to
study the bases of R
m
.
We consider the case where : = 2: and R
m
is nothing but C
n
. Given a family 1 :=

1
. . . . .
2n
of vectors in C
n
, we form the :(2:) matrix := (
1
. . . . .
2n
). Prove that
1 is an R-basis if, and only if, the (2:) (2:) matrix
_

_
is non-singular, where denotes the matrix with complex conjugate entries.
172. Prove Schurs Pfaan identity
Pf
__
c
j
c
i
c
i
+ c
j
__
1i,j2n
=

i<j
c
j
c
i
c
j
+ c
i
.
See Exercise 119 for a hint.
100
173.
Loo-Keng Hua.
Check the easy formula valid whenever the inverses
concern regular : : matrices:
(1
n
+
1
)
1
+ (1
n
+ )
1
= 1
n
.
Deduce Hua Identity
(1 + 1
1
1)
1
+ ( + 1)
1
= 1
1
.
Hint: transport the algebra structure of M
n
(/) by the linear map  1. This
procedure is called isotopy ; remark that the multiplicative identity in the new structure
is 1. Then apply the easy formula.
174. In M
n
(/), we dene the Jordan product by
1 :=
1
2
(1 + 1).
Of course, we assume that the characteristic of the eld / is not equal to 2. We warn
the reader that the bullet is not an associative product. We notice that the square
coincides with
2
.
(a) Prove Jordan Identity

2
( 1) = (
2
1).
(b) Deduce that there is no ambiguity in the denition of
m
when : N. In other
words, generates an associative as well as commutative bullet-algebra.
(c) For every matrices . 1, we dene two linear maps l
A
and \
A,B
by
l
A
(1) := 2 ( 1)
2
1
and
\
A,B
:= 4l
AB
2(l
A
l
B
+ l
B
l
A
).
Prove Thedys Identity:
l
V
A,B
(C)
= \
A,B
l
C
\
A,B
.
Note: One should not confuse Pascual Jordan, a German physicist, with the French
mathematician Camille Jordan, whose name has been given to a canonical form of
matrices over an algebraically closed eld. Amazingly, simple Euclidean (P.) Jordan
algebras obey a spectral theorem, where every element is diagonalizable with real
101
eigenvalues ; hence their (C.) Jordans form is trivial. Somehow, we can say that
there is an exclusion principle about Jordan concepts.
Besides Jordan and Thedy identities, there are two more complicated identities, due
to Glennie, valid in M
n
(/). It has been an important discovery, in the theory of
Jordan algebras, that Glennies and Thedys identities do not follow from Jordans.
As a matter of fact, there exists a Jordan algebra, namely that of 3 3 Hermitian
matrices over Cayleys octonions (a non-associative division algebra), in which all
these three identities are violated. This exceptional object is called Albert algebra.
175. Let 1. Q M
n
(R) be rank-one projectors. Prove that the matrices r1 + Q are diago-
nalizable with real eigenvalues for every r. R if, and only if, either 0 < Tr(1Q) < 1
or 1Q = Q1 = 0
n
or Q = 1.
176. In M
n
(C), we dene endomorphisms 1
A
and 1
A
of the underlying Jordan algebra) when M
n
(C) by
1
A
() :=
1
2
( + ). 1
A
() := .
Show that
1
exp(tA)
= exp(2t1
A
).
177. Let M
n
(/) and j /[A] be given. Show that the minimal polynomial divides
j (that is j() = 0
n
) if, and only if, there exists a matrix C M
n
(/[A]) such that
j(A)1
n
= (A1
n
)C(A).
178. (Delvaux, Van Barel) Let :. : 1 and GL
n
(/), 1 GL
m
(/), G. H M
mn
(/) be
given. Let us dene 1 := G
T
1H and o := 1
1
H
1
G
T
.
(a) Show that the following matrices are equivalent within M
m+n
(/):
_
1
1
H
G
T

_
.
_
1
1
0
0 1
_
.
_
o 0
0
_
.
(b) Deduce the equality
rk1 rko = : :.
179. (Higham & al.) Let  GL
n
(/) be given. We dene the classical group G GL
n
(/)
by the equation

T
 = .
Let j /[A] be given, with j , 0. Let us form the rational function )(A) :=
j(A
1
)j(A)
1
.
(a) Prove (again) that for every G, the following identities hold true:
j() = j(
1
)
T
. j(
T
) = j(
1
).
102
(b) Deduce that
( G) =()() G).
whenever j() is non-singular.
(c) Exemple: every Mbius function preserves a complex group dened by an equation

 = .
180. A matrix pencil is a polynomial
1(A) := A
m

0
+ A
m1

1
+ + A
m1
+
m
.
whose coecients are : : matrices. It is a monic pencil if
0
= 1
n
scalar eld is C and the matrices
k
are Hermitian, we say that the pencil is Hermitian.
Finally, a Hermitian pencil is hyperbolic if the roots of the polynomial
1
u
(A) := 1(A)n. n
are real and simple (notice that 1
u
R[A]) for every non-zero vector n C
n
.
(a) Let H
n
and 1 M
n
(C) be given. We assume that there exist r. C
n
such
that
r. r < 0 < . and 1r. r = 1. = 0.
Show that there exists a non-zero vector . C
n
such that .. . = 1.. . = 0.
Hint: Apply the ToeplitzHausdor Theorem from Exercise 21 to + i1.
(b) Let 1(A) be a hyperbolic Hermitian pencil. We denote
1
(n) < <
n
(n) the
roots of 1
u
when n ,= 0. Recall that the
j
s are smooth functions. We denote also

j
:=
j
(o
n1
)
the image of the unit sphere under
j
. This is obviously a compact interval [

j
.
+
j
]
of R.
i. Check that

j+1
and
+
j

+
j+1
.
ii. Show that 1

u
(
j
(n)) is not equal to zero, that its sign c
j
does not depend on n,
and that c
j
c
j+1
= 1.
iii. Assume that

j+1

+
j
. Let us choose t [

j+1
.
+
j
]. Prove that there exists a
unit vector . such that 1(t).. . = 1

(t).. . = 0. Reach a contradiction.

iv. Deduce that the root intervals
j
are pairwise disjoint.
Nota: A Hermitian pencil is weakly hyperbolic if the roots of every 1
u
are real,
not necessarily disjoint. For such pencils, two consecutive root intervals intersect at
most at one point, that is

+
j

j+1
.
103
181. We prove here the converse of Exercise 127, when the scalar eld / is innite. We recall
that in an innite eld, a polynomial function vanishes identically on /
n
if, and only if,
the associated polynomial is zero. Actually, if a product of polynomial functions vanishes
identically, one of the polynomials at least is zero.
We thus assume that, for every matrix 

equivalent to  (

= 1Q with 1. Q
non-singular), there holds det(

) = det(1

), where


=
_

_
.
Using the rank decomposition, we may assume that  = diag1
n
. J, where J is a
quasi-diagonal : : matrix.
(a) Choosing 1 an Q appropriately, show that for every A. 1 M
n
(/), there holds
det(A1 +J) = (det A)(det 1 ). Deduce that det(1
n
+J1) = 1 for every non-singular
1.
(b) Deduce that the polynomial A det(1
n
+ JA) 1 (a polynomial in :
2
indetermi-
nates) is zero, and thus that the polynomial A Tr(JA) is zero.
(c) Show at last that the rank of  is at most :.
Nota: This, together with Hilberts Nullstellensatz, implies that for every minor j of
of size : (:. 2:], viewed as a polynomial of the entries, there exist an integer :
such that j
m
belongs to the ideal spanned by the polynomials
P,Q
:= (det

)(det 1

)
(det 1

)(det C

) with

:= 1Q. Clearly, : 2, but the least

integer :(:) is not known, to our knowledge.
182. Recall that if \ is a vector space of dimension :, a complete ag in \ is a set of subspaces
1
0
= 0 1
1
1
n
= \ , with dim1
k
= / for each index /. Let 1 be a subset of
1. . . . . :, say that 1 = j
1
< < j
r
. Given a complete ag 1 and an index subset
1 of cardinality :, we dene

P
(1) = 1 A
r
[ dim1
k
1 /.
with A
r
the set of subspaces of \ of dimension :. Let be an : : Hermitian matrix,
and take \ = C
n
.
(a) Let 1 A
r
be given. Show that the quantity
r

k=1
r
k
. r
k

does not depend on the choice of a unitary basis r

1
. . . . . r
m
of 1. We call it the
trace of over 1, and denote Tr
L
.
104
(b) Let c
1
c
n
be the eigenvalues of , counting with multiplicities. Let B =
(n
1
. . . . . n
n
) the corresponding unitary basis and 1

() be the complete ag spanned

by B: 1
k
= n
k
1
k1
. Let 1 := (j
1
. . . . . j
r
) be as above. Show that

pP
c
p
= sup
L
P
(F(A))
Tr
L
.
Hint: To show that every such Tr
L
is less than or equal the left-hand side, nd an
adapted basis of 1. To show the inequality leftright, choose an appropriate 1.
183. (From G. Pisier).
Let 2 M
n
(C) be given, whose entries are unit numbers: [.
jk
[ = 1. We dene the linear
map over M
n
(C)
1 :  1() := 2 . (1())
jk
:= .
jk
:
jk
.
(a) Show that
|1()|
1
= ||
1
. |1()|

= ||

. |1()|
2

:||
2
.
(b) Deduce that
|1()|
p
:

||
p
. j [1. ]. := min1,j. 1,j

.
Hint: Use RieszThorin Interpolation Theorem.
(c) Given  M
n
(C), let /:() be its absolute value: c
ij
:= [:
ij
[. Prove that
|/:()|
p
:

||
p
. j [1. ].
Hint: Find a 2 adapted to this .
(d) Dene := exp(2i,:) and M
n
(C) by
jk
:=
jk
. Show that

= :1
n
.
Deduce that
|/:()|
2
=

:||
2
.
and therefore the constant

: is optimal in the inequality
|/:()|
2
C
2
||
2
.
Nota: The analogous problem in M
n
(R) is less trivial. One can show that

:
is optimal if and only if there exists a unitary matrix l with entries of constant
modulus, namely :
1/2
. In the real case, this means that the entries are :
1/2
.
This amounts to the existence of a Hadamard matrix. Such matrices exist only for
a few values of :. For instance, they do not for : = 3. 5. 6. 7. The determination of
sizes : for which a Hadamard matrix exists is still an open question.
105
(e) More generally, we show now that :

is the optimal constant in the inequality

|/:()|
p
C
p
||
p
.  M
n
(C).
i. Check that if j 2, then
|r|
2
|r|
p
:
1
p

1
2
|r|
2
.
while if j 2, then
|r|
2
|r|
p
:
1
p

1
2
|r|
2
.
ii. Deduce that ||
p
:
1
, and therefore
|/:()|
p
:

||
p
.
Conclude.
iii. Deduce in particular that ||
p
= :
1
.
184. Let . 1 H
n
be given, such that + 1 is positive denite.
(a) Show that there exists a Hermitian matrix, denoted by 21, such that
inf
yC
n
(r ). r +1.
equals 21r. r for every r C
n
.
We call 21 the inf-convolution of and 1.
(b) Check that 2 is symmetric and associative.
(c) Compute 21 in closed form. Check that
21 = 1( + 1)
1

(surprizingly enough, this formula is symmetric in and 1, and it denes a Hermi-

tian matrix).
(d) If and 1 are non-degenerate, show that
21 = (
1
+ 1
1
)
1
.
185. (T. Yamamoto.)
Given M
n
(C), we denote by :
1
() :
n
() its singular values, and by
1
(). . . .
its eigenvalues ordered by non-increasing modulus:
[
1
()[ [
n
()[.
In this list, it is useful that an eigenvalue with multiplicity / corresponds to / consecutive
eigenvalues of the list. In this exercise, we prove
lim
m+
:
k
(
m
)
1/m
= [
k
()[. / = 1. . . . . :.
106
Remark that for / = 1, this is nothing but the fundamental Lemma of Banach algebras,
namely
lim
m+
|
m
|
1/m
= ().
where the operator norm is | |
2
.
(a) Prove that there exist subspaces G (of dimension : / + 1) and H (of dimension
/) that are invariant under , such that the spectrum of the restriction of to G
(respectively to H) is
k
(). . . . .
n
() (resp.
1
(). . . . .
k
()). Hint: Use Jordan
reduction.
(b) Verify the formula
(25) :
k
() = sup
dimF=k
inf
x

F
|r|
2
|r|
2
.
where r

1 means that r runs over the non-zero vectors of 1. Hint: Both sides
are unitarily invariant. Use the decomposition in singular values.
(c) Deduce the bounds
_
_
([
H
)
1
_
_
1
:
k
() |[
G
| .
(d) Apply these bounds to
m
and conclude.
186. (R. C. Thompson.) Let 1 be a principal ideal domain. Let  M
nm
(1) be given,
with : := min:. :. Let

M
pq
(1) be a submatrix of , with j + :. Show
that 1
p+qr
(

) divides 1
r
() (recall that 1
k
() is the g.c.d. of all minors of size /
of ). Hint: If j + = : + 1, have a look to Exercise 10 of Chapter 2.
Let / be a eld and M
n
(/) be given. Let

M
pq
(/[A]) be a submatrix of
A1
n
, with j + :. We denote the invariant factors of 

by
1
[
2
[ . Deduce
that the product
1

p+qn
divides the characteristic polynomial of .
187. The rst part is an exercise about polynomials. The second part is an application to
matrices, yielding Dunford decomposition. The eld / has characteristic zero.
(a) Let 1 /[A] be monic of degree :. Suppose rst that 1 splits, say
1(A) =
s

j=1
(A c
j
)
m
j
. .
i. Show that there exists a unique polynomial Q of degree :, such that for every
, = 1. . . . . :, one has
Q(c
j
) = c
j
and Q
()
(c
j
) = 0. 1 / :
j
1.
107
ii. Dene the polynomial
(A) :=
s

j=1
(A c
j
).
Show that c
j
is a root of Q of order :
j
at least. Deduce that 1 divides Q.
iii. If 1 does not split, let 1 be an extension of / in which 1 splits. Let Q and
be dened as above. Show that /[A] and Q /[A].
(b) Let  M
n
(/) be given. We apply the construction above to 1 = 1
M
, the
characteristic polynomial. We let 1[A) := A Q(A), and we dene 1 := Q()
and  := 1().
i. What is the spectrum of  ? Deduce that  is nilpotent.
ii. Show that (1) = 0
n
. Deduce that 1 is diagonalisable in a suitable extension
of /.
iii. Deduce the Dunford decomposition:  writes as 1+ for some diagonalisable
1 and nilpotent , with [1. ] = 0
n
. Both 1 and  have entries in /, though
1 could be diagonalisable only in a suitable extension, that one containing all
the eigenvalues of .
188.
Leonhard Euler.
Let denote the Euler indicator, (:) is the num-
ber of integers less than : that are prime to :.
We recall the formula

d|n
(d) = :.
In the sequel, we dene the : : matrices G (for
gcd), and 1 (for divisibility) by
p
ij
:= gcd(i. ,). := diag(1). . . . . (:). d
ij
:=
_
1 if i[,.
0 else.
(a) Prove that 1
T
1 = G.
(b) Deduce the Smith determinant formula:
det(( gcd(i. ,) ))
1i,jn
= (1)(2) (:).
(c) Compute the invariant factors of G as a matrix of M
n
(Z), for small values of :. Say
up to : = 10.
108
189. (from D. Ferrand.) Let be a permutation of 1. . . . . : and 1

be the associated
permutation matrix.
(a) We denote by c
m
the number of cycles of length : in . Show that the characteristic
polynomial of 1

is

m1
(A
m
1)
cm
.
(b) Let and be two permutations. We assume that 1

and 1

are similar in
n
(/),
the eld / having characteristic zero. Show that for every :, there holds c
k
() =
c
k
(). Deduce that and are conjugated as permutations.
190. Let  M
nm
() be given. Verify that det
T
 equals the sum of the squares of the
minors of  of size :.
191. One begins with the following observation: if . 1 GL
n
(/) are such that
2
= 1
2
,
then  :=
1
1 is conjugated to its inverse
1
. Verify!
We prove now the converse, and even slightly better; namely, if  and
1
are similar,
then there exist . 1 GL
n
(/) such that
2
= 1
2
= 1
n
and  :=
1
1.
(a) Show that both the assumption and the conclusion are invariant under conjugation.
(b) Assume : = 2:. We dene o. 1 GL
n
(/) by
o =
_
1
m
0
m
J(4; :) 1
m
_
. 1 =
_
1
m
1
m
0
m
1
m
_
.
where J(c; :) stands for the Jordan block with eigenvalue c. Check that o and 1
are involutions, and show that o
1
1 has only one eigenvalue and one eigendirection.
Conclude that the result holds true for J(1; 2:) (notice that J(1; :) is always similar
to its inverse).
(c) We keep the notations of the previous question. Show that the intersection of
ker(o
T
+ 1
n
) and ker(1
T
+ 1
n
) is a line. Deduce that there is a hyperplane H
that is stable under both o and 1. Show that the restrictions of o and 1 to H
are involutions, and that of o
1
1 is similar to J(1; 2: 1). Hint: the latter is
the restriction of J(1; 2:) to a stable hyperplane; it cannot be something else than
J(1; 2:1).
(d) We have thus proven that for every :, J(1; :) satises the claim. Use this to prove
that J(1; :) satises it too.
(e) Let c / be given, with c ,= 0. 1. Check that
_
0
m
J(c; :)
J(c; :)
1
0
m
_
is an involution. Deduce that diagJ(c; :). J(c; :)
1
satises the claim. Conclude
that diagJ(c; :). J(c
1
; :) satises it too.
109
(f) If the characteristic polynomial of  splits on / (for instance if / is algebraically
closed), prove the claim. Hint: Apply Jordan decomposition.
(g) In order to solve the general case, we use the second canonical form in Frobenius
reduction. Let 1 be a power of an irreducible monic polynomial over /.
i. Show that the inverse of the companion matrix 1
P
is similar to the companion
matrix of

1, the polynomial dened by

1(A) =
1
1(0)
A
degP
1(1,A).
ii. Show that diag1
P
. 1
1
P
is the product of two involutions. Hint: Mimic the
case of J(c; :).
iii. Conclude. Hint: Verify that if the list of elementary divisors of an invertible
matrix  is j
1
. . . . . j
r
, then the elementary divisors of
1
are j
1
. . . . . j
r
. Mind
that one must treat the cases (A 1)
s
apart.
192. Let  M
n
(R) have the following properties:
is irreducible,
For every pair i ,= ,, one has :
ij
0,
c
T
 = 0, where c
T
= (1. . . . . 1).
(a) Show that = 0 is a simple eigenvalue of , associated with a positive eigenvector
\ 0, and that the other eigenvalues have a negative real part.
(b) Let us denote 1 := diag1,
1
. . . . . 1,
n
. Prove that the symmetric matrix 1 +

T
1 is negative semi-denite, its kernel being spanned by \ .
193. (Tan Lei). The following is an easy but not so well-known consequence of Perron
Frobenius Theorem.
Let be a non-negative : : matrix. Then () < 1 if, and only if, there exists a
positive vector r 0, such that r < r.
194. In Le Verriers method, and even in Fadeevs variant (see Exercise 67), the complexity of
the computation of the characteristic polynomial of an : : matrix  in characteristic
0 is :
4
if we use the naive way to multiply matrices. It is :
+1
if we now multipling
matrices in C(:

) operations. Here is an improvement, found by F. Preparata and D.

Sarwate in 1978.
We still compute the Newton sums of the eigenvalues, which are equal to the traces of
the powers
k
, / = 0. .... :1. However we do not compute all the matrix powers. Let
: be the smallest integer larger than or equal to

: (: =

: if : is a square).
110
Urbain Le Verrier.
(a) What is the complexity of the calculation of the
powers , . . . ,
r
?
(b) What is the complexity of the calculation of the
powers 
2r
, . . . ,
r(r1)
?
(c) How many operations do we need to compute the
Newton sums o
0
. . . . . o
m1
once we now the powers
computed above ? Hint: To compute Tr(1C), one
does not need to compute 1C.
(d) Prove that the complexity of the computation of
the characteristic polynomial is at most C(:
+1/2
).
195.
In the theory of dierential equations, one
says that a matrix M
n
(R) is stable if
its eigenvalues lie in the open left half-plane
. ; 1. < 0. It is strongly stable if 1
is stable for every diagonal non-negative ma-
trix 1. The lack of strong stability yields
what is called a Turing instability. We pro-
pose here necessary conditions for to be
strongly stable.
Let us dene  = , so that the spectrum
of  + 1 has a positive real part for all 1
as above.
Alan Turing (with von Neumann).
(a) Let i
1
< < i
r
be indices between 1 and :. Prove that the principal submatrix

_
i
1
i
r
i
1
i
r
_
has its spectrum in 1. 0. Hint: take d
j
= 0 if , is an index i
s
, d
j
= otherwise.
Take the limit as + and use Schurs complement formula.
(b) Deduce that every principal minor of  must be non-negative.
(c) Show that the polynomial 1(A) := det(A1
n
+ ) has positive coecients. Hint:
This real polynomial has roots of positive real parts. Deduce that for every size
1 : :, there exists a principal minor of order : in  which is positive.
111
(d) What does all that mean for the principal minors of ?
R. A. Satnoianu and P. van den Driessche
(2005) provided an example which shows
that these necessary conditions are not suf-
cient when : 4. They turn out to be so
for : = 1. 2. 3.
Schurs complement formula is associated
with the Banachiewiczs inversion formula
(Corollary 8.1.1).
196. Let M
n
(C) be given, with characteristic polynomial A
n
c
0
A
n1
c
n1
. Let
be the solution of the dierential equation

(n)
= c
0

(n1)
+ + c
n1
.
with the initial conditions (0) =

(0) = =
(n2)
(0) = 0 and
(n1)
(0) = 1 (this is the
fundamental solution of the ODE). Finally, dene matrices
0
. . . . .
n1
by
j
= /
j
()
with
/
0
(.) = 1. /
1
(.) = . c
0
. . . .
/
j
(.) = ./
j1
(.) c
j1
. . . . /
r
(.) = 1
A
(.).
Prove that
exp(t) =
(n)
(t)
0
+
(n1)
(t)
1
+ +(t)
n1
.
Nota. The sequence
0
. . . . .
n1
is the
FibonacciHorner basis of the algebra spanned by
. It is actually a basis only if 1
A
is the minimal
polynomial of . The interest of the formula is
that it is valid even if the minimal polynomial has
a lower degree.
Leonardo Fibonacci.
112
197. We apply the Jacobi method to a real 33 matrix . Our strategy is that called optimal
choice.
(a) Let (j
1
.
1
), (j
2
.
2
), . . . , (j
k
.
k
), . . . be the sequence of index pairs that are chosen at
consecutive steps (recall that one vanishes the o-diagonal entry of largest modulus).
Prove that this sequence is cyclic of order three: It is either one of the sequences
. . . . (1. 2). (2. 3). (3. 1). (1. 2). . . ., or . . . . (1. 3). (3. 2). (2. 1), (1. 3). . . .
(b) Assume now that has simple eigenvalues. At each step, one of the three o-diagonal
entries is null, while the two other ones are small, since the method converges. Say
that they are 0. r
k
.
k
with 0 < [r
k
[ [
k
[ (if r
k
vanishes then we are gone because
one diagonal entry is an eigenvalue). Show that
k+1
r
k
and r
k+1
2r
k

k
,,
where is a gap between two eigenvalues. Deduce that the method is of order
= (1 +

5),2, the golden ratio (associated notably with the Fibonacci sequence),
meaning that the error c
k
at step / satises
c
k+1
= C(c
k
c
k1
).
This behaviour ressembles that of the secant method for the resolution of algebraic
equations.
Illustration for the golden ratio (left) and the Fibonacci sequence (right).
198. (Pusz and Woronowicz). Let . 1 H
+
n
two given positive semi-denite matrices. We
show here that among the positive semi-denite matrices A H
+
n
such that
H(A) :=
_
A
A 1
_
0
2n
.
there exists a maximal one. The latter is called the geometric mean of and 1, and is
denoted by #1. Then we extend properties that were well-known for scalars.
(a) We begin with the case where is positive denite.
i. Prove that H(A) 0
2n
is equivalent to A
1
A 1 (see Exercise 6, Chapter
8).
113
ii. Deduce that
1/2
A
1/2
(
1/2
1
1/2
)
1/2
. Hint: The square root is oper-
ator monotone over R
+
. See the Additional Exercise 74.
iii. Deduce that among the matrices A H
+
n
such that H(A) 0
2n
, there exists a
maximal one, denoted by #1. Write the explicit formula for #1.
iv. If both . 1 are positive denite, prove that (#1)
1
=
1
#1
1
.
(b) We now consider arbitrary elements . in H
+
n
.
i. Let c 0 be given. Show that H(A) 0
2n
implies A ( + c1
n
)#1.
ii. Prove that c ( + c1
n
)#1 is non-decreasing.
iii. Deduce that #1 := lim
0
+( + c1
n
)#1 exists, and that it is the largest
matrix in H
+
n
among those satisfying H(A) 0
2n
. In particular,
lim
0
+
( + c1
n
)#1 = lim
0
+
#(1 + c1
n
).
The matrix #1 is called the geometric mean of and 1.
(c) Prove the following identities. Hint: Dont use the explicit formula. Use instead
the denition of #1 by means of H(A).
#1 = 1#;
If  GL
n
(C), then (#1)

= (

)#(1

).
(d) Prove the following inequality between harmonic, geometric and arithmetic mean:
2(
1
+ 1
1
) #1
1
2
( + 1).
Hint: Just check that
H
_
2(
1
+ 1
1
)
_
0
2n
and H
_
1
2
( + 1)
_
0
2n
.
In the latter case, use again the fact that :

: is operator monotone.
(e) Prove that the geometric mean is operator monotone:
(
1

2
and 1
1
1
2
) (
1
#1
1

2
#1
2
).
and that it is a operator concave, in the sense that for every (0. 1), there holds
(
1
+ (1 )
2
)#(1
1
+ (1 )1
2
) (
1
#1
1
) + (1 )(
2
#1
2
).
Note that the latter property is accurate, since the geometric mean is positively
homogeneous of order one. Note also that the concavity gives another proof of
the arithmetico-geometric inequality, by taking
1
= 1
2
= ,
2
= 1
1
= 1 and
= 1,2.
114
(f) Prove the identity between arithmetic, harmonic and geometric mean:
_
2(
1
+ 1
1
)
1
_
#
+ 1
2
= #1.
Hint: Use the fact that # is the unique solution in H
+
n
of the Ricatti equation
A
1
A = . Use it thrice.
199. (Continuation.) Each positive denite Hermitian matrix denes a norm |r|
A
:=

r. If  M
pq
(C) and
1
HDP
q
,
2
HDP
q
, we denote by ||
A
1
A
2
the norm
sup|r|
A
1
; |r|
A
2
= 1.
(a) Show that the dual norm of the Hermitian norm | |
A
is | |
A
1.
(b) Prove the following interpolation result (compare with the RieszThorin Theorem):
For every matrix  M
pq
(C) and every
1
. 1
1
HDP
q
,
2
. 1
2
HDP
q
, we
have
||
A
1
#B
1
A
2
#B
2
||
1/2
A
1
A
2
||
1/2
B
1
B
2
.
Hint: Again, use the denition of the geometric mean, not the formula.
Comment: In the terminology of interpolation theory, one writes for . 1 HDP
n
(C
n
; #1) = [(C
n
; ). (C
n
; 1)]
1/2
.
where 1,2 is the interpolation parameter. Recall that
[(C
n
; ). (C
n
; 1)]
0
= (C
n
; 1). [(C
n
; ). (C
n
; 1)]
1
= (C
n
; ).
More generally, [(C
n
; ). (C
n
; 1)]

can be computed for every diadic = :2

k
by
means of iterated geometric mean. For instance
[(C
n
; ). (C
n
; 1)]
3/4
= [(C
n
; ). [(C
n
; ). (C
n
; 1)]
1/2
]
1/2
= [(C
n
; ). (C
n
; #1)]
1/2
.
(c) Following the idea presented above, show that there exists a unique continuous curve
: H(:) for : [0. 1], with the property that H(0) = 1, H(1) = and
H
_
: + t
2
_
= H(:)#H(t). 0 :. t 1.
This curve is dened by the formula
H(:) =
1/2
(
1/2
1
1/2
)
1s

1/2
.
We denote [. 1]
s
= H(:). Verify that [. 1]
1s
= [1. ]
s
.
115
200. In periodic homogenization (a chapter of Applied Partial Dierential Equations) of elliptic
PDEs, one has a continuous map r (r) from R
n
into SDP
n
, which is -periodic,
being a lattice. The homogenized matrix

is dened as follow. For every vector c R
n
,
we admit that there exists a unique (up to an additive constant) solution u : R
n
R of
the problem
div((r)u) = 0 in R
n
. u(r + ) = u(r) + c . .
(Notice that the last property implies that u is periodic, and its average equals c.)
Then we dene

c :=< u .
where < denotes the average over 1 .
In this exercise we consider the simple case where : = 2 and depends only on the rst
coordinate r
1
. In particular, one can take spanned by 1c
1
(1 the period of r
1
(r
1
))
and by c
2
.
(a) Let u be as above. Prove that u,r
2
is a constant. More precisely, show that
u,r
2
c
2
.
(b) Likewise, prove that
c
11
u
r
1
+ c
12
c
2
is a constant and compute that constant.
(c) Finally, prove the following formula

=
_
[c
11
] [c
12
]
[c
12
] < (det ),c
11
+
<a
12
/a
11
>
2
<1/a
11
>
_
where [)] :=< ),c
11
, < 1,c
11
. In particular, verify that det

= [det ].
201. We recall that the symplectic group Sp
n
(/) is dened as the set of matrices  in M
2n
(/)
which satisfy
T
J
n
 = J
n
, where
J
m
:=
_
0
n
1
n
1
n
0
n
_
.
(a) Let
=
_
1
C 1
_
be a symplectic matrix. Check that 1
T
and
T
C are symmetric, and that
T
1
C
T
1 = 1
n
.
(b) Let A Sym
n
(/) be given and  be as above. Show that (CA +1)
T
(A +1) is
symmetric. Deduce that if CA +1 is non-singular, then  A := (A +1)(CA +
1)
1
is symmetric.
116
(c) Show that the matrices 1(A) := C
T
(A +1)
T
C and o(A) :=
T
(CA +1)
T
C
are symmetric for every symmetric A.
(d)
Rowan Hamilton.
From now on, we choose / = R. We
say that  is a Hamiltonian matrix if
1
T
and
T
C are positive denite.
i. Let A SPD
n
be given. Show that 1(A) and o(A) are positive denite.
ii. Show also that
1
C SPD
n
.
iii. Deduce that 1 is similar to the product of three positive denite symmetric
matrices.
iv. Conclude that 1 is positive denite (see Exercise 6 of Chapter 7).
To summarize, a Hamiltonian matrix  acts over SPD
n
by A  A (Siegel,
Bougerol).
(e) Prove that the set of Hamiltonian matrices is a semi-group: The product of two
Hamiltonian matrices is Hamiltonian (Wojtkowski).
202. For every M
nm
(C) and every t R, show that
: : = Tr exp(t

) Tr exp(t

).
Comment: This is a special case of a formula giving the index of a Fredholm operator
1, thus in innite dimension:
ind1 = Tr exp(t11

) Tr exp(t1

1). t 0.
Notice that in general the dierence exp(t11

) exp(t1

1) does not make sense.

203. Let :. : 2 be two integers. If
1
. . . . .
r
M
n
(/) are given, one denes
1
r
(
1
. . . . .
r
) :=

Sr
c()
(1)

(r)
.
where o
r
is the group of permutations of 1. . . . . : and c : o
r
1. +1 is the signature.
(a) Verify that 1
r
: M
n
(/)
r
M
n
(/) is an alternate :-linear map.
(b) We consider the case / = R and we endow M
n
(R) with the Frobenius norm. We
thus have a Euclidean structure, with scalar product . 1 = Tr(1
T
).
117
i. Show that the supremum (:. :) of
|1
r
(
1
. . . . .
r
)|
|
1
| |
r
|
over
1
. . . . .
r
,= 0
n
is reached. We choose an :-uplet (
1
. . . . . 
r
) at which
this maximum is obtained. Check that one is free to set |
j
| = 1 for all ,.
ii. Let
_
c /
c d
_
SL
2
(R) be given. Show that
|1
r
(
1
.
2
. 
3
. . . . .
r
)| = |1
r
(c
1
+ /
2
. c
1
+ d
2
. 
3
. . . . .
r
)|.
Deduce that if (:. :) ,= 0, then |c
1
+ /
2
| |c
1
+ d
2
| 1.
iii. Derive from above that (:. :) ,= 0 implies that 
i
.
j
= 0 for every pair
i ,= ,.
iv. Conclude that 1
r
0
n
for every : :
2
+ 1
(c) We go back to a general eld of scalars /. Prove that for every : :
2
+1 and every

1
. . . . .
r
M
n
(/), one has
1
r
(
1
. . . . .
r
) = 0
n
.
Hint: Apply the principle of algebraic identities. Use the fact that R is innite.
Comment. The vanihing of 1
r
is called a polynomial identity over M
n
(/). The
above result is far from optimal. The Theorem of Amitsur and Levitzki tells us that
1
r
vanishes identically over M
n
(/) if, and only if, : 2:. See Exercise 289.
204. This exercise yields a lemma of Minkowski.
(a) Show that, if either j = 2 and 2, or j 3 and 1, then j

+ 2.
(b) Dene := j

, with j a prime number and 1 if j is odd, 2 if j = 2. Deduce

from above that for every 2 , /, j
2+vp(k)
(
p
(/) is the power to which j divides
/, its j-valuation) divides
_
/
,
_
j
j
.
(c) Let be as above and 1 M
n
(Z) be such that j does not divide 1 in M
n
(Z). Let
/ 2 be an integer and form := 1
n
+ 1. Show that

k
1
n
+ /1 mod j
2+vp(k)
.
(d) Deduce that if is as above and if
k
= 1
n
, then = 1
n
.
(e) More generally, let M
n
(Z) be given. Prove that if : divides 1
n
and if

k
= 1
n
for some integers : 3 and / 1, then = 1
n
. In other words, the kernel
of the homomorphism GL
n
(Z) GL
n
(Z,:Z) is torsion free.
118
(f) Show that the statement is false when : = 2. Find a matrix 1
2
+2M
2
(Z) such
that
2
= 1
2
and ,= 1
2
.
205. Let M
n
(/), 1 M
m
(/) and  M
nm
(/) be such that  = 1. It is well-
known that if : = : and  is non-singular, then the characteristic polynomials of and
1 are equal: 1
A
= 1
B
. Prove that gcd1
A
. 1
B
has a factor of degree : = rk . Hint:
Reduce to the case where  is quasidiagonal.
206. Let / be a eld. Given two vectors A. 1 in /
3
, we dene the vector product as usual:
A 1 :=
_
_
r
2

3
r
3

2
r
3

1
r
1

3
r
1

2
r
2

1
_
_
.
Prove the following identity in M
3
(/):
A(1 2)
T
+ 1 (2 A)
T
+ 2(A 1 )
T
= det(A. 1. 2) 1
3
. A. 1. 2 /
3
.
207. Let / be a eld and 1 j :. : be integers.
(a) Let  M
nm
(/) be given, with rk = j. Show that there exist two matrices
A M
np
(/) and 1 M
pm
(/) such that  = A1 .
(b) We write such a rankj matrix in block form:
=
_
1
C 1
_
.
with M
pp
(/). If is non-singular, show that 1 = C
1
1.
(c) We assume that  as rank j and that det(J + ) = 0, where
J =
_
0
p
0
0 1
np
_
.
Show that the block is singular. Deduce that there exists a non-zero vector
. /
np
such that either (J + )2 = 0, or (J +
T
)2 = 0, where
2 :=
_
0
.
_
.
208. (Continuation.) We assume that : = /j and  has rank j at most. We can therefore
factorize  in the form
 =
_
_
_

1
.
.
.

k
_
_
_
_
1
1
1
k
_
.
119
Let A
1
. . . . . A
k
be indeterminates, and dene A := diagA
1
1
p
. . . . . A
k
1
p
. Show that
det(A + ) = det
_
A
1
A
k
1
p
+

A
j
1
j

j
_
.
where

A
j
denotes A
1
A
j1
A
j+1
A
k
. Hint: As usual, Schur formula is useful.
209. We show here that H (det H)
1/n
is concave over HPD
n
.
(a) Recall that, given H and 1 in HPD
n
, the product H1 is diagonalizable with real,
positive, eigenvalues (see Exercise 258).
(b) Deduce that
(det H)
1/n
(det 1)
1/n

1
:
Tr H1.
Hint: Use the arithmetic-geometric inequality.
(c) Show that
(det H)
1/n
= min
_
1
:
Tr H1 ; 1 HPD
n
and det 1 = 1
_
.
(d) Deduce concavity.
210. The notation comes from a nonlinear electrodynamics called the BornInfeld model.
Max Born.
In the canonical Euclidian space R
3
(but R
n
works as
well), we give ourselves two vectors 1 and 1, satisfying
|1|
2
+ (1 1)
2
1 +|1|
2
.
Prove the following inequality between symmetric ma-
trices
11
T
+ 11
T
(1 +|1|
2
)1
3
.
211. Fix two integers 0 : : 1. We give ourselves complex numbers c
jk
for every
1 ,. / : such that [/ ,[ : (2: + 1 diagonals). We assume that c
kj
= c
jk
.
120
Prove that we can complete this list of entries so as to make a matrix HDP
n
if, and
only if, every principal submatrix of size : + 1,
_
_
_
c
jj
c
j,j+m
.
.
.
.
.
.
.
.
.
c
j+m,j
c
j+m,j+m
_
_
_
is positive denite.
212. We denote
A :=
_
1 1
0 1
_
. 1 :=
_
1 0
1 1
_
.
(a) Let  SL
2
(Z) be a non-negative matrix (that is an element of SL
2
(N)). If  ,= 1
2
,
show that the columns of  are ordered: (:
11
:
12
)(:
21
:
22
) 0.
(b) Under the same assumption, deduce that there exists a matrix

SL
2
(N) such
that either  =

A or  =

1 . Check that Tr 

Tr . Under which
circumstances do we have Tr 

< Tr  ?
(c) Let  SL
2
(N) be given. Arguing by induction, show that there exists a word u
0
in two letters, and a triangular matrix 1 SL
2
(N), such that  = 1u
0
(A. 1 )
SL
2
(N).
(d) Conclude that for every  SL
2
(N), there exists a word u in two letters, such that
= u(A. 1 ).
Comment. One can show that every element of SL
2
(Z), whose trace is larger than 2, is
conjugated in SL
2
(Z) to a word in A and 1 . This word is not unique in general, since if
 u
0
(A. 1 )u
1
(A. 1 ), then  u
1
(A. 1 )u
0
(A. 1 ) too.
213. Let .  M
n
(/) be given.
(a) Show that there exists a non-zero pair (c. /)

/
2
(

/ the algebraic closure of /) such

that det(c + /) = 0.
(b) Let (c. /) be as above and r be an element of ker(c +/). Show that (
 )rr = 0. Deduce that det(  ) = 0 for every .  M
n
(/).
(c) We assume that  and  commute to each other. Show that det(  )
can be computed as a nested determinant (see exercise 120).
(d) When  = 1
n
, show that det( 1
n
1
n
) = 0 also follows from the Cayley
Hamilton theorem. Hint: Use the previous question.
214. Let  M
m
(/) and  M
n
(/) be given. Prove that
det   = (det )
n
(det )
m
.
Hint: Use again Exercise 120.
121
Likewise, let  M
pq
(/) and  M
rs
(/) be given, with j: = :. Check that 
is a square matrix. Show that rk( ) (minj. )(min:. :), and deduce that
det( ) = 0 when j ,= . The case j = is covered by the identity above.
215. (P. Finsler, J. Milnor.) Let 1 : Sym
n
(R) R be a C
1
-function. Let us dene the
symmetric matrix  by
:
ij
:=
1
c
ij
(0
n
).
Let j

be the smallest eigenvalue of . Prove that

j

= liminf
_
1() 1(0)
Tr
[ 0. Tr 0
_
.
216. Let H be a Hermitian matrix, such that for every 1 HPD
n
, there holds
det(H + 1) det H.
Prove that H is positive semi-denite.
217. If Sym
n
(R), we denote
A
the associated quadratic form. In the sequel, and 1
are too real symmetric matrices.
(a) Assume that there exists a positive denite matrix among the linear combinations
of and 1 (the pencil spanned by and 1). Prove that there exists a 1 GL
n
(R)
such that both 1
T
1 and 1
T
11 are diagonal. Hint: A classical result if 1 itself
is positive denite.
A
(r) =
B
(r) = 0 implies r = 0, and that : 3.
i. Show that 1(1) (ker 1) = 0.
ii. Let () be the determinant of +1. Using a basis of ker 1, completed as a
basis of R
n
, show that the degree of equals the rank of 1. Deduce that there
exists a non-degenerate matrix in this pencil.
(c) We keep the assumption that
A
=
B
= 0 implies r = 0. From the previous
question, we may assume that 1 is non-degenerate.
i. Let us dene
2(r) :=

A
(r) + i
B
(r)
_

A
(r)
2
+
B
(r)
2
C. r ,= 0.
Show that there exists a dierentiable real-valued map r (r) over R
n
0,
such that 2(r) = exp(i(r)) for every r ,= 0 in R
n
.
ii. Let r be a critical point of . Show that
A
(r)1r =
B
(r)r. Show that there
exists such a critical point r
1
. Show that
B
(r
1
) ,= 0.
122
iii. We dene 1
1
:= (1r
1
)

. Show that the restriction of 1 to 1

1
is non-degenerate.
Prove that a critical point r
2
of the restriction of to 1
1
0 again satises

A
(r
2
)1r
2
=
B
(r
2
)r
2
, as well as (r
1
)
T
1r
2
= 0 and
B
(r
2
) ,= 0.
iv. Arguing by induction, construct a sequence r
1
. . . . . r
n
of vectors of R
n
with the
properties that r
j
| 1r
j
, and (r
j
)
T
1r
k
vanishes if and only if , ,= /.
v. Conclusion: There exists a 1 GL
n
(R) such that both 1
T
1 and 1
T
11 are
diagonal.
(d) We are now in the position that both and 1 are diagonal, and still
A
=
B
= 0
implies r = 0. We wish to show that there exists a linear combination of and 1
that is positive denite
i. We argue by contradiction. Suppose that none of the vectors diag + jdiagB
is positive (in the sense of Chapter 5) when (. j) run over R
2
. Show that there
exists a hyperplane in R
n
, containing diag and diagB, but no positive vector.
Hint: Apply HahnBanach.
ii. Deduce that there exists a non-negative, non-zero vector R
n
such that

j

j
c
jj
=

j

j
/
jj
= 0.
iii. Show that this implies = 0.
In conclusion, we have found the equivalence (as long as : 3) of the conditions:
If
A
(r) =
B
(r) = 0, then r = 0,
There exists a positive denite linear combination of
A
and
B
,
and they imply a third one
There exists a basis of R
n
, orthogonal for both
A
and
B
.
(e) Provide a counter-example when : = 2: Find and 1 such that
A
=
B
= 0
implies r = 0, but there does not exist a basis simultaneously orthogonal for and
1. In particular, combinations of and 1 cannot be positive denite.
218. This may be a new proof of Gardings Theorem.
(a) Let 1 : R
+
be a positive, homogeneous function of degree , over a convex
cone of R
N
. We assume that 1 is quasi-concave over ; by this we mean that
at every point r , the restriction of the Hessian D
2
1(r) to ker d1(r) is non-
positive. Prove that 1
1/
is concave over . Hint: Use repeatedly Euler Identity
for homogenous functions.
(b) We now focus to R
N
H
n
(and thus  = :
2
), with 1() = det  and = HPD
n
.
i. If  HPD
n
is diagonal, compute explicitly d1() and D
2
1(). Then check
that 1 is quasi-concave at .
ii. Extend this property to all  HPD
n
, using the diagonalisability through a
unitary conjugation.
iii. Deduce a particular case of Gardings Theorem:  (det )
1/n
is concave
over HPD
n
.
123
Gardings Theorem is that if 1 is a hyperbolic polynomial, homogeneous of degree
:, and is its forward cone, then 1
1/n
is concave over . This may be proved in
full generality with the argument above, together with the fact that the quadratic
form 2 D
2
1(A)2
2
is of signature (1.  1). See an other proof below.
Notice that the rst concavity result given above can be written even for non-smooth
functions 1, thus without invoquing rst- an second-order dierential. Prove that
if G : R
+
is homogeneous of degree one, and if 1 := r [ G(r) 1 is
convex (by homogeneity, this amounts to quasi-concavity), then G is concave.
219. Here is an other proof of the concavity of det
1/n
over HPD
n
.
(a) Given non-negative real numbers c
1
. . . . . c
n
, we denote
k
(c) the /-th elementary
symmetric polynomial, which is a sum of
_
:
/
_
monomials. Prove that

k
(c)
_
:
/
_

n
(c)
k/n
.
Hint: Use the arithmetico-geometric inequality.
(b) Prove that for every 1 HPD
n
, one has
det(1
n
+ 1)
_
1 + (det 1)
1/n
_
n
.
(c) Deduce the inequality
(det(H + 1))
1/n
(det H)
1/n
+ (det 1)
1/n
for every H. 1 HPD
n
. Conclude.
220. We consider the Hermitian norm | |
2
over C
p
and C
q
. We denote by B the unit ball
(the set of linear contractions) in M
pq
(C). Recall that a contraction is a map satisfying
|)(r) )()|
2
< |r |
2
whenever ,= r.
(a) Show that  M
pq
(C) is a contraction if, and only if, ||
2
< 1. Deduce that


is also a contraction.
(b) Let H H
q
and 1 GL
q
(C) be given. Show that 1

H1
1
< 1
q
is equivalent to
H < 1

1.
(c) Given a matrix l U(j. ), written in block form
l =
_
1
C 1
_
.
we dene a map 1 over M
pq
(C) by
1(2) := (2 + 1)(C2 + 1)
1
.
Show that 1 maps B into itself.
124
(d) Show that the set of maps 1 form a group (denoted by ) as l runs over U(j. ),
and that the map l 1 is a group homomorphism.
(e) Show that for every 2 given in B, there exists an 1 as above, such that 1(2) = 0
pq
.
Deduce that the group acts transitively over B.
221. (Loo-Keng Hua.) This exercise uses the previous one, and in particular has the same
notations. We dene the following function over B B:
(\. 2) :=
[ det(1
q
\

2)[
2
det(1
q
\

\) det(1
q
2

2)
.
(a) Of course, (0
pq
. 2) 1 for every contraction 2, with equality only if 2 = 0
pq
.
(b) Show that if l U(j. ) and 1 is dened as above, then
(1(\). 1(2)) = (\. 2).
We say that is invariant under the action of .
(c) Deduce Huas Inequality: For any two contractions \ and 2, one has
det(1
q
\

\) det(1
q
2

2) [ det(1
q
\

2)[
2
.
with equality if and only if \ = 2. Hint: Use transitivity ; then it is enough to
treat the case \ = 0
pq
.
In other words, (\. 2) 1, with equality only if \ = 2. The quantity
1(\. 2) :=
det(1
q
\

2)
det(1
q
\

\)
1/2
det(1
q
2

2)
1/2
.
whose square modulus is (\. 2), is the Bergman kernel of the symmetric domain B.
222. We use the notations of Exercise 146. We assume 1 = C. The exterior algebra
k
1 is
naturally endowed with a Hermitian structure, in which a unitary basis is given by the
vectors c
i
1
c
i
k
with 1 i
1
< < i
k
:, c
1
. . . . . c
n
being the canonical basis of
1 = C
n
.
(a) Prove that
(k)
1
(k)
= (1)
(k)
.
(b) Prove that
_

(k)
_

= (

)
(k)
.
(c) Deduce that if l U
n
, then l
(k)
is unitary too.
(d) Let :
1
. . . . . :
n
denote the singular values of a matrix . Prove that [Tr[ :
1
+
+ :
n
.
(e) Let
1
. . . . .
n
denote the eigenvalues of . Deduce for every 1 / : the inequality
[
k
()[
k
(:).
where
k
is the elementary symmetric polynomial if degree / in : arguments. The
case / = 1 has been established in the previous question. The case / = : is trivial.
Hint: Apply the case / = 1 to
(k)
, and use Exercise 146.
125
(f) Use this result to prove that
[ det(1
n
+ )[ det(1
n
+[[).
where [[ :=

is the non-negative symmetric part in the polar decomposition.

223. (L. Dines.) Let . 1 Sym
n
(R) be given. Show that the range of the map r
(r
T
r. r
T
1r) is a convex subset of R
2
.
Compare with the ToeplitzHausdor Lemma (Exercise 21).
224. Let M
n
(/) be a block triangular matrix. Show that is nilpotent if, and only if, its
diagonal blocks are nilpotent.
225. A subspace \ of M
n
(/) is nilpotent if every element of \ is nilpotent. Let \ be a nilpotent
subspace.
(a) If \ , prove that Tr
2
= 0.
i. If char/ ,= 2, deduce that Tr(1) = 0 for every . 1 \ .
ii. More generally, check that
Tr(1) = (Tr )(Tr 1) +
2
() +
2
(1)
2
( + 1).
and deduce that Tr(1) = 0 for every . 1 \ , even in characteristic 2.
Hereabove,
2
() denotes the second elementary symmetric polynomial in the
eigenvalues of .
(b) Let l denote the subspace of upper triangular matrices, and l
+
that of strictly
upper triangular. Likewise, we denote 1 and 1

for lower triangular matrices. One

has
M
n
(/) = 1

l = 1 l
+
.
Let be the projection over 1

, parallel to l. We denote 1 := \ ker and

1 := (\ ).
i. Show that 1 l
+
.
ii. If  1 and 1 1, prove that Tr(1) = 0. In other words, 1
T
and 1 are
orthogonal subspaces of l
+
, relatively to the form .  := Tr(
T
).
iii. Deduce that
dim1 + dim1 diml
+
.
(c) In conclusion, show that the dimension of a nilpotent subpace of M
n
(/) is not larger
than
:(: 1)
2
.
126
Nota. Every nilpotent subspace of dimension :(: 1),2 is conjugated to l. How-
ever, it is not true that every nilpotent subspace is conjugated to a subspace of l.
For instance, let : = 3: Prove that the space of matrices
_
_
0 0 r
0 0
r 0
_
_
is nilpotent, but is not conjugated to a space of triangular matrices. Hint: the
kernels of these matrices intersect trivially.
226. We use the scalar product over M
n
(C), given by .  = Tr(

). We recall that

the corresponding norm is the SchurFrobenius norm | |
F
. If 1 GL
n
(C), we denote
1 = l[1[ the polar decomposition, with [1[ :=

1 and l U
n
. The Aluthge
transform (1) is dened by
(1) := [1[
1/2
l[1[
1/2
.
(a) Check that (1) is similar to 1.
(b) If 1 is normal, show that (1) = 1.
(c) Show that |(1)|
F
|1|
F
, with equality if, and only if, 1 is normal.
(d) We dene
n
by induction, with
n
(1) := (
n1
(1)).
i. Given 1 GL
n
(C), show that the sequence
_

k
(1)
_
kN
is bounded.
ii. Show that its limit points are normal matrices with the same characteristic
polynomial as 1 (Jung, Ko & Pearcy, or Ando).
iii. Deduce that when 1 has only one eigenvalue j, then the sequence converges
towards j1
n
.
Comment: The sequence does converge for every initial 1 M
n
(C), according to
J. Antezana, E. R. Pujals and D. Strojano.
(e) If 1 is not diagonalizable, show that these limit points are not similar to 1.
227. (After C. de Lellis & L. Szekelyhidi Jr.)
If r R
n
, we denote r r := rr
T
. We also use the standard Euclidian norm. The
purpose of this exercise is to prove that the convex hull of
1 :=
_
(. o) [ R
n
. [[ = 1 and o =
1
:
1
n
_
equals
C =
_
(. o) [ o Sym
n
(R). Tr o = 0 and
1
:
1
n
o
_
.
127
Left: Mmmh ! This Camillo de Lellis is not
the mathematician.
(a) Show that every (. o) in C satises
[[ 1 and o
_
1
1
:
_
1
n
.
(b) Check that 1 C.
(c) Prove that C is a convex compact subset of R
n
Sym
n
(R).
(d) Let (. o) C be given, such that

1
:
1
n
,= o.
i. Show that [[ < 1.
ii. Let j be the largest eigenvalue of o. Show that j 11,:. In case of equality,
show that there exists a unit vector u such that
o = u u
1
:
1
n
.
iii. In the latter case (j = 11,:), show that = u for some (1. 1). Deduce
that (. o) is not an extremal point of C.
iv. We now assume on the contrary that j < 1 1,:. Let  denote the kernel of
o +
1
:
1
n
.
If  ker o

for a symmetric, trace-less o

, show that (. o + co

) C for [c[
small enough.
If dim : 2, deduce again that (. o) is not an extremal point of C.
v. We still assume j < 11,:, and we now treat the case where  is a hyperplane.
Show that there exists a vector . ,= 0 such that
o = + . .
1
:
1
n
.
128
Show that there exists a non-zero pair (. ) R
2
such that, dening u = .
and : = . u+u. +. ., one has (. o) (u. :) C. Deduce that (. o)
is not an extremal point in C.
(e) Deduce that every extremal point of C belongs to 1.
(f) Conclude, with the help of Krein-Milmans Theorem.
(g) Show that in particular, 0
n
is a relative interior point of the convex set C (that is,
an interior point of C as a subset of the ane space spanned by C).
228. We recall that the Pfaan of a 4 4 alternate matrix is c
12
c
34
+ c
13
c
42
+ c
14
c
23
.
Show that Alt
4
(R) GL
4
(R) has two connected components, each one homeomorphic to
o
2
o
2
R
2
, with o
2
the two-dimensional sphere.
229. Let  M
n
(/) be given in block form
 =
_
1
C 1
_
.
where the diagonal blocks have square shape. Recall that if is non-singular, then its
Schur complement is dened as

c
:= 1 C
1
1.
Let us restrict ourselves to the case / = C and to Hermitian matrices. If  is positive
denite, then so is , and the Schur complement is well-dened.
Is instead  is positive semi-denite, show that 1(1) 1(), and thus there exists
a (not necessarily unique) rectangular matrix A such that 1 = A. Show that the
product A

A does not depend on the particular choice of A above. Deduce that the
map
c
extends continuously to the closure of HPD
n
, that is to the cone of positive
semi-denite matrices. Finally, show that det  = det det
c
.
230. After R. B. Bapat, we dene a doubly stochastic :-tuple as an :-uplet = (
1
. . . . .
n
) of
Hermitian semi-positive denite matrices
j
H
+
n
(yes, the same :) with the properties
that
for every , = 1. . . . . :, Tr
j
= 1,
and moreover,

n
j=1

j
= 1
n
.
The set of doubly stochastic :-tuples is denoted by T
n
.
(a) Given as above and \ U
n
, we dene (\ ) M
n
(R) by
(\ )
jk
:=
_

j
_

j
.
where
1
. . . . .
n
are the columns of \ .
Show that (\ ) is a doubly-stochastic matrix.
129
(b) Conversely, let = (
1
. . . . .
n
) be an :-uplet of : : complex matrices. Dene
(\) as above. Show that, if (\) is doubly stochastic for every unitary \ , then
is a doubly stochastic :-uplet.
(c) Check that the subset of T
n
made of those s such that every
j
is diagonal, is
isomorphic to the set of doubly stochastic matrices.
(d) Remark that T
n
is a convex compact subset of (H
n
)
n
. If each
j
has rank one,
prove that is an extremal point of T
n
.
Nota: there exist other extremal points if : 4. However, the determination of all
the extremal points of T
n
is still an open question.
231. (Continuation.) We examine some of the extremal points = (
1
. . . . .
n
) of T
n
.
(a) If =
1
2
(1 + C) with 1. C T
n
, prove that 1(1
j
) and 1(C
j
) are contained in
1(
j
) for every ,.
(b) If is not extremal, show that one may choose 1. C as above, such that 1(1
i
) ,=
1(
i
) for some index i.
(c) In the situation described above, assume that
i
has rank two. Prove that 1
i
= r
i
r

i
and C
i
=
i

i
, where r
i
.
i
is a unitary basis of 1(
i
).
(d) We now assume that every
j
has rank two, and that for every pair , ,= /, the planes
1(
j
) and 1(
k
) are not orthogonal. Prove that all the matrices 1
j
and C
j
have
the forms r
j
r

j
and
j

j
respectively, with r
1
. . . . . r
n
and
1
. . . . .
n
two unitary
bases of C
n
.
(e) We set : = 4 and choose two unitary bases
1
. . . . .
4
and u
1
. . . . . u
4
of C
4
. We
dene

1
=
1
2
(
1

1
+ u
3
u

3
).
2
=
1
2
(
2

2
+ u
4
u

4
).

3
=
1
2
(u
1
u

1
+ u
2
u

2
).
4
=
1
2
(
3

3
+
4

4
).
Check that T
4
. Find a choice of the s and us such that there does not exist
a unitary basis r
1
. . . . . r
4
with r
j
1(
j
). Deduce that is an extremal point.
232. (From M. H. Mehrabi.) Let . 1 M
n
(R) be such that [. 1] is non-singular and verify
the identity

2
+ 1
2
= [. 1].
for some R.
Show that
:=
i
_
1 +
2
is a (2:)-th root of unity.
130
233. We are given three planes 1
1
, 1
2
and 1
3
in the Euclidian space R
3
, of respective equations
.
j
r = 0. We are searching an orthogonal basis
1
.
2
.
3
such that
j
1
j
for each ,.
Prove that such a basis exists if, and only if,
:= (det(.
1
. .
2
. .
3
))
2
4(.
1
.
2
)(.
2
.
3
)(.
3
.
1
)
is non-negative. When is positive, there exist two such bases, up to scaling.
234. Let  M
n
(/) be given. Use Theorem 2.3.1 of the book to calculate the Pfaan of the
alternate matrix
_
0
n


T
0
n
_
.
Warning. Mind that the Pfaan of the same matrix when  = 1
n
is (1)
n(n1)/2
.
235. In a vector space \ of dimension :, we are given : subspaces 1
j
, , = 1. . . . . :. We
examine the question whether there exists a basis
1
. . . . .
n
of \ with
j
1
j
for every
,.
(a) Check the obvious necessary condition: for every index subset J,
(26) dim
_
+
jJ
1
j
_
[J[.
In the sequel, we want to prove that it is also a sucient condition. We shall argue
by induction over :.
(b) Let us assume that the result is true up to the dimension : 1.
From now on we assume that dim\ = : and that the property (26) is fullled.
i. Prove the claim for the given set 1
1
. . . . . 1
n
, in the case where there exists a
index subset J such that
dim
_
+
jJ
1
j
_
= [J[. 1 [J[ : 1.
Hint: Apply the induction hypothesis to both
\ := +
jJ
1
j
and 2 := \,\.
ii. There remains the case where, for every J with 1 [J[ : 1, one has
dim
_
+
jJ
1
j
_
[J[ + 1.
Select a non-zero vector
1
1
1
. Dene 1
j
= (/
1
+ 1
j
),/
1
, a subspace of
\ := \,/
1
. Check that 1
2
. . . . . 1
n
and \ satisfy the assumption (26). Apply
the induction hypothesis.
131
(c) Conclude that a necessary and sucient condition for the existence of a basis

1
. . . . .
n
of \ with
j
1
j
for every , = 1. . . . . :, is that (26) is fullled for
every index subset J.
236. We consider dierential equations over M
n
(C), of the form
(27)
d
dt
= [(t). ].
Hereabove, (t) can be given a priori, or be determined by (t). As usual, the bracket
denotes the commutator.
(a) Show that
d
dt
det  = Tr
_


T
d
dt
_
.
where

 is the adjugate of . Deduce that t det (t) is constant.
(b) Dene 1(t), the solution of the Cauchy problem
d1
dt
= 1. 1(0) = 1
n
.
Show that (t) = 1(t)
1
(0)1(t). In particular, the spectrum of (t) remains
equal to that of (0) ; one speaks of an isospectral ow.
(c) First example: take :=

. Show that t |(t)|

F
(Frobenius norm) is constant.
(d) Second example: take := [

. ]. Show that t |(t)|

F
is monotonous non
increasing. Deduce that the only rest points are the normal matrices.
237. This problem and the following one examine the dimension of the commutant of a given
matrix M
n
(/) when / is algebraically closed. We begin with the equation
(28) 11 = 1 C.
where 1 = J(0; j) and C = J(0 : ) are nilpotent matrices in Jordan form. The unknown
is 1 M
pq
(/). We denote by e
1
. . . . . e
q
and f
1
. . . . . f
p
the canonical bases of /
q
and
/
p
, respectively. We thus have Ce
j
= e
j1
, and likewise with 1 and f
j
.
(a) Show that for every solution 1 of (28), there exist scalars
1
. . . . .
q
such
1 e
j
=
1
f
j
+
2
f
j1
+ +
j
f
1
.
(b) If j < , explain why
1
. . . . .
qp
must vanish.
(c) Deduce that the solution space of (28) has dimension min(j. ).
238. (Continuation.) Let M
n
(/) be given. We denote com() the space of matrices
A M
n
(/) such that A = A.
132
(a) If

) are conjugated, and thus have

the same dimension.
(b) Assuming that / is algebraically closed, we thus restrict to the case where =
diag
1
. . . . .
r
, where
j
has only one eigenvalue
j
and , ,= / implies
j
,=
k
.
We decompose A blockwise accordingly:
A = (A
jk
)
1j,kr
.
Show that A commutes with if, and only if,
j
A
jk
= A
jk

k
for every ,. /. In
particular, , ,= / implies that A
jk
= 0: A is block-diagonal too. Therefore,
dimcom() =
r

j=1
dimcom(
j
).
(c) We are thus left with the special case where has only one eigenvalue , and has
invariant polynomials (A )
m
1
. . . . . (A )
mn
with
:
1
:
n
and :
1
+ + :
n
= :.
i. Show that the commutant of does not depend on , but only on :
1
. . . . . :
n
.
ii. Use the previous exercise to nd
dimcom() =
n1

j=0
(2, + 1):
nj
.
(d) Deduce that for every M
n
(/), one has
dimcom() :.
239. (After C. Hillar & Jiawang Nie.) Let o Sym
n
(Q) be given. We assume that o 0
n
in
the sense of quadratic forms. The purpose of this exercise is to prove that o is a sum of
squares (
j
)
2
with
j
Sym
n
(Q).
Joseph-Louis Lagrange.
Recall that if : = 1, this follows from the La-
granges Theorem: every positive integer is the
sum of four squares of integers.
We recall the key fact that in an Abelian ring, the
product of two sums of squares is a sum of squares.
We denote by j the minimal polynomial of o,
which we write
j(A) = A
s
c
s1
A
s1
+ + (1)
s
c
s
.
133
(a) Check that j has simple roots. In particular, c
0
and c
1
cannot vanish both.
(b) Verify that the coecients c
k
are non-negative. Deduce that c
1
,= 0.
(c) We decompose (1)
s1
j(A) = (A
2
)A :(A
2
). Check that 1 := (o
2
) is a sum
of squares, and that it is invertible. Show then that 1
2
:(o
2
) is a sum of squares.
Conclude.
240. Let . 1 M
n
(C) be given. We assume that there exists a non-singular matrix 1 such
that
1 = 11.

1 = 11

.
Prove that and 1 are unitarily similar: there exists a l U
n
such that l = l1.
Hint: Using the polar decomposition to 1, show that we may assume 1 HPD
n
, and
then 1 diagonal.
This statement is a part of the proof of Specht Theorem: and 1 are unitarily similar
if, and only if, the equality Tr u(.

) = Tr u(1. 1

) holds true for every word u in two

letters. The rest of the proof involves representation theory and is beyond the scope of
our book.
241. (From C. Villani.) Let t 1(t) Sym
n
(R) be a continuous function over [0. 1]. We
denote J
0
(t) and J
1
(t) the matrix-valued solutions of the dierential equation
d
2
J
dt
2
+ 1(t)J = 0.
uniquely determined by the Cauchy data
J
0
(0) = 1
n
. J

0
(0) = 0
n
. J
1
(0) = 0
n
. J

1
(0) = 1
n
.
We wish to prove that o(t) := J
1
(t)
1
J
0
(t) is symmetric, whenever J
1
(t) is non-singular.
(a) Let n
j
(t) (, = 1. 2) be two vector-valued solutions of the ODE
(29) n

+ 1(t)n = 0.
Verify that t n

1
(t). n
2
(t) n

2
(t). n
1
(t) is constant.
(b) For a solution n of (29), show that
J
0
(t)n(0) + J
1
(t)n

(0) = n(t).
(c) For t [0. 1), let us dene the space 1
t
of the solutions of (29) such that n(t) = 0.
Show that it is an :-dimensional vector space. If J
1
(t) is non-singular, verify that
an alternate denition of 1
t
is the equation
n

(0) = o(t)n(0).
(d) Deduce the symmetry of o(t) whenever it is well-dened.
134
242. This is a sequel of Exercise 26, Chapter 4 (see also Exercise 153 in this list). We recall
that denotes the unit sphere of M
2
(R) for the induced norm | |
2
. Also recall that is
the union of the segments [:. :] where : 1 := SO
2
(R) and : o, the set of orthogonal
symmetries. Two distinct segments may intersect only at an extremity.
We construct the join J(1. o) as follows: in the construction above, we replace the
segments [:. :] by the lines that they span. In other words, J(1. o) is the union of the
(ane) lines passing through one rotation and one orthogonal symmetry. Of course,
J(1. o).
(a) Prove that J(1. o) is the algebraic set dened by the equation
||
2
F
= 1 + (det )
2
.
where ||
2
F
:= Tr(
T
) (Frobenius norm).
(b) If J(1. o) O
2
(R), show that belongs to a unique line passing through a
rotation and an orthogonal symmetry (these rotation and symmetry are unique).
(c) Show that for every matrix J(1. o), one has ||
2
1.
(d) Find a dieomorphism from a neighborhood of 1
2
to a neighborhood of 0
2
, which
maps the quadrics of equation (1 + 2)
2
= 421 onto J(1. o).
243. (Continuation.)
(a) Interpret the equation of J(1. o) in terms of the singular values of .
(b) More generally, show that the set USV
n
of matrices  M
n
(R), having : = 1 as
a singular value, is an irreducible algebraic hypersurface.
(c) Show that the unit sphere of (M
n
(R). | |
2
) is contained in USV
n
. Therefore USV
n
is the Zariski closure of the unit sphere.
244. Let . 1 M
2
(/) be given, where the characteristic of / is not 2. Show that [. 1]
2
is a
scalar matrix 1
2
. Deduce the following polynomial identity in M
2
(/):
(30)
_
[. 1]
2
. C

= 0
2
.
Remark. Compared with 1
4
(. 1. C. 1) = 0
2
(Amitsur & Levitzkii Theorem, see Exer-
cise 289), the identity (30) has one less argument, but its degree is one more.
245. In M
n
(C), prove the equivalence between
det  = 0.
and
There exists a matrix M
n
(C) such that . GL
n
(C) for every . C.
135
Hint: Use the rank decomposition (Theorem 6.2.2); show that  is equivalent to a
nilpotent matrix in M
n
(C).
246. Let / be a eld and
1 =
_
1
C 1
_
be an orthogonal matrix, with and 1 square.
Prove that
det 1 = det .
Hint: multiply 1 by
_

T
C
T
0 1
_
.
Extend this result to elements 1 of a group O(j. ).
247. We endow M
n
(C) with the induced norm | |
2
. Let G a subgroup of GL
n
(C) that is
contained in the open ball 1(1
n
; :) for some : < 2.
(a) Show that for every  G, there exists an integer j 1 such that
p
= 1
n
.
Hint: The eigenvalues of elements of G must be of unit modulus and semi-simple
(otherwise G is unbounded); they may not approach 1.
(b) Let . 1 G be s.t. Tr() = Tr(1) for all  G. Prove that = 1.TODO
(c) Deduce that G is a nite group.
(d) On the contrary, let 1 be a rotation in the plane (: = 2) of angle , Q. Prove
that the subgroup spanned by 1 is innite and is contained in 1(1
2
; 2).
248. Let .
1
. . . . . .
n
C have positive real parts. Prove that the Hermitian matrix with
entries
c
jk
:=
1
.
j
+ .
k
is positive denite.
Hint: Look for a Hilbert space 1 and elements )
1
. . . . . )
n
1 such that
c
jk
= )
j
. )
k
.
249. Let : N

be given. We denote 1
m
:
m
the :-th power in M
n
(C). Show that
the dierential of 1
m
at is given by
D1
m
() 1 =
m1

j=0

j
1
m1j
.
Deduce the formula
Dexp() 1 =
_
1
0
c
(1t)A
1c
tA
dt.
136
250. Let ), an entire analytic function, be given.
If 1 = diag(d
1
. . . . . d
n
M
n
(C), we dene a matrix )

(1) M
n
(C) by
)

(1)
jk
=
_
_
_
)

(d
j
). / = ,.
f(d
j
)f(d
k
)
d
j
d
k
. / ,= ,.
where we identify
)(/) )(c)
/ c
:= )

(c).
if / = c.
If ) = 1
m
(notations of the previous exercise), check that
(31) D)(1)1 = )

(1) 1.
where 1 denotes the Hadamard product.
Prove that (31) holds true for every analytic function ) (DaletskiKrein Formula).
Hint: Use polynomial approximation.
251. (B. G. Zaslavsky & B.-S. Tam.) Prove the equivalence of the following properties for real
: : matrices :
Strong PerronFrobenius. The spectral radius is a simple eigenvalue of , the only
one of this modulus ; it is associated with positive left and right eigenvectors.
Eventually positive matrix. There exists an integer / 1 such that
k
0
n
.
252. (M. Goldberg.) In a nite dimensional associative algebra / with a unit, every element
has a unique minimal polynomial (prove it). Actually, associative may be weakened into
power-associative: the powers c
k
are dened in a unique way. You certainly think that if
B is a sub-algebra and c B, then the minimal polynomial of c is the same in / and B.
So try this ....
Here / = M
n
(/). Select a matrix  ,= 1
n
. 0
n
such that 
2
= . What is its minimal
polynomial (it is the one in the usual, matricial, sense) ?
Then consider
B := / =  ; M
n
(/).
Check that B is a subalgebra of /, and that  is the unit element of B. What is its
minimal polynomial in B ?
The explanation of this paradox lies in the notion of subalgebra. The equality of minimal
polynomials is guarranted if the subalgebra and the algebra have the same unit, which is
not the case hereabove.
137
253. (C. A. Bergers theorem, proof by C. Pearcy.) Recall (see Exercise 21) that the numerical
n
(C) is the non-negative real number
u() := max[r

r[ ; r C
n
.
The numerical radius is a norm, which is not submultiplicative. We show that it satises
however the power inequality.
In what follows, we use the real part of a square matrix
Re :=
1
2
( +

).
which is Hermitian and satises
r

(Re)r = 1(r

r). r C
n
.
(a) Show that u() 1 is equivalent to the fact that Re(1
n
.) is semi-denite
positive for every complex number . in the open unit disc.
(b) From now on, we assume that u() 1. If [.[ < 1, verify that 1
n
. is non-
singular. Hint: The numerical radius dominates the spectral one.
(c) If  GL
n
(C) has a non-negative real part, prove that Re (
1
) 0
n
. Deduce
that Re(1
n
.)
1
0
n
whenever [.[ < 1.
(d) Let : 1 be an integer and be a primitive :-th root of unity in C. Check that
the formula
1
1 A
m
=
1
:
m1

k=0
1
1
k
A
can be recast as a polynomial identity.
Deduce that
(1
n
.
m

m
)
1
=
1
:
m1

k=0
(1
n

k
.)
1
.
whenever [.[ < 1.
(e) Deduce from above that
Re(1
n
.
m

m
)
1
0
n
.
whenever [.[ < 1. Going backward, conclude that for every complex number in
the open unit disc, Re(1
n

m
) 0
n
and thus u(
m
) 1.
(f) Finally, prove the power inequality
u(
m
) u()
m
.  M
n
(C). : N.
Nota: A norm which satises the power inequality is called a superstable norm.
It is stable if there exists a nite constant C such that |
m
| C ||
m
for every
M
n
(/) and every : 1. Induced norms are obviously superstable.
138
(g) Let 4 be a given constant. Prove that () := u() is a submultiplicative
norm over M
n
(C) (Goldberg & Tadmor). Use the matrices
=
_
0 1
0 0
_
. 1 =
_
0 0
1 0
_
to show that this becomes false for < 4.
254. (J. Duncan.) We denote r. the usual sesquilinear product in C
n
. To begin with,
let  GL
n
(C) be given. Let us write  = lH = 1l the left and right polar
decomposition. We thus have H =

 and 1 =



.
(a) Prove that l

H =

1 l.
(b) Check that
r. =

H r. l

1 . r. C
n
.
Deduce that
[r. [
2
Hr. r 1. .
(c) More generally, let a rectangular matrix M
nm
(C) be given. Prove the gener-
alized CauchySchwarz inequality
[r. [
2

r. r

. . r. C
n
.
Hint: Use the decompositions
C
m
= ker

1(

). C
n
= ker

1().
then apply the case above to the restriction of from 1(

) to 1().
255. (a) Let . 1 M
n
(C), with normal. If 1 commutes with , prove that 1 commutes
with

. This is B. Fugledes theorem. Hint: Use the spectral theorem for normal
(b) More generally, let
1
.
2
be normal and 1 rectangular. Assume that
1
1 = 1
2
.
Prove that

1
1 = 1

2
(Putnams theorem). Hint: Use the matrix
_

1
0
0
2
_
and apply Fugledes theorem.
(c) Let . 1 M
n
(C) be given. Assume that the span of and 1 is made of normal
matrices. Prove that [. 1] = 0
n
(H. Radjavi & P. Rosenthal). Hint: Use the
matrices C = + 1 and 1 = + i1 to prove [

. 1] = 0
n
, then apply Fugledes
theorem.
139
256. Let / be a eld of characteristic zero. We consider two matrices . 1 M
n
(/) satisfying
[[. 1]. ] = 0
n
.
In other words,
2
1 = 0
n
, where
:  [. ]
is a linear operator over M
n
(/).
(a) Check that is a derivation:
() = () + ().
Therefore obeys a Leibniz formula.
(b) Deduce that for every : 1, one has
(32)
m
(1
m
) = :! (1)
m
.
(c) Deduce from (32) that
m
(1
j
) = 0
n
whenever : ,.
(d) Using the CayleyHamilton Theorem, infer that
n
(1
n
) = 0
n
.
(e) Back to (32), establish that [. 1] is nilpotent (N. Jacobson).
(f) Application: let M
n
(C) satisfy [[.

]. ] = 0
n
. Prove that is normal.
Nota: If / = C or R, one can
deduce from (32) and Proposition
4.4.1 that (1) = 0, which is
the required result.
Left: Gottfried Leibniz.
257. (Thanks to A. Guionnet) Let : :( 1) be two integers. If \ U
n
, we decompose
blockwise
\ =
_
1 Q
1 1
_
.
140
with 1 M
n
(C). Notice that 1 is the matrix of a contraction.
If . C, of unit modulus, is not an eigenvalue of 1

, we dene
\(.) := 1 + .Q(1
nm
.1)
1
1 M
m
(C).
Show that \(.) is unitary.
If : = 2:, build other such rational maps from dense open subsets of U
n
to U
m
.
258. It seems that I have taken for granted the following fact:
If H HPD
n
and / H
n
are given, then the product H/ is diagonalizable
with real eigenvalues. The list of signs (0. ) of the eigenvalues of H/ is the
same as for those of /.
Here is a proof:
(a) Show that H/ is similar to

H /

H.
(b) Use Sylvesters inertia for the Hermitian form associated with /.
259. (L. Mirsky.)
For a Hermitian matrix H with smallest and largest eigenvalues

:(H) :=
+

.
Show that
:(H) = 2 max [r

H[.
where the supremum is taken over the pairs of unit vectors r. C
n
that are orthogonal:
r

= 0.
260. For a given GL
n
(C), we form  :=
1

. Let (. r) be an eigen-pair: r = r.
(a) Show that either [[ = 1, or r

r = 0.
(b) Let us assume that r

r ,= 0. Prove that is a semi-simple eigenvalue.

(c) Find an GL
2
(C) such that the eigenvalues of  are not on the unit circle.
(d) Show that there exists matrices GL
n
(C) without a bilateral polar decomposition
= HQH, where as usual Q U
n
and H HPD
n
.
261. If M
n
(C) is given, we denote :() R
n
+
the vector whose components :
1
:
2

:
n
are the singular values of .
Warning. This exercise involves two norms on M
n
(C), namely the operator norm | |
2
and the SchurFrobenius norm | |
F
.
(a) Using von Neumanns inequality (16), prove that for every matrices . 1 M
n
(C),
we have
|:() :(1)|
2
| 1|
F
.
141
(b) Deduce the following property. For every semi-denite positive Hermitian matrix H,
the projection (with respect to the distance d(. 1) := | 1|
F
) of 1
n
+ H over
the unit ball of | |
2
is 1
n
.
262. Let GL
n
(C) be given, and l1\

be an SVD of . Identify the factors Q and H of

the polar decomposition of , in terms of l. \ and 1.
Let us form the sequence of matrices A
k
with the rule
A
0
= . A
k+1
:=
1
2
(A
k
+ A

k
).
Show that A
k
has the form l1
k
\

with 1
k
diagonal, real and positive. Deduce that
lim
k+
A
k
= Q
and that the convergence is quadratic.
263. Let H H
n
be positive semi-denite.
(a) Prove that H exp(H) is Hermitian and satises the inequality
H exp(H)
1
c
1
n
.
(b) Deduce that the solutions of the ODE
dr
dt
+ Hr = 0
satisfy
|Hr(t)|
2

1
ct
|r(0)|
2
.
Nota. This result extends to evolution equations in Hilbert spaces. For instance,
the solutions of the heat equation (a partial dierential equation)
n
t
=
x
n. r R
d
. t 0
satisfy the inequality
_
R
d
[
x
n(r. t)[
2
dr
1
ct
_
R
d
[n(r. 0)[
2
dr.
264. Let . 1 HPD
n
be given. Prove that for every vector / C
n
, we have
/

(;1)/

1/.
where ;1 is the geometric mean of and 1 (see Exercise 198).
Hint: Use the explicit formula
;1 =
1/2
_

1/2
1
1/2
_
1/2

1/2
.
142
265. (Bataille.) Let . 1 M
n
(/) be such that
2
1 = . We assume moreover that and
1 have the same rank.
(a) Show that ker = ker 1.
(b) Prove 11 = 1.
(c) Show that 1 and have the same rank and deduce /
n
= 1(1) ker .
(d) Finally, prove that 1
2
= 1.
266. Let us denote by A the set of nilpotent matrices in M
n
(/). We also denote G
n
the set
of polynomials j /[A] of degree less than :, such that j(0) = 0 and j

(0) ,= 0. In other
words, j G
n
if and only if
j(A) = c
1
A + + c
n1
A
n1
. c
1
,= 0.
For j. G
n
, we dene j as the unique element : G
n
such that
:(A) j((A)). mod A
n
.
(a) Verify that (G
n
. ) is a group, and that (j. ) j() is a group action over A.
(b) Apply this to prove that if / has characteristic 0, then for every , = 1. 2. . . . and
every  A, the matrix 1
n
+ admits a ,-th-root in 1
n
+A, and only one in this
class.
(c) Denote this ,-th-root by (1
n
+ )
1/j
. If / = C, prove that
lim
j+
(1
n
+ )
1/j
= 1
n
.
267. (M. Cavachi, Amer. Math. Monthly 191 (2009)) We consider a matrix GL
n
(Z)
with the property that for every / = 1. 2. . . ., there exists a matrix
k
M
n
(Z) such that
= (
k
)
k
. Our target is to prove that = 1
n
.
(a) Show that the distance of the spectrum of
k
to the unit circle tends to zero as
/ +. Deduce that the sequence of characteristics polynomials of the
k
s takes
nitely many values.
(b) Prove that there exists two integers (1 ),[/ such that , ,= /, while
j
and
k
have
the same characteristic polynomials. Show that their roots actually belong to the
unit circle, and that they are roots of unity, of degree less than or equal to :.
(c) Show that in the previous question, one may choose , and / such that / is divisible
by :!,. Deduce that the spectrum of
j
reduces to 1.
(d) Verify that, with the terminology of Exercise 266, belongs to 1
n
+A and
j
=
1/j
.
(e) Verify that in the previous question, one may choose , arbitrarily large.
(f) For , as in the previous question and large enough, show that
j
= 1
n
. Conclude.
143
268. For a subgroup G of U
n
, prove the equivalence of the three properties
(P1) G is nite,
(P2) there exists a / 1 such that 
k
= 1
n
for every  G,
(P3) the set Tr [  G is nite.
More precisely,
(a) Prove that (11) =(12) =(13).
(b) Let us assume (13). Choose a basis 
1
. . . . .
r
of the subspace spanned by G,
made of elements of G. Every element  of G writes
=
r

j=1

j

j
.
Express the vector in terms of the vector of components Tr(

j
). Deduce that
can take only nitely many values. Whence (11).
(c) The assumption that G U
n
is crucial. Find an innite subgroup 1 of GL
n
(C), in
which the trace of every element equals :. Thus 1 satises (13) though not (11).
269. This is about the numerical range, dened in Exercise 21. We recall that the numerical
range 1() of a matrix M
n
(C) is a convex compact set, which contains the spectrum
of .
(a) Let C be given, such that () < [[. Show that there exists a conjugate 1
1
1
such that , 1(1
1
1). Hint: Use the Householder Theorem.
(b) Use the case above to show Hildebrants Theorem: the intersection of 1(1
1
1),
as 1 runs over GL
n
(C), is precisely the convex hull of the spectrum of . Hint:
separate this convex hull from an exterior point by a circle.
270. (Suggested by L. Berger.) Here is a purely algebraic way to solve the problem raised in
Exercise 267.
(a) Show that det = 1.
(b) Let j be a prime number. Show that (
k
)
o(p,n)
1
n
. mod j, where o(j. :) is the
order of GL
n
(Z,jZ).
(c) Deduce that 1
n
. mod j. Conclude.
271. Given H HPD
n
(R), prove the formula

n
det H
=
_
C
n
c
z

Hz
d..
144
where we identify C
n
to R
2n
. Hint: One may split . = r + i, H = o + i where
o SPD
n
and is skew-symmetric, and apply Exercise 75 to the integral with respect
to r rst, then to the integral with respect to .
Extend the formula above to non-Hermitian matrices such that H+H

is positive denite.
Hint: Use holomorphy.
272. (R. Bellman.) Let
1
. . . . .
r
M
nm
(C) be strict contractions, meaning that

j
<
1
m
. According to Exercise 221, this implies that for every pair 1 i. , :, the matrix
1
m

j
is non-singular.
The purpose of this exercise is to prove that the Hermitian matrix 1 whose (i. ,) entry is
1
det(1
m

j
)
is positive denite. This generalizes Loo-Keng Huas inequality
(a) Let . C
n
be given. Show that the matrix of entries .

j
. is positive semi-
denite.
(b) With the help of Exercise 21, Chapter 3, prove that the matrix of entries (.

j
.)

is positive semi-denite for every / N. Deduce the same property for the matrix
of entries exp(.

j
.).
(c) Express the matrix 1 as an integral, with the help of Exercise 271. Conclude.
n
(R) be given. We consider the transformation
over M
n
(C) (we warn the reader that we manipulate both elds R and C)
1 1

:= ( + 1)
1
(1 + 1).
In the sequel, we shall use the 2: 2: matrix
1 :=
_
1

_
.
In order to ensure that 1

is dened for every 1 but the elements of a dense open subset

of M
n
(C), we assume a priori that 1 is non-singular.
(a) We are interested only in those transformations that map symmetric matrices 1 onto
symmetric matrices 1

. Show that this is equivalent to the identity

1J1
T
= J. J :=
_
0
n
1
n
1
n
0
n
_
.
for some R

.
145
(b) Prove the identity
1

= ( + 1)
1
(1 1)( + 1)

.
Deduce that if 0, and if the imaginary part of 1 is positive denite, then the
imaginary part of 1

is positive denite. Nota: Complex symmetric matrices with

positive denite imaginary form the Siegel domain 1
n
.
(c) Let 1 M
n
(C) be given. Check that
A :=
_
1
n
1
1
n
1
_
is non-singular if, and only if, the imaginary part of 1 is non-singular. Write in
closed form its inverse.
(d) From now on, we assume that 0. We are looking for xed points (1

= 1) of
the transformation above, especially whether there exists such a xed point in the
Siegel domain. Remark that we may assume = 1, and therefore 1 is a symplectic
matrix.
i. Let 1 Sym
n
(C) be a xed point. We dene A as above. Show that
A1A
1
=
_
0
n
0
n

_
for some  M
n
(C).
ii. We assume now that this xed point is in the Siegel domain. Find a matrix
1 M
n
(C) such that
1 :=
_
1 0
n
0
n
1
_
A
is symplectic. Show that 1 11
1
is symplectic and has the form
_
 0
n
0
n

_
.
iii. Show that  is unitary. Deduce a necessary condition for the existence of a
xed point in the Siegel domain: the eigenvalues of 1 have unit modulii.
Nota: Frobenius showed that this existence is equivalent to i) the eigenvalues of
1 have unit modulii, ii) 1 is diagonalizable. The uniqueness of the xed point
in 1
n
is much more involved and was solved completely by Frobenius. Let us
mention at least that if 1 has simple eigenvalues, then uniqueness holds true.
This existence and uniqueness problem was posed by Kronecker. See a detailed,
albeit non-technical account of this question and its solution, in T. Hawkins
article in Arch. Hist. Exact Sci. (2008) 62:2357.
274. We are interested in matrices  Sym
3
(R) with :
jj
= 1 and [:
ij
[ 1 otherwise. In
particular, there exist angles
k
such that :
ij
= cos
k
whenever i. ,. / = 1. 2. 3.
146
(a) Prove that det  = 0 if, and only if, there exists signs such that

3
2Z.
(b) We give ourselves a non-zero vector r R
3
. We ask whether there exists a matrix
as above (obviously with det  = 0) such that r = 0.
i. Prove the necessary condition that [r
i
[ [r
k
[ + [r
j
[ for every pairwise distinct
i. ,. /.
ii. Prove that this condition is also sucient. Hint: Reduce the problem to the
case where r is non-negative. Then there exists a triangle whose edges have
lengths r
1
. r
2
. r
3
.
275. (F. Holland) Let
1
. . . . .
r
HPD
2
be given. The eigenvalues of
j
are denoted

1
(
j
)
2
(
j
).
Let Q
1
. . . . . Q
r
U
2
be given. Prove the inequality
det
_
r

j=1
Q

j
Q
j
_

_
r

j=1

1
(
j
)
_ _
r

j=1

2
(
j
)
_
.
Hint: Use the Weyl Inequalities for the eigenvalues of

r
j=1
Q

j
Q
j
.
276. (Wiegmann) Let  be a complex, normal matrix.
(a) If the diagonal entries of  are its eigenvalues (with equal multiplicities), show that
 is diagonal. Hint: compute the Frobenius norm of .
(b) More generally, consider a block form of , with diagonal blocks

, / = 1. . . . . :.
Let us assume that the union of the spectra of the diagonal blocks equal the spectrum
of , with the multiplicities of equal eigenvalues summing up to the multiplicity as
an eigenvalue of . Prove that  is block-diagonal.
277. Let two matrices . 1 M
n
(/) be given. We say that (. 1) enjoys the property L if
the eigenvalues of +j1 have the form
j
+j
j
(, = 1. . . . . :) for some xed scalars

j
.
j
. Necessarily, these scalars are the respective eigenvalues of and 1.
(a) If and 1 commute, show that (. 1) enjoys property L.
(b) Let us assume that / has characteristic zero, that is diagonalizable and that (. 1)
enjoys property L. Up to a conjugation (applied simultaneously to and 1), we
may assume that is diagonal, of the form diagc
1
1
m
1
. . . . . c
r
1
mr
with c
1
. . . . . c
r
pairwise distinct. Let us write 1 blockwise, with the diagonal blocks 1

of size
:

.
Prove that (Motzkin & Taussky)
det(A1
n
j1) =
r

=1
det((A c

)1
m

j1

).
147
Hint: Isolate one diagonal block of A1
n
j1. Then compute the determinant
with the help of Schurs complement formula. Then look at its expansion about the
point (c

. 1. 0). One may simplify the analysis by translating c

to zero.
(c) We now assume that / = C, and and 1 are normal matrices. If (. 1) enjoys
property L, prove that and 1 commute (Wiegmann.) Hint: Use Motzkin &
Taussky result, plus Exercise 276.
278. (a) Let u
1
. . . . . u
n1
/
n
and x
0
. . . . . x
n
/
n
be given. Prove that
n

=0
det(u
1
. . . . . u
n1
. x

) det

= 0.
where

A

denotes the matrix whose columns are x

0
. . . . . x
n
/
n
, x

being omitted.
(b) Deduce the following formula for matrices .  M
n
(/):
det() =
n

k=1
det
N
k
det 
k
M
.
where
N
k
denotes the matrix obtained from  by replacing its last column by the
/-th column of , and 
M
k
denotes the matrix obtained from  by replacing its
/-th column by the last column of .
(c) More generally, prove Sylvesters Lemma: given 1 ,
1
< < ,
r
:, then
det() equals the sum of those products det

det 

where

is obtained by
exchanging : columns of  by the columns of  of indices ,
1
. . . . . ,
r
. There are
_
:
:
_
choices of the columns of , and the exchange is made keeping the order
between the columns of , respectively of .
279. Recall that the Hadamard product of two matrices . 1 M
pq
(/) is the matrix 1
M
pq
(/) of entries c
ij
/
ij
with 1 i j and 1 , . If M
n
(/) is given blockwise
=
_
c
11

12

21

22
_
.
and if c
11
is invertible, then the Schur complement
22

21
c
1
11

12
is denoted [c
11
and
we have the formula det = c
11
det([c
11
).
(a) Let . 1 M
n
(/) be given blockwise as above, with c
11
. /
11
/

(and therefore

22
. 1
22
M
n1
(/).) Prove that
( 1)[c
11
/
11
=
22
(1[/
11
) + ([c
11
) 1. 1 :=
1
/
11
1
21
1
12
.
148
(b) From now on, and 1 are positive denite Hermitian matrices. Show that
det( 1) c
11
/
11
det (
22
(1[/
11
)) .
Deduce Oppenheims Inequality:
det( 1)
_
n

i=1
c
ii
_
det 1.
Hint: Argue by induction over :.
(c) In case of equality, prove that 1 is diagonal.
(d) Verify that Oppenheims Inequality is valid when and 1 are only positive semi-
denite.
(e) Deduce that
det( 1) det det 1.
See Exercise 285 for an improvement of Oppenheims Inequality.
280. Let : Sym
m
(R) Sym
n
(R) (or as well : H
m
H
n
) be linear. We say that is
positive if 0
m
implies () 0
n
, and that it is unital if (1
m
) = 1
n
.
(a) Let be positive and unital. If is positive semi-denite (resp. denite), prove
that
_
1
n
()
() (
2
)
_
0
n
or, respectively,
_
(
1
) 1
n
1
n
()
_
0
n
Hint: Use a spectral decomposition of .
(b) Deduce that ()
2
(
2
) or, respectively, ()
1
(
1
).
281. (Exercises 281 to 284 are taken from P. Halmos, Linear algebra. Problem book, MAA
1995.) Let . 1 M
n
(C) be given. We prove here that if , 1 and 1 are normal, then
1 is normal too.
(a) Let us dene C := [1.

]. Expand C

C and verify that Tr C

C = 0. Deduce that
1 commutes with

.
(b) Let QH be the polar decomposition of . Recall that, since is normal, Q and H
commute. Prove that 1 commutes with H. Hint: H is a polynomial in H
2
.
(c) Deduce the formula Q

(1)Q = 1. Conclude.
282. Let . 1 M
n
(R) be unitary similar (in M
n
(C)) to each other:
l U
n
s.t. l = l1.
149
(a) Show that there exists an invertible linear combination o of the real and imaginary
parts of l, such that o = o1 and

o = o1

, simultaneously.
(b) Let QH be the polar decomposition of o. Prove that and 1 are actually orthog-
onally similar:
Q = Q1.
283. Let
1
. . . . .
r
M
n
(/) and j
1
. . . . . j
r
/[A] be given. Prove that there exists a
polynomial j /[A] such that
j(
j
) = j
j
(
j
). , = 1. . . . . :.
Hint: This is a congruence problem in /[A], similar to the Chinese remainder lemma.
284. Prove that for every matrix M
n
(R) with : 2, there exists an invariant plane
R
n
(dim = 2 and ).
285. (S. Fallat & C. Johnson.) Let and 1 be ::, Hermitian positive semi-denite matrices.
According to Exercise 279, both
_
n

i=1
c
ii
_
det 1 and
_
n

i=1
/
ii
_
det
det 1.
Thanks to the Hadamard inequality, they also have the lower bound det det 1. In order
to nd a more accurate inequality than Oppenheims, as well as a symmetric one, it is
thus interesting to compare
det 1 + det det 1 vs
_
n

i=1
c
ii
_
det 1 +
_
n

i=1
/
ii
_
det .
Using induction over the size :, we shall indeed prove that the latter is less than or equal
to the former.
(a) If either : = 2, or 1 is diagonal, or
1 =
_
_
_
1 1
.
.
.
.
.
.
1 1
_
_
_
.
show that actually
det 1 + det det 1 =
_
n

i=1
c
ii
_
det 1 +
_
n

i=1
/
ii
_
det .
150
(b) We turn to the general case. To begin with, it is enough to prove the inequality for
positive denite matrices . 1.
In the sequel, and 1 will thus be positive denite.
(c) We decompose blockwise
=
_
c
11
r

_
. 1 =
_
/
11

_
.
Let 1 := e
1
e
T
1
be the matrix whose only non-zero entry is )
11
= 1. Prove that

:=
det
det

1
is positive semi-denite (actually, it is singular). Hint: Use Schurs complement
formula.
(d) Apply the Oppenheim inequality to estimate det

1.
(e) Using Exercise 2, deduce the inequality
det 1
_
n

i=1
c
ii
_
det 1 +
det
det

_
/
11
det

_
n

i=2
c
ii
_
det 1
_
.
(f) Apply the induction hypothesis. Deduce that
det 1
_
n

i=1
c
ii
_
det 1 +
_
n

i=1
/
ii
_
det det det 1
+
det
det

(det 1 /
11
det 1

)
_
det

i=2
c
ii
_
.
Conclude.
286. (S. Sedov.) Let r be an indeterminate. For : 1, let us dene the snail matrix o
n
(r) by
o
n
(r) :=
_
_
_
_
_
1 r r
n1
r
4n5
r
n
.
.
.
.
.
.
r
3n3
r
2n2
_
_
_
_
_
.
where the powers 1, r, . . . , r
n
2
1
are arranged following an inward spiral, clockwise.
(a) Prove that for : 3,
det o
n
(r) = r
4n
2
9n+12
_
1 r
4n6
_ _
1 r
4n10
_
det o
n2
(r).
Hint: Make a combination of the rst and second rows. Develop. Then make a
combination of the last two rows.
151
(b) Deduce the formula
det o
n
(r) = (1)
(n1)(n2)/2

f(n)
n2

k=0
_
1 r
4k+2
_
.
where the exponent is given by
)(:) =
1
3
(2:
3
6:
2
+ 13: 6).
287. Let j 2 be a prime number. We recall that F
p
denotes the eld Z,jZ. Let M
n
(F
p
)
be given. Prove that is diagonalizable within M
n
(F
p
) if, and only if
p
= .
Hint: The polynomial A
p
A vanishes identically over F
p
and its roots are simple.
288. Let R be a commutative ring containing the rationals (was it a eld, it would be of
characteristic zero), and let M
n
(R) be given. Let us assume that Tr() = Tr(
2
) =
= Tr(
n
) = 0. Prove that
n
= 0
n
.
Hint: Begin with the
case where R is a eld
and use the Newtons
sums.
Left: Isaac Newton.
289. We present the proof by S. Rosset of the AmitsurLevitzki Theorem (for users of the 2nd
edition, it is the object of Section 4.4).
(a) We need the concept of exterior algebra (see Exercise 146). Let c
1
. . . . . c
2n
be a
basis of /
2n
. Then the monomials c
j
1
c
jr
with ,
1
< < ,
r
(the sequence
may be empty) form a basis of the exterior algebra (/
2n
). This is an associative
algebra, with the property that for two vectors c. ) /
2n
, one has c ) = ) c.
We denote by R the subalgebra spanned by the 2-forms c
i
c
j
. Check that R is a
commutative sub-algebra.
(b) If
1
. . . . .
2n
M
n
(/), let us dene
:=
1
c
1
+ +
2n
c
2n
M
n
(R).
i. Show that for every / 1,

i
1
<<i

(
i
1
. . . . .
i

)c
i
1
c
i

.
152
where the standard polynomial o

in non-commutative indeterminates A
1
. . . . . A

is dened by
o

(A
1
. . . . . A

) :=

c()A
(1)
A
()
.
Hereabove, the sum runs over the permutations of 1. . . . . /, and c() denotes
the signature of .
ii. When / is even, show that Tr o

(1
1
. . . . . 1

) = 0, for every 1
1
. . . . . 1

M
n
(/).
iii. If / has characteristic zero, deduce that
2n
= 0
n
. Hint: Use the previous
exercise.
(c) Whence the Theorem of Amitsur & Levitzki: for every
1
. . . . .
2n
M
n
(/), one
has
o
2n
(
1
. . . . .
2n
) = 0
n
.
Hint: First assume that / has characteristic zero. Then use the fact that o
2n
is a
polynomial with integer coecients.
(d) Prove that o
2n1
(
1
. . . . .
2n1
) does not vanish identically over M
n
(/). Hint:
Specialize with the matrices 1
11
. 1
12
. . . . . 1
1n
. 1
21
. . . . . 1
n1
, where 1
m
is the matrix
whose (i. ,)-entry is one if i = : and , = /, and zero otherwise.
290. Prove the following formula for complex matrices:
log det(1
n
+ .) =

k=0
(1)
k+1
/
Tr(
k
).
k
.
Hint: Use an analogous formula for log(1 + c.).
291. Let . 1 H
n
be such that
det(1
n
+ r + 1) det(1
n
+ r) det(1
n
+ 1). r. R.
(a) Show that for every / N,
Tr((r + 1)
k
) r
k
Tr(
k
) +
k
Tr(1
k
).
Hint: Use Exercise 290.
(b) Infer
2 Tr(11) + 4 Tr(
2
1
2
) = 0.
(c) Deduce that 1 = 0. This is the theorem of Craig & Sakamoto. Hint: Set
A := 1. Use the fact that (A + A

)
2
+ 2A

A is semi-positive denite.
292. We recall that the function H (H) := log det H is convex over HPD
n
. We extend
to the whole of H
n
by posing (H) = + otherwise. This extension preserves the
convexity of . We wish to compute the Legendre transform

(1) := sup
HHn
Tr(H1) (H).
153
(a) Check that

(l

1l) =

(b) Show that

In particular,

(1) is innite, unless 1 is negative denite. Hint: Use diagonal

matrices only.
(c) Show that, for every positive denite Hermitian matrices H and H

, one has
log det H + log det H

+ : Tr(HH

).
Hint: HH

is diagonalizable with positive real eigenvalues.

(d) Conclude that

(1) : + (1).
(e) Let (H) := (H) :,2. Verify that

(H) = (H). Do you know any other

convex function on a real space having this property ?
293. Let H
n
be given. Show, by an explicit construction, that the set of matrices H H
n
satisfying H 0
n
and H admits a least element, denoted by
+
:
H H
n
. (H 0
n
and H ) (H
+
).
Let 1 be an other Hermitian matrix. Deduce that the set of matrices H H
n
satisfying
H 1 and H admits a least element. We denote it by 1 and call it the
supremum of and 1. We dene the inmum by 1 := (() (1)).
Prove that
1 + 1 = + 1.
294. Let J M
2n
(R) be the standard skew-symmetric matrix:
J =
_
0
n
1
n
1
n
0
n
_
.
Let o Sym
2n
(R) be given. We assume that dimker o = 1.
(a) Show that 0 is an eigenvalue of Jo, geometrically simple, but not algebraically.
(b) We assume moreover that there exists a vector r ,= 0 in R
2n
form
T
o, restricted to r. Jr

is positive denite. Prove that the eigenvalues

of Jo are purely imaginary. Hint: Use Exercise 258 in an appropriate way.
(c) In the previous question, o has a zero eigenvalue and may have a negative one. On
the contrary, assume that o has one negative eigenvalue and is invertible. Show that
Jo has a pair of real, opposite eigenvalues. Hint: What is the sign of det(Jo) ?
295. Let M
n
(C) be given. Denoting its minimal polynomial by
A
, let us dene a
dierential operator
1
A
:=
A
_
d
dt
_
.
The degree of
A
is denoted by :.
154
(a) Prove that there exists functions )
j
(t) for , = 0. . . . . : 1, such that
exp(t) = )
0
(t)1
n
+ )
1
(t) + + )
r1
(t)
r1
. t R.
(b) Prove that t )
j
(t) is (

. Hint: This requires proving a uniqueness property.

(c) Applying 1
A
to the above identity, show that these functions satisfy the dierential
equation
1
A
)
j
= 0.
Deduce that
)
j
=

c
t
j
j
(t).
where the

s are the distinct eigenvalues and j

j
are polynomials.
(d) Determine the initial conditions for each of the )
j
s. Hint: Use the series dening
the exponential.
296. (After Krishnapur.) Let M
nm
(C) be given. We assume that : : and denote the
columns of by C
1
. . . . . C
m
, that is C
j
:= e
j
with e
1
. . . . . e
m
the canonical basis of
C
m
. Let also
1
. . . . .
m
be the singular values of .
Besides, we dene H
i
the subspace of C
n
spanned the C
j
s for , ,= i. At last, we denote
by d
i
the distance of C
i
to H
i
.
(a) To begin with, we restrict to the case where has full rank: rk = :. Check that

is non-singular.
(b) Let us dene the vector \
j
:= (

)
1
e
j
. Show that \
j
C
i
=
j
i
for all i. , =
1. . . . . :.
(c) Let us decompose \
j
=
j1
e
1
+ +
jn
e
n
. Show that \
j
is orthogonal to H
j
and
that its norm equals
jj
d
j
.
(d) Deduce the identity
jj
d
2
j
= 1.
(e) Prove the Negative second moment identity

2
j
=

j
d
2
j
.
Hint: Compute the trace of (

)
1
.
(f) What do you think of the case where rk < : ?
155
297. Here is another proof of Fugledes Theorem (see Exercise 255), due to von Neumann.
John von Neumann.
(a) Given . 1. C. 1 M
n
(C), check the iden-
tity
Tr([. 1][C. 1]+[. C][1. 1]+[. 1][1. C]) = 0.
(b) Deduce that if [. 1] = 0
n
and if either or
1 is normal, then [. 1

] = 0
n
.
298. (After J. von Neumann.) An embedding from M
m
(C) to M
n
(C) is an algebra homomor-
phism ) with the additional properties that )(1
m
) = 1
n
and )(

) = )()

for every
M
m
(C).
(a) Prove that ) sends GL
m
(C) into GL
n
(C).
(b) Deduce that the spectrum of )() equals that of .
(c) Likewise, prove that the minimal polynomial of )() divides that of . In particular,
if is semi-simple, then so is )().
(d) If )() = 0
n
and is Hermitian, deduce that = 0
m
. Use this property to prove
that if )() = 0
n
, then  = 0
m
(injectivity). Deduce that : :.
(e) Let 1 M
m
(C) be a unitary projector: 1
2
= 1 = 1

. Prove that )(1) is a unitary

projector.
(f) Let r C
m
be a unit vector. Thus )(rr

) is a unitary projector, whose rank is

denoted by /(r). Prove that / is independent of r. Hint: An embedding preserves
conjugacy.
(g) Show that : divides :. Hint: Decompose 1
m
as the sum of unitary projectors rr

.
(h) Conversely, let us assume that : = :j. Prove that there exists an embedding from
M
m
(C) to M
n
(C). Hint: Use the tensor product of matrices.
299. Let M
n
(C) be a nilpotent matrix of order two:
2
= 0
n
. This exercise uses the
standard operator norm | |
2
.
(a) Using standard properties of this norm, verify that ||
2
2
|

+ 

|
2
for
every  M
n
(C).
(b) When / is a positive integer, compute (

)
k
in close form. Deduce that
|

|
2
2
1/k
||
2
2
.
156
(c) Passing to the limit as / +, prove that
(33) ||
2
= |

|
1/2
2
.
300. Let 1 M
n
(C) be a contraction in the operator norm: |1|
2
1. Let us form the matrix
o H
nN
:
o :=
_
_
_
_
_
1
n
1

1
N
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 1

1
N
1 1
n
_
_
_
_
_
.
We also dene the matrix 1 M
nN
(C) by
1 :=
_
_
_
_
_
_
_
_
1
n
0
n
0
n
1
.
.
.
.
.
.
0
n
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
n
0
n
1 1
n
_
_
_
_
_
_
_
_
.
Express o in terms of 1. If (1
n
1) = r, show that
r

or = ||
2
2
|1|
2
2
.
Deduce that o is positive semi-denite.
301. Let H
1
. . . . . H
r
H
n
be positive semi-denite matrices such that
r

j=1
H
j
= 1
n
.
Let also .
1
. . . . . .
r
C be given in the unit disc: [.
j
[ 1.
Prove that _
_
_
_
_
r

j=1
.
j
H
j
_
_
_
_
_
2
1.
Hint: Factorize
_
_
_

r
j=1
.
j
H
j
0
n

0
n
0
n

.
.
.
.
.
.
_
_
_
=

diag.
1
1
n
. . . . . .
r
1
n

with  appropriate.
157
302. (Sz. Nagy.) If  M
n
(C) is similar to a unitary matrix, show that
(34) sup
kZ
|
k
|
2
< .
Conversely, let  GL
n
(C) satisfy (34). Show that the eigenvalues of  belong to the
unit circle and are semi-simple. Deduce that  is similar to a unitary matrix.
303. Let
1
. . . . . 
r
M
n
(C) satisfy the canonical anticommutation relations

i

j
+
j

i
= 0
n
.

i

j
+

j

i
=
j
i
1
n
. 1 i. , :.
In particular, each
i
is nilpotent of order two.
Prove the identity
_
_
_
_
_
r

j=1
.
j

j
_
_
_
_
_
2
= |.|
2
. . C
r
.
Hint: Use Exercise 299.
304. (S. Joshi & S. Boyd.) Given M
n
(C) and a (non trivial) linear subspace
2
\ of C
n
,
we dene
G([\ ) := sup |r| [ r \. |r| = 1 . H([\ ) := inf |r| [ r \. |r| = 1 .
where we use the canonical Hermitian norm. Obviously, we have H([\ ) G([\ ),
thus

V
() :=
G([\ )
H([\ )
1.
Notice that when is singular,
V
() may be innite.
Given an integer / = 1. . . . . :, we wish to compute the number

k
() := inf
V
() [ dim\ = / .
(a) What is
V
() when \ = C
n
?
(b) Show that
k
is unitary invariant: if 1 = l\ with l and \ unitary, then
k
(1) =

k
(). Deduce that
k
() depends only upon the singular values of .
(c) From now on, we assume that is non-singular. From the previous question, we
may assume that is diagonal and positive. We denote (0 <)
n

1
its
singular values, here its diagonal entries. Show rst that
G([\ )
nk+1
. H([\ )
k
.
Deduce that
(35)
k
() max
_

nk+1

k
. 1
_
.
2
This exercise is unchanged when replacing the scalar eld by R.
158
(d) We wish to show that (35) is actually an equality. Let c
1
. . . . . c
n
be the canonical
basis of C
n
. Show that given i , and [
i
.
j
], there exists a unit vector
Span(c
i
. c
j
) such that || = .
(e) Let  be least integer larger than or equal to
n+1
2
, that is the integral part of 1+:,2.
Show that each plane Span(c
N
. c
N+
) with / = 1. . . . . : contains a unit vector

such that |

| =
N
. Deduce that
nN+1
() = 1. In other words,
k
() = 1
for every / :  + 1.
Nota: This can be recast as follows. Every ellipsoid of dimension : 1 contains a
sphere of dimension d, the largest integer strictly less than :,2.
(f) Assume on the contrary that / :+1. Using the same trick as above, construct
a linear space \ of dimension / such that
W
() =
nk+1
,
k
. Conclude.
305. Among the class of Hessenberg matrices, we distinguish the unit ones, which have 1s
below the diagonal:
=
_
_
_
_
_
_
_
_

1
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1
_
_
_
_
_
_
_
_
.
(a) Let  M
n
(/) be a unit Hessenberg matrix. We denote by
k
the submatrix
obtained by retaining the rst / rows and columns. For instance, 
n
=  and

1
= (:
11
). We set 1
k
the characteristic polynomial of
k
.
Show that (B. Kostant & N. Wallach, B. Parlett & G. Strang)
1
n
(A) = (A :
nn
)1
n1
(A) :
n1,n
1
n2
(A) :
2n
1
1
(A) :
1n
.
(b) Let Q
1
. . . . . Q
n
/[A] be monic polynomials, with deg Q
k
= /. Show that there
exists one and only one unit Hessenberg matrix  such that, for every / = 1. . . . . :,
the characteristic polynomial of
k
equals Q
k
. Hint: Argue by induction over :.
Nota: The list of roots of the polynomials 1
1
. . . . . 1
n
are called the Ritz values of .
306. (Partial converse of Exercise 21.) Let and 1 be 2 2 complex matrices, which have
the same spectrum. We assume in addition that
det[

. ] = det[1

. 1].
Prove that and 1 are unitarily similar. Hint : Prove that they both are unitarily
similar to the same triangular matrix.
Deduce that two matrices in M
2
(C) are unitary similar if, and only if they have the same
numerical range.
159
307. Let 1 M
n
(C) be a projection. Let us dene
:= 1(1

1 + (1
n
1)

(1
n
1))
1
1

.
Prove that vanishes over (1(1))

. Then prove that r = r for every r 1(1). In

other words, is the orthogonal projection onto 1(1). Hint: If r 1(1), then r = 1r,
thus you can nd the solution of
(1

1 + (1
n
1)

(1
n
1)) = 1

r.
308. Compute the conjugate of the convex function over H
n
H
_
(det H)
1/n
. if H 0
n
.
+. otherwise.
Hint: Use Exercise 209. Remark that the conjugate of a positively homogeneous function
of degree one is the characteristic function of some convex set.
309. (a) Let M
n
(C) and C be given. Show that if is not in the numerical range
\(), then 1
n
is invertible, and its inverse veries
|(1
n
)
1
|
2

1
dist(. \())
.
(b) Conversely, let us consider compact subsets A of the complex plane, such that
(36) |(1
n
)
1
|
2

1
dist(. A)
. , A.
Obviously, such an A is contained in the resolvant set C Sp(). If c 0 tends to
zero, show that
|1
n
+ c|
2
= 1 + c sup1. [ . \() + C(c
2
).
Deduce that
sup1. [ . \() sup1. [ . A.
Finally, prove that \() is the convex hull of A. In particular, if = J(2; 0) is
a 2 2 Jordan block, then A contains the circle C
1/2
. When is normal, Sp()
satises (36).
310. Let / be a eld of characteristic zero. We consider a matrix M
n
(/). If A M
n
(/),
we dene the linear form and the linear map

X
() := Tr(A). ad
X
() = A A.
(a) We assume that is nilpotent.
160
i. If  = , show that  is nilpotent.
A
ker
A
.
iii. Deduce that there exists a matrix 1 such that
A
=
B
A
.
iv. Show that = 1 1.
(b) Conversely, we assume instead that there exists a 1 M
n
(/) such that = 1
1.
i. Verify that for / N, 1
k

k
1 = /
k
.
ii. Deduce that is nilpotent. Hint: ad
B
has nitely many eigenvalues.
(c) If = J(0; :) is the nilpotent Jordan block of order :, nd a diagonal 1 such that
[1. J] = J.
311. (From E. A. Herman.) Let M
n
(C) be such that
2
= 0
n
. We denote its rank, kernel
and range by :,  and 1.
(a) By computing |( +

)r|
2
2
, show that ker( +

) =  1

.
(b) Verify that 1  and 

. Deduce that  1

is of dimension : 2: and
that the rank of
1
2
( +

) (the real part of ) is 2:.

(c) Show that  is an isotropic subspace for the Hermitian form r r

( +

)r,
contained in 1(+

of +

are both equal to :.

(d) Example: Take : = 2: and
:=
_
0
r
1
0
r
0
r
_
. 1 GL
r
(C).
Find the equations of the stable and the unstable subspaces of +

. Hint: the
formula involves

11

.
312. Let 1 be a hyperplane in H
n
. We use the scalar product . 1 := Tr(1). Prove the
equivalence of the following properties.
Every nonzero matrix 1 1 has at least one positive and one negative eigenvalues.
1

is spanned by a positive denite matrix.

313. (After von Neumann, Halperin, Aronszajn, Kayalar & Weinert.) We equip C
n
with the
standard scalar product, and
n
(C) with the induced norm. Let 
1
and
2
be two
linear subspaces, and 1
1
, 1
2
the corresponding orthogonal projections. We recall that
1

j
= 1
j
= 1
2
j
. We denote  :=
1

2
and 1 the orthogonal projection onto .
Finally, we set Q := 1
n
1 and
j
:= 1
j
Q. Our goal is two prove
lim
m+
(1
2
1
1
)
m
= 1
(von NeumannHalperin Theorem), and to give a precise error bound.
161
(a) Show that 1
j
1 = 1 and 11
j
= 1.
(b) Deduce that
j
= 1
j
1. Verify that
j
is an orthogonal projection.
(c) Show that (
2

1
)
m
= (1
2
1
1
)
m
1.
(d) Deduce that
|(1
2
1
1
)
m
1|
2
= |(
1

1
)
2n1
| = |
1

1
|
2n1
.
(e) Let r C
n
be given. If
1
r =
2
r = r, show that r = 0. Deduce that |
1

1
| < 1
and establish the von NeumannHalperin Theorem.
314.
A Hadamard matrix is a matrix  M
n
(Z)
whose entries are 1s and such that
T
 = :1
n
.
The latter means that
1

n
is orthogonal.
We construct inductively a sequence of matrices
H
m
M
2
m(Z) by
H
0
= (1). H
m+1
=
_
H
m
H
m
H
m
H
m
_
.
Verify that each H
m
Nota: It is unknown whether there exists or not
a Hadamard whose size is not a power of 2.
315. (M. Rosset and S. Rosset.) Let be a Principal Ideal Domain. We dene the vector
product A 1 in
3
with the same formul as in R
3
. For instance, the rst coordinate
is r
2

3
r
3

2
.
(a) We admit for a minute that the map : (A. 1 ) A 1 is surjective from
3

3
into
3
. Prove that every matrix  M
2
() with zero trace is a commutator
1C C1 with 1. C M
2
().
(b) We now prove the surjectivity of . Let 2 =
_
_
c
/
c
_
_

3
be given. Show that there
exists A =
_
_
r

c
_
_
such that cr+/ +c. = 0 and gcdr. . . = 1. Then Bezout gives
a vector l =
_
_
n

u
_
_
such that nr + +u. = 1. Set 1 = 2 l. Then 2 = A 1 .
316. Let / be a eld, l1
n
(/) be the set of upper triangular matrices with 1s along the diagonal.
It is called the unipotent group.
(a) Prove that l1
n
(/) is a subgroup of GL
n
(/).
162
(b) If G is a group, 1(G) is the group generated by the commutators rr
1

1
. It is
a normal subgroup of G. Show that 1(l1
n
(/)) consists of the matrices such that
:
i,i+1
= 0 for every i = 1. . . . . : 1.
(c) Let G
0
= l
n
(/) and G
1
= 1(l1
n
(/)). We dene G
r
by induction; G
r+1
is the group
generated by the commutators where r G
0
and G
r
. Describe G
r
and verify
that G
n
= 1
n
. One says that the group l1
n
(/) is a nilpotent.
317. (Follow-up of the previous exercise. Thanks to W. Thurston.) We take / = Z,3Z and
: = 3.
(a) Show that every element  ,= 1
3
of l1
3
(Z,3Z) is of order 3.
(b) What is the order of l1
3
(Z,3Z) ?
(c) Find an abelian group of order 27, with 26 elements of order 3.
(d) Deduce that the number of elements of each order does not characterizes a group in
a unique manner.
318. Let j 3 be a prime number.  GL
n
(Z) be of nite order (
r
= 1
n
for some : 1),
and such that  1
n
mod j.
(a) Prove that  = 1
n
. Hint: If  ,= 1
n
, write  = 1
n
+ j

with not divisible b

j. Likewise, : = j

/ with 0 and / not divisble by j. Verify that for each / 2,

j
k
_
:
/
_
is divisible by j
++1
. Deduce that 
r
1
n
+ /j
+
mod j
++1
. Conclude.
(b) Let G be a nite subgroup of GL
n
(Z). Deduce that the reduction mod j is injective
over G. Yet, this reduction is not injective over GL
n
(Z).
(c) The result is false if j = 2: nd a matrix 1
n
+ 2 of nite order, with ,= 0
n
.
319. (After C. S. Ballantine.) We prove that every Hermitian matrix H with strictly positive
trace can be written as H = 1 + 1 with . 1 HPD
n
.
(a) We rst treat the case where the diagonal entries /
jj
are strictly positive. Prove that
such a pair (. 1) exists with diagonal. Hint: Induction over :. Choose c
n
0
small enough.
(b) Conclude, with the help of Exercise 131.
(c) Conversely, if . 1 HPD
n
are given, prove that Tr(1 + 1) 0.
320. Let o be a nite set of cardinal :. Let 1
1
. . . . . 1
m
be subsets of o with the properties
that
every 1
j
has an odd cardinal,
163
for every i ,= ,, 1
i
1
j
has an even cardinal.
(a) Let us form the matrix with : columns (indexed by the elements of o) and :
rows, whose entry c
jx
equals 1 if r 1
j
, and 0 otherwise. Prove that
T
= 1
m
.
Hint: Yes, 1
m
! Leave some room to your imagination.
(b) Deduce : :.
321. Let . 1 M
n
(C) satisfy the relation [. 1] = . Let us dene the following matrix-
valued functions of t R:
A(t) := 1c
tA
. 1 (t) = c
tB
. 2(t) = c
t(A+B)
c
tA
c
tB
.
(a) Find a dierential equation for A. Deduce that
[1. c
tA
] = tc
tA
.
(b) Find a dierential equation for 1 . Deduce that
c
tB
= c
t
c
tB
.
(c) Find a dierential equation for 2. Deduce that
c
t(A+B)
= c
(t)A
c
tB
c
tA
. (t) := 1 (t + 1)c
t
.
Remarks.
From Exercise 256, we know that is nilpotent.
This result can be used in order to establish an explicit formula for the semi-
group generated by the FokkerPlanck equation
t
) =
v
) +
v
).
322. Let M
n
(R). We denote 1 the vector whose coordinates are ones. Prove the equiva-
lence of the following properties.
The semi-group (
t
:= c
tA
)
t0
is Markovian, meaning that 
t
is stochastic (
t
0
n
and 
t
1 = 1).
The o-diagonal entries of are non-positive, and 1 = 0.
For every A R
n
, we have
(A A) 2A (A).
where we use the Hadamard product, (A 1 )
j
= r
j

j
.
323. (S. Boyd & L. Vandenberghe). The spectral radius of a non-negative matrix M
n
(R)
is an eigenvalue (PerronFrobenius Theorem), which we denote
pf
().
164
(a) When 0
n
, prove that
log
pf
() = lim
k+
1
/
log 1
T

k
1.
where 1 is the vector whose entries are all ones.
(b) Let . 1 be positive matrices and let us dene C by c
ij
:=
_
c
ij
/
ij
.
i. Show that for every : 1,
(C
m
)
ij

_
(
m
)
ij
(1
m
)
ij
.
ii. Deduce that
pf
(C)
_

pf
()
pf
(1) .
(c) More generally, show that log
pf
() is a convex function of the variables log c
ij
.
Nota: This is not what is usually called log-convexity.
324. Given 1 M
n
(C), assume that the only subspaces 1 invariant under both 1 and 1

(that is 11 1 and 1

1 1) are C
n
and 0. Let us denote by 1 the sub-algebra of
M
n
(C) spanned by 1 and 1

, i.e. the smallest algebra containing 1 and 1

.
(a) Verify that ker 1 ker 1

= 0.
(b) Construct a matrix H 1 that is Hermitian positive denite.
(c) Deduce that 1
n
1.
(d) We show now that 1 does not admit a proper two-sided ideal J (we say that the
unit algebra 1 is simple).
i. Suppose that a two-sided ideal J is proper. Choose  in J. Prove that
(Tr )1
n
J. Hint: use CayleyHamilton Theorem.
Deduce that Tr 0 over J.
ii. If  J, verify that 

J. Deduce that  = 0
n
and conclude. Hint:

1.
Comment. This can be used to prove the following statement, which interpolates the
unitary diagonalization of normal matrices and the AmitsurLevitski theorem that the
standard noncommutative polynomial o
2n
vanishes over M
n
(/).
Let us say that a matrix M
n
(C) is :-normal if the standard polynomial in 2:
non-commuting variables o
2r
vanishes identically over the sub-algebra spanned
by and

. In particular, every is :-normal, whereas is 1-normal if and

only if it is normal.
Then is :-normal if and only if there exists a unitary matrix l such that
l

l is block-diagonal with diagonal blocks of size :: with : :.

The reader will prove easily that a matrix that is unitarily similar to such a block-diagonal
matrix is :-normal.
165
325. We consider here the linear equation  = 1 in  M
nm
(/), where M
n
(/)
and 1 M
m
(/) are given. The solution set is a vector space denoted o(. 1).
(a) If 1 /[A], verify that 1() = 1(1). If the spectra of and 1 are disjoint,
deduce that  = 0.
(b) When = J(0; :) and 1 = J(0; :), compute the solutions, and verify that the
dimension of o(. 1) is min:. :.
(c) If is conjugate to

and 1 conjugate to 1

, prove that o(

. 1

) is obtained from
o(. 1) by applying an equivalence. In particular, their dimensions are equal.
(d) In this question, we assume that / is algebraically closed
i. Let
1
. . . . .

be the union of the spectra of and 1. If i /, we denote

(A
i
)

ij
the elementary divisors of , and (A
i
)

ik
those of 1. Using a
canonical form

and 1

, prove that the dimension of o(. 1) equals the sum

of the numbers

i
:=

j,k
min
ij
.
ik
.
ii. Deduce that the dimension of the solution set of the matrix equation  = 1
equals

deg[g.c.d.(j
i
.
j
)].
where j
1
. . . . . j
n
are the invariant factors of and
1
. . . . .
m
are those of 1.
(e) Show that the result above persists when / is an arbitrary eld, not necessarily
algebraically closed. This is the CecioniFrobenius Theorem. Hint: o(. 1) is
dened by a linear system in /
nm
. Its dimension remains the same when one replaces
/ by a eld 1 containing /.
326. A symmetric matrix o Sym
n
(R) is said compatible if it is of the form c/
T
+ /c
T
with
c. / 1
n
. Prove that o Sym
n
(R) is compatible if and only if
either o = 0
n
,
or o is rank-one and non-negative,
or o has rank two and its non-zero eigenvalues have opposite signs.
327. (After R. A. Horn and C. R. Johnson (I).) Let M
n
(C) be given.
(a) Suppose that is similar to a matrix 1 M
n
(R). Prove that is similar to

.
Hint: 1
T
is similar to 1.
(b) Conversely, we assume that is similar to

.
i. Verify that is similar to . Deduce that the spectrum of is invariant
under conjugation, and that when is a non-real eigenvalue, the Jordan blocks
corresponding to

have the same sizes as those corresponding to .
ii. Deduce that is similar to a matrix 1 M
n
(R).
166
328. (After R. A. Horn and C. R. Johnson (II).) Let M
n
(C) be given. We assume that
is similar to

: = 1
1

1.
(a) Show that there exists R such that c
i
1 + c
i
1

is non-singular.
(b) Deduce that there exists o H
n
GL
n
(C) such that = o
1

o.
329. (After R. A. Horn and C. R. Johnson (III).) Let M
n
(C) be given.
(a) We assume that is similar to

1

o
with o H
n
. Verify that

o is Hermitian. Deduce that has the form H1 where

H and 1 are Hermitian, one of them being non-singular.
(b) Conversely, assume that has the form H1 where H and 1 are Hermitian, one of
them (say H for deniteness) being non-singular. Verify that is similar to

.
So far, we have shown that is similar to a matrix 1 M
n
(R) if and only if it is of the
form H1 where H and 1 are Hermitian, one of them being non-singular.
330. (After R. A. Horn and C. R. Johnson (IV).) We now assume that = H1 where H and
1 are Hermitian. We warn the reader that we allow both H and 1 to be singular.
(a) To begin with, we assume that H =
_
H

0
0 0
_
where H

H
p
is non-singular. Let
_
1

_
be the block form of 1 matching that of H so that
=
_
H

0 0
_
.
i. Let be a non-real eigenvalue of . Show that is an eigenvalue of

:= H

are the same.

ii. Deduce that the Jordan blocks of corresponding to

have the same sizes
as those corresponding to . Hint: According to the previous exercises,

is
similar to a matrix 1

M
p
(R).
iii. Deduce that is similar to a matrix 1 M
n
(R).
(b) Prove the same result in the general case. Hint: Diagonalize H.
Summary: The following properties are equivalent to each other for every M
n
(C).
is similar to a matrix 1 M
n
(R),
is similar to

,
is similar to

via a Hermitian transformation,

There exist H. 1 H
n
such that = H1 and one of them is non singular,
There exist H. 1 H
n
such that = H1.
167
331. If M
n
(R) and r (0. +), verify that det(r1
n
+
2
) 0. Deduce that if : is odd,
then 1
n
cannot be written as
2
+ 1
2
with . 1 M
n
(R). Note: On the contrary, if
: is even, then every matrix  M
n
(R) can be written as
2
+1
2
with . 1 M
n
(R).
332. If S
n
, we denote by 1

the permutation matrix associated with . A nite sum

of permutation matrices is obviously a matrix  M
n
(N), whose sums of rows and
columns are equal. We shall prove the converse statement: If  M
n
(N) has equal sum
o for rows and columns, then  is a nite sum of permutation matrices.
Let 1, J be two sets of indices 1 i. , :. If the bloc
IJ
is identically 0, prove that
the sum of the entries of the opposite bloc 
I
c
J
c equals (: j )o. If o 1, deduce
that j + :.
Let us recall (Exercise 9) that this property implies that there exists a permutation
such that :
i(i)
,= 0 for every 1 i :. Then argue by induction over o.
333. Let  M
n
(R) be a given non-negative matrix. If S
n
, let us denote
:

:=
n

i=1
:
i(i)
.
Finally, we dene
o := max
Sn
:

.
We assume that for every entry :
ij
, there exist S
n
such that (i) = , and :

= o.
(a) Find a positive linear form ) over M
n
(R) such that :

= )(1

) for every
S
n
, where 1 is the Hadamard product.
(b) Rephrase the assumption in the following way: There exists a subset A of S
n
such
that :

= o for every A, and

Q :=

X
1

0
n
.
(c) Let S
n
be given.
i. Verify that Q1

is a sum of / 1 permutation matrices, with / := [A[. Hint:

Use exercise 332.
ii. Deduce that :

o, and therefore = o.
334. Recall that a numerical function ) : (c. /) R is operator monotone if whenever : 1,
. 1 H
n
are given with the spectra of and 1 included in (c. /), we have
( 1) =()() )(1)).
where H 1 is understood in the sense of Hermitian forms. More generally, ) is operator
monotone of grade : if this holds true at xed size : :. We prove here the Loewners
Theorem, under some regularity assumption.
168
(a) Check that if : :, then operator monotonicity of grade : implies operator mono-
(b) We already know that if both H and 1 are 0
n
, their Hadamard product H 1
is 0
n
too. Here is a converse: Let 1 H
n
be given. If 1 H 0
n
whenever
H H
n
is 0
n
, prove that 1 0
n
.
(c) We assume that ) ((c. /) and recall the assumptions of Exercise 250: If 1 =
diag(d
1
. . . . . d
n
M
n
(C), we dene a matrix )

(1) M
n
(C) by
)

(1)
jk
=
)(d
j
) )(d
k
)
d
j
d
k
.
where we identify
)(/) )(c)
/ c
:= )

(c).
if / = c.
If ) is operotor monotone of grade :, prove that )

(1) 0
n
whenever d
1
. . . . . d
n

(c. /). Hint: Use DaletskiKrein formula.
(d) If : 2, deduce that either ) is constant, or it is strictly increasing.
(e) We assume that ) (
3
(c. /) and : 2. We may take : = 2. Compute 1
T
)

(1)1
when
1 =
_
1
1
d
2
d
1
0
1
d
2
d
1
_
.
Deduce that 2)

3)
2
. In other words, either ) is constant or ()

)
1/2
is concave.
335. The relative gain array of a square matrix GL
n
(/) is dened as () :=
T
,
where the product is that of Hadamard (entrywise). It was studied by C. R. Johnson &
H. Shapiro). Many questions about it remain open, including that of the range of .
(a) If is triangular, show that () = 1
n
.
(b) Verify that 1 is an eigenvalue of (), associated with the eigenvector c whose
components are all ones. Open problem: Is this the only constraint ? In other
words, if c = c, does there exist a matrix such that () =  ?
(c) If is not necessarily invertible, we may dene () :=

, with

the cofactor
matrix. Therefore we have () =
1
det A
() when is non-singular.
Show that there exists a homogeneous polynomial in the entries of , such that
det () () det .
(d) If : = 3 and is symmetric,
=
_
_
c .
. / r
r c
_
_
.
169
verify that
() = (c/c r.)
2
+ c/r
2

2
+ /c
2
.
2
+ cc.
2
r
2
(cr
2
)
2
(/
2
)
2
(c.
2
)
2
.
Deduce that () = 0 when is rank-one, or when a row of vanishes. However
det does not divide ().
336. (The symmetric positive denite case.) We continue with the study of the relative gain
array. We now restrict to matrices o SPD
n
.
(a) Show that (o) 1
n
(Fiedler), with equality only if o = j1
n
for some j. Hint:
Write o =

T
j
in an orthonormal basis. Compute o
T
and (o). Then use
the inequality
a
b
+
b
a
2.
(b) Following the same strategy, show that
(o)
1
2
_
(o) +
1
(o)
_
1
n
.
where (o) is the condition number of o.
(c) Deduce the inequality
((o))
1
2
_
(o) +
1
(o)
_
.
(d) Draw the conclusion: if o SPD
n
, then
lim
m+

(m)
(o) = 1
n
.
Verify that this convergence has order 2 at least, like in a Newtons method.
337. (S. Maciej, R. Stanley.)
Let us recall that the permanent of M
n
(/) is dened by the formula
per :=

Sn
n

i=1
c
i(i)
.
It is the same formula as for the determinant, with the exception that the coecient c()
has been replaced by +1. We consider matrices with real entries c
ij
[0. 1]. We assume
that there are : zeroes among the entries, with : :. We wish to bound the permanent
of .
(a) Verify that the maximum of the permanent is achieved, at some matrix whose :
2
:
other entries are 1s.
170
(b) Prove the formula
per =

1j<kn
per
jk
per1
jk
.
where
jk
denotes the block obtained by retaining only the rst two rows and the
,-th and /-th colums, whereas 1
jk
is the block obtained by deleting the rst two
rows and the ,-th and /-th colums.
(c) Let us assume that has : zeroes and :
2
: ones, and that two zeroes of belong
to the same row.
i. Show that has a row of ones.
ii. Wlog, we may assume that c
11
= c
12
= 0 and the second row is made of ones.
We dene

form by switching c
11
and c
21
. Thus the upper-left block of

is
_
1 0
0 1
_
.
Using the formula above, show that per < per

.
(d) If per is maximal among the admissible matrices, deduce that the zeroes of are
on : distinct rows and : distinct columns.
(e) We may therefore assume that c
ii
= 0 for i = 1. . . . . : and c
ij
= 0 otherwise. Prove
that
per =
m

=0
(1)

_
:
/
_
(: /)! .
Deduce that
per :!
_
1
:
2:
_
.
338. Let Z be given. We assume that is not the square of an integer. We consider the
set 1

of matrices M
n
(Z) whose characteristic polynomial is A
2
. If c. / Z and
/[ c
2
, we denote

(a,b)
:=
_
c c
/ c
_
1

. c :=
c
2
/
.
Finally, we say that two matrices . 1 M
n
(Z) are similar in Z if there exists 1
GL
2
(Z) such that 1 = 11.
(a) If (c. /) is above and Z, verify that 
(a,b)
,
(a,b)
, 
(a+b,b)
,
(a,(a
2
)/b)
are
similar in Z.
(b) Let  1

be given. We dene () as the minimal / 0 such that  is similar

to
(a,b)
. Prove that this denition makes sense.
(c) Show that there exists an c Z such that [c[
1
2
(), such that  is similar to

(a,b)
.
171
(d) Compare [ c
2
[ with ()
2
. Deduce that
()
_
. if 0.
_
4[[,3. if < 0.
(e) Finally, show that 1

is the union of nitely many conjugation classes in Z.

339. Let / be a eld and 1 /[A] be a monic polynomial of degree :.
(a) When is the companion matrix 1
P
diagonalizable ?
(b) Show that the Euclidian algorithm can be used to split 1 into factors having simple
roots, in nitely many elementary operations.
(c) Deduce an explicit construction of a diagonalizable matrix
P
M
n
(/) whose char-
acteristic polynomial is 1 (diagonalizable companion).
Nota: When / R and the roots of 1 are real, Exercise 92 gives an alternate
construction.
340. Elaborate a test which tells you in nite time whether a matrix M
n
(/) is diagonal-
izable or not. Hints: is diagonalizable if and only if 1() = 0
n
for some polynomial
with simple roots, One may construct explicitly the factor 1 of the characteristic poly-
nomial 1
A
, whose roots are simple and are those of 1
A
(see Exercise 339).
341. Let M
n
(C) be a strictly diagonally dominant matrix. We denote
:
i
:=
[c
ii
[

j=i
[c
ij
[
< 1. i = 1. . . . :.
Let us recall that is non-singular, and denote 1 :=
1
.
(a) Let i ,= / be indices. Show that
[/
ik
[ :
i
max[/
jk
[ ; , ,= i.
(b) Deduce that for every i ,= /, we have [/
ik
[ :
i
[/
kk
[. Hint: Fix / and consider the
index , that maximizes [/
jk
[.
Remark that
1
is not necessarily strictly diagonally dominant.
342. We come back to the relative gain array dened in Exercise 335:
() :=
T
.
We consider the strictly diagonally dominant case and use the notations of Exercise 341.
(a) Verify that

j=i
[()
ij
[ :
i
(max
j=i
:
j
)[()
ii
[.
172
(b) Deduce that () is strictly diagonally dominant too. Denoting :() := max
i
:
i
, we
have :(()) :()
2
.
(c) We now consider the iterates
(k)
:=
(k)
(). Show that
(k)
is strictly diagonally
dominant and that :(
(k)
) 0. Deduce that
(k)
= 1
(k)
(1
n
+ 1
(k)
), where 1
(k)
is
diagonal and 1
(k)
0
n
.
(d) Show that
(k+1)
= 1
(k)

(k)
_
1
(k)
_
1
where 
(k)
1
n
. Deduce that 1
(k+1)
1
n
.
(e) Finally prove (Johnson & Shapiro)
lim
k+

(k)
() = 1
n
for every strictly diagonally dominant matrix.
343. We continue the analysis of the relative gain array dened in Exercise 335, following
Johnson & Shapiro. We consider permutation matrices and linear combinations of two of
them. We recall that permutation matrices are orthogonal: 1
T
= 1.
(a) Assume that C is the permutation matrix associated with an :-cycle. If . C is
such that .
r
,= 1, verify that
(1
n
.C)
1
=
1
1 .
r
(1
n
+ .C + + .
n1
C
n1
).
Deduce that
(1
n
.C) =
1
1 .
n
(1
n
.
n
C).
(b) Use the formula above to prove that

(k)
(1
n
.C) =
1
1 .
n
k
(1
n
.
n
k
C).
Deduce that if [.[ < 1 (respectively [.[ 1) then
(k)
(1
n
.C) converges towards
1
n
(resp C) as / +.
(c) If .
n
= . ,= 1, deduce that
1
1z
(1
n
.C) is a xed point of .
(d) If GL
n
(C) and 1 is a permutation matrix, verify that (1) = 1().
(e) Show that if 1 and Q are permutation matrices and 1 + Q is non-singular, then
1
2
(1 + Q) is a xed point of . Hint: reduce to the case where 1 = 1
n
, then work
blockwise to deal only with cycles.
344. The QR method for the calculation of the eigenvalues of a complex matrix was designed
in 196162 by J. Francis. Not only he proved the convergence when the eigenvalues have
distinct moduli, but he recognized the necessity of shifts:
shifts help to enhance the convergence by reducing the ratio
n
,
n1
, where
n
is
the smallest eigenvalue,
173
complex shifts help to discriminate pairs of distinct eigenvalues that have the same
modulus. This problem is likely to happen for matrices with real entries, because of
complex conjugate pairs.
We describe below a few basic facts about shifts. The algorithm works as follows. We
0
C and make the Q1
factorization

0
1
n
= Q
0
1
0
.
Then we re-combine
1
:= 1
0
Q
0
+
0
1
n
. More generally, if
j
is an iterate, we choose

j
C, decompose
j

j
1
n
= Q
j
1
j
and recompose
j+1
:= 1
j
Q
j
+
j
1
n
. Thus the
standard QR algorithm (without shift) corresponds to choices
j
= 0.
We still denote 1
j
:= Q
0
Q
j1
and l
j
:= 1
j1
1
0
.
(a) Verify that
j+1
= Q

j
Q
j
and then
k
= 1

k
1
k
.
(b) Show that 1
k
l
k
is the QR factorization of the product (
k1
1
n
) (
0
1
n
).
(c) We consider the case where M
n
(R). We choose
1
=
0
. Show that 1
2
O
n
(R)
and deduce that
2
M
n
(R).
Nota: J. Francis found a way to perform the two rst iterations at once, by using
only calculations within real numbers. Therefore the shifted QR method does not
need to dive into the complex numbers when has real entries. See D. Watkins,
American Math. Monthly, May 2011, pp 387401.
345. A square matrix is said to be non-derogatory if for every eigenvalue , one has dimker(
1
n
) = 1.
(a) Let J = J(0; :) be the basic Jordan block. If 1 M
r
(/) commutes with J, show
that 1 is a polynomial in J.
(b) More generally, if 1 commutes with a non-derogatory matrix , show that there
exists j /[A] such that 1 = j(). Hint: Jordanization, plus polynomial interpo-
lation.
346. (Bezerra, R. Horn.) This is a follow-up of the previous exercise. Let M
n
(C) be
non-derogatory, and suppose that 1 M
n
(C) commutes with both and

. Show that
1 is normal. Hint: Use Schurs Theorem that a matrix is unitarily trigonalizable.
347. If
1
. . . . .
r
R
n
have non-negative entries, then the matrix
1

T
1
+ +
r

T
r
is symmetric
positive semidenite and has non-negative entries. A natural question is whether the
converse holds true: given a symmetric matrix o, positive semidenite with non-negative
entries, do there exist vectors
j
0 such that o =
1

T
1
+ +
r

T
r
? According to P. H.
Diananda, and to M. Hall & M. Newman, this is true for : 4. The following example,
174
due to Hall, shows that it is false when : 5:
o =
_
_
_
_
_
_
4 0 0 2 2
0 4 3 0 2
0 3 4 2 0
2 0 2 4 0
2 2 0 0 4
_
_
_
_
_
_
.
(a) Verify that o is positive semidenite. In particular, det o = 0, thus o has a non-
trivial kernel. Compute a generator of the kernel
(b) Suppose that o was a
1

T
1
+ +
r

T
r
for some non-negative vectors
j
.
i. Show that o can be written as a sum o
1
+ o
2
with
o
1
=
_
_
_
_
_
_
4 0 0 2 2
0 0 0 0 0
0 0 0 0 0
2 0 0 c 0
2 0 0 0 /
_
_
_
_
_
_
. o
2
=
_
_
_
_
_
_
0 0 0 0 0
0 4 3 0 2
0 3 4 2 0
0 0 2 r 0
0 2 0 0
_
_
_
_
_
_
.
and c. /. r. 0 and both matrices are positive semidenite.
ii. Show that necessarily, c = / = 0 and r = = 4. Hint: Use the kernel of o.
(c) Let 1
n
denote the cone of : : symmetric matrices with non-negative entries. Let
Sym
+
n
denote the cone of positive semidenite matrices. Finally C
n
denotes the
cone of symetric : : matrices o having the property that for R
n
,
( 0) =(
T
o 0).
Show that if : 5, then Sym
+
n
+ 1
n
C
n
. Hint: argue by duality.
348. (a) Parametrization: Given (c. /. c. d) such that c
2
+ /
2
+ c
2
+ d
2
= 1, verify that the
matrix
_
_
c
2
+ /
2
c
2
d
2
2/c 2cd 2/d + 2cc
2/c + 2cd c
2
/
2
+ c
2
d
2
2cd 2c/
2/d 2cc 2cd + 2c/ c
2
/
2
c
2
+ d
2
_
_
is orthogonal.
(b) Interpretation: Let H be the skew eld of quaternions, whose basis is (1. i. j. k). If
= c +/i +cj +dk H, we denote := c /i cj dk its conjugate. We identify
the Euclidian space R
3
with the imaginary quaternions = /i + cj + dk. In other
words, R
3
i = . We have : = : . Finally, the norm over H is
|| :=

c
2
+ /
2
+ c
2
+ d
2
.
One has |

| = || |

|.
175
Suppose that H has unit norm. Verify that
1
= . We consider the linear map
: 1
q
: := :
1
. Show that 1
q
(R
3
) R
3
. Verify that the restriction 1
q
of 1
q
to
R
3
is an isometry.
Prove that 1
q
is a continuous group homomorphism from the unit sphere of H
into O
3
(R). Deduce that its image is included into SO
3
(R).
(c) Conversely, let 1 be a rotation of R
3
with axis n and angle . Show that 1 = 1
q
for := cos

2
+n sin

2
. The morphism 1
q
is thus onto.
349. Let  M
n
(/) be given. We denote 
(j)
the ,-th principal minor,

(j)
:= 
_
1 . . . ,
1 . . . ,
_
.
We assume that these principal minors are all nonzero. Recall that this ensures a factor-
ization  = 1l.
(a) For any xed pair (i. ,), use the DesnanotJacobi formula (Dodgson condensation
formula, exercise 24) to establish the identity

_
1 . . . / 1 i
1 . . . . . . /
_

_
1 . . . . . . /
1 . . . / 1 ,
_

(k)

(k1)
=
k1

k
.
where

k
:=

_
1 . . . / i
1 . . . / ,
_

(k)
.
(b) Deduce the formula
n

k=1

_
1 . . . / 1 i
1 . . . . . . /
_

_
1 . . . . . . /
1 . . . / 1 ,
_

(k)

(k1)
= :
ij
.
(c) Deduce that in the factorization  = 1l, we have
/
ij
=

_
1 . . . , 1 i
1 . . . . . . ,
_

(j)
. n
ij
=

_
1 . . . . . . i
1 . . . i 1 ,
_

(i1)
.
(d) In particular, if  is totally positive, meaning that all minors 
_
i
1
. . . i
r
,
1
. . . ,
r
_
are
strictly positive whenever i
1
< < i
r
and ,
1
< < ,
r
, then the non-trivial entries
of 1 and l are strictly positive. Comment: Actually, all the minors
1
_
i
1
. . . i
r
,
1
. . . ,
r
_
. respectively l
_
i
1
. . . i
r
,
1
. . . ,
r
_
.
with ,
1
< < ,
r
i
1
< < i
r
(resp. i
1
< < i
r
,
1
< < ,
r
) are strictly
positive.
176
350. Let T
n
M
n
(R) be the set of : : tridiagonal bi-stochastic matrices.
(a) Verify that every  T
n
is of the form
_
_
_
_
_
_
_
_
_
_
_
1 c
1
c
1
0 0
c
1
1 c
1
c
2
c
2
.
.
.
0 c
2
1 c
2
c
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. c
n1
0
.
.
. c
n1
1 c
n1
c
n
c
n
0 0 c
n
1 c
n
_
_
_
_
_
_
_
_
_
_
_
.
In particular,  is symmetric.
(b) Verify that T
n
is a convex compact subset of M
n
(R), dened by the inequalities
0 c
i
1. c
i
+ c
i+1
1.
(c) Prove that the extremal points of T
n
are those matrices  for which (c
1
. . . . . c
n
) is
a sequence of 0s and 1s, in which two 1s are always separated by one or several 0s.
We denote T
n
the set of those sequences.
(d) Find a bijection between T
n
T
n1
and T
n+1
. Deduce that the cardinal of T
n
is the
:-th Fibonacci number.
351. (After Y. Benoist and B. Johnson.) In Exercise 330, we proved that every real matrix
 M
n
(R) can be written as the product H1 of (possibly complex) Hermitian matrices.
Of course, one has || |H| |1|. We show here that if : = 3 (and therefore also if
: 3), there does not exist a nite number c
n
such that (H. 1) can always be chosen so
that |H| |1| c
n
||. Thus this factorization is unstable.
So let us assume that for some nite c
3
, the following property holds true: for every
M
3
(R), there exist H. 1 H
3
such that  = H1 and |H| |1| c
3
||.
We recall that the condition number of a non-singular matrix 1 is
(1) := |1| |1
1
| 1.
(a) Let  be given. Show that () c
3
(), where () is dened as the inmum
of (1) where 1
1
1 =
T
.
(b) Let B := (
1
.
2
.
3
) be a basis of R
3
and let (u
1
. u
2
. u
3
) be the dual basis. If

i
=
i
for all i, verify that
T
u
i
= u
i
.
(c) Chosing pairwise distinct eigenvalues
i
, deduce that (B) c
3
(), where (B)
is the inmum of (1) for which 1
i
| u
i
.
(d) Deduce that for every basis B, one has (B) c
3
.
177
(e) We choose the basis B

given by

1,
=
_
_
1
0
0
_
_
.
2,
=
_
_
1
c
0
_
_
.
3,
=
_
_
1
0
c
_
_
.
Compute the dual basis. If the inequality above is true, prove that there exists a
subsequence (1

)
m0
that converges to some 1
0
GL
3
(C), such that 1

i,
| u
i,
.
Conclude.
352. Let H H
n
be a Hermitan matrix, with indices of inertia (:

. 0. :
+
). Hence H is
non-degenerate.
(a) Let us write H blockwise as
_
H

A
A

H
+
_
.
where H

has size :

. If H

is denite negative, prove that its Schur comple-

ment H
+
A

H
1

A is positive denite.
(b) Deduce that if 1 C
n
is a subspace of dimension :

on which the form r r

Hr
is negative denite (1 is a maximal negative subspace for H), then the form

H
1
is positive denite on 1

(1

is a maximal positive subspace for H

1
).
353. We consider the method of Jacobi for the approximate calculation of the spectrum of a
Hermitian matrix H. We take the notations of Section 13.4 of the second edition.
(a) Recall that the equation t
2
+ 2t 1 = 0 admits two roots t. t

, with t = tan and

[

4
.

4
). Verify that the other root corresponds to t

= tan

with

= +

2
.
(b) We call the inner angle and

the outer angle. Show that if the choice of angle

or

leads to an iterate 1 or 1

, then 1

is conjugated to 1 by a rotation of angle

2
in the (j. )-plane.
(c) Deduce that if we x the list of positions (j
k
.
k
) to be set to zero at step /, the
choice of the angles
k
is irrelevant because, if
(k)
and 1
(k)
are two possible iterates
at step /, then they are conjugated by a sign-permutation matrix. Such a matrix is
a permutation matrix in which some 1s have been replaced by 1s.
354. (After X. Tuni.) We show here that it is not possible to dene a continuous square root
map over GL
2
(C).
(a) Let GL
2
(C) be given. If has two distinct eigenvalues, prove that there are
exactly four matrices A such that A
2
= .
(b) If instead is not semi-simple, verify that there are exactly two matrices A such
that A
2
= .
178
(c) We suppose that there exists a square root map
1/2
over GL
2
(C), which is
continuous. Without loss of generality, me may assume that
_
1 1
0 1
_
1/2
=
_
1
1
2
0 1
_
.
For r [0. ), prove the formula
_
c
2ix
1
0 1
_
1/2
=
_
c
ix
(1 + c
ix
)
1
0 1
_
.
Hint: Use a continuity argument.
(d) By letting r , conclude.
355.
Carl Friedrich Gau.
Let Sym
n
be a positive semi-
denite matrix. We assume that its
diagonal part 1 is positive denite. If
/ 1(), we consider the GaussSeidel
method to solve the system r = /:
(1 1)r
(m+1)
= 1
T
r
(m)
+ /.
Remark that 11 is non-singular from
the assumption. The iteration matrix
is G = (1 1)
1
1
T
.
(a) Dene
(m)
= r
(m)
r, where r is some solution of r = /. Verify that
(m+1)
=
G
(m)
. In order that
(m)
converges for every choice of the initial data
(0)
, prove
that it is necessary and sucient that
if = 1 is an eigenvalue of G, then it is semi-simple,
the rest of the spectrum of G is of modulus < 1.
(b) Verify that ker(G1
n
) = ker .
(c) Show that G commutes with (1 1)
1
. Deduce that (1 1)
1
1() is a G-
invariant subspace.
(d) Prove that R
n
= ker (1 1)
1
1(). Hint: if
T
(1 1) = 0, then
T
(1 +
) = 0, hence = 0.
(e) Prove that the spectrum of the restriction of G to (11)
1
1() has modulus < 1.
Conclude. Hint: follow the proof of Lemma 20 in Chapter 12 of the 2nd edition.
356. (After C. Hillar.) A matrix M
n
(C) is completely invertible if its numerical range
\() does not contain 0. In particular, it is invertible, because \() contains the
spectrum of .
We suppose that is completely invertible.
179
(a) Show that there exists R and c 0 such that \(c
i
) is contained in the
half-plane 1. c.
(b) Let us denote 1 := c
i
. Verify that 1 + 1

2c1
n
. Deduce that 1

+ 1
1
is
positive denite.
(c) Prove that there exists 0 such that \(1
1
) is contained in the half-plane
1. .
(d) Conclude that
1
is completely invertible. Therefore a matrix is completely invert-
ible if and only if its inverse is so.
357. I shall not comment the title of this exercise, but it is related to the following fact: if
SO
n
(R) is given in block form
 =
_
1
C 1
_
where and 1 are square matrices, not necesssarily of the same size, prove that det =
det 1. Hint: Find block-triangular matrices 1 and l such that 1 = l.
The same identity holds true if  SO(j. ).
358. Let us say that a lattice
/ =
d
j=1
Zv
j
of R
d
has a ve-fold symmetry if there exists a matrix M
d
(R) such that
5
= 1
d
,
,= 1
d
and / = /.
(a) Verify that the lattice Z
5
has a ve-fold symmetry.
(b) If / = /, verify that is similar to a matrix 1 M
d
(Z). Deduce that its
eigenvalues are algebraic integers of degree d.
(c) Suppose d = 3 and / has a ve-fold symmetry . Show that is diagonalizable in
M
3
(C), with eigenvalues 1, and , where is a primitive root of unity of order 5.
Deduce that there does not exist a 3-dimensional lattice with a ve-fold symmetry.
(d) In the same vein, if / is a 4-dimensional lattice with a ve-fold symmetry , show
that
4
= 1
4

2

3
.
(e) Let ,= be the roots of A
2
A 1. Let us form the matrix
:=
1
2
_
_
_
_
1 0
1 0
1 0
1 1 1 1
_
_
_
_
.
Dene
v
1
=
_
_
_
_
1
0
0
0
_
_
_
_
. v
2
= v
1
. v
3
= v
2
. v
4
= v
3
.
180
Show that the v
j
s span a lattice that has a ve-fold symmetry.
359. (After O. Taussky and H. Zassenhaus.) We show here that if M
n
(/), then there
exists a symmetric matrix o GL
n
(/) such that
T
= oo
1
(compare with Exercises
327330). Of course, we already know that there exists a (possibly non-symmetric) matrix
1 GL
n
(/) such that
T
= 11
1
.
(a) Let us begin with the case where the characteristic and minimal polynomials of
coincide. We recall that then, the subspace Com() of matrices commuting with
equals
1() [ 1 /[A].
and that its dimension equals :.
i. Dene the subspace CT() M
n
(/) by the equation  =
T
. Dene also
the subspace ST() M
n
(/) by the equations
o =
T
o and o = o
T
.
Verify that ST() CT() 1 Com().
ii. Show that dimST() :.
iii. Deduce that every matrix  such that  =
T
is symmetric.
(b) We now drop the condition about the characteristic and minimal polynomials of .
Using the case above, show that there exists a symmetric and non-singular o such
that o =
T
o. Hint: Apply the case above to the diagonal blocks of the Frobenius
form of .
Remark. This situation is interesting in innite dimension too. For instance, let us
take a dierential operator 1 = 1
2
+ 1 c, where c is a (
1
-bounded function, that is
1n = n

+(cn)

. As an unbounded operator over 1

2
(R), 1 has an adjoint 1

= 1
2
c1,
that is 1 =

. There are a lot of self-adjoint operators o satisfying o1 = 1

o.
For instance,
o. = (.

+ .
works whenever (r) and (r) solve the linear dierential equations

= c and = c

+ cst .
More generally, the analysis above suggests that for every 1 R[A], o := 1(1) is a
self-adjoint operator satisfying o1 = 1

o.
However, the situation can be very dierent from the nite-dimensional one, just because
it may happen that an operator  is not conjugated to

. This happens to the deriva-

tion 1 =
d
dx
within the algebra of dierential operators. We recover conjugacy by leaving
the realm of dierential operators: the symmetry
o
0
: (r )(r)) (r )(r))
satises o
0
1 = 1

o
0
= 1o
0
. Then, as above, every o = o
0
1(1) with 1 R[A] is
self-adjoint and satises o1 = 1

o.
181
360. (Continuation.) We investigate now the real case: / = R. We ask under which condition
it is possible to choose o symmetric positive denite such that o =
T
o.
(a) Show that a necessary condition is that has a real spectrum.
(b) Prove that if is similar to 1, and if there exists SPD
n
such that 1 = 1
T
,
then there exists o SPD
n
such that o =
T
o.
(c) If is a Jordan block, describe explicitly the solutions of o =
T
o. Verify that
some of them are positive denite.
(d) Deduce that a necessary and sucient condition is that has a real spectrum.
361. Let : 2 be a given integer. We identify below the convex cone 1 spanned by matrices
of the form
2
, with M
n
(R) running over skew-symmetric matrices.
(a) If is skew-symmetric with real entries, verify that
2
is symmetric, positive
semi-denite, and that its non-zero eigenvalues have even multiplicities.
(b) Show that 1 is contained in the set
_
o Sym
n
[ 0
n
o
Tr o
2
1
n
_
.
(c) Let us dene
C
1
:=
_
c R
n
[

j
c
j
= 1 and 0 c
j

1
2
. ,
_
.
i. Show that C
1
is a compact, convex subset of R
n
, and that its extremal points
have the form
1
2
(e
i
+e
j
) for some i < ,.
ii. Deduce that C
1
is the convex hull of
C
1
:=
_
1
2
(e
i
+e
j
) [ 1 i < , :
_
.
iii. Finally, show that
_
c R
n
[ 0 c
j

1
2

k
c
k
. ,
_
= conv ((e
i
+e
j
) [ 0. 1 i < , :) .
(d) Using this last result, show that actually,
1 =
_
o Sym
n
[ 0
n
o
Tr o
2
1
n
_
.
182
362. (After Feit & Higman.) Let  M
n
(/) be given, and j /[A] a non-zero polynomial
such that j() = 0 (that is, a multiple of the minimum polynomial of ). Let : be the
multiplicity of / as a root of j, and dene the polynomial (A) = j(A),(A )
m
.
Prove that the multiplicity of as an eigenvalue of  equals
Tr ()
()
.
Hint: decompose /
n
into 1(( )
m
) and ker( )
m
.
363. Let : R R be a continuous convex function. If r. R
n
are such that r , prove
that
n

j=1
(
j
)
n

j=1
(r
j
).
Hint: Use Proposition 6.4 in the 2nd edition.
Application: Let  M
n
(C) be given, with eigenvalues
j
and singular values
j
. If :
is a positive real number, deduce from above and from Exercise 69 the inequality
n

j=1
[
j
[
s

j=1

s
j
.
364. (Golden, Wasin So, Thompson.)
(a) Let H. 1 HPD
n
be given. Using the previous exercise, prove that for every
integer : 1,
Tr
_
(H1)
2m
_
Tr
_
(1H
2
1)
m
_
.
(b) If : = 2, prove that the equality holds above if, and only if [H. 1] = 0
n
.
(c) Let . 1 be Hermitian matrices. Show that the sequence (n
k
)
k1
dened by
n
k
= Tr
_
(c
A/2
k
c
B/2
k
)
2
k
_
is non-increasing.
(d) Deduce that
Tr(c
A
c
B
) Tr c
A+B
.
Hint: UseTrotters formula
lim
m+
(c
A/m
c
B/m
)
m
= c
A+B
.
(e) In the equality case, that is if Tr(c
A
c
B
) = Tr c
A+B
, show that Tr((c
A
c
B
)
2
) =
Tr(c
A
c
2B
c
A
). Then deduce that [c
A
. c
B
] = 0
n
, and actually that [. 1] = 0
n
.
183
365. Let t (t) be a continuous map with values in the cone of real : : matrices with
non-negative entries. We denote by 1 the vector space of solutions of the dierential
equation
dr
dt
= (t)r.
We also dene e = (1. . . . . 1)
T
.
(a) Verify that the map r |r(0)|
1
is a norm over 1.
(b) If 0 is given, we denote r

1 the solution satisfying r

() = e. Verify that
r

(t) 0 for every t [0. ].

(c) Let us dene

= r

,|r

(0)|
1
. Show that the family (

)
>0
is relatively compact
in 1, and that it has a cluster point as +.
(d) Deduce HartmanWintners Theorem: There exist a non-zero solution (t) of the
ODE such that (t) 0 and

(t) 0 for every t 0. In particular, (t) admits a

limit as t +.
(e) When is constant, give such a solution in close form.
366. Let 1 be a non-void compact convex subset in nite dimension. If r 1, the Hahn
Banach Theorem ensures that there exists at least one convex cone (actually a half-space)
with apex r, containing 1. The set of all such convex cones admits a smaller one, namely
the intersection of all of them. We call it the supporting cone of 1 at r, and denote it
C
K
(r). If there is no ambiguity, we just write C(r).
We admit the following properties, which are classical in convex analysis:
1 is the intersection of its supporting cones C(r) when r runs over the extremal
points of 1,
If r 1, C(r) is the smallest cone with apex at r, containing all the extremal
points of 1.
In what follows, we determine C
K
(1
n
) when 1 is the convex hull of SO
n
(R) in M
n
(R).
We assume that : 2.
(a) We look for those matrices Q M
n
(R), such that Tr Q(1 1
n
) 0 for every
1 SO
n
(R).
Using the one-parameters subgroups of SO
n
(R), show that for every skew-
symmetric matrix , one has
Tr(Q) = 0. Tr(Q
2
) 0.
Show that Q is symmetric.
184
We denote
1

2

n
the eigenvalues of Q. Using Exercise 361, show
that
1
+
2
0. In other words,
[
1
[
2

n
.
Hint: Use the extremal points of the cone C
1
.
(b) Conversely, let Q Sym
n
(R) be such that
1
+
2
0. Using DacorognaMarechals
Inequality (17), prove that Tr(Q1) Tr Q for every 1 SO
n
(R).
(c) Deduce that C(1
n
) is the set of all matrices such that Tr(Q) Tr Q for every
symmetric Q whose least eigenvalues satisfy
1
+
2
0.
(d) Show that C(1
n
) is the set of matrices  whose symmetric part o =
1
2
( +
T
)
satises
_
1 +
Tr o :
2
_
1
n
o 1
n
.
Hint: Use again Exercise 361.
367. (From Zhiqin Lu.) In this exercise and in the next one, one can replace the scalar eld R
by C, to the price that A
T
be replaced by A

.
Let A M
n
(R) be given. We form the matrix 1 =
_
0
n
A
0
n
0
n
_
M
2n
(R). We dene the
linear map
1
X
: \ M
n
(R) [1
T
. [1. \ ]].
(a) Show that 1
X
2n
(R), for the standard Euclidian product \. \ =
Tr(\
T
\ ).
(b) Verify that the set of block-diagonal matrices
_
1 0
n
0
n
C
_
is an invariant subspace for
1
X
.
The restriction of 1
X
to this subspace will be denoted l
X
.
(c) Let :
1
:
n
be the singular values of A. Show that the spectrum of l
X
consists in the numbers j = :
2
i
+:
2
j
, with multiplicities equal to the number of pairs
(i. ,) such that this equality holds.
For instance, if all the numbers :
2
i
and :
2
j
+ :
2
k
are pairwise distinct (except for the
trivial :
2
j
+:
2
k
= :
2
k
+:
2
j
) then :
2
i
is simple and :
2
j
+:
2
k
has multiplicity two for , ,= /.
368. (Continuation of the previous one.) We present below the proof by Zhiqin Lu of the
B ottcherWenzel Inequality
|[A. 1 ]|
2
F
2|A|
2
F
|1 |
2
F
. A. 1 M
n
(R).
For A M
n
(R), one denes o
X
:  [A
T
. [A. 1 ]].
(a) Show that o
X
185
(b) Verify that the ratio
[X,Y ]
2
F
Y
2
F
is maximal if and only if 1 ,= 0 belongs to the eigenspace
1 associated with the largest eigenvalue of o
X
.
(c) If 1 1, show that [A
T
. 1
T
] 1. Show also that [A
T
. 1
T
] is not colinear to 1 ,
unless A = 0 or 1 = 0. Deduce that dim1 2.
(d) Show that there exists 2 ,= 0 in 1 such that 2
1
=
_
2 0
n
0
n
2
_
is orthogonal to the
the main eigenvector of l
X
(that associated with :
2
1
).
(e) Show that
sup
Y =0
|[A. 1 ]|
2
F
|1 |
2
F
= l
X
2
1
. 2
1

and deduce that

sup
Y =0
|[A. 1 ]|
2
F
|1 |
2
F
(:
2
1
+ :
2
2
)|1 |
2
F
.
Then conclude.
Remark: This proof gives a little more when : 3, because we now that A
(A) =
_
:
2
1
+ :
2
2
is a norm (use Exercise 162), with | |
F
. We do have
|[A. 1 ]|
F

2 (A)|1 |
F
.
369. (With the help of P. Migdol.) Let : 2 be given. If  M
n
(C), we denote :() the
:() = sup
x
2
=1
[r

r[.
Recall that : is norm over M
n
(C). We also dene the real and imaginary parts of  by
1 =
1
2
( + 

) H
n
=
1
2i
(

).
(a) Prove that
:() = sup
R/2Z
|1(c
i
)|
2
.
(b) Let . 1 M
n
(C) be given. We apply the previous question to  = [. 1]. Let
be such that :() = |1(c
i
)|
2
.
i. Let us denote A = 1(c
i
), 1 = (c
i
), 2 = 11 and 1 = 1. Check that
:([. 1]) = |[A. 1] + [1. 2]|
2
and deduce
:([. 1]) 2(|A|
2
|1|
2
+|1 |
2
|2|
2
).
ii. Conclude that
(37) :([. 1]) 4:():(1). . 1 M
n
(C).
186
(c) By picking up a convenient pair . 1 M
n
(C), show that the constant 4 in (37) is
the best possible. In other words,
sup
A,B=0
:([. 1])
:():(1)
= 4.
370. Let be Hermitian positive denite, with Choleski factorization 11

.
(a) Show that in the polar factorization 1 = Hl, one has H =

.
(b) Show that in the QR-factorization

= Q1, one has 1 = 1

.
371. We consider real symmetric matrices; the Hermitian case could be treated the same way.
We recall that the eigenvalues and eigenvectors of a matrix are smooth functions of its
entries so long as the eigenvalues are simple (see Theorem 5.3 of the second edition). We
also recall that a functional calculus is available over Hermitian matrices with continuous
functions: if ) : 1 R is continuous and H = l

1l where l is unitary and 1 =

diag(c. /), then )(H) = l

)(1)l with )(1) = diag()(c). )(/)), whenever the spectrum

of H is contained in 1. This construction does not depend upon the way we diagonalize
H.
Loewners theory is the study of operator monotone functions. A numerical function )
over an interval 1 is operator monotone if whenever the real symmetric matrices . 1 have
their spectrum included in 1, then 1 implies )() )(1). This notion depends
upon the size : of the matrices under consideration, the class of operator monotone
functions getting narrower as : increases.
Finally, we recall that if  GL
n
(R), then 1 if and only if 
T
1
T
.
(a) If ) is operator monotone, show that ) is monotone in the classical sense. Hint:
Take = 1
2
.
(b) Let o() be a smooth curve such that o() has simple eigenvalues. Up to a
unitary conjugation, we may assume that o(
0
) = diag(c
1
. . . . . c
n
). Writing o() =
1()diag(
1
().
2
())1()
T
with
j
(
0
) = c
j
, 1(
0
) = 1
n
and 1() orthogonal,
compute the derivatives of
j
and of 1 at
0
. Deduce that
d
d

0
)(o()) = H o

(
0
).
is the Hadamard product of o

(
0
) with the matrix H = H(c) whose entries are
/
ij
=
_
)

(c
j
). if i = ,.
f(a
j
)f(a
i
)
a
j
a
i
. if i ,= ,.
(c) If ) is operator monotone over : : matrices whose spectrum belong to 1, deduce
from above that H(c) 0
n
for every c
1
. . . . . c
n
1.
187
(d) Conversely, we suppose that the matrices H(c) above are non-negative. Let . 1
Sym
n
(R) have their spectra in 1, with 1 positive denite. We admit that there
exists a smooth curve o() such that o(0) = , o() = 1, o() has simple
eigenvalues for every (0. 1) and o

() 0
n
. This follows from the fact that
the set of symmetric matrices C such that C 1 is open, and the subset of
matrices with a multiple eigenvalue has codimension 2 in Sym
n
(R).
Prove that
d
d
)(o()) 0
2
for all (0. 1). Deduce that )() )(1).
(e) By continuity, extends this result to the case where 1 0
n
.
Hence a C
1
-function ) is operator monotone over : : matrices (property (OM:))
if and only if the matrices H(c) are non-negative for every c
1
. . . . . c
n
in the domain
of ).
(f) If / : verify that (OM:) implies (OM/).
(g) Show that (OM:) amounts to saying that for all 1 / :, and for every c
1
. . . . . c
k

1, then
det H(c) 0.
(h) Show that (OM2) amounts to saying that )

0 and
1

is concave.
(i) If 1 = R and ) is operator monotone, show that ) is ane.
(j) Consider the case 1 = (0. +) and )(t) = t

. Show that ) is operator monotone if

and only if [1. 1].
372. Let / be a eld and 1 a linear subspace of M
n
(/). We assume that every element of 1
is singular. We wish to prove that there exist 1. Q GL
n
(/) such that 11Q =: 1

has
the property that every  1

has a zero entry :

nn
.
(a) This is true if : = 1.
(b) Suppose that  ,= 0
n
. Show that there exist non-singular 1
1
. Q
1
and 1 : < :
such that
J
r
=
_
1
r
0
0 0
nr
_
1
1
1Q
1
=: 1
1
.
(c) Let 1 M
nr
(/) be the subpace of matrices  such that there exists a matrix
=
_

_
1
1
.
Prove that every element of 1 is singular. Hint: Apply Schurs complement formula
to t + J
r
.
(d) Then argue by induction over :.
373. (From Smyrlis.) We address the following practical problems:
188
Let : = 2j + 1 be an odd integer, and let : coins be given. The facial value of each
coin is an integer (in cents, say). Suppose that, whenever we remove one coin, the
: 1 remaining ones can be arranged into two sets of j coins, the total value of the
sets being equal to each other. Prove that all coins have the same value.
Let : = 2j + 1 be an odd integer, and let : complex numbers .
j
be given. Suppose
that, whenever we remove one number, the :1 remaining ones can be arranged into
two sets of j numbers, which share the same isobarycenter. Prove that all numbers
coincide.
Of course the rst problem is a special case of the second, a natural integer being a
complex number. Yet, we shall prove the general case after proving the special case.
(a) We begin with the rst problem. Show that there exists a matrix M
n
(Z) and
r Z
n
with r 0 such that r = 0, the diagonal entries of vanish, and the other
entries of any ith line are 1, summing up to 0.
(b) Prove that the coordinates of r all have the same parity.
(c) Let us dene = r r
1
(1. . . . . 1)
T
. Verify that = 0 and
1
= 0. Prove that the
coordinates of have to be even. Finally, prove = 0 and conclude.
(d) One turns towards the second problem and denote . = (.
1
. . . . . .
n
)
T
. Then there
exists a matrix with the same properties as above, and . = 0. In particular, 0
is an eigenvalue of .
(e) Using the answer to the rst problem, prove that the kernel of , when acting
over Q
n
, is one-dimensional. Deduce that its kernel, when acting over C
n
, is one-
dimensional. Conclude.
374. This gives an alternate argument about the rank of in the previous exercise.
Let M
n
(Z) (: needs not be odd) be such that its diagonal entries are even, whereas
its o-diagonal entries are odd. Prove that rk : 1. Hint: After deleting the rst
row and the last column, compute the determinant modulo 2.
375. (After A. Bezerra & H.-J. Werner.) Recall that an : : matrix  is idempotent if

2
= . In particular, it satises
/
n
= 1() ker . ker(1
n
) = 1(). ker  = 1(1
n
).
Let and 1 be : : idempotent matrices
(a) Show that
ker(1
n
1) = 1() (1(1) (ker ker 1)).
(b) Deduce that
dimker(1
n
1) = dim(1() 1(1)) + dim((1() + 1(1)) ker ker 1).
Hint: Use twice the identity dim(A + 1 ) + dim(A 1 ) = dimA + dim1 .
189
376. This is to show that when H = +

, then the spectrum of does not tell us much

about that of H, apart from TrH = 21(Tr), and the obvious fact that the spectrum of
H is real.
Thus, let H be a Hermitian :: matrix. Show that there exists a matrix =
_
1
n
TrH
_
1
n
+
with  nilpotent, such that +

= H. Therefore the spectrum of is a singleton.

Hint: Use Exercise 131 above, which is Exercise 9 of Chapter 5 in the 2nd edition.
377. Let and 1 M
n
(C) be given. We assume that they span a two-dimensional subspace.
We wish to prove that there exist non-trivial factors :
j
+ t
j
1 (1 ,  = 2
n
1)
such that
N

j=1
(:
j
+ t
j
1) = 0
n
.
Here non-trivial means that (:
j
. t
j
) ,= (0. 0).
(a) Let . 1 M
n
(/) be given, with : = rk1. Using the rank decomposition, we write
1 =
r

i=1
r
i
c
T
i
.
Show that rk(11) :, and that rk(11) < : if and only if the : : matrix 1
dened by j
ij
:= c
T
i
1r
j
is singular.
(b) Show, by induction, that for every 1 : :, there exists a product of 2
nr
1
factors whose rank is at most :.
378. (Thanks to F. Brunault.) We prove here that if : 2 and 1 C[A] is such that 1()
is diagonalizable for every M
n
(C), then 1 is a constant polynomial.
(a) If 1 C[A], . C and 1 M
n
(C), verify
1(.1
n
+ 1) = 1(.)1
n
+ 1

(.)1 +
1
2!
1

(.)1
2
+ .
where the sum is nite. Hint: this is Taylor expansion for a polynomial in a com-
mutative algebra.
(b) If u C,  is nilpotent and u1
n
+  is diagonalizable, prove that  = 0
n
.
(c) Given 1 C[A], suppose that 1() is diagonalizable for every M
n
(C) for some
: 2.
i. If  is nilpotent, show that for every . C,
1

(.) +
1
2!
1

(.)
2
+ = 0
n
.
ii. Chosing  ,= 0
n
above, deduce that det(1

(.)1
n
+ 

) = 0 for some nilpotent

.
190
iii. Conclude.
Remark: In this analysis, C can be replaced by any eld of characteristic 0.
379. (Continuing.) On the contrary, consider some nite eld / = F
p
m.
(a) Given : 2, prove that there exists a 1
n
/[A] that is divisible by every polynomial
j /[A] of degree :.
(b) Show that for all M
n
(/), one has 1
n
() = 0
n
; hence the matrix 1
n
() is
diagonalizable!
(c) We consider the case : = 1, that is / = F
p
. Show that we may take
1
n
(A) =
n

m=1
(A
p
m
A).
Its degree is j
p
n
1
p1
.
380. We begin with some geometry over closed convex cones in R
n
. Let 1 be such a cone.
We assume in addition that 1 (1) = 0. By HahnBanach Theorem, it is not hard
to see that there exists a compact convex section 1
0
of 1 such that 1 = R
+
1
0
. For
instance, if 1 = (R
+
)
n
, then the unit simplex works.
If r. R
n
, we write r when r 1.
(a) If r. 1, we dene
(r. ) = sup 0 [ r . (r. ) = infj 0 [ jr .
We may have = 0 or = +. We set
(r. ) = log
(r. )
(r. )
.
i. Verify that (r. ) [0. +], (r. ) = (. r) and (r. .) (r. ) + (. .).
ii. Suppose that the interior l of 1 is non-void. Show that is a distance over
the (projective) quotient of l by the following relation: r if there exists
t (0. +) such that = tr. This is called the Hilbert distance.
(b) Suppose now that a matrix M
n
(R) is given, such that 1 1. Let us dene
:= sup(r. ) [ r. 1.
i. Let r. 1 be given, and = (r. ), = (r. ). Show that (r. )
and (r. ), and therefore (r. ) (r. ). In summary, induces
a non-expansive map over l, .
191
ii. Dene (remark that j 0
= (( r). (r )). j = (( r). (r )).
Show that
j +
j + 1
(r. ).
+
+ 1
(r. ).
Deduce that
(r. ) log
+
+ 1
j + 1
j +
=
_
(x,y)
0
)

(t) dt. )(t) := log

+ c
t
j + c
t
.
iii. Verify that
: (0. +) log
+ :
j + :
is maximal at : =

j.
iv. Deduce that
(r. ) / (r. ). / := tanh c
/4
.
In particular, if < +, then induces a contraction over l, .
(c) Deduce an other proof of a part of the PerronFrobenius Theorem: if is strictly
positive, then it has one and only one positive eigenvector.
It is possible to recover the full PerronFrobenius Theorem with arguments in the
same vein. This proof is due to G. Birkho.
381. Let 1 be an abelian ring and M
n
(1) be given. The left annihilator Ann

() is
the set of 1 M
n
(1) such that 1 = 0
n
; it is a left-submodule. Likewise, the right
annihilator Ann
r
() is the set of 1 M
n
(1) such that 1 = 0
n
.
(a) If 1 is a principal ideal domain, show that there exists a non-singular matrix Q
(depending on ) such that
Ann

()
T
= Q Ann
r
().
Deduce that Ann

() and Ann
r
() have the same cardinality.
(b) (After K. Ardakov.) We choose instead 1 = /[A. 1 ],(A
2
. A1. 1
2
), where / is a
nite eld.
i. Verify that 1 = / m, where m = /r / is a maximal ideal (the unique one)
satises m
2
= (0). We have [1[ = [/[
3
and [m[ = [/[
2
.
ii. Let c. / 1 be such that cr + / = 0. Show that c. / m.
iii. Let us dene
=
_
0 r
0
_
.
Describe its left- and right-annihilators. Verify that
[/[
8
= Ann

() ,= Ann
r
() = [/[
10
.
192
382. Let Sym
d
be the space of d d real symmetric matrices, where d 2 is given. When
endowed with the scalar product o. 1 = Tr(o1), this becomes a Euclidian space of
dimension  =
d(d+1)
2
, isomorphic to R
N
. The space Q of quadratic forms over Sym
d
is
therefore isomorphic to Sym
N
(!!). If Q and \ is a subspace of Sym
d
, the trace of
over \ is well-dened, and it equals

i
(o
i
).
where o
1
. . . . . is any orthonormal basis of \ .
By duality, the space of linear forms over Q is isomorphic to Q itself, through 1() =
Tr(
q
), where Sym
N
and
q
Sym
n
is the matrix associated with .
(a) Let 1 be a linear form over Q. Show that there exist a unitary basis o
1
. . . . . o
N
and
numbers
i
such that
1() =

i
(o
i
). Q.
In particular, the numbers
i
are unique up to a permutation, and given a number
, the subspace \

of Sym
d
spanned by the o
i
s such that
i
= is unique. Hint:
Diagonalize the matrix associated with 1.
Finally, show that
1() =

Tr
_
[
V
_
. Q.
(b) We consider the linear representation of O
d
over Sym
d
by
(l. o) l
T
ol.
It induces a representation of O
d
over Q by
(l. )
U
.
U
(o) = (l
T
ol).
We suppose that a linear form 1 over Q is invariant under this action:
1() = 1(
U
). Q. l O
d
.
Show that there exist numbers . such that
1() = (1
d
) + Tr . Q.
(c) Let 1 be the (convex) cone of semi-positive quadratic forms, 1 Q. Show that
the extremal rays of 1 are spanned by the forms
H
: o (Tr(Ho))
2
.
193
(d) Let 1() = 0 be the equation of a supporting hyperplane of 1 at
I
d
. By convention,
1 0 over 1. We thus have 1(
I
d
) = 0. Let us dene
/() :=
_
1(
U
) dj(l).
where j is the Haar measure over O
d
. Verify that /() = 0 is the equation of a
supporting hyperplane to 1 at
I
d
, and also that / is invariant under the action of
O
d
.
(e) Deduce that
/() = (d Tr (1
d
)). Q.
for some positive constant .
383. Let 1
n
be the set of :: symmetric real matrices that can be written as a sum

where the vectors

R
n
are non-negative.
(a) Show that 1
n
is a convex cone, stable under the Hadamard product.
(b) As an example, let c = (c
1
. . . . . c
n
) and

/ = (c
1
. . . . . c
n
) be two sequences of real
numbers, with 0 < c
1
< < c
n
and 0 < /
n
< < /
1
. We form the symmetric
matrix o(c.

/) whose entries :
ij
are given by c
min(i,j)
/
max(i,j)
. Show that o(c.

/) 1
n
384. Let us order S
n
, for instance by lexicographic order. Given a matrix M
n
(/), we
form a matrix 1 M
n!
(/), whose rows and columns are indexed by permutations, in the
following way:
j

:=
n

i=1
c
(i)(i)
.
Notice that j

depends only upon

1
. Thus 1 is a kind of circulant matrix.
(a) Show that det and
per :=

Sn
c
i(i)
are eigenvalues of 1, and exhibit the corresponding eigenvectors.
(b) If / = R and is entrywise non-negative, prove that per is the Perron eigenvalue
of 1, that is 1 0
n
and per is the spectral radius of 1.
385. This exercise is at the foundation by Frobenius of the representation theory. Let G be a
nite group, with : = [G[. If (A
g
)
gG
are indeterminates, Dedekind formed the matrix
M
n
(C[A
e
. ]) whose entries are
c
gh
= A
gh
1. p. / G.
Remark that if G is cyclic, then is a circulant matrix.
Let \ := C[G] and be the regular representation, where a basis of \ as a linear space
is G, and we have (p)/ = p/ (G acts on itself by left-multiplications).
194
(a) Show that
1 :=

gG
A
g
(p)
acts linearly on C[A
e
. ] \ =: \, and its matrix in the basis G is .
(b) By decomposing \ into irreducible representations \
1
. , show that \ decomposes
as the direct sum of \
1
. , in which each factor is stable under 1.
(c) Deduce that the polynomial 1(A
e
. ) := det splits as
1 =

i
1
n
i
i
where each 1
i
is a homogeneous polynomial of degree :
i
= dim\
i
.
Remark: One may prove that each 1
i
is irreducible, and that they are pairwise
prime.
386. Let H HPD
n
be given. Using Cholesky factorization, nd an other proof of Hadamard
Inequality
det H
n

i=1
/
ii
.
This gives also the equality case (H must be diagonal).
387. We show here that if c
1
. . . . . c
n
are integers, then

1i<jn
c
j
c
i
, i
is an integer.
(a) Dene j
j
(A) := A(A1) (A, +1) Z[A], where j
0
(A) = 1 (empty product)
and j
1
(A) = A. Show that there exists an innite triangular matrix 1 = (t
ij
)
0i,j
with 1s on the diagonal, such that the basis j
0
. j
1
. j
2
. . . . is the image of the basis
1. A. A
2
. . . . under 1.
(b) Deduce that
det
_
_
_
_
_
1 1
j
1
(r
1
) j
1
(r
2
)
j
2
(r
1
) j
2
(r
2
)
.
.
.
.
.
.
_
_
_
_
_
=

1i<jn
(r
j
r
i
).
Hint: There is a Vandermonde determinant.
(c) Let us denote
_
c
/
_
=
c(c 1) (c / + 1)
/!
=
j
k
(c)
/!
.
195
Prove that
det
_
_
_
_
_
1 1
_
x
1
1
_ _
x
2
1
_

_
x
1
2
_ _
x
2
2
_
.
.
.
.
.
.
_
_
_
_
_
is an integer. Conclude.
388. (After E. Starling.) Let / be a eld whose characteristic is not 2. Let . 1 M
n
(/) be
such that
2
= 1
2
= 1
n
. If + 1 is non-singular, dene
- 1 = ( + 1)
1
( 1 + 21
n
).
(a) Show that
/
n
= 1
+
() 1

() = 1
+
(1) 1

(1).
where 1

() denotes the eigenspace of  associated with the eigenvalue 1.

(b) If +1 is non-singular, show that dim1

() = dim1

(1). In other words, and

1 have the same spectrum.
(c) Show that
1
+
( - 1) = 1
+
(1). 1

( - 1) = 1

().
(d) Deduce that ( - 1)
2
= 1
n
.
389. Let o M
n
(R) be symmetric, with non-negative entries. Concerning the diagonal entries,
we even assume :
ii
0 for every i = 1. . . . . :.
(a) Let us dene the subset of R
n
1 := r 0 [
n

i=1
r
i
= 1.
We consider a minimizing sequence (r
k
)
kN
of r
1
2
r
T
or over 1. Show that each
coordinate sequence (r
k
j
)
kN
is bounded below by some c 0. Deduce that the
sequence is bounded, and therefore it has a cluster point in 1.
(b) Deduce that r
1
2
r
T
or achieves its inmum over 1, at some point r

1.
(c) Using r

, show that there exists a vector A R

n
such that A (oA) = 1, where
is the Hadamard product and 1 = (1. . . . . 1)
T
.
(d) Deduce that there exists a diagonal matrix 1 0
n
such that 1o1 is bi-stochastic.
390. Recall that an idempotent matrix  M
n
(/) represents a projector: 
2
= . Let

1
. . . . .
r
be idempotent matrices and dene =
1
+ +
r
.
(a) If
i

j
= 0
n
for every pair i ,= ,, verify that is idempotent.
196
(b) Conversely, we suppose that is idempotent. Using the trace, show that 1() =

r
i=1
1(
i
). Deduce that
i

j
for every i ,= ,.
391. This a third proof (due to M. Rosenblum) of the PutnamFugledes Theorem that if
.  M
n
(C) are normal matrices and if 1 = 1, then 1

=

1. The rst
proof is given in Exercise 255 and the second one in Exercise 297.
(a) Verify that if ) : C C is an entire function, then 1)() = )()1.
(b) Deduce that the holomorphic function
1(.) := c
izM

1c
izN

: C M
n
(C)
is bounded. Hint: Remark that c
izM

c
i zM
is unitary, hence bounded. This is where
the assumption of normality enter in the play.
(c) Deduce that 1 is constant. Conclude by calculating 1

(0).
392. This exercise gives a version of Lemma 20, Section 13.3 (2nd edition), where the assump-
tion that matrices are Hermitian positive denite is replaced by entrywise poisitive.
We borrow it from Nonnegative matrices in the mathematical sciences, by A. Berman &
R. J. Plemmons, chapter 5.
Let .  M
n
(R) be non-singular. We form  =  and H =
1
, so that H
is the iteration matrix of the scheme
r
k+1
= r
k
+ /
in the resolution of r = /.
We assume that H 0 (which could be veried by a discrete maximum principle). Show
that the iteration is convergent, that is (H) < 1, if and only if
1
0
n
. Hint: If
(H) < 1, prove that
1
 =

k1
H
k
. Conversely, if
1
is non-negative, apply
PerronFrobenius and give a relation between the eigenvalues of
1
 and those of H.
393. (After K. Costello & B. Young.)
Let 1 be a nite set of cardinal : 2 and denote T(1) its Boolean algebra, made of all
subsets of 1. Let us dene a square matrix 1, whose rows and columns are indexed by
the non-void subsets of 1:
/
ij
=
_
1 if 1 J = .
0 if 1 J ,= .
We shall prove that 1 is invertible and compute its inverse matrix.
(a) Show that there exists only one permutation of T(1) satisfying (1) 1 = for
every 1 1.
197
(b) Let  be the square matrix, whose rows and columns are indexed by all the subsets
of 1, dened by
:
ij
=
_
0 if 1 J = .
1 if 1 J ,= .
Deduce that det  = 1.
(c) Let be the square matrix, whose rows and columns are indexed by all the subsets
of 1, dened by
c
ij
=
_
1 if 1 J = .
0 if 1 J ,= .
Show that has rank 2
n
1. Deduce that 1 is invertible. Hint:  = ee
T

where e is dened below.

(d) We chose an order in T(1) such that the last element is 1 itself. Denote
=
_
_
_
_
_
0
.
.
.
0
1
_
_
_
_
_
. e =
_
_
_
_
_
1
.
.
.
1
1
_
_
_
_
_
.
Verify 1 = e.
(e) We identify 1
1
through an induction over 1. If G = 1 c with c , 1, and if an
order has been given in T(1), then we order T(G) as follows: rst list the non-void
subsets of 1, then c, then the non-void subsets of 1 augmented of c. If 1 and C
denote the matrices associated with 1 and G respectively, verify that
C =
_
_
1 0 1
0
T
1 e
T
1 e ee
T
_
_
.
(f) If 1 is invertible, show that C is too, with inverse
C
1
=
_
_
0 11

T
0
T
1
1
1
1
_
_
.
Conclude.
394. Let  M
n
(R) be bistochastic, with singular values
1

n
.
(a) Show that
1
= 1.
(b) If J denotes the matrix whose all entries equal 1, what are the singular values of

1
n
J ?
198
395. Let 1 M
n
(C) be given. Prove that 1

1 and 11

are unitarily similar. Hint: Use the

singular value decomposition.
More generally, if 1 M
nm
(C) with : :, prove that there exists an isometry \
M
nm
(C) (that is \

\ = 1
m
) such that 1

1 = \

(11

)\.
396. (After J.-C. Bourin & E.-Y. Lee.)
Let H H
+
n+m
a Hermitian matrix, positive semi-denite, given blockwise
H =
_
A
A

1
_
. H
+
n
. 1 H
+
m
.
(a) Show that there exists a decomposition
H = 1

1 + o

o. 1 =
_
C 1
0 0
_
. o =
_
0 0
1

1
_
.
Hint: Use the square root of H.
(b) Deduce that there exist unitary matrices l. \ , such that
H = l
_
0
0 0
_
l

+ \
_
0 0
0 1
_
\

.
Hint: Use Exercise 395.
Remark: This decomposition implies a lot of inequalities between sums of eigen-
values of . 1 and H, respectively, following A. Horn.
(c) Arguing by induction, show that there exist vectors r
1
. . . . . r
n
S
n1
(the unit
sphere) such that
H =
n

j=1
/
jj
r
j
r

j
.
397. (After J.-C. Bourin.)
Usually, one uses the convexity of the numerical range to prove that for a given ma-
trix  M
n
(C), there exists a unitarily similar l

l that has a constant diagonal.

However, one may prove directly the latter property.
(a) Show that  is unitarily similar to a matrix 1 such that i ,= , implies :
ji
= :
ij
.
Hint: Consider the so-called real part
1
2
( + 

).
(b) If r C
n
is such that [r
j
[ = :
1/2
, show that r

1r =
1
n
Tr  + i(r. r) where
is a sesquilinear form to be determined.
(c) Deduce that there exists a unit vector such that

 =
1
n
Tr .
(d) Show that  is unitarily similar to a matrix of constant diagonal.
199
398. (After D. R. Richman.)
Let / be a eld and : 2 an integer.
(a) Consider a matrix  M
n
(/) of Hessenberg form
 =
_
_
_
_
_
_
_
_
_
_
_
1
.
.
.
.
.
. 1 C
.
.
.
.
.
.
.
.
. 1
.
.
.
.
.
. 1
c
1
c
n
_
_
_
_
_
_
_
_
_
_
_
.
If A is an indeterminate, we form +AJ M
n
(/[A]), where J = diag(1. . . . . 1. 0).
Prove that det( +AJ) = 0 if and only if c
1
= = c
n
= 0. Hint: Induction over
:.
(b) Let j /[A] be a given monic polynomial. Let  be given as above. Deduce that
there exists a vector . /
n
such that the characteristic polynomial of  + c
n
.
T
equal j. Hint: Linearity of the determinant as a function of a row.
399. (After D. R. Richman.)
(a) Let o be an integral ring and j a prime number such that jo = 0. If 1 M
n
(o),
prove that Tr(1
p
) = (Tr 1)
p
. Hint: Work in a splitting eld of the characteristic
polynomial.
(b) Let 1 be an integral ring and :. / 2 integers. If  M
n
(1) is the sum of /-th
powers of matrices 1
j
M
n
(1), prove that for every prime factor j of /, there exists
an r 1 such that Tr  r
p
mod j1.
Nota: Richman gives also a sucient condition for the Waring problem to have a
solution. For instance, when j is prime and j :, the fact that Tr  is a /-th power
mod j1 implies that  is a sum of j-th powers. As an example, every  M
n
(Z)
with : 2 is a sum of squares of integral matrices.
200
Index of scientists (mostly mathematicians)
Albert, 102
Aluthge, 12, 127
Amitsur, 15, 118, 135, 152, 165
Ando, 15, 127
Antezana, 127
Ardakov, 192
Arnold, 83
Aronszajn, 161
Artin, 72
Atiyah (Sir M.), 80
Balian, 7, 56
Ball, 20
Ballantine, 163
Banach, 4, 7, 26, 54, 64, 69, 123, 184
Banachiewicz, 112
Bapat, 129
Barbot, 54
Bataille, 143
Bellman, 14, 145
Benoist, 177
Berger
C. A., 138
L., 144
Bergman, 125
Berman, 197
Bezerra, 174, 189
Birkho, 4, 21, 67, 90, 192
Bochnak, 72
Born, 120
B ottcher, 18, 185
Bougerol, 117
Bourin, 199
Boyd, 96, 158, 164
Brouwer, 91
Brunault, 190
Cauchy, 13, 20, 45, 53, 75, 80, 139
Cavachi, 143
Cayley, 50, 102, 121, 140, 165
Cecioni, 16, 166
Choleski, 37, 187
Cholesky, 195
Churchill, 40
Coste, 49, 72
Costello, 197
Craig, 15, 153
Dacorogna, 71, 85, 185
Daletski, 137, 169
Dedekind, 194
de la Harpe, 55, 67
de Lellis, 127
Delvaux, 102
de Oliveira, 67
Desnanot, 9, 30, 86, 87, 176
Diaconis, 96
Diananda, 174
Dines, 126
Dodgson (see Lewis C.), 30, 176
Domokos, 72, 73
Duncan, 139
Dunford, 11, 107
Euler, 92, 108, 123
Fallat, 150
Farkas, 9, 90
Feit, 183
Ferrand, 109
Fibonacci, 112
Fiedler, 170
Finsler, 122
Flanders, 97
Fokker, 164
Francis, 17, 173, 174
Fredholm, 117
Frobenius, 4, 6, 13, 16, 18, 42, 56, 63, 80,
90, 96, 110, 137, 146, 147, 164, 166,
181, 194, 197
201
Fuglede, 13, 15, 18, 139, 156, 197
Garding, 45, 46, 123
Gaubert, 87
Gauss, 17, 179
Gershgorin, 9, 82
Ghys, 55
Glennie, 102
Goldberg, 137, 139
Golden, 183
Gram, 72, 76
Green, 8, 77
Gro, 23
Guedj, 61
Guionnet, 140
Haagerup, 67
Haar, 55, 194
Hadamard, 16, 18, 105, 137, 148, 150, 162,
164, 169, 194, 195
Hahn, 7, 54, 123, 184
Hall
M., 174
Ph., 21
Halmos, 149
Halperin, 161
Hamilton, 15, 50, 121, 140, 165
Hankel, 72
Hansen, 47
Hardy, 24, 62
Hartman, 184
Hausdor, 28, 103
Hawkins, 146
Heisenberg, 10, 100
Herman, 161
Hessenberg, 5, 15, 30, 159
Higham, 102
Higman, 183
Hilbert, 8, 61, 66, 72, 76, 104, 191
Hildebrant, 144
Hillar, 34, 133, 179
Hodge, 17
Homan, 68
H older, 8, 20
Holland, 147
Horn
Alfred, 10, 99, 199
Roger, 6, 16, 51, 166, 167, 174
Horner, 50, 112
Householder, 86, 144
Hua, Loo-Keng, 10, 12, 14, 101, 125, 145
Hwang, Suk-Geun, 96
Ilyushechkin, 72, 73
Infeld, 120
Jacobi, 4, 9, 11, 17, 29, 30, 34, 86, 87, 113,
176, 178
Jacobson, 140
Jeries, 32
Jensen, 6, 48
John, 27
Johnson
Bill, 177
Charles R., 16, 34, 150, 166, 167, 169,
173
Jordan
Camille, 41, 66, 101, 110, 161, 166
Pascual, 10, 101, 102
Joshi, 158
Jung, Il Bong, 127
Kalman, 5, 39
Karaev, 66
Kawashima, 68
Kayalar, 161
Key, 36
Knuth, 63
Ko, Eungil, 127
K onig, 90
Kostant, 159
Kovacs, 77
Krattenthaler, 82
Krein, 129, 137, 169
Krishnapur, 155
Krzek, 17
Kronecker, 146
Lagrange, 14, 133
202
Lax, 6, 31, 46, 51, 72
Lee, Eun-Yung, 199
Legendre, 15, 153
Lei, Tan, 110
Leibniz, 140
Lemmens, 65
Le Verrier, 6, 11, 50, 110
Levitski, 165
Levitzki, 15, 118, 135, 152
Lewis Caroll, 30
Lipschitz, 8
Loewner, 16, 18, 53, 68, 89, 168, 187
Logofet, 32
Lopatinski, 8, 80
Lorentz, 43, 45
Lu, Zhiqin, 185
Lyapunov, 35
Maciej, 170
Majda, 44
Marechal, 85, 185
Markov, 16, 164
May, 32
Mazur, 4, 26
Mehrabi, 130
Migdol, 186
Milman, 129
Milnor, 122
Minkowski, 118
Mirsky, 141
Moore, 97
Morrison, 19
Motzkin, 147
Mbius (Mobius), 103
Nash, 5, 39, 40
Nevanlinna, 53
Newell, 72
Newman, 174
Newton, 72, 110, 152, 170
Nie, Jiawang, 133
Oppenheim (Sir A.), 14, 149, 150
Parlett, 159
Parrott, 7, 54, 60, 61
Pearcy, 127, 138
Pedersen, 47
Penrose, 97
Perron, 4, 6, 13, 16, 96, 110, 137, 164, 197
Perthame, 87
Piche, 82
Pick, 7, 62
Pisier, 105
Planck, 164
Plemmons, 197
Poincare, 53, 92
P olya, 76
Potter, 74
Preparata, 110
Pujals, 127
Pusz, 113
Putnam, 139, 197
Pyo, Sung-Soo, 96
Ricatti, 115
Richman, 200
Riesz, 4, 105, 115
Ritz, 15, 159
Rosenblum, 24, 197
Rosenthal, 139
Rosset
Myriam, 162
Shmuel, 15, 152, 162
Rota, 85, 86
Roth, 97
Rovniak, 24
Roy, 72
Sakamoto, 15, 153
Sarwate, 110
Satnoianu, 112
Schur, 5, 6, 10, 12, 31, 42, 48, 51, 63, 77, 80,
82, 87, 94, 100, 111, 129, 148, 151,
188
Schwarz, 13, 20, 45, 80, 139
Sedov, 151
Seidel, 17, 179
203
Shapiro, 16, 169, 173
Sherman, 19
Shizuta, 68
Siegel, 14, 117, 146
Silver, 77
Smith, 11, 108
Smyrlis, 188
So, Wasin, 183

Solc, 17

Solcov a, 17
Specht, 134
Staline, 40
Stanley, 170
Starling, 196
Starzhinskii, 40
Stone, 73
Strang, 85, 86, 159
Strassen, 83
Strojano, 127
Sturm, 29
Sun, Jun, 96
Sylvester, 14, 148
Sz. Nagy, 158
Szekelyhidi, 127
Szeg o, 62, 76
Tam, Bit-Shun, 137
Tartar, 41, 69, 72
Taussky, 17, 147, 181
Taylor, 71
Thedy, 10, 101
Thomann, 44
Thompson, 107, 183
Thorin, 4, 105, 115
Thurston, 163
Tian, Yongge, 36
Toepliz, 4, 9, 15, 24, 28, 86, 103
Trenkler, 23
Trotter, 89
Tuni, 178
Turing, 111
Ulianov, 32
Van Barel, 102
van den Driessche, 40, 112
Vandenberghe, 164
Vandermonde, 21, 195
van Gaans, 65
Villani, 134
von Neumann, 79, 15, 57, 58, 68, 69, 73,
85, 95, 141, 156, 161
Wallach, 159
Waring, 19
Weinberger, 31
Weinert, 161
Wenzel, 18, 185
Werner, 189
Weyl, 4, 5, 147
Wiegmann, 147, 148
Wielandt, 68, 97, 98
Williams, 77
Wimmer, 40, 97
Wintner, 184
Wojtkowski, 117
Woronowicz, 113
Xiao, Lin, 96
Yakubovich, 40
Yamamoto, 106
Young, 197
Zariski, 13, 135
Zaslavsky, 137
Zassenhaus, 17, 181
204
Index of stamps
Banach, Stefan (Poland), 25
Born, Max (Rep. de Guinee), 119
Brouwer, Luitzen (Holland), 90
Cauchy, Augustin (France), 75
de Lellis, Camillo (Vatican state), 127
Euclid (Vatican state), 46
Euler, Leonhard (Switzerland), 91, 107
Fibonacci, Leonardo (Dominica), 111
Gau, Carl Friedrich (East Germany), 178
golden ratio (Macao), 112
Hadamard, Jacques (Rep. de Guinee), 161
Hamilton, Rowan (Eire), 49, 116
Heisenberg, Werner (Germany), 99
Hilbert, David (D. R. of Congo), 75
Hua, Loo-Keng (China), 100
Lagrange, Joseph-Louis (France), 132
Le Verrier, Urbain (France), 110
Leibniz, Gottfried (Romania), 139
Lorentz, Hendrik (Holland), 44
Lyapunov, Alexandr M. (USSR), 34
Newton, Isaac
(France), 71
(Paraguay), 151
Poincare, Henri (France), 52
Turing, Alan (Portugal), 110
von Neumann, John
(Hungary), 84, 155
(Portugal), 110
(Tatarstan), 94
(USA), 56
205