Diagonals of operators Majorization, a Schur–Horn theorem and zero-diagonal idempotents
A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Department of Mathematical Sciences McMicken College of Arts and Sciences University of Cincinnati, June 2016
Author: Jireh Loreaux Chair: Gary Weiss, Ph.D. Degrees: B.A. Mathematics, 2010, Committee: Jonathan Brown, Ph.D. University of Cincinnati Victor Kaftal, Ph.D. Costel Peligrad, Ph.D. Shuang Zhang, Ph.D. Abstract
This dissertation investigates two related but mostly independent topics. A diagonal of an operator is the sequence which appears on the diagonal of the matrix representation with respect to a prescribed orthonormal basis.
The classical Schur–Horn Theorem provides a characterization of the set of diagonals of a
n n selfadjoint operator on C (or R ) in terms of its eigenvalues. The relation which provides this description is called majorization. New kinds of majorization (p-majorization and approximate p-majorization) are defined herein, and it is shown that they characterize the diagonals of positive compact operators with infinite dimensional kernel. When the kernel is finite dimensional the diagonals lie between two closely related classes of sequences: those whose zero set differs in cardinality from the zero set of the eigenvalues by p and are also p-majorized (or in the case of the second set, approximately p-majorized) by the eigenvalue sequence. These results extend the work of Kaftal and Weiss in 2010 which describes precisely the diagonals of those positive compact operators with kernel zero.
Zero-diagonal operators are those which have a constant zero diagonal in some or- thonormal basis. Several equivalences are obtained herein for an idempotent operator to be zero-diagonal including one which is basis independent, answering a 2013 question of Jasper about the existence of nonzero zero-diagonal idempotents. These techniques are extended to prove that any bounded sequence of complex numbers appears as the diagonal of some idempotent, and that the diagonals of finite rank idempotent operators are those complex-valued absolutely summable sequences whose sum is a positive integer.
ii c 2016 by Jireh Loreaux. All rights reserved.
Acknowledgments
The author was supported by the University of Cincinnati Department of Mathematical
Sciences, the Charles Phelps Taft Research Center, the Choose Ohio First Scholarship
Program, as well as the following conferences: Non-Commutative Geometry and Operator
Algebras; Southeastern Analysis Meeting; Great Plains Operator Theory Symposium;
International Conference on Operator Theory; East Coast Operator Algebras Symposium;
Wabash Modern Analysis Seminar; and Singular Traces: Theory and Applications hosted by the Centre International de Rencontres Math´ematiques.
I would like to thank my advisor, Gary Weiss, for his continual prodding, thorough feedback and valuable insights, without which I might never have completed this disserta- tion. I am appreciative of the candor with which he communicates both commendation and exhortation. I am also indebted to him for our many disagreements over writing style and mathematical reasoning which resulted in more economical, precise and lucid exposition. Moreover, his support is a primary reason I remained at the University of
Cincinnati for graduate school.
I would like to acknowledge Victor Kaftal for devoting his time and energy to instruct me on the subject of von Neumann algebras even when no formal course could be offered within the department. I must also thank Shuang Zhang for siphoning time from his busy schedule as Department Head to teach me the basics of C∗-algebras.
iv I owe my gratitude to my committee for bothering to read, edit and critique this dissertation. I thank the constituents of the functional analysis seminar for listening to me ramble seemingly unceasingly concerning these results.
I am grateful to Daniel Beltit¸˘afor introducing me to the work of Fan, Fong and Herrero, without which my research on diagonals of idempotents may never have flourished. I am indebted to John Jasper for suggesting to me at GPOTS 2013 several of the major problems of study in this dissertation. I thank Jean-Christophe Bourin for communicating that Theorem 4.2.7 is a corollary of his pinching theorem [Bou03, Theorem 2.1].
My friends in the department, especially Sasmita Patnaik, Marcos Lopez, Michael
Snook, Dewey Estep, Seth Ball and Catalin Dragan deserve special mention for listening to my complaints and providing ample distraction from work when the need arose. Their laughter and companionship were essential to slogging through coursework and research daily. I owe Michael Snook a great deal for the many fruitful discussions concerning libre software, GNU/Linux, cryptography, and of course, LATEX. I would not be where I am today were it not for my parents, Richard and Debra Loreaux, who instilled in me a passion for mathematics and made great sacrifices to provide me with the best opportunities to pursue this career. Their love, support and encouragement have been mainstays during my journey long before I entered the University of Cincinnati.
And finally, the most influential people in my life: my wife, Katie, and my daughters
Dara and Anastacia. Words cannot adequately express Kate’s role in this arena. She pushed me onward when I felt as though I wanted to quit; she never doubted me for a second and encouraged me daily; most importantly, she has never given up on me in spite of my many failings. As for my daughters, they have inspired me and made sure
I always had my hands full. Their smiles and laughter provide me with an unending source of joy and happiness.
“Great are the works of the Lord; They are studied by all who delight in them.”
Psalm 111:2
v This dissertation is dedicated in memoriam to my father
$N$N$$N$$N$$$N$N$$$$$$$$$$$$$$$$$$$$$@$$@$$$@$@$@$@$@$@$$@@$@@$@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@p} N$$N$Q$N$$$N$$$N$$$N$$$$$$N$$$$$$$$$$$$$$$$@$$$$@$$@$@$@$@$$@$@$@$@$@$$@$@$@$@$@$@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@pl $N$Q@$$$$N$$N$N$$$N$$$$$N$$$$$$$$$$$$$@$$$$$$$@$@$@$@$$@$$$@$@$@@$@@@$@@$@$@@$@@@@@@$@$@$@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@pl $N$$$N$N$$$Q@N$$N$$$$N$$$$$$$$$$$$$$@$$$@$@$@$$@$@$@$$@$@@@$@$@$@@$@$@@$@@@$@$@@$@$@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$l N$$N$$$N$N$$$$$N$$$$$$$$Q@$$$$$$$$$@$$$@$$$$$$@$$$$@$@$@$$$@$@$@$@@@$@$@@$@@@@@@@@@@@@$@@@@$@@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@pl $$N$$Q$$N$$N$$$$$$$N$$$$$$$$$$$@$@$$$$@$$$@$$@$$@@$$@$@$$@@@$@@$@$@@@@$@$@$@$@$@$$@$@@N@@@$@@@$@@@$@@@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$l $N$$$$$$$N$$N$$N$$N$$$$$$$$$$$$$$$$@$@$$@$$@$$@$$$@$@@@$@$$$@$@$@@$@$@@@$@$@$@$@$@QQ@$@$Q$@$@$@$$$Q@$$@@$@$$$$@@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$} $$N$N$$$$$N$$N$$$N$$$$$$$$$$$$@$@$$$$$@$$@$$$@$@$@$@$$$@@$@@$@@$@$@@$$$$@$@$$@$$$Q@QQQEQ@Q@$QQQ@Qp@pQ@Q@$Q@$Q@Q@$@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@} $N$N$$$$N$$$$$$N$$$$$$$$$$$$$$$$$$$@$@$$$$$@$$$$@$$$$@@$@$@$@@$@$@@$@@@@$$Q$QQggagZgcoOaza2oAOo222avgSQQQQQQ@Q@@@N@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$] $N$$$$$N$$$N$$$$$$$$$$$$$$$@$@$$$@$$$$$@$$@$@$@$$$@@$@$@@@$@@$@$@$@@@$$$$N@QQXa2jp/=-='' 7jj:""--eZ@jz2ZQQQQQ@@$$@@$@Q@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$N$N$$$$$$N$N$$$$$$$$$$$$$$$$$$$@$@$$$@$@$@$@@$@$@@$@$@$@@@$@@@@QQAXgajp<^--7 ;j:::: 2 .;4u/o\zzgQQQ$Q@@@$X@$@@@$@Q@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$] $N$N$$$$$$Q@Q@$$$$$$$N$$$$@$@$$@@$$$$$@$@$@$@$$$$@$@$$@@$$$@Q@$@QXgvv"-=>7: 7-%-@zzzQpQ$QxZ@Q@$@Q$Q@@@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$N$$$$$$$$$$$$$$$$@$$$$@$$@$@$@$$$@$$@$@@@$@$@$@N@Q@Q@QQ0>5"-j' ''70@jgQzDgQZQQQQXQ@$@$@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $N$N$$$$$$$$$$$$$$$$$$$$$$$$@$$$@$@$$$@$@$$@$@$@$@$$@$$$Q$QAv^`7a 7--MuaaxuzAQgXQQ@@Q@Q@@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$N$$$$N$$$$$$$$$$$$$$@$$$$@$$$$$@$@$@@$@$@$@$$@$$$QQQrT# ' 7(pTT@$OjzgzzzgQQQ$Q@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@$] $$$N$$$$$$$$$$$$$$$$$$$$$$$@$$$@$@$@$@$@$$@$@$$$@$Q$gQST-7 7 : j7"=-5-=MzggQ@$@@Q@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$$$$$$$$$$$$@$@$@$$@$$@$$$$$@$@$@$@@@$NQQ@X>' "jR4GA@$@@@@@@@$@@@@$@@@@$@@@@$@@@@$@@@@$@@@] $N$$$$$$$$$$$$$$$$$$$@$$$$$$$$$@$$@@$@$@$$$@$$$$$@S* .3@zQQ@@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$$$$$$$@$$$$$$$@$$@$@$$@$$$@$@$@$$Q@$$D"' 7{Jz$@@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$@$$$$@$$$$$@$$$$$@$$$@$$$@$@$@$$$@@$Ay` . . 73zX@Q@N@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$@N$$@$$$$$$@$$$$@$$$@$$@$@$@$$$@@$@NQP q ''7"vQQ@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$$$$$@$$@$$$$@$$$$@$$@$$@$@$@$@$$@D} 74zQ@$@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$@$@$$@$$$@$@$$@$@$$@$@$@$@$$@$$Xzr! (7@@$@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$@$$@$$$$@$$@$$$$$@$$$@$$@$@$@$@@$@QQv' '7j\gQ@@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$$$$$$$@$$@$@$$@$@$$@$@$@$$@$$@Ar .7"ggQ@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$@$$$@$@$$$@$$$@$$@$$@$@$@$@@@$pa" 7@ZQ@$@@@@@@@@@@@@@@@@@@@@@@@] @$$$$$@$@$$@$$$$$@$$$@$$@$$@$@$@$@$@@AZy! 7dX@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$$$@$@$$$$@$@$@$@$@$@$@$@$NQQxp; ' : ;-{Q$@@@@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$@$$$$$@$@$@$@$@$$@$@$@$@@$$Qr; 2 .Q __: qz; . 74@@@$@@@@@@@@@@@@@@@@@@@@] $$$$$$$@$$$$$@$@$@$$$$@$@$@$@$@$@$$$@x$. ; "$ Z7 .1j:r,::j . : ' .{CQ@$@@@@@@@@@@@@@@@@@@@@] $$$$$$$$$@$@$$@$$$$@@$@$@$@@@$@$@@$@Qzl ,. :;- q.-Z.zgQQy"/#"Q_nj, _ \d@@@@@@@@@@@@@@@@@@@@@@] $$$$$@$@$$$$@@$$@@$$$@$@$@$$$@@$@@@$@y @_2\\2j _;2,jzzj:Q@Q$}# ;';="7#; - .p ; ' 72Q@@@@@@@@@@@@@@@@@@@@@] $$@$$$$$$$@$$$$@$$$@$@$@$@@@$@$@@$@$Qr '"jjzu::Q:::QQ:QQQj$jjj@p.;: ;.-' .7.Qqg..@j/ q , ; :; . (Z@@@@@@@@@@@@@@@@@@@@@] $@$$@$$@$@$$@$@$@$@$$@$@@$$@$@@$@@@$Qr q..:Q:Q@QQQ@::Q@N@jj@j@jjjjQQ__Q$r jqq,7:::.2c.. @z;q_,q;Q; ,z ' (@Q@@@@@@@@@@@@@@@@@@@] $$$$$$$$$$@$@$$$@$@$@$@$@$@$@@$$@$@$pl .._1Q:Q@@@Q$@@@@@@@@@jjjjjj2jjjj2j2j$QQQy/Q;\32$dA} : ' Z; 1Q:g:jta' {Q@@@@@@@@@@@@@@@@@@@] $@$$$@$$@$$@$$@@$$@$$@$@$@@$@$@$@@$Qr aq1u:Q@@j@@j@@@@j@j@jj@jj2jjj2j22jj@2j22jQ\@QQ$@@pp}...:j :jQQ$$Q/j q 7 {Q@@@@@@@@@@@@@@@@@@@] @$$$@$$@$$@$$@$$$@$@$@$@$@$@$@$@@@@$a) :;jza@@@j@@@@Q@@j@@@j@@jj2jjjj2222222j22jj2jjjjjjj2QjQQQ:..Qj$jjj@Q:j 7' d@$@@@@@@@@@@@@@@@@@] $$$@$$$$$@$$$$$@@$@$@$@$@@$$$$@$@$$$a ''@jaQQ@j@QN@@j\Qj@jjj@jjjj\j22jjj2jj2jjjj2222u2jjj2j2ajj\j22j2jjj@\Q:, Q;, q, (@$@@@@@@@@@@@@@@@@@] $@$$$@$@$$$@@$@$$@$$@$$@$$@@$@$@@@@$r : 7"2:@@@Qj@@@@jjj@j@j@j@jjjjj222jj222j222jjjj@jjjj2jjjjj2jj2jjjjjjjjQp_,{z2, .g dQQ@@@@@@@@@@@@@@@@@] $$$@$$@$@$@$$@$@@$@$@$@$@$@$@$@$@$@Q) q#uQG@@j@@j@j@@@@jjj@jjjjjj2jj2222j22j2j2jjjjjj2jjj222jj2j2jjjjj@jjj\Q@2QQ:,qNg -Z@@@@@@@@@@@@@@@@@@] $@$$$@$$$@$@$@$@$@$@$@$@$@$@$@@$@$@$) a(u:QQ@$$@@@@Qj@@j@@jjjjjjjjjja2juz2a2j2j2jj2jj222222jjjjjj2jjjjjjj\j@@Qj\QQ._;,j {1@@@@@@@@@@@@@@@@@@] $$$@$$$@$@$@@$@@$@$$@@$@@$@$@$@@@@@A, :qja:Q$@Q$@@@Qj@@@jjjjjjjj222u2a22222juz2u2j2jjjj2jj2j2jjjjjjjjjjj@jj@jj@jjjjjNQr, ,7 (aq$@@@@@@@@@@@@@@@@] @$@$$@$@$$@$@$@$@$@$@$@$@@@@$@@@$@$Q} j:{z2ga:$@@@@j@@jjjjjj2222222a2aauaa2au2juc22222jj2222j2jjjjj@jjjjjj@j@@jj@j$@Q$QL_: . {1Q@@@@@@@@@@@@@@@@@] $@$$@$$$$@$@$@$$@$@@$@@$@@$@@@$@@@$$} :.2222zu:Q@@@jj@jjjjj2j2j2j22aaaazzaaua22222j222j2222222j2jj2jjjj2@j@j@j@2jj@jjj@@:Qa" quQ@@@@@@@@@@@@@@@@@] @N@$$$@$@$@$@$@$@@$@$@$@@$@@$@$@@$@X; :jgu22a2QQ@@jjQ@@@j@jjj22222zuauaaaaaaaaz2aaa2222222222222j2jj2j2jjjjjj@j@jjjj@j@$pQ:;; #12Q@@@@@@@@@@@@@@@@@] $$$$@$$$$@$@$@$@@$$@@$@@$@@$@@@@@$Qx' :ujza22u:Q@QQKQu22j::sQ$j@@$jj2za2auazaaazaaau22222u222z2a222j2j2j222jjjj2jj@@j@jQjQQ:Qj, 'qjZ@@@@@@@@@@@@@@@@@] @$@$$$@$@$$$@$@$$@$@$@@$@@$@@@$@@$Q+ .:j2aa2jj::g:Zy\\"c2c:"C=74Ha--RK@j22aa2azaaaaajuu22aaazaaaua2222j2222jjjjjjjjjj@@j$QQ:2}: '1QgQ@@@@@@@@@@@@@@@@] $$$@$@$$@$@@$$$@$$@$@$@$@$@@$@@@$$Ab; .2jj22zjuj\{"jj:}\rjjj>:''; .7#-"oG@jj2ja2za2a2az2a2gzaz2a22zuz2a22j22j2j2@jjj@$@@Q:yja$: .2@g@@@@@@@@@@@@@@@@@] @@$$@$$@$@$$$@@$@$@$@$@@$@@@@@$ptQQQ, :222az2zj2u2--Q-;jj"{\.2:.q__:__,; z-3jQj22aaaaaaaau22aaa2u22ua2ua222j2jjjjjjjj@QjQ@Qnr2j7 .u@zQ@@@@@@@@@@@@@@@@] $$$@$$$@$@$@$@$@$@$$@$@$@$@$@$@I;"G@p} :j2gza:gr2co;Q,q"j2.QQj2j2j2jj2QjQ:L_./{aZ@@jj222z222jjj22\a222jjj2ajjj2jjj2j2jjj@@@j@Qnjj:j "QjaQ@@@@@@@@@@@@@@@@] @$@$@$$$@$@$@$@$@$@@$@@$@@$@$@@IQ:23Qi .a:g:gza2a2j.,:j\a:Q@@Q@@@j@j@j$j@@j$N:::zzKj@jjC2222jj$j@@@$QKsBwY:jzKQpA@QjK@@$@jj@@Q:j""@. q\1jz@@@@@@@@@@@@@@@@@] $$@$@$$@$@$@$$@$@@$@$@$@@$@@@@@xjNQr/:, jag:::az:2c:jjj:jj2""""#7: -"===4AQQ:yuZQ$@j2j2j22\@jpQ:i\.Q.:j :77-7oHaS:2ZsGSM@@@Qa};": q\qjQ@$@@@@@@@@@@@@@@@] @$@$@$@$@$$$@@$@$$@$@@@$@$@$@$$x$@p\Q:}a .nz:Q:yqQQQQgr:"$p; ., _ Q. :#:"2""ZQjjjuu222j@QQQZ}\j\_____q,z;7$\#\02ur22G@$Qa"j@: juqgQ@@@@@@@@@@@@@@@@@] $$$@$$@$@$@$@$@$@@@@$@$@@$@@$@Nzr@:q@py; '_Qgg::Q:QQA:g"}-Qj 7:: jaL, .,..jZ\:Qj22222j$QQKorjc:QQj@j2j222\__:ju\2j:2j;Kga:p@p {jujg@@@@@@@@@@@@@@@@@@] @$@$@$@$@$$@$@$@$$@$@@@$@@$@$@@upQQ@@$z; .ungz:QQ$NQ@Qnr\j{;_{22r N ;__i,,#RR->! 24:I: :j@jjjjjjpp&y-:{j::@@@@jj222aaazj@Q;S2_\_j2"j:::@ 73j:jQQ@@@@@@@@@@@@@@@@@] $$$@$$@$@$@$@$@$@@$@@$@@@$@@@$$j}g$@@\}'; @:2azzoQQQQ@K::2zuZQczaz22QQQ_Qp.p2:.___j12:$jQ:{@$@j2jpQo@j;7#j:ja::7=--"ju22jjj\:Q2",:;:{!:::g ;:\jujQ@@@@@@@@@@@@@@@@@@] @$@$@$@$@$@$@$@@$@$@@@$@@@$@$@$2$@j@j@Q, :jajajzQ$@Q@:QQQQQQ$@Q$::gQQ:N@@@QQ1aj:jj2:';":"qQjjj$\@Ij {K:T7q__Q$ (-@@@@Q$gj_7 g:$"::j j:jj:2Z@@@@@@@@@@@@@@@@@@] $@$@$@$@$@$@$@$@$@@$@$@@@$@@@$@rQz2j@j$i_p.ajj{juzgQ@Qp@@@QQQQg:QK$$AQj@j@j@@@QQua{#.\j:":1Q@j@jjjn"2p'2n_\'"7@u} ! ,7-&::a2y:7q;"q{:. a_{"2qQQ@@@@@@@@@@@@@@@@@@] @$$$@$$@$@$@@@@$@@$@$@@$@$@$@@$Iquz2jQQj:"2jj"j\zQQQQQj@$@@jj$pQpQQ:QQQQpKKQMA:K222\\1jj2gQQ$Q@@@@:;;N.{\pp_.;/j__.' q1Z, 7-ur";:2j\";q:' 7ja2QzQ@@@@@@@@@@@@@@@@@@@] $$@$$@$@$$@$@$@$@@@$@@$@@@@@@$@r@2jj2\pQr"jj"j\2uQNQ@@Q@@j@@@jj2jjj@jjQpQQQQ:z:2:a212a2ujzz::QQ@$Ny@Z:jpoj@@:zz:Qj;Q:\7-#Q. 7{j/":{\:"\}j 2.:::jQ@@@@@@@@@@@@@@@@@@@@] @$$@$$@$$@$@$@$@@$@@$@@@$@$@$@QIZ@QAQE@f"2jjj":2:QQQ@QQ@@@@j@j@@j@jjj$jjj@jQ$NQ@QQQ@Q@Q;auac:QpQppc\Qqj;:(ZQ:::2S#@@Q@QQ_L\., :2\2j\1:r:; j,""jQzQ@@@@@@@@@@@@@@@@@@@@] $$@$$@$@$@$@@@@@$@@$@@$@@$@@$@@I"Q$:2Qrj""2"::j2aQQQ$Q@@@@j\j@jjj2jjjj2jjjjj2jjjj2j@Qp2juc:QQj@j@j@n;:j2\:;d@jjjppQQ:;uK&:Qa;L,q#"{zz:;2jp2 .--zj"j@Q@@@@@@@@@@@@@@@@@@@@@] @$@$@$@$@@$@$@$@$@$@@$@@$@@$@$@f:1jj@Q\:{:u":"aa:QQQ$$@@@@jj@jj222222a222jj2jjjjjpn=22\uQQ@@@j@j@j@Q;"1zau2\"CZ@jrjpjpQQQQ@N$QjQ:_.1dQ:Q\, : 7!._yj\Q2Q@$@@@@@@@@@@@@@@@@@@@] $@$$$@@$$$@$$@$@$@@@$@$@$@$@@@$QQQjj@2::j:2:jja::aZQQ$@@jjjjjjj22222aaa2222j@jjErj\2u2QQ@Qjjjjjjj@@jp/jQ@@QQQ:\ao4K@$@@j@j@jQ@Q@\pQ;uQQb\$ ' _Q:"j\:jQ@@@@@@@@@@@@@@@@@@@@@@] @$$@$@$$@$@$@@$@@@$@$@@$@@@@@$@@2QQ@$n::\jj"j2u:&2zQ@@@@jjj@jj2ju222222jj$@$$ssrjz:g:QQ@jj22jj2jjjjj@i"z@j@@@@@@Q:::::::ZQ$QQ@jj;$$Q::&r:: {j$\jj1qgQ@@@@@@@@@@@@@@@@@@@@@@] $@$@$@$@$@$@$@@@$@@@$@$@@@$@@@@$rQNQ@::/jjjjjzgz2g::QQ$jjjjjjjjj2j2jjj@@Q$:Q:a2Q"ou2gzajAjjagzaga2jjQp::u@2jjjjjQ2j@jQQ@Q@@@Q@@QQ@@@@Qaj@' qq$:j\jqzQ@@@@@@@@@@@@@@@@@@@@@@@] @$$@$@$@$@@$@@$@@$@$@$@$@$@$@$@@I@ngr@:j2;jjzzQg2z:QQ@j@j@$j2jj22jjj$@Q:QQj$::jpz7#--""jju@jjaaZcjjp:2":(Qjj2a2j2jjjjj22jjjjjj@@@@pp@@s{$ q:jA\j:1jzN@@@@@@@@@@@@@@@@@@@@@@@] $@$$@$@$@$@$@@@@$@@$@@$@@@@@@@@@nS:Q::\j2jaugzzQ2zzQQ$@@@jjjjj$@@@@QQp@@j@K&Q:jj;,@"rj,,71z@j@j@@B"jjj:j/1Qjjaaaau222jj2jjjj@j$@@QQ@QQ::, q/($2jj2qzQ@@@@@@@@@@@@@@@@@@@@@@@@] @$@$@$$@@$@@$@$@$@$@@$@@$@@$@@@@Q;::::\\jjugg:::QQQQQQ$@jjjjj@@$n::Qpj@@@Q:g2j2;/;Q"{2j:\Q\azQ:n\7Q7az::Z:uuQ22azaz22222jjjjj@jQ$@@pQ:aq: _.:;1i:\"2qzQ@@@@@@@@@@@@@@@@@@@@@@@@] @$@$@@$@$@@$@$@@@@$@@@@$@$@@@@$$Q]\}@:j22atz2ZQQQ@$ggQN@jj@$K5"j:QjQjjQQpA:ggz:Qyj.:1ju21a.:#"} q,j#- ./z:QgZ@22a22222@j2@j@@@@QQ::gj:: .Qj$zajjj:jqg@@@@@@@@@@@@@@@@@@@@@@@@@] $$$$$@$@$@$@$@@$$@@$@$@@@@@@@@@@$Q_"@\{z2Z:Q2::Q@@$z:@Qjj@@0_uzQ@@jj$@$QE:Q:Q;1zz:222z:I:\a;a2j"{# Z:2rzgQQQ:Q@j22222222jjj@@jQ@Kzz;;: Jaa$Aj:j\jQzQ@@@@@@@@@@@@@@@@@@@@@@@@@] @$@@@$@$@@$@$@$@@$@$@@$@$@@$@$@@$Qgp:::azg::QQQN@@[email protected]@j@@QQ::QQQQQQz::Qj$QpQ:::::\jjpN:;ag,:u22G@jQ:Z$@j222j2jjj@@@ppp2aj#Q {@jAa"j"\QQQ@@@@@@@@@@@@@@@@@@@@@@@@@@] $@$$$@$@$@$@@$@@$@$@@$@@@@@@@@@@@QQp:q\;z::QQQQ@jQ:Z$@@jp::QQj@@QZKspjoy=9=VuQQQ@@@@$@@@@@@Q$:n\";j:;jj2ajcQ@Q@QQ@j2jjjj$j@@$Q@Qg2j># {@pyj{{{Qz@@@@@@@@@@@@@@@@@@@@@@@@@@@@] @$@@@$@$@$@$$@$@@$@@@@$@@@$@@@@$@$Qr::jazQQQQQQj@@:QQ@jjpQ@@@@QNyjZj;: :jj ,z#j:"7VY@jjQQjQQ@pQg:zgnc2Q\j:Qj@@pQ$Qj@@j@@@j@@QQaCj:p .;;r2jj2:jQ@@@@@@@@@@@@@@@@@@@@@@@@@@@@] @$$$$$@$@$@$@$@$@$@$@$@@$@@@$@@@@NQrZ:"2u:::Q@jj@@QQ@\@j@@@jjQ@QQn_. _,: .j\..__}::j/5>qjpHAAQNQQQQ:uQg2azQ@j@j:QQj@j@j@@@QQ::z":pj {Eo2\{jj1Q@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] @$$@$@$@$@$@$@$@$@@$@@$@@@@@@@@@@QQa;::ja:QQQQ@@@@Q:$@@jN@jjj@j@jNQ2, -Mjp1gzazQZr{gz_.@@-Z#Q@:7:--"3c&z21ZQ@jQAQK$@@@j@@@$@Qzj2:-@. {y2j:jjqQQ@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] $$@@$$$@$@$@$@$@@$@$@$@$@$@$@$@@$@$z}:"jaQQgd@@jjQQQ@@\@@jjj@@j@@@@::i'Z-<@aaagQZcgQQ@g21gga_,__z 'z::.#"oz::QjQ:ru@@@p@@QQQ\aj{":Q ; (2jjjjqgQ@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@] @$@$$@$@$@@$@$@$@@$@@@@@@@@@@@@@@@QQ\zg:j2:g:QjQ@@@Q@j@@j@j@@@$Q@@QQQ:2;,,z 7 Abstract ii Copyright iii Acknowledgments iv Dedication vi Notation ix 1. Introduction 1 1.1. What are diagonals of operators? ...... 1 1.2. Information exchange between operators and diagonals ...... 4 1.3. Description of results ...... 9 1.4. Notation ...... 11 2. Background 13 2.1. Majorization and extensions of the Schur–Horn Theorem ...... 13 2.1.1. Recent History ...... 15 2.1.2. Main contribution ...... 22 vii 2.2. Diagonals of idempotents and frames ...... 23 2.2.1. Projections and idempotents ...... 23 2.2.2. General techniques for diagonals of non-selfadjoint operators . . . 26 2.2.3. Jasper’s problem and the relation to frames ...... 29 3. Schur–Horn theorems 31 3.1. Preliminaries on stochastic matrices ...... 33 3.2. p-majorization (sufficiency) ...... 36 3.2.1. Discussion on majorizations ...... 47 3.3. Approximate p-majorization (necessity) ...... 54 3.4. E( (A)) convexity ...... 62 U 4. Diagonals of idempotents 66 4.1. Zero-diagonal idempotents ...... 67 4.2. Diagonals of the class of idempotents and applications ...... 84 4.3. Diagonals of the class of finite rank idempotents ...... 91 Bibliography 96 A. Prerequisite material 102 A.1. Hilbert spaces ...... 102 A.2. Bounded operators on Hilbert spaces ...... 104 A.3. Trace-class and Hilbert–Schmidt operators ...... 107 B. A primer on zero-diagonal operators 108 B.1. Zero-diagonal operators are “trace-zero” operators ...... 109 B.2. The shape of the set of traces ...... 118 Index 125 viii Notation Functions R Real numbers Tr The standard trace on B( ), Nonnegative real numbers H R+ page 4 C Complex numbers diag The normal -isomorphism h ∗ from `∞ to , pages 2, 3 Ah Operators EA Conditional expectation onto , A Mn(K) n n matrices over a field K page 2 × W (T ) Numerical range of the operator B( ) Bounded operators on a Hilbert H T , page 6 space, page 1 W (T ) Essential numerical range of the Ideal of compact operators e K operator T , page 111 + Positive elements of σ(T ) Spectrum of the operator T , K K page 98 1 Ideal of trace-class operators, L page 100 R Tr T Set of traces of an operators T , { } page 27 2 Ideal of Hilbert–Schmidt L Number classes operators, page 100 A masa (maximal abelian Positive integers A N -subalgebra) in B( ), page 2 ∗ H Z Integers The diagonal masa in B( ) Ah H with respect to the basis h, Z+ Nonnegative integers page 2 ix (A) Unitary orbit of the operator A, `2 Complex-valued U page 19 square-summable sequences, page 97 (A) Partial isometry orbit of the V operator A, page 19 `∞ Bounded complex-valued sequences RA Range projection of the operator A, page 12 ξ, η Sequences related by majorization Re T Real part of T , page 98 ξ∗ Decreasing rearrangement of a Im T Imaginary part of T , page 98 sequence ξ c+, page 11 ∈ 0 [S, T ] Commutator ST TS, page 5 s(A) Singular value sequence sn(A) − h i of the compact operator A, Relations page 11 Majorization, pages 7, 14, 16 Set operations ≺ p p-majorization, page 37 ≺ co(S) Convex hull of a set S in a vector space -p Approximate p-majorization, page 54 S◦ Interior of a set S Strong majorization, page 15 4 S Closure of a set S Weak majorization, page 15 ≺≺ Vectors Sequences Hilbert space, pages 1, 95 H dn A sequence h i e, f, g, h Orthonormal sets in H c0 Complex-valued sequences converging to zero, page 11 (x, y) Inner product of x, y , page ∈ H 95 + c0 Nonnegative sequences converging to zero, page 11 span x closed linear span of x ∗ c0 Nonnegative nonincreasing [x, y] The line segment joining x, y sequences converging to zero, paramaterized by tx + (1 t)y − page 11 for 0 t 1 ≤ ≤ `1 Complex-valued absolutely x y The rank-one operator ⊗ summable sequences z (z, y)x 7→ x Chapter 1 Introduction 1.1. What are diagonals of operators? Consider a separable complex Hilbert space and the -algebra B( ) of bounded H ∗ H (continuous) linear operators. Later we will be concerned primarily with the case when is infinite dimensional, but for now we will allow both finite and infinite dimensional H . For convenience, denote by N := dim the Hilbert dimension of (as opposed to H H H the algebraic dimension). Let ( , ) : C denote the inner product on . If the · · H × H → H reader is unfamiliar with these topics please consult Appendix A before continuing for a brief overview of relevant concepts. N Given an operator T B( ) and an orthonormal basis h = hi , we can define ∈ H { }i=1 the matrix representation [T ]h with respect to the basis h whose ij-th entry is given by (T hj, hi). Then the diagonal of T with respect to the basis h is the (finite or infinite) ∞ bounded sequence (T hi, hi) ` . This is a natural way to view the diagonal sequences h i ∈ of an operator. In fact, when N < this is precisely the standard matrix representation ∞ from linear algebra. There is another, perhaps more elegant, way to view the diagonals (i.e., diagonal sequences) of an operator T . Fix an orthonormal basis h for , and consider the H 1 operators of the form[1] N X diag ai := ai(hi hi), h h i ⊗ i=1 ∞ for scalar sequences ai ` . The matrix representation of diag ai with respect to h i ∈ h h i the basis h is a1 0 0 ··· 0 a2 0 ··· 0 0 a3 ··· . . . . . . . .. and for this reason such operators are called diagonal operators. It is clear that they form an abelian -subalgebra of B( ), and in fact is maximal with respect to these ∗ Ah H Ah properties, and hence it is called a masa[2] (maximal abelian selfadjoint subalgebra) of B( ). H There is a natural map EA : B( ) h defined as h H → A N X EA (T ) := (T hi, hi)(hi hi). h ⊗ i=1 The map EA is the unique trace-preserving conditional expectation onto . Moreover, h Ah this map is faithful (i.e., T 0 and E(T ) = 0 imply T = 0) and normal (i.e., Tn T ≥ ↑ in the strong operator topology implies E(Tn) E(T )). With respect to the matrix ↑ representation [T ]h, the conditional expectation EAh consists of changing the off-diagonal entries of T to zero. Thus we may also refer to the collection of EAh (T ) as h varies over all orthonormal bases of as the diagonals of T . This paradigm has the added benefit H (over sequences) that we never leave the setting of B( ). Another benefit to viewing the H [1] For vectors x, y ∈ H, the tensor product x ⊗ y is the rank-one operator on B(H) which acts on a vector z ∈ H by (x ⊗ y)z = (z, x)y. Thus hi ⊗ hi is the rank-one projection onto the vector hi. [2] The cautious reader may note that there are masas A in B(H) which cannot be obtained as Ah for some orthonormal basis h. However, all atomic masas (i.e., those which are type I as von Neumann algebras) do arise in this manner, and thus these are natural masas to consider because B(H) is a type I factor. Moreover, these are the only masas of B(H) which admit trace-preserving conditional expectations. 2 diagonals of T in this manner is that this concept can be naturally extended to other von Neumann algebras, but such a discussion is outside our scope. We may refine this view as follows: Given an orthonormal basis h and an operator U B( ), the collection Uh forms an orthonormal basis if and only if U is unitary (i.e., ∈ H UU ∗ = U ∗U = 1). Furthermore, if h0 is any other orthonormal basis, there is a unitary U satisfying h = Uh0. In this context, for each unitary U we find N ∗ X ∗ EA (UTU ) = (UTU hi, hi)(hi hi) h ⊗ i=1 N X ∗ ∗ 0 0 = (TU hi,U hi)(Uh Uh ) i ⊗ i i=1 (1.1.1) N X = (T h0 , h0 )U(h0 h0 )U ∗ i i i ⊗ i i=1 ∗ = UEAh0 (T )U . Letting (T ) denote the unitary orbit of T under conjugation (i.e., UTU ∗ U unitary ), U { | } we find that every diagonal of T is unitarily equivalent to a member of EA ( (T )). Thus h U we may view the diagonals of T instead as elements in this latter set. Moreover, the -isomorphism diag : `∞ given by ∗ h → Ah N X diag ai = ai(hi hi) h h i ⊗ i=1 takes the diagonal sequences of T coming from the matrix representations of T to the diagonals viewed as elements of EA ( (T )). Furthermore, (1.1.1) allows us to conclude h U ∗ ∗ UEA ( (T ))U = EA (U (T )U ) = EA ( (T )). h0 U h U h U Therefore the diagonals of T in one basis are unitarily equivalent to the diagonals in any other basis, so there is no reason for preferential treatment of bases. 3 We now have several equivalent ways of viewing diagonals of operators: diagonal sequences arising from matrix representations; • conditional expectations onto diagonal masas ; • Ah elements of the expectation of the unitary orbit. • We will choose whichever description is most convenient for the given situation. In particular, in Chapter 3 we will view diagonals as EA ( (T )), the image under the h U conditional expectation of the unitary orbit. In Chapter 4 we will view diagonals in terms of matrix representations with respect to the relevant orthonormal basis. 1.2. Information exchange between operators and diagonals In this section we describe several facts concerning diagonals of an operator T B( ) ∈ H which are suggestive of their importance. Details are often omitted for the sake of clarity of exposition. For now, let be finite dimensional. H (i) The trace of T is the linear functional on B( ) defined as the sum of the diagonal H entries N X Tr T := (T hi, hi), i=1 which is independent of the choice of orthonormal basis h. Indeed, if e is any other orthonormal basis then expanding h in terms of e and applying properties of the inner product we have N N N N ! X X X X (T hi, hi) = T (hi, ej)ej , (hi, ek)ek i=1 i=1 j=1 k=1 N N N X X X = (hi, ej)(hi, ek)(T ej, ek), i=1 j=1 k=1 4 and switching the order of summation N N N N X X X X (T hi, hi) = ((ek, hi)hi, ej)(T ej, ek) i=1 j=1 k=1 i=1 N N X X = (ek, ej)(T ej, ek) (1.2.1) j=1 k=1 N N N X X X = δj,k(T ej, ek) = (T ei, ei). j=1 k=1 i=1 This independence of basis is equivalent to the statement Tr[T,U] = 0 for any unitary operator U, where [T,U] := TU UT . Indeed, if Ue = h, then − N N N N X X X X (UT hi, hi) = (UT ei, ei) = (UT Uhi, Uhi) = (T Uhi, hi), i=1 i=1 i=1 i=1 by (1.2.1) and U ∗U = I. Moreover, this implies Tr[T,S] = 0 for any other operator S[3], since S may be written as a linear combination of four unitary operators. It is a straightforward exercise to show that Tr X = 0 if and only if 0,..., 0 is a h i diagonal of X (see Lemma 4.1.9). If an operator X is a commutator (i.e., X = [T,S]), then by the above remarks Tr X = 0. Sh¯oda’sLemma [Sh¯o37]proves the converse, that X is a commutator if Tr X = 0. The proof proceeds as follows: If Tr X = 0, then there is a basis e in which X has zero diagonal. Choose S to be a diagonal operator in the basis e with distinct diagonal entries s1, . . . , sn. Then let T be the operator whose entries in xij the basis e are given by tij := when i = j and tii = 0. Then a matrix computation si−sj 6 shows X = ST TS. − Fact 1.2.1. An operator X has 0,..., 0 as a diagonal in some basis if and only if it h i is a commutator, i.e., X = [S, T ]. [3] In fact, the trace is the unique (up to a scalar multiple) faithful positive linear functional which vanishes on commutators. It is often this property which is used to define traces in more general settings. 5 Now remove the restriction that be finite dimensional. H (ii) The diagonal entries of T (with respect to any given orthonormal basis h) are also elements of the numerical range W (T ) = (T h, h) (h, h) = 1 . A nonobvious fact { | } concerning the numerical range is the Toeplitz–Hausdorff Theorem [Hau19; Toe18], which states that W (T ) is a convex subset of C. It is also true that W (T ) σ(T ), which ⊇ is obvious when is finite dimensional since the spectrum consists of the eigenvalues H and these are always elements of the numerical range. Hence the convex hull of the spectrum σ(T ) is also included in the closure of the numerical range. When T is normal (i.e., when [T,T ∗] = 0 so T commutes with its adjoint) this can be strengthened to ∗ W (T ) = co σ(T ). In particular, if T 0 (i.e., T = T is selfadjoint and σ(T ) R+), ≥ ⊆ then W (T ) = co σ(T ) R+. The more interesting fact is that the converse is true. That ⊆ is, if W (T ) R+ then T 0. Because every element of the numerical range appears ⊆ ≥ on the diagonal of T in some basis, and the diagonal entries are always members of the numerical range, we have: Fact 1.2.2. The diagonal of T is a nonnegative sequence for every choice of orthonormal basis h if and only if T 0. ≥ (iii) The ideal of compact operators is the closure in the operator norm of the finite K rank operators. It is the only proper norm-closed ideal of B( ). A standard exercise in H operator theory is to show that for any orthonormal basis h and any compact K , ∈ K one has Khi 0. This implies that the diagonal entries of K in any basis converge k k → to zero. The converse seems to the author to be less well-known. That is, if K is an operator whose diagonal entries in any basis converge to zero, then K . ∈ K A similar result holds for the ideal of trace-class operators 1 which consists of those L compact operators K such that K := (K∗K)1/2 has summable eigenvalues when ∈ K | | repeated according to multiplicity. Precisely, if K 1 then for any orthonormal basis h, ∈ L 6 the diagonal (Khi, hi) is absolutely summable. Conversely, if in any basis the diagonal h i of K is absolutely summable, then K is trace-class. Both of these results are subsumed by the generic statements: Fact 1.2.3. If is a two-sided ideal (not necessarily closed) of B( ) then I H (i)If E (K) for every basis h, then K . h ∈ I ∈ I (ii)If is arithmetic-mean closed[4] and K , then E (K) for any basis h (that I ∈ I h ∈ I is, is diagonally invariant). I In other words, if is arithmetic-mean closed then an operator lies in if and only if I I every diagonal lies in . This subsumes the previous results since and 1 are arithmetic- I K L mean closed. The condition in (ii) that be arithmetic-mean closed is necessary in I general. Indeed, for the finite rank operators, witness any nonzero rank-one operator ξ ξ with ξ `2 of infinite support. ⊗ ∈ (iv) In light of the above examples, it is natural to ask which sequences are the diagonals of a specific operator. The vast majority of the work on this question in the literature is concerned solely with selfadjoint operators, with the original finite dimensional result being proved in two parts by Schur [Sch23] and Horn [Hor54], see the latter for details. For finite dimensional , they characterized the diagonals of a selfadjoint operator X B( ) H ∈ H in terms of a relation called majorization. N Definition 1.2.4. Given ξ, η R we say that ξ is majorized by η, denoted ξ η if ∈ ≺ n n N N X ∗ X ∗ X X ξ η , 1 n N, and ξi = ηi, i ≤ i ≤ ≤ i=1 i=1 i=1 i=1 where ξ∗, η∗ denote the monotone nonincreasing rearrangements of ξ, η. With this definition we can state the classical Schur–Horn Theorem. [4] For details see [KW11, Theorem 4.5]. Briefly, I is arithmetic-mean closed if I = a(Ia), where Ia and aI are the arithmetic mean and pre-arithmetic ideals generated respectively by: operators with s-numbers the arithmetic means of the s-numbers of operators from ideal I, and operators whose arithmetic means of their s-numbers are s-numbers of operators in I. 7 Theorem 1.2.5 (Schur–Horn, [Hor54; Sch23]). Let be finite dimensional and let H N X B( ) be a selfadjoint operator with eigenvalue sequence λ R . The diagonals of ∈ H ∈ X are characterized by N EAh ( (X)) = diag d d R , d λ . U { h | ∈ ≺ } Since the proof of the classical Schur–Horn Theorem in 1954 there have been several attempts to extend it to the infinite dimensional setting. We will cover these in detail in Section 2.1. (v) Instead of determining those sequences which are the diagonals of a specific operator, we can ask which are the diagonals of a class of operators, that is, the union of the diagonals of the individual operators in that class. There is a variety of material on this subject which we briefly review now. In the same paper [Hor54] in which he characterizes the diagonal sequences of a fixed selfadjoint matrix in Mn(C), Horn identifies the diagonals of the class of rotation matrices n as the convex hull of the points in R whose coordinates are either 1 or 1 with an even − number of negative entries. He then uses this to identify the diagonals of the classes of orthogonal matrices and of unitary matrices. The diagonals are those real-valued, respectively complex-valued, sequences whose sequence of absolute values lie in the convex n hull of the points in R whose coordinates are either 1 or 1. See [Hor54, Theorems − 8-11]. Fong shows in [Fon86] that any bounded sequence of complex numbers appears as the diagonal of a nilpotent operator in B( ) of order four (N 4 = 0), thus characterizing H diagonals of the broader classes of nilpotent (N k = 0 for some k) and also quasinilpotent k 1/k operators (limk→∞ N = 0). In this paper Fong remarks that a finite complex-valued k k sequence appears as the diagonal of a nilpotent matrix in Mn(C) if and only if its sum is zero. 8 More recently, Giol, Kovalev, Larson, Nguyen and Tener [Gio+11] classified the diagonals of idempotent matrices in Mn(C) as those whose sum is a positive integer less than n, along with the constant sequences 0,..., 0 and 1,..., 1 (see Theorem 2.2.2 h i h i below). In Section 2.2 we provide additional detail on the subject of diagonals of idempotents. 1.3. Description of results In Chapter 2 we explore the history and significance of the topics encountered in this dissertation, as well as providing a more detailed explanation of our main theorems. The remainder of the work is divided into several independent pieces. In Chapter 3, the results of which have been published in [LW15], we prove an infinite dimensional Schur–Horn theorem. As stated in Theorem 1.2.5, the classical Schur–Horn Theorem provides a characterization of the diagonals of a selfadjoint operator acting on a finite dimensional Hilbert space in terms of its eigenvalues. More specifically, it says that the diagonal sequences are precisely those which are majorized by the eigenvalue sequence. The work in Chapter 3 is concerned with extending the Schur–Horn Theorem to the infinite dimensional setting for positive compact operators. This extends recent work of Kaftal and Weiss [KW10] which provided the first exact characterization of the diagonals of a positive compact operator with kernel zero. One key aspect of Chapter 3 is the definition and investigation of new types of majorization, called p-majorization (Definition 3.2.2) and approximate p-majorization (Definition 3.3.1). In some sense, these types of majorization are -close, and in fact they coincide when p = 0 or p = . Our first main theorem (Theorem 3.2.4) asserts that ∞ sequences which are p-majorized by the eigenvalues of a positive compact operator and which have at most p fewer zeros than the eigenvalue sequence are diagonal sequences of that positive compact operator. Our second main theorem (Theorem 3.3.4) asserts a partial converse to the first by stating that a diagonal sequence of a positive compact 9 operator which has at least p fewer zeros than the eigenvalue sequence are approximately p-majorized by the eigenvalue sequence. Thus this bounds the diagonal sequences of a positive compact operator between two sets which are somehow -close when the kernel is finite dimensional, and provides a complete characterization of the diagonal sequences when the kernel is infinite dimensional (Corollary 3.3.6). Our last main theorem in this chapter (Corollary 3.4.1) proves that the diagonals of a positive compact operator with infinite dimensional kernel form a convex set. Chapter 4, the results of which have been published in [LW16], is devoted to the study of diagonals of idempotent operators. We begin by answering a question due to Jasper concerning the existence of nonzero idempotents with zero diagonal. We not only show that there exist nonzero idempotents with zero diagonal, we provide several equivalent conditions for an idempotent to have zero diagonal in some basis (Theorem 4.1.5). One of these conditions, that the idempotent cannot be written as the sum of a projection and a Hilbert–Schmidt operator, is especially useful because it is basis independent. Since the proof of Theorem 4.1.5 is essentially nonconstructive, we also provide a more hands-on approach to constructing a basis in which an idempotent has zero diagonal (Algorithm 4.1.11). These techniques then generalize to allow us to prove that any bounded complex-valued sequence appears as the diagonal of some idempotent (Theorem 4.2.7). Because Theorem 4.2.7 bears no particular resemblance to the finite dimensional case (Theorem 2.2.2 due to [Gio+11]), we also investigated the diagonals of the class of finite rank idempotents. Ultimately, we prove that the diagonals of finite rank idempotents are precisely those absolutely summable complex-valued sequences whose sum is a positive integer (Theorem 4.3.3), which is a natural analogue of the finite dimensional case. 10 1.4. Notation For a set S, let S denote its cardinality. will always denote a separable complex | | H Hilbert space of either finite or infinite dimension. Lowercase Fraktur letters (e.g. h, e, b) will always denote either orthonormal sets or orthonormal bases for the Hilbert space in question, and the corresponding lowercase roman glyph (e.g., h, e, b) will represent an element of this orthonormal basis. An atomic masa (i.e., arising as diagonal operators with respect to a fixed orthonormal basis) will be denoted by , or more simply by when the basis h is understood. Ah A Uppercase roman letters such as T, S, X will generally denote operators in B( ). The H ideal of compact operators in B( ) will be denoted as . H K Let c0 denote the space of all complex-valued sequences converging to zero, and let + ∗ c0 ,c0 denote the cones of nonnegative sequences converging to zero and nonnegative nonincreasing sequences converging to zero. For a sequence ξ c+, let ξ∗ c∗ denote ∈ 0 ∈ 0 the monotonization of ξ, or rather the monotonization of ξ supp ξ when ξ is not finitely | supported. That is, ξ∗ := ξ∗ where ξ∗ denotes the j-th largest element of ξ. Notice h j i j that if ξ is finitely supported, then ξ∗ ends in zeros. However, if ξ has infinite support, then ξ∗ has no zeros, and in this case the monotonization ξ∗ reflects neither the zeros of ξ nor their multiplicity. This notion of monotonization coincides with the decreasing rearrangement[5] of a function from measure theory by taking the counting measure on N and restricting the resulting step function to be defined on N. The symbol d will often denote a diagonal sequence of some operator. For a compact [6] operator K , the singular value sequence s(K) = sn(K) is the sequence ∈ K h i eigenvalues of (K∗K)1/2 listed in decreasing order repeated according to multiplicity [5]For a measure space (X, µ) and a complex-valued measurable function f which is finite almost everywhere, the decreasing rearrangement f ∗ : [0, ∞) → [0, ∞) is given by ∗ −1 f (t) = inf{x ∈ R+ | µ(|f| (x, ∞)) ≤ t}. [6]also known as s-numbers 11 (with the zero eigenvalue necessarily omitted if K has infinite rank). Equivalently, the singular numbers can be defined as sn(K) := inf F K , F ∈Fnk − k where n denotes the set of finite rank operators whose rank is strictly less than n. F The range projection RA for an operator A B( ) is the orthogonal projection onto ∈ H ⊥ ∗ ⊥ ran A = ker A . Thus for a selfadjoint operator A, RA is the projection onto ker A and ⊥ hence Tr RA = dim ker A, and in general Tr RA = rank A. Throughout this dissertation ⊥ we opt for using Tr RA and Tr RA instead of rank A and dim ker A. 12 Chapter 2 Background 2.1. Majorization and extensions of the Schur–Horn Theorem The classical Schur–Horn Theorem (Theorem 1.2.5) characterizes the diagonals of a selfadjoint operator on a finite dimensional Hilbert space in terms of the majorization preorder (Definition 1.2.4) on finite real-valued sequences. This suggests the more general problem: Problem 2.1.1 (Schur–Horn Problem). Characterize the diagonals (or an appropriate substitute thereof) of selfadjoint operators (on an infinite dimensional space, in a von Neumann algebra, etc.) in terms of majorization (or a suitable modification thereof). Since the advent of the Schur–Horn theorem, there has been substantial progress towards developing infinite dimensional analogues. This was perhaps started by the work of Markus [Mar64] and Gohberg and Markus [GM64], and more recently the topic was revived by A. Neumann in [Neu99]. However, Neumann studied an approximate Schur–Horn phenomenon, i.e., the operator-norm closure of the diagonals of bounded selfadjoint operators (equivalently, the `∞-norm closure of the diagonal sequences), which Arveson and Kadison deemed too coarse a closure [AK06, Introduction paragraph 3]. Instead, they studied the expectation of the trace-norm closure of the unitary orbit of a 13 trace-class operator and then proved approximate Schur–Horn analogues for trace-class operators in B( ) (the type I∞ factor). They also formulated a Schur–Horn conjecture H [1] for type II1 factors, but discussion of this topic is outside the scope of this paper . Kaftal and Weiss in [KW10] extended Arveson and Kadison’s work by proving exact Schur–Horn theorems for positive compact operators and pinpointing the diagonals of those with either finite rank or zero kernel. Majorization plays an essential role in Schur–Horn phenomena, but also has given rise to new kinds of majorization inside and outside this arena. The Riemann and Lebesgue majorizations of Bownik–Jasper are applied to Schur–Horn phenomena in [BJ15a; BJ15b; Jas13b]. Similarly, the uniform Hardy–Littlewood majorization of Kalton–Sukochev [KS08] is used ubiquitously in Lord–Sukochev–Zanin [LSZ13] as an essential tool to study traces and commutators for symmetrically normed ideals. In this paper we also develop new kinds of majorization essential to our work (see Figure 2.1 and Figure 2.2 and the accompanying description). + The following Definition 2.1.2 of majorization on c0 from [LW15] agrees with most of the literature, but it is a departure from that of [KW10] which did not include this equality condition. When Kaftal–Weiss needed an equality-like condition, they used instead the more restrictive Definition 2.1.3 of strong majorization. Definition 2.1.2 ([LW15, Definition 1.2]). Let ξ, η c+. One says that ξ is majorized ∈ 0 by η, denoted ξ η, if for all n N, ≺ ∈ n n ∞ ∞ X ∗ X ∗ X X ξ η and ξj = ηj. j ≤ j j=1 j=1 j=1 j=1 [1] For work on II1 and II∞ factors, see the work of Argerami and Massey [AM07; AM08; AM13], Bhat and Ravichandran [BR14] and a recent unpublished work of Ravichandran [Rav14]. 14 Definition 2.1.3 ([KW10, Definition 1.2]). Let ξ, η c+. One says that ξ is strongly ∈ 0 majorized by η, denoted ξ η, if for all n N, 4 ∈ n n n X X X ξ∗ η∗ and lim inf (η∗ ξ∗) = 0. j ≤ j n j − j j=1 j=1 j=1 Note that when ξ η, one has ξ `1 if and only if η `1. In the η-summable case, ≺ ∈ ∈ majorization as defined above in Definition 2.1.2 is equivalent to strong majorization in Definition 2.1.3. However, in the nonsummable case, the latter is clearly a stronger constraint than the former. Strong majorization is not an essential tool in our main theorems, but we thought it important to emphasize the distinction between these various definitions of majorization. The reasons for our Definition 2.1.2 departure from that of Kaftal–Weiss are convenience, efficiency of notation and unification of cases. This notation allows us to state in a more unified way the results for both trace-class and non-trace-class operators simultaneously without splitting the conclusions into cases (compare Theorem 3.1.3 to the two cases in [KW10, Corollary 5.4]). 2.1.1. Recent History Gohberg and Markus In [GM64], Gohberg–Markus almost characterized the diagonals of a selfadjoint trace- class operator K. To state their result we need some additional notions of majorization. Definition 2.1.4 ([GM64, pages 202–203]). Let ξ, η c+. One says that ξ is weakly ∈ 0 majorized by η, denoted ξ η, if for all n N ≺≺ ∈ n n X X ξ∗ η∗. j ≤ j j=1 j=1 15 This allows us to naturally define majorization of real-valued sequences in `1. For a real-valued sequence ξ, let ξ+ := max ξj, 0 and ξ− := ( ξ)+. h { }i − Definition 2.1.5 ([GM64, pages 202–203]). Let ξ, η `1 be real-valued. One says ∈ that ξ is majorized by η, denoted ξ η, if both ξ+ η+ and ξ− η−, and also ≺ ≺≺ ≺≺ P∞ P∞ 1 ξ = 1 η. Note that Definition 2.1.2 and Definition 2.1.5 agree on their common domain of pairs of nonnegative sequences in `1. Thus, while we may have overloaded the notation , ≺ there can be no misunderstanding of its meaning in a given context. In this context Gohberg and Markus prove the following theorem. Theorem 2.1.6 ([GM64, Theorem 1]). Let T 1 be a selfadjoint trace-class operator ∈ L in B( ) with eigenvalue sequence λ c0 repeated according to multiplicity. Then any H ∈ diagonal d of T satisfies d c0 and d λ, and conversely, if d c0 is majorized by λ, ∈ ≺ ∈ then d 0 is a diagonal of T for some size of 0. ⊕ The reason this is almost a characterization is that the size of the zero 0 is not specified, and it is easy to show that not all d majorized by λ with arbitrarily sized zeros appear as diagonals. Indeed, if λ0 is a sequence consisting only of the nonzero values of a sequence [2] λ of infinite support, then λ0 λ and λ0 0 is a diagonal of T if and only if the size ≺ ⊕ ⊥ of 0 equals Tr RT . Theorem 2.1.6 can, however, be easily enhanced to provide an exact characterization of the diagonals of positive trace-class operators with kernel zero through the use of the upcoming elementary Proposition 3.2.1. The result is then a special case of the forthcoming Theorem 2.1.14. [2] To prove this statement, it suffices to show that if λ0 ⊕ 0 is a diagonal of T then the unit vectors producing the values of λ0 are, in fact, eigenvectors for T . This can be established by induction on λ0 (proceeding in decreasing order) and using the fact that if |T x, x| = kT k then x is an eigenvector of T . 16 Arveson and Kadison In [AK06], Arveson and Kadison consider positive trace-class operators and prove an approximate Schur–Horn Theorem for this class. It is approximate in the sense that instead of characterizing the diagonals of a positive trace-class operator, i.e., the expectation of the unitary orbit, they characterize the expectation of the trace-norm closure of the unitary orbit. They first provide an alternative way to view the trace-norm closure of the unitary orbit of a positive trace-class operator. Proposition 2.1.7 ([AK06, Proposition 3.1]). Let A 1 be a positive trace-class ∈ L+ operator with s(A) the singular values of A. k·k1 (i) (A) = S(A) := B + s(B) = s(A) . U { ∈ K | } k·k (ii) B (A) 1 if and only if A 0 and B 0 are unitarily equivalent for an infinite ∈ U ⊕ ⊕ dimensional 0. k·k (iii) If A is finite rank, (A) 1 = (A). U U Their approximate Schur–Horn Theorem for positive trace-class operators is: Theorem 2.1.8 ([AK06, Theorem 4.1]). Let be a masa in B( ) of diagonal operators A H with respect to an orthonormal basis h and let A 1 be a positive trace-class operator. ∈ L+ Then k·k1 + EA( (A) ) = diag d d c , d s(A) . U { h | ∈ 0 ≺ } For finite rank positive operators, this provides a complete description of their diagonal sequences as encapsulated in the following corollary. Corollary 2.1.9. Let be a masa in B( ) of diagonal operators with respect to an A H orthonormal basis h and let A B( )+ be a positive finite rank operator. Then ∈ H + EA( (A)) = diag d d c , d s(A) . U { h | ∈ 0 ≺ } 17 Proof. Apply Proposition 2.1.7 (iii) and Theorem 2.1.8. They then prove one inclusion of the Schur–Horn Theorem for positive compact operators: Theorem 2.1.10 ([AK06, Theorem 4.2]). Let be a masa in B( ) of diagonal A H operators with respect to an orthonormal basis h and let A + be a positive compact ∈ K operator. Then ∗ + EA(L AL) L 1 = diag d d c , d s(A) . { | k k ≤ } { h | ∈ 0 ≺≺ } This indeed has part of the Schur–Horn Theorem as a corollary because it implies that diagonals of a positive compact operator A are weakly majorized by the singular value sequence s(A). In addition, the sum of the diagonal sequence is always equal to the sum of the singular values since A is positive and the trace is preserved under conjugation by unitaries (i.e., change of bases). Therefore, the diagonals of A are majorized by the singular values s(A). Symbolically, this is written as + EA( (A)) diag d d c , d s(A) . (2.1.1) U ⊆ { h | ∈ 0 ≺ } When A + has finite rank, Corollary 2.1.9 shows the sets in (2.1.1) are equal, thus ∈ K providing an exact Schur–Horn theorem for finite rank positive operators on an infinite dimensional Hilbert space[3]. It not hard to see that when A + has infinite rank and finite dimensional kernel, ∈ K then the inclusion is proper (reference Proposition 3.2.1). The inclusion is also proper when A has instead infinite dimensional kernel (see Theorem 3.3.4). Thus, the Schur–Horn [3]While similar, this result is inherently different from the classical Schur–Horn Theorem about selfadjoint operators on a finite dimensional space since the diagonal are infinite sequences instead of finite sequences. 18 Problem for positive compact operators is reduced to describing which proper subset of the sequences majorized by the singular values corresponds to the diagonal sequences. Kaftal and Weiss In [KW10], Kaftal and Weiss provided an exact[4] extension of the Schur–Horn Theorem to a certain subset of positive compact operators. That is, in terms of majorization they characterize precisely the expectation of the unitary orbit of strictly positive compact operators and the expectation of the partial isometry orbit (Definition 2.1.11 below) for all positive compact operators. They then ask for, but leave as an open question, a characterization of the expectation of the unitary orbit of positive compact operators with infinite rank and nonzero kernel. Definition 2.1.11 ([KW10, Page 3152] Partial Isometry Orbit). Given an operator A B( ), the partial isometry orbit of A is the set ∈ H ∗ ∗ (A) = V AV V B( ),V V = RA RA∗ . V { | ∈ H ∨ } Notice this extends to partial isometries the standard notion of unitary orbits (A) = U UAU ∗ unitary U ( ) . { | ∈ U H } At first, the partial isometry orbit may seem like a strange collection to consider, but our next result shows that for A + it is just another description for more familiar ∈ K objects. Recall that for A +, S(A) := B + s(B) = s(A) . ∈ K { ∈ K | } Proposition 2.1.12. If A + is a positive compact operator, then ∈ K k·k (A) = S(A) = (A) . V U [4]i.e., precise characterizations without taking closures of any kind 19 k·k If in addition A 1 , then these sets are also equal to (A) 1 . Furthermore, if A in ∈ L+ U fact has finite rank, then all these sets coincide with the unitary orbit (A). U Proof. We begin by establishing the first equality: (A) = S(A). Let V AV ∗ (A). V ∈ V By the conditions on V we know that V ∗V AV ∗V = A. Since V ∗V = 1, by standard k k inequalities concerning the singular value sequences [GK69, Page 27] we obtain s(A) = s(V ∗V AV ∗V ) s(V AV ∗) s(A). ≤ ≤ This proves (A) S(A). V ⊆ To prove the reverse inclusion, take B S(A). Since A, and therefore also B, are ∈ positive and compact, they have respective orthonormal eigenvectors e = en , f = fn { } { } corresponding to the eigenvalues sn(A) = sn(B), taking only those n for which these are positive. Define V to be the partial isometry which maps span e isometrically onto ∗ span f via V en = fn for all n N. Then V V = RA since it is the projection onto ∈ ∗ ∗ span en n∈ , and so V AV (A) and a computation guarantees V AV = B. This { } N ∈ V proves the inclusion S(A) (A). ⊆ V k·k We now prove the second equality: S(A) = (A) . This proof mimics the proof of U Proposition 2.1.7 (i) from [AK06] with only slight modifications. The singular value sequences are known to satisfy sn(A) sn(B) A B (see, for example, [GK69, | − | ≤ k − k Corollary 2.3]), and therefore S(A) is norm closed. Along with the inclusion (A) U ⊆ k·k (A) = S(A), it follows trivially that (A) S(A). For the reverse inclusion, let V U ⊆ B S(A). For each n N we may write ∈ ∈ 0 0 A = An + An and B = Bn + Bn, 20 0 0 where s(An) = s(Bn) = s1(A), . . . , sn(A), 0,... and A = B = sn+1(A). It is h i k nk k nk ∗ clear that there are unitaries Un for which UnAnUn = Bn. Thus B UnAU = k − nk 0 0 ∗ k·k k·k B UnA U 2sn+1(A) 0, and hence B (A) . This proves S(A) (A) . k n − n nk ≤ → ∈ U ⊆ U The additional result when A 1 is precisely Proposition 2.1.7 (i), and when A has ∈ L+ finite rank we apply Proposition 2.1.7 (iii). In light of Proposition 2.1.12, the following Schur–Horn type result of Kaftal–Weiss is a natural extension of Theorem 2.1.8 by Arveson–Kadison to positive compact operators. Here we present a reformulation of their theorem using our notation for majorization. Theorem 2.1.13 ([KW10, Proposition 6.4]). Let A +. Then ∈ K E( (A)) = B + s(B) s(A) . V { ∈ D ∩ K | ≺ } The common underpinning between Theorem 2.1.8 and Theorem 2.1.13 is S(A). Since S(A) does not encode the dimension of the kernel of A both theorems sidestep the effect of this information on the diagonal sequences of A. The techniques of Kaftal–Weiss, however, allow for the analysis of special cases, as evidenced by the following theorem. Theorem 2.1.14 ([KW10, Proposition 6.6]). Let A + with RA = I. Then ∈ K E( (A)) = E( (A)) B RB = I . U V ∩ { ∈ D | } The question of precisely what is E( (A)) for all A + has only been partially U ∈ K answered. In particular, it was answered when A has finite rank by Corollary 2.1.9 or when RA = I (that is, when A is strictly positive) by the previous theorem. We should note that Corollary 2.1.9 can also be deduced from Theorem 2.1.13 and the finite rank fact that (A) = (A), which itself can be found in Proposition 2.1.12 (or see U V Remark 3.2.5 for a direct approach). 21 2.1.2. Main contribution Corollary 2.1.9 and Theorem 2.1.14 left open the case when A has infinite rank and nonzero kernel. We attempt here to narrow this gap. Indeed, we characterize E( (A)) U when A has infinite dimensional kernel. When A has finite dimensional kernel, we give a necessary condition for membership in E( (A)) (which we conjecture is also sufficient), U and we give a sufficient condition for membership in E( (A)) (which we know not to be U necessary when 0 < Tr R⊥ < , see Example 3.2.7 below which also appears in [KW10, A ∞ Proposition 6.10, Example 6.11]). Moreover, these two conditions are somehow “-close” in an admittedly vague sense[5], thereby significantly restricting the range of sequences for which it is unknown whether or not they are diagonals. These main results are embodied in Theorem 3.2.4, Theorem 3.3.4 and Corollary 3.3.6[6]. Both of these membership conditions involve new kinds of majorization, which here we call p-majorization and herein we introduce approximate p-majorization (for 0 p , Definition 3.2.2 and ≤ ≤ ∞ Definition 3.3.1 below). There is a natural hierarchy of these new types of majorization which the diagram in Figure 2.1 describes succinctly. All of these implications are natural (see the discussions following Definition 3.2.2 and Definition 3.3.1) except the two corresponding to the dashed arrows, which are proved in Proposition 3.2.8 and are only applicable when both sequences in question are in c+ `1. A linear interpretation of 0 \ the diagram in Figure 2.1 is presented below it in Figure 2.2. -1 -2 - ··· ∞ and ≺ ≺ 64 1 2 ≺ ≺ ··· ≺∞ Figure 2.1.: Hierarchy of majorization [5]Compare Definition 3.2.2 to Definition 3.3.1. [6]We refrain from stating these results in their entirety in the introduction because doing so would require definitions which do not appear until Chapter 3, namely p-majorization (Definition 3.2.2) and approximate p-majorization (Definition 3.3.1). 22 -1 1 -2 - ( and 4) ≺ ≺ ··· ∞ ≺∞ ≺ 6 Figure 2.2.: Linear hierarchy of majorization 2.2. Diagonals of idempotents and frames 2.2.1. Projections and idempotents As mentioned in the previous section, a large portion of the work concerning diagonal sequences of operators deals with the selfadjoint case. In Chapter 4 we study diagonal sequences of idempotents, so diagonals of projections (selfadjoint idempotents) are of particular relevance to us. These were characterized by Kadison in Theorem 2.2.1 [Kad02a; Kad02b]. We find this theorem especially interesting because it straddles the fence between a characterization of diagonals of a single operator in a specified class (Section 1.2 item (iv)) and of a class of operators (Section 1.2 item (v)). Indeed, although it is stated as a characterization of the diagonals of the class of projections, it can easily be adapted to identify the diagonals of any fixed projection. This is because two projections P,P 0 B( ) are unitarily equivalent if and only if Tr P = Tr P 0 and ∈ H 0 Tr(1 P ) = Tr(1 P ) and thus share the same diagonals. Moreover, for dk a diagonal − − h i sequence for P , these quantities are precisely the sum of the diagonal entries dk and the sum of 1 dk, respectively. Then one can apply the four finite/infinite cases in the next − theorem. Theorem 2.2.1 ([Kad02a; Kad02b]). Given an infinite sequence dk [0, 1]N with h i ∈ X X a = dk and b = (1 dk), − dk<1/2 dk≥1/2 2 ∗ then there is a projection P B( ) (i.e., P = P = P ) with diagonal dn if and only ∈ H h i if one of the following mutually exclusive conditions holds: (i) either a or b is infinite; 23 (ii) a, b < and a b Z. ∞ − ∈ The requirement that 0 dk 1 for all k N is clearly necessary since 0 P I and ≤ ≤ ∈ ≤ ≤ the diagonal entries of P are elements of its numerical range. The second condition, that a b Z, is less obvious but can be viewed as a kind of index obstruction to an arbitrary − ∈ sequence in [0, 1]N appearing as the diagonal of a projection. Indeed, in [Arv07] Arveson provided details on this index obstruction and showed that it applies more generally to any normal operator with finite spectrum (each element of infinite multiplicity) that consists of the vertices of a convex polygon. Since we study diagonals of idempotents in B( ), which when not projections are H non-selfadjoint, we are especially interested in diagonals of non-selfadjoint operators. One particularly relevant result in this direction is the aforementioned characterization of diagonals of idempotent matrices in Mn(C) by Giol, Kovalev, Larson, Nguyen and Tener [Gio+11]. n Theorem 2.2.2 ([Gio+11, Theorem 5.1]). A finite sequence dk C appears as the h i ∈ diagonal of an idempotent D Mn(C) if and only if one of the following three mutually ∈ exclusive conditions holds. (i) dk = 0 for all k (in which case D = 0); (ii) dk = 1 for all k (in which case D = I); P (iii) dk 1, . . . , n 1 . ∈ { − } Since Tr D 1, . . . , n 1 for any nonzero, non-identity idempotent matrix (as ∈ { − } is well-known, see for instance Lemma 4.1.1), this theorem says that this is the only requirement for a sequence to appear as the diagonal of some idempotent. We prove a natural extension of this result to finite rank idempotents on an infinite dimensional space in the next theorem. H 24 Theorem 4.3.3 ([LW16, Theorem 4.3]). The diagonals of the class of nonzero finite rank idempotents consist precisely of those absolutely summable sequences whose sum is a positive integer. Giol, Kovalev, Larson, Nguyen and Tener proved Theorem 2.2.2 because of its relevance to frame theory. Because of a similar frame-theoretic question (characterizing inner prod- ucts of dual frame pairs), Jasper asked for a characterization of diagonals of idempotents in B( ). For a key test case, Jasper posed the following two operator-theoretic questions H (private communication, May 2013 [Jas13a]): Question 2.2.3. If an idempotent has a basis in which its diagonal is absolutely summable, is it finite rank? Question 2.2.4. If an idempotent has a basis in which its diagonal consists solely of zeros (i.e., is a zero-diagonal operator in the terminology of Fan [Fan84]), is it finite rank? If we restrict the idempotents to be selfadjoint (i.e., projections), then they are positive operators and the answer to each question is certainly affirmative since the trace is preserved under conjugation by a unitary operator (i.e., a change of basis). In fact, for projections, having an absolutely summable (or even summable) diagonal is a characterization of those projections with finite rank since rank P = Tr P . Moreover, the only projection with a zero diagonal is the zero operator for this same reason. Hence, a negative answer to either of these questions for the entire class of idempotents would be a notable departure from the case of projections and would therefore suggest that the classification of their diagonals is potentially harder than one might na¨ıvely expect. As it turns out, Larson constructed a nonzero (and necessarily infinite rank) idempotent that lies in a continuous nest algebra which has zero diagonal with respect to this nest[7] [7] An operator T has zero diagonal with respect to the nest if PλTPλ = 0 for some linearly ordered set of projections {Pλ}λ∈Λ inside the nest such that with respect to the decomposition of the identity P I = Λ Pλ every element of the nest is block upper-triangular. 25 [Lar85, Proof of Theorem 3.7]. However, the existence of an idempotent with zero diagonal with respect to a nest algebra certainly depends on the order type of the nest to some extent. Indeed, the nest algebra consisting of the upper triangular matrices with respect to some basis en n∈ for has order type ω (the first infinite ordinal), and a { } N H simple computation shows that the only idempotent with zero diagonal inside this nest algebra is the zero operator (proceed by induction on the superdiagonals). Leaving the realm of nest algebras, we can ask two questions: Which idempotents are zero-diagonal? • Which idempotents have an absolutely summable diagonal? • As it turns out, both of these questions have the same answer, which we provide in Theorem 4.1.5, the statement of which can be found below on page 28. 2.2.2. General techniques for diagonals of non-selfadjoint operators The techniques for analyzing diagonals of non-selfadjoint operators seem to differ greatly from those used for selfadjoint operators. For example, the techniques used in determining diagonals of selfadjoint operators often rely heavily on majorization and keeping track of the explicit changes of the basis (or equivalently, the unitary operators) involved in the construction. In contrast, the Toeplitz–Hausdorff Theorem, that the numerical range W (T ) of a bounded operator T is convex, is one of the central tools in the work of Fan, Fong and Herrero [Fan84], [FFH87], [FF94] to determine diagonals of non-selfadjoint operators. Indeed, they frequently use the nonconstructive version of the Toeplitz– Hausdorff Theorem despite the existence of constructive versions in which a formula is specified for the vector yielding the prescribed value of the quadratic form (T f, f). The Fan, Fong and Herrero results relevant to us here are restated below. The first is an infinite dimensional analogue of the finite dimensional result that an n n matrix × has trace zero if and only if it is zero-diagonal. 26 Theorem 2.2.5 ([Fan84, Theorem 1]). If T B( ) and there exists some basis ∈ H ∞ ej for for which the partial sums { }j=1 H n X sn := (T ej, ej) j=1 have a subsequence converging to zero, then T is zero-diagonal. Fan’s proof of [Fan84, Theorem] has a flaw (see Appendix B for details), but it can be circumvented with a slight modification. Appendix B is devoted to reproving Theorem 2.2.5 in order both to fix the flaw and to provide more detail than [Fan84]. ∞ Definition 2.2.6 ([FFH87]). Let T B( ) and let e = ej be a basis for . ∈ H { }j=1 H Pn Suppose the partial sums sn = (T ej, ej) converge to some value s C. Then we j=1 ∈ say that Tre T := s is the trace of T with respect to the basis e. The set of traces of T , denoted R Tr T , is then the set of all such traces Tr T as e ranges over all orthonormal { } e bases for which Tre T is defined (i.e. lim sn exists and is finite). Observe that in order to make sense of this definition, it is essential both that these trace values are finite and that we must order e by N. A curious fact about the set R Tr T from Definition 2.2.6 is that it may take on only { } four different shapes: the plane, a line, a point or the empty set. It is no coincidence that these shapes coincide with those obtainable as the limits of convergent rearrangements of a series of complex numbers (i.e., the L´evy–SteinitzTheorem extending the Riemann Rearrangement Theorem to complex numbers [L´ev05;Ste13]). Theorem 2.2.7 ([FFH87, Theorem 4]). Suppose T B( ). Then there are four ∈ H possible shapes that R Tr T can acquire. More specifically, R Tr T is: { } { } iθ 1 (i) the plane C if and only if for all θ R, (Re e T )+ / (the trace-class); ∈ ∈ L iθ 1 iθ 1 (ii) a line if and only if for some θ R, both (Re e T )± / but (Im e T ) ; ∈ ∈ L ∈ L (iii) a point if and only if T 1; ∈ L 27 iθ 1 iθ 1 (iv) the empty set if and only if for some θ R, (Re e T )+ / but (Re e T )− . ∅ ∈ ∈ L ∈ L In fact, their proof of Theorem 2.2.7 shows that given T B( ) there exists a basis ∈ H e for , e ordered by N, for which every element of R Tr T can be obtained from a H { } basis which is a permutation of e. For the next theorem Fan–Fong utilize the previous two theorems to provide intrinsic (i.e., basis independent) criteria for when a bounded operator is zero-diagonal. Theorem 2.2.8 ([FF94]). An operator T is zero-diagonal if and only if for all θ R, ∈ iθ iθ Tr(Re e T )+ = Tr(Re e T )−. We neither use nor cite this theorem elsewhere, however, it seems interesting to include it because it shares its intrinsic nature with our Theorem 4.1.5 (i). Theorem 4.1.5 ([LW16, Theorem 2.5]). For D B( ) an infinite rank idempotent ∈ H the following are equivalent: (i) D is not a Hilbert–Schmidt perturbation of a projection; (ii) the nilpotent part[8]of D is not Hilbert–Schmidt; (iii) R Tr D = C; { } (iv) D is zero-diagonal; (v) D has an absolutely summable diagonal; (vi) D has a summable diagonal (i.e., R Tr D = ). { } 6 ∅ An interesting general fact is that (v) and (vi) of Theorem 4.1.5 are actually equivalent for any bounded operator, not merely idempotents (see Proposition 4.1.12). [8]We have not yet defined the nilpotent part of an idempotent D, but it is a natural object defined in Lemma 4.1.1 that gives a canonical decomposition for idempotents. In particular, we can write I 0 D = ( T 0 ) and we call T the nilpotent part of D. 28 2.2.3. Jasper’s problem and the relation to frames The tools developed in Chapter 4 for the proof of Theorem 4.1.5 led to a complete answer to Jasper’s original problem: Problem 2.2.9. Characterize the diagonals of the class of idempotent operators. In short, there are no restrictions on the diagonals of the class of idempotent operators. ∞ Theorem 4.2.7 ([LW16, Theorem 3.7]). Every dn ` admits an idempotent D h i ∈ ∈ B( ) whose diagonal is dn with respect to some orthonormal basis. H h i We mentioned briefly that Jasper’s problem had its origins in frame theory. Below we give a short description of the frame-theoretic equivalent[9] of Problem 2.2.9. We begin with the definition of a frame. Definition 2.2.10. A collection x = xj j∈J of vectors in is said to be a frame if the { } H P so-called frame operator S = xj xj satisfies aI S bI for some positive scalars j∈J ⊗ ≤ ≤ a, b. The analysis operator (or frame transform) is the operator θ : B( ) `2(J) defined x H → P 2 by θ = ej xj where e = ej j∈J is the standard basis on ` (J). The synthesis x j∈J ⊗ { } ∗ operator is just the adjoint θx of the frame transform. Then the frame operator is simply ∗ ∗ S = θx θx and the so-called Gram matrix (or Grammian) G = θxθx is the operator on 2 ` (J) whose ij-th matrix coefficient (with respect to e) is (xj, xi)`2(J). ∗ Definition 2.2.11. Given a frame x, a frame y is said to be dual to x if θy θx = I, i.e., if the synthesis operator of y is a left inverse for the analysis operator of x. The canonical ∗ −1 dual frame is the frame associated to the analysis operator θx(θx θx) . [9]The equivalence of these two problems was originally described to us by Jasper, but a fairly concise explanation can be found on this MathOverflow post [10113]. 29 Another way to view dual frames is via idempotents. Consider that y is dual to x if and ∗ ∗ ∗ ∗ ∗ ∗ only if θxθy is idempotent. Indeed, if they are dual then (θxθy )(θxθy ) = θx(θy θx)θy = θxθy , ∗ ∗ ∗ ∗ and conversely, if θxθy is idempotent, then θxθy = θxθy θxθy and multiplying on the right ∗ ∗ by θx and on the left by θy we obtain I = θxθy . Furthermore, if Px denotes the projection ∗ −1 ∗ θx(θx θx) θx , then Px is the range projection of θx. Moreover, it is the range projection of any idempotent θ θ∗ where y is dual to x. This guarantees that the map y θ θ∗ is a x y → x y bijection between the dual frames of x and the set of idempotents with range projection Px. ∗ One can notice that the diagonal of the idempotent θxθy is precisely the sequence of inner products (xj, yj) . Therefore a characterization of diagonals of idempotents h i provides a characterization of inner products of dual frame pairs, thereby establishing the correspondence. 30 Chapter 3 Schur–Horn theorems Most of the results of this chapter have been published in [LW15]. In this chapter we pursue an exact Schur–Horn theorem for positive compact operators with infinite rank and infinite dimensional kernel. We also provide some information in the case when the operator has infinite rank and finite dimensional kernel. We begin in Section 3.1 with a brief primer on the closely related subject of stochastic matrices, which are ubiquitous throughout the work of [KW10] from whence flow many of our results. In Section 3.2 we start by providing a necessary condition on the kernel dimension of diagonals of a positive operator. However, the main thrust of the section is the definition of p-majorization (for p Z+ , Definition 3.2.2) and the proof that this ∈ ∪ {∞} type of majorization is sufficient to guarantee membership in the diagonal sequences of an operator when the difference in the dimension of the kernels is less than or equal to p (Theorem 3.2.4). We then follow a short tangent on the relationship between p-majorization and strong majorization (Definition 2.1.3), and how this may be exploited in the determination of diagonals. As a partial converse of Theorem 3.2.4, in Section 3.3 we define approximate p- majorization (for p Z+ , Definition 3.3.1) and show that this type of majorization ∈ ∪ {∞} is necessary for membership in the diagonal sequences of an operator when the difference 31 in the dimension of the kernels is at least p (Theorem 3.3.4). Since -majorization and ∞ approximate -majorization are equivalent, these theorems establish a characterization ∞ of diagonals of positive compact operators with infinite rank and infinite dimensional kernel (Corollary 3.3.6). We conclude by proving that in the setting of Corollary 3.3.6 the diagonals form a convex set. Convexity is closely related to the Schur–Horn Theorem (see discussion in Section 3.4) and in all the cases in which the diagonals of a positive compact operator are known, they form a convex set. However, it is in general false that the diagonals of an arbitrary operator (even a normal operator on a three-dimensional space) constitute a convex set. In fact, given Kadison’s characterization of diagonals of projections, one can see that the diagonals of an infinite rank projection with infinite dimensional kernel are not convex as the next example shows. Example 3.0.1. In the notation of Theorem 2.2.1, note that there is a projection with diagonal d such that a + b < and a b = 1. Moreover, the sequence d0 = d0 given by ∞ − h ji 0 if dj < 1/2, 0 dj := 1 if dj 1/2 ≥ is obviously the diagonal of a projection. If we consider the convex combination ϕ = λd + (1 λ)d0 for 0 < λ < 1, and − X X aλ = ϕk and bλ = (1 ϕk), − ϕk<1/2 ϕk≥1/2 then we compute X X aλ = ϕk = λdk = λa, ϕk<1/2 dk<1/2 32 and X X bλ = (1 ϕk) = (1 λdk 1 + λ) = λb. − − − ϕk≥1/2 dk≥1/2 Thus aλ bλ = λ(a b) = λ / Z, and hence ϕ is not the diagonal of a projection by − − ∈ Theorem 2.2.1. 3.1. Preliminaries on stochastic matrices Stochastic matrices play a central role in investigating Schur–Horn theorems due to the following definition and lemma. Definition 3.1.1 (Stochasticity). A matrix P with positive entries is called substochastic if its row and column sums are bounded by 1; • column-stochastic if it is substochastic and its column sums equal 1; • row-stochastic if it is substochastic and its row sums equal 1; • doubly stochastic if it is both row- and column-stochastic; • unistochastic if it is the Schur-product of a unitary matrix with its complex • conjugate (the Schur-product of two matrices A = (aij) and B = (bij) is the matrix (aijbij), that is, it is the entrywise product of A, B); orthostochastic if it is the Schur-square of an orthogonal matrix, i.e., of a unitary • with real entries. And the connection between expectations of orbits and stochastic matrices is: Lemma 3.1.2 ([KW10, Lemmas 2.3, 2.4]). Let ξ, η `∞ and for any contraction ∈ 2 L = (Lij) B( ), let Qij := Lij for all i, j. Then ∈ H | | ξ = Qη if and only if diag ξ = E(L diag η L∗). 33 Furthermore, (i) Q is substochastic; (ii) L is an isometry if and only if Q is column-stochastic; (iii) L is a co-isometry (L∗ is an isometry) if and only if Q is row-stochastic; (iv) L is unitary if and only if Q is unistochastic; (v) L is orthogonal if and only if Q is orthostochastic. For completeness we repeat the straightforward short proof. ∞ Proof. Given ξ, η ` , notice that for any n N, ∈ ∈ ∗ ∗ (E(L diag η L )en, en) = (L diag η L en, en) ∞ ∞ X X = diag η) L¯njej, L¯nkek j=1 k=1 ∞ ∞ X X = ηjL¯njej, L¯nkek j=1 k=1 ∞ X 2 = Lnj ηj = (Qη)n. | | j=1 Notice now that ∞ ∞ X X ∗ ∗ ∗ 2 Qij = LijL = (LL ei, ei) = L ei 1 for every i, (3.1.1) ji k k ≤ j=1 j=1 and similarly ∞ X 2 Qij = Lej 1 for every j. (3.1.2) k k ≤ i=1 (i) Immediate from (3.1.1) and (3.1.2). (ii) If L is an isometry, then it is immediate from (i) and the equality cases of (3.1.2) that Q is column-stochastic. Conversely assume that Q is column-stochastic and hence 34 ∗ Lej = 1 for all j by (3.1.2). Then L Lej, ej = 1 for all j and thus it follows that k k h i E(I L∗L) = 0. Since E is faithful and I L∗L 0 because L is a contraction by − − ≥ hypothesis, it follows that L∗L = I. (iii) Apply (ii) to L∗. (iv) Immediate from (ii) and (iii). (v) Immediate from (iv) and the fact that L has real entries. Lemma 3.1.2 provides a dictionary which translates between the realms of sequences and stochastic matrices and of diagonals of operators (expectations of orbits). Many of the results in [KW10] are stated and proved in terms of these stochastic matrices. We reformulate here one of their theorems more relevant to our study using our notation for majorization. Theorem 3.1.3 ([KW10, Corollary 5.4]). If ξ, η c∗, then ∈ 0 ξ = Qη for some orthostochastic matrix Q ξ η. ⇐⇒ ≺ Note that via Lemma 3.1.2, the above Theorem 3.1.3 along with Proposition 3.2.1 imply Theorem 2.1.14 which characterizes diagonals of positive compact operators with zero kernel. In fact, because Theorem 3.1.3 is about orthostochastic matrices as opposed to unistochastic matrices, we can conclude that Theorem 2.1.14 is valid even when the underlying Hilbert space is real instead of complex. Moreover, all our theorems in this H chapter are applicable to real Hilbert spaces since Theorem 3.2.4 is the only one for which this fact is not easily verified, but its extension to real Hilbert spaces is handled in Remark 3.2.6. 35 3.2. p-majorization (sufficiency) A result which was known to Kaftal and Weiss [KW10, Proposition 6.6 (i)], and almost certainly to others, is a necessary condition for membership in the expectation of the unitary orbit of a positive operator. Namely, the dimensions of the kernels of operators in the range of the expectation of the unitary orbit of a positive operator cannot increase from the dimension of the kernel of the operator itself. Proposition 3.2.1 ([LW15, Proposition 2.1]). If A B( )+ and B E( (A)), then ∈ H ∈ U ker B ker(UAU ∗) for some U ( ), and hence also Tr R⊥ Tr R⊥. ⊆ ∈ U H B ≤ A Vector proof. Since B E( (A)), B = E(UAU ∗) for some unitary U ( ), and since ∈ U ∈ U H ⊥ ∗ ⊥ Tr RA = dim ker A = dim ker UAU = Tr RUAU ∗ , the required trace inequality follows from the inclusion ker B ker UAU ∗, which itself follows from A 0 and ⊆ ≥ ker B = ker E(UAU ∗) ∗ = span en E(UAU )en, en = 0 { | h i } ∗ = span en UAU en, en = 0 { | h i } ∗ 1/2 2 = span en (UAU ) en = 0 { | k k } ker(UAU ∗)1/2 = ker UAU ∗ ⊆ Projection proof. Since A 0, R⊥ is the largest projection P so that P AP = 0. Since ≥ A B E( (A)), there is some unitary U ( ) so that B = E(UAU ∗). Notice also that ∈ U ∈ U H R⊥ since B . Therefore, B ∈ D ∈ D ∗ ⊥ ∗ ⊥ ∗ ⊥ ∗ ⊥ E(U(U RBU)A(U RBU)U ) = E(RBUAU RB) ⊥ ∗ ⊥ ⊥ ⊥ = RBE(UAU )RB = RBBRB = 0. 36 ∗ ⊥ ∗ ⊥ ∗ Then since E is faithful and U(U RBU)A(U RBU)U is positive, it is zero, and thus conjugating by U ∗ shows (U ∗R⊥U)A(U ∗R⊥U) = 0. By the maximality of R⊥, U ∗R⊥U B B A B ≤ ⊥ RA and therefore Tr R⊥ = Tr(U ∗R⊥U) Tr R⊥. B B ≤ A Before we proceed with our analysis, we need to introduce next a concept similar to [KW10, Definition 6.8(ii)] which here we call p-majorization[1]. Roughly speaking, it is majorization along with eventual p-expanded majorization. And this leads us to the definition below of -majorization which is both new and fruitful. ∞ Definition 3.2.2. Given ξ, η c+ and 0 p < , we say that ξ is p-majorized by η, ∈ 0 ≤ ∞ denoted ξ p η, if ξ η and there exists an Np N such that for all n Np, we have ≺ ≺ ∈ ≥ the inequality n+p n X X ξ∗ η∗. k ≤ k k=1 k=1 And -majorization, denoted ξ ∞ η, means ξ p η for all p N. ∞ ≺ ≺ ∈ Note that ξ 0 η is precisely the statement that ξ η (recall Definition 2.1.2, which ≺ ≺ 0 includes equality of the sums). One also observes that if ξ p η and p p, then ≺ ≤ ξ p0 η (we use often the special case that p-majorization implies 0-majorization, i.e., ≺ majorization). For this reason, ξ p η for infinitely many p is equivalent to ξ p η for all ≺ ≺ p < , in which case we say that ξ is -majorized by η and we write ξ ∞ η. Although ∞ ∞ ≺ not immediately obvious, it is in fact possible to have ξ ∞ η even when η does not have ≺ 1 finite support. For example, let ξ := η1, η1, η2, η2,... (hint: Np = p). 2 h i One should also take note that p-majorization is actually strictly stronger than p0- 0 + majorization when p < p. That is, there exist sequences ξ, η c for which ξ p0 η but ∈ 0 ≺ ξ p η. From the remarks of the previous paragraph, it suffices to exhibit ξ, η when 6≺ [1]The author, along with Gary Weiss, introduced the concept of p-majorization in [LW15, Definition 2.2]. The reader should note that the asterisks (indicating monotonization) which appear in the displayed inequality of Definition 3.2.2 are missing from [LW15, Definition 2.2]. This is a typo in [LW15]. An identical statement holds for [LW15, Definition 3.1] and Definition 3.3.1. 37 p = p0 + 1. To produce such sequences, start with any 0 < η c∗ and define ∈ 0 (p) ξ := η1/p,..., η1/p, η2, η3,... . h| {z } i p times (p) (p) Even though ξ is not necessarily monotone, it is not difficult to verify that ξ p0 η ≺ (p) but ξ p η. 6≺ Remark 3.2.3 ([LW15, Remark 2.3]). Because (A) (A), Theorem 2.1.13 for A V ⊇ U ∈ + implies that s(B) s(A) is a necessary condition for B E( (A)); Proposition 3.2.1 K ≺ ∈ U shows Tr R⊥ Tr R⊥ is another necessary condition for membership in E( (A)) and with B ≤ A U majorization is equivalent to membership in E( (A)) when Tr R⊥ = 0 by Theorem 2.1.14. U A It was natural in [KW10] to ask how the role of majorization is impacted by the dimension of these kernels. In particular here we enhance this program by asking, what role, if any, p = Tr R⊥ Tr R⊥ plays in relating membership in E( (A)) to majorization (with the A − B U convention that p = 0 when Tr R⊥ = = Tr R⊥. Equivalently, p is the minimal n for A ∞ B which Tr R⊥ Tr R⊥ + n. This is the strategy that guides our program. A ≤ B The result below appears in [KW10] as Lemma 6.9 for p < , the proof of which ∞ utilizes orthostochastic matrices. We provide a different proof which instead utilizes expectations of unitary orbits because it leads to a straightforward extension to the p = case. See Remark 3.2.6 for when an orthostochastic matrix can be produced to ∞ implement the construction. ⊥ ⊥ Theorem 3.2.4 ([LW15, Theorem 2.4]). Let A, B +, B and Tr R Tr R . ∈ K ∈ D B ≤ A ⊥ ⊥ If for some 0 p , s(B) p s(A) and Tr R Tr R + p, ≤ ≤ ∞ ≺ A ≤ B then B E( (A)). ∈ U Proof. Without loss of generality we may assume that A (i.e., A, B are simultaneously ∈ D diagonalized) because (A) = (UAU ∗) for every U ( ). We may further reduce U U ∈ U H ⊥ ⊥ to the case when Tr RB = 0 via a splitting argument. Because dim ker B = Tr RB < 38 Tr R⊥ = dim ker A, without loss of generality we can assume S := ker B ker A. Then A ⊆ ⊥ with respect to = S S , one has A = A1 A2, B = B1 B2, A1 = 0 = B1, and B2 H ⊕ ⊕ ⊕ ⊥ ∗ has zero kernel. Thus once we find a unitary U on S for which E(UA2U ) = B2, then I U satisfies E((I U)A(I U)∗) = B. ⊕ ⊕ ⊕ Case 1: A has finite rank. This is a direct consequence of Corollary 2.1.9. However, below we give an alternate proof which shows directly that (A) = (A) (even if A is not necessarily selfadjoint), U V and then applies Theorem 2.1.13. If A has finite rank, then (A) = (A), even if A is not necessarily selfadjoint. Indeed, U V the elements of (A), by Definition 2.1.11, have the form V AV ∗ for some partial isometry V ∗ ∗ ∗ for which V V = RA RA∗ (:= P ), so PA = AP = A. Then V AV = (VP )A(VP ) ∨ and VP is also a partial isometry with (VP )∗(VP ) = P . But this partial isometry VP is finite rank since RA and RA∗ , and hence also P , are finite rank, and so VP can be extended to a unitary U for which V AV ∗ = (VP )A(VP )∗ = UAU ∗. This shows (A) (A), and hence equality. That the conclusion of Theorem 3.2.4 holds in this case V ⊆ U is then covered by Theorem 2.1.13, since A 0, (A) = (A), and since s(B) p s(A) ≥ U V ≺ implies s(B) s(A). ≺ ⊥ ⊥ Case 2: A has infinite rank and Tr RB = Tr RA (the p = 0 case). ⊥ With the earlier reduction that Tr RB = 0 and using p = 0, Case 2 is a direct consequence of Theorem 2.1.14 and Theorem 2.1.13. ⊥ ⊥ Case 3: A has infinite rank and Tr RB < Tr RA (the most complicated case). Since Tr R⊥ Tr R⊥ + p, if necessary, by passing to a smaller p we may assume A ≤ B Tr R⊥ = Tr R⊥ + p, even if Tr R⊥ = (in which case p = ). Since Tr R⊥ = 0, one has A B A ∞ ∞ B Tr R⊥ = p, and since Tr R⊥ < Tr R⊥, one has 1 p . A B A ≤ ≤ ∞ ∞ We now employ another splitting. First let ej denote the basis that simultaneously { }j=1 p diagonalizes A, B, and then assign them different names, that is, let fj be the { }j=1 collection of ej’s such that (Aej, ej) = 0. Since A is diagonalized with respect to the 39 ∞ p ∞ basis ej the collection fj forms an orthonormal basis for ker A. Let gj { }j=1 { }j=1 { }j=1 ∞ ∞ ⊥ consist of the remainder of the set ej , which means gj is a basis for ker A. { }j=1 { }j=1 ⊥ Then = 1 2, where 1 = ker A and 2 = ker A. Let Np := n N n p and H H ⊕ H H H { ∈ | ≤ } ∞ ∞ then define diagH ⊕H : ` (Np) ` (N) by 1 2 × → D ϕ1 0 0 0 ··· ··· ...... ...... 0 ϕp 0 0 diag ϕ, ρ = ··· ··· (3.2.1) H1⊕H2 h i 0 0 ρ1 0 ··· ··· . . . . . . .. . . ρ . 2 . 0 0 0 .. ··· ··· p ∞ ⊥ Then because of the way in which we chose fj and gj and because Tr R = 0, { }j=1 { }j=1 B if we let η0 := s(A) and ξ = s(B) then A, B are 0 A = diagH1⊕H2 0, η and B = diag ξ, ∞ 0 where 0 ` (Np) is the zero sequence and by hypothesis ξ p η . ∈ ≺ The heuristic idea of the proof is the following in descriptive informal language. First construct a sequence ξ0 which is a sparsely compressed version of ξ but sufficient to retain majorization by η0, i.e., ξ0 η0. Next apply Case 2 to obtain a special unitary U for ≺ which E(U(diag 0, η0 )U ∗) = diag 0, ξ0 . Finally, apply another unitary to decompress h i h i the diagonal 0, ξ0 to the diagonal ξ. h i p 0 p Now inductively choose sequences of nonnegative integers Nm and N { }m=1 { m}m=0 with the following properties: 0 0 0 0 0 (i) 0 = N < N + 1 < N1 < N < N + 1 < N2 < N < if p = , 0 0 1 1 2 ··· ∞ 0 (or < Np < N if p < ). ··· p ∞ 40 (ii) For each m Np, whenever n Nm (m 1), one has ∈ ≥ − − n+m n X X 0 ξk η . ≤ k k=1 k=1 0 (iii) ξNm + ξN ξNm−1 for each m Np. m ≤ ∈ For transparency and brevity, we only loosely describe the construction of the pair of p 0 p 0 sequences Nm and N . The construction proceeds in pairs, Nm,N . We { }m=1 { m}m=0 m 0 0 may choose Nm to satisfy property (ii) since ξ p η and hence ξ m η because m p. ≺ ≺ ≤ For this we use the fact that property (ii) is an eventual property in the sense that if it holds for some Nm it holds for any larger Nm. Moreover, because ξ > 0 and ξ 0, ξ has ↓ infinitely many strictly decreasing jumps, i.e., for infinitely many j one has ξj ξj+1. If necessary, increase Nm so that it satisfies this condition. Then since ξ 0, we may choose → 0 Nm to satisfy property (iii). To construct the entire pair of sequences, simply iterate this 0 process while simultaneously ensuring Nm+1 > Nm + 1, which we can guarantee because property (ii) is an eventual property. 0 0 0 Next since N < Nm < N < Nm+1 < N for all 1 m < p, one has for p = 0 m m+1 ≤ ∞ 0 0 0 0 N + 1 = 1 N1 < N N2 1 < N 1 N3 2 < N 2 0 ≤ 1 ≤ − 2 − ≤ − 3 − ≤ · · · 0 0 Nm (m 1) < N (m 1) Nm+1 m < N m · · · ≤ − − m − − ≤ − m+1 − ≤ · · · 0 Np (p 1) < N (p 1) for p < . · · · ≤ − − p − − ∞ When p < , if we set N 0 = for convenience of notation, then, regardless of whether ∞ p+1 ∞ p < or p = , these inequalities partition N as ∞ ∞ p G 0 0 N = Nm (m 1),Nm+1 m − − − m=0 41 0 0 with each Nm (m 1) Nm−1 (m 2),Nm (m 1) and m Np. − − ∈ − − − − ∈ Next define the sequence ξ0 which shifts and alters ξ at one point in each interval 0 0 Nm−1 (m 2),Nm (m 1) : for each m Np (or for each m Np+1 if p < , in − − − − ∈ ∈ ∞ which case the last interval is [N 0 (p 1), )), set p − − ∞ ξN + ξN 0 if k = Nm (m 1); m m − − 0 ξk = 0 0 ξk+m−1 if N(m−1) (m 2) k < Nm (m 1) but − − ≤ − − k = Nm (m 1). 6 − − 0 0 This partition of N ensures that ξ is well-defined. Property (iii) guarantees that ξ is monotone decreasing. And property (ii) allows us to conclude that ξ0 η0 which ≺ will follow from equations (3.2.2)–(3.2.4). We omit their straightforward natural proofs for the sake of clarity of exposition. Equations (3.2.2)-(3.2.3) have natural proofs by induction and (3.2.4) is merely one case for when p < . ∞ For all m Np, ∈ k k+m−1 X 0 X 0 ξ = ξj for N (m 2) k < Nm (m 1), (3.2.2) j m−1 − − ≤ − − j=1 j=1 and k k+m−1 X 0 X 0 ξ = ξN 0 + ξj for Nm (m 1) k < N (m 1). (3.2.3) j m − − ≤ m − − j=1 j=1 When p < , the last interval requires separate consideration. That is, if N 0 (p 1) ∞ p − − ≤ 0 0 0 k < N p = and N (p 1) j k one has ξ = ξj+p and so p+1 − ∞ p − − ≤ ≤ j k k+p X 0 X 0 ξ = ξj for k [N (p 1), ) (3.2.4) j ∈ p − − ∞ j=1 j=1 42 It is now simple to prove that ξ0 η0 using condition (ii) and equations (3.2.2)–(3.2.4). ≺ 0 Indeed, notice that for m Np, if Nm−1 (m 2) k < Nm (m 1) and since also ∈ − − ≤ − − 0 m 1 Np 0 and ξ η , and considering separately the cases m = 1 and m > 1, − ∈ ∪ { } ≺ one has k k+m−1 k X 0 X X 0 ξj = ξj ηj. (3.2.5) (3.2.2) ξ≺≤η0 j=1 j=1 or (ii) j=1 0 And for m Np, if Nm (m 1) k < Nm (m 1), one has ∈ − − ≤ − − k k+m−1 k+m k X 0 X X X 0 ξj = ξN 0 + ξj ξj ηj. (3.2.6) (3.2.3) m ≤ ≤ j=1 j=1 j=1 (ii) j=1 Finally, if p < and N 0 (p 1) k < N 0 p = , one has ∞ p − − ≤ p+1 − ∞ k k+p k X 0 X X 0 ξj = ξj ηj. (3.2.7) (3.2.4) ≤ j=1 j=1 (ii) j=1 0 0 Recalling that N is the disjoint union of N (m 2),Nm (m 1) for m Np, (m−1) − − − − ∈ (3.2.2–3.2.4) and (3.2.5–3.2.7) imply that for all k N ∈ k k k X X 0 X 0 ξj ξ η . ≤ j ≤ j j=1 (3.2.2–3.2.4) j=1 (3.2.5–3.2.7) j=1 P∞ P∞ 0 P∞ 0 0 Passing to the limit as k yields ξj = ξ = η since ξ η . Hence → ∞ j=1 j=1 j j=1 j ≺ ξ0 η0. ≺ By Case 2 and our earlier reduction (first paragraph of proof) applied to A = 0 0 0 diag 0, η and B = diag 0, ξ , we obtain a unitary U of the form IH W H1⊕H2 h i H1⊕H2 h i 1 ⊕ for which ∗ 0 0 E(UAU ) = B = diagH1⊕H2 0, ξ . 43 0 p ∞ Since U has this form, for all f, f fj and g gj one has ∈ { }j=1 ∈ { }j=1 0 = UAU ∗f, f 0 = UAU ∗f, g = UAU ∗g, f . h i h i Then for all m Np define Vm on span fm, gNm−(m−1) given by the unitary 2 2 matrix ∈ { } × a b m m Vm = (3.2.8) bm am − where s s 0 ξNm ξNm am = and bm = . 0 0 ξNm + ξNm ξNm + ξNm 0 0 ∗ Then Cm = 0 ξ0 is the compression of UAU to span fm, gN −(m−1) and if Nm−(m−1) { m } one interprets Vm as acting on this same subspace, one computes ξN ∗ m ∗ VmCmVm = (3.2.9) ξN 0 ∗ m because ∗ am bm 0 0 am bm VmCmVm = − 0 bm am 0 ξ bm am − Nm−(m−1) 0 bm(ξN + ξN 0 ) am bm m m − = 0 0 am(ξNm + ξNm ) bm am 2 0 0 bm(ξNm + ξN ) ambm(ξNm + ξN ) = m m 2 0 0 ambm(ξNm + ξNm ) am(ξNm + ξNm ) ξN m ∗ = . ξN 0 ∗ m 44 0 p Next let = span fm, g and let V be the unitary defined by H H { Nm−(m−1)}m=1 p M V = IH0 Vm. ⊕ m=1 From the above computations, one can see that the diagonal operator E(V UAU ∗V ∗) = Π∗(diag ξ)Π, for an appropriate permutation Π of the basis en n∈ . But conjugation by operators { } N which permute the basis corresponding to E, in particular Π, commutes with E, and ∗ ΠVU is unitary, so E((ΠVU)A(ΠVU) ) = diag ξ. Remark 3.2.5. In the proof of Case 1, we only used the facts that A had finite rank and s(B) s(A). Although not needed in this case, the other hypotheses hold automatically ≺ and for edification we explain why. Even though s(B) p s(A) for some p 0 implies ≺ ≥ s(B) s(A), one has the stronger converse: when s(A) has finite support, s(B) s(A) ≺ ≺ implies s(B) p s(A) for every p 0, i.e., s(B) ∞ s(A). Indeed, let Np be the largest ≺ ≥ ≺ index for which s(A) has a nonzero value. Then for all k Np, we have ≥ k+p k+p k X X X sj(B) sj(A) = sj(A) ≤ j=1 j=1 j=1 and therefore s(B) p s(A). Since p is arbitrary, s(B) ∞ s(A). This shows that the ≺ ≺ second inequality in the hypotheses is satisfied for p = since its right-hand side is ∞ infinite. The first inequality is satisfied since Tr R⊥ = because A is finite rank. A ∞ Remark 3.2.6 ([LW15, Remark 2.5] Orthostochasticity). In the above proof, if A is already diagonalized with respect to the basis en n∈ , then all the unitary operators { } N either are orthogonal with respect to this basis, or can be chosen as such. Indeed, Π and V are orthogonal, and U = IH W where W , coming from Theorem 2.1.14, can 1 ⊕ 45 be chosen to be orthogonal by [KW10, Corollary 6.1, NS(ii0) and S(ii0)]. Thus if is a H real Hilbert space and A is a positive compact operators, then it is diagonalizable via an orthogonal matrix, hence Theorem 3.2.4 is valid even when is real instead of complex. H + A consequence of this combined with Lemma 3.1.2 is that if ξ, η c , ξ p η and if ∈ 0 ≺ −1 −1 −1 ξ (0) η (0) p + ξ (0) , ≤ ≤ then there exists an orthostochastic matrix Q for which ξ = Qη. Example 3.2.7 ([KW10, Proposition 6.10, Example 6.11]). The following example, in conjunction with Lemma 1.6, will show that the condition s(B) p s(A) is not necessary ≺ for B E( (A)) and 0 < p := Tr R⊥ Tr R⊥ < . In particular, the converse of ∈ U A − B ∞ Theorem 3.2.4 fails for this case, but is true when Tr R⊥ Tr R⊥ is either infinite or A − B undefined as proved by Corollary 3.3.6. A counterexample is the following. Let η˜ = 0, η where 0 < η c∗. Let Q be an h i ∈ 0 orthostochastic matrix with the property that Qij = 0 if and only if i > j > 1 (so that otherwise Qij > 0 and rows and columns sum to one), e.g., [KW10, Example 6.11]. Then choosing ξ := Qη˜ we claim that ξ 1 η. One can see this from the calculation 6≺ n n n ∞ X ∗ X X X ξ ξi = Qijη˜j i ≥ i=1 i=1 i=1 j=1 n ∞ X X = Qijηj−1 i=1 j=2 n n ∞ n X X X X = Qijηj−1 + Qijηj−1 j=2 i=1 j=n+1 i=1 n−1 ∞ n n−1 X X X X = ηj + Qijηj−1 > ηj, (3.2.10) j=1 j=n+1 i=1 j=1 where the latter inequality follows since η > 0. 46 This example can be easily extended to create similar orthostochastic examples (e.g., Lp−1 1/2 1/2 Q˜ = Q ) for when 1 < p < and 0 < ξ p η, but ξ = Q˜η˜ for 1 1/2 1/2 ⊕ ∞ 6≺ η˜ = 0, η1, 0, η2,..., 0, ηp, ηp+1,... with p zeros. Then since Q˜ is orthostochastic, by h i Lemma 3.1.2 one has that diag ξ = U(diagη ˜)U ∗ for some orthogonal matrix U with 2 Q˜ij = Uij . Therefore p-majorization for any 1 p < is not necessary to characterize | | ≤ ∞ E( (A)). U To verify ξ p η observe that for sufficiently large n, one has 6≺ n p−1 n−2p+2 X ∗ X X ∗ ξ = ηi + (Q 0, ηp, ηp+1,... )i i h i i=1 i=1 i=1 p−1 n−2p+1 X X > ηi + ηi+p−1 (3.2.10) i=1 i=1 p−1 n−p n−p X X X = ηi + ηi = ηi. i=1 i=p i=1 3.2.1. Discussion on majorizations The Figures 2.1–2.2 at the end of the introduction show the interconnections between various types of majorization. Here through Proposition 3.2.10 we begin a discussion of some of these interconnections. Next we exhibit a relationship between strong majorization (recall Definition 2.1.3) and -majorization. As stated, Proposition 3.2.8 may seem 4 ∞ to apply to summable sequences, but in fact the hypotheses, majorization and not strong majorization, negate that possibility as addressed just after the proof. Proposition 3.2.8 ([LW15, Proposition 2.7]). If ξ, η c+ and ξ η, then ∈ 0 ≺ ξ η = ξ ∞ η. 64 ⇒ ≺ 47 Proof. It suffices to show that if ξ η, then the lim inf condition implies ξ p η for every ≺ ≺ p N. Indeed, suppose that ∈ n X lim inf (η∗ ξ∗) = > 0. n→∞ k − k k=1 + ∗ Then since ξ c one can choose N N for which ξ < for all k N. Then for all n ∈ 0 ∈ k 2p ≥ sufficiently large for which both n N and Pn (η∗ ξ∗) > one has that ≥ k=1 k − k 2 n+p n n X X X ξ∗ < ξ∗ + p < η∗. k k · 2p k k=1 k=1 k=1 Note that Proposition 3.2.8 applies only to ξ, η / `1, since if either ξ or η is summable, ∈ majorization implies both are summable, and in this case majorization and strong majorization are equivalent (see Definition 2.1.3, succeeding comment). Furthermore, the converse of Proposition 3.2.8 fails because there exist sequences ξ, η c∗ `1 for which ∈ 0 \ ξ ∞ η hence also ξ η, and yet ξ η, as exhibited in the following example. ≺ ≺ 4 Example 3.2.9 ([LW15, Example 2.8]). In this exposition so far we have not consid- ered any hands-on examples of -majorization. In refuting this possible converse of ∞ Proposition 3.2.8 we provide one, but first we explain our natural motivation for it. ⊥ Motivation. Suppose A + has Tr RA = = Tr R . Then there is some basis with ∈ K ∞ A respect to which A = diag s1(A), 0, s2(A), 0,... . It is a natural question to ask if there h i exists some B E( (A)) with Tr R⊥ = 0. The answer is yes by Theorem 3.2.4, but ∈ U B there is a more straightforward way to see this in this case. Indeed, if we let Rj denote L the 2 2 rotation by π/4 acting on the subspace span e2j−1, e2j , and U = Rj, then × { } ∗ 1 ∞ B := E(UAU ) = diag D2s(A), where D2ϕ denotes the 2-ampliation of ϕ ` , i.e., 2 ∈ D2ϕ = ϕ1, ϕ1, ϕ2, ϕ2,... . h i 48 ∗ 1 Example. Fix any η c and choose ξ := D2η ∞ η, where -majorization is easily ∈ 0 2 ≺ ∞ verified by taking Np = p. Then k P2n X j=n+1 ηj if k = 2n, n N; (ηj ξj) = ∈ − j=1 P2n−1 1 ηj + ηn if k = 2n 1, n N. j=n+1 2 − ∈ From this it follows that k k X k ηk (ηj ξj) η k . (3.2.11) 2 ≤ − ≤ 2 d /2e j=1 1 From the second inequality in (3.2.11) it follows that if lim inf kηk = 0, then ξ = 2 D2η 4 η −1 (e.g. for η = ((k + 1) log(k + 1)) , lim kηk = 0). However, the first inequality of 1 −1 (3.2.11) shows the inverse, that if lim inf kηk > 0, then ξ = D2η η (e.g. for η = k , 2 64 1 lim kη2k = 2 ). This example provides the failure of the converse of Proposition 3.2.8. ∗ 1 Moreover one has, for η c , D2η η if and only if lim inf kηk = 0. ∈ 0 2 4 The previous example shows that for appropriate η c∗, there exist ξ c∗ which are ∈ 0 ∈ 0 counterexamples to the converse of Proposition 3.2.8. However, with more work one can + ∗ show for every sequence η c , there is some ξ c with ξ ∞ η and ξ η, as the next ∈ 0 ∈ 0 ≺ 4 proposition shows. Though it will not be used later, we present it here for completeness. Proposition 3.2.10 ([LW15, Proposition 2.9]). For every η c+ there exists some ∈ 0 ∗ ξ c with ξ ∞ η and ξ η. ∈ 0 ≺ 4 Proof. Without loss of generality, we may assume η c∗ since the definitions of - ∈ 0 ∞ majorization and strong majorization depend only on its monotonization η∗. 1 If lim inf kηk = 0, set ξ = 2 D2η. By the example preceding this proposition, we have ξ ∞ η and ξ η. ≺ 4 It therefore suffices to assume lim inf kηk > 0. For this we adopt the following conventions. Let juxtaposition denote concatenation of finite sequences. Let `(ϕ) denote 49 P P`(ϕ) Pb the length of a finite sequence ϕ. Let ϕ be an abbreviation of j=1 ϕj and a ϕ Pb an abbreviation for j=a ϕj. Finally, if ϕ is a sequence (not necessarily finite) and b a, b dom ϕ with a b, let ϕ denote the finite subsequence ϕa, ϕa+1, . . . , ϕb . ∈ ≤ |a h i ∗ If lim inf kηk > 0 and η c , then η > 0. To construct ξ we proceed inductively. Let ∈ 0 0 k be any positive sequence converging to zero and let N = 1. Then let N1 be the h i 0 η 0 N0 smallest positive integer for which 0 < ηN . Let p1 be the largest positive integer 1 ≤ 2 j ηN0 k 0 0 0 for which p1ηN1 ηN , i.e., p1 := η 2 and (p1 + 1)ηN1 > ηN = η1. Then setting ≤ 0 N1 ≥ 0 p1 (1) N1 z }| { ξ = η ηN , . . . , ηN , |2 h 1 1 i this choice of p1 guarantees that N1 X X (1) 0 η ξ = η1 p1ηN < ηN . (3.2.12) ≤ − − 1 1 1 (1) N1 Denote the difference between the lengths of ξ and ηj by h i1 (1) M1 := `(ξ ) N1 = p1 1 1 − − ≥ 0 0 (1) P N1−1 and then exploiting η 0, choose N1 > `(ξ ) for which η N 0 −M < 1. Then one has → 1 1 0 0 N1−1 N1−1 0 X X (1) N1−1 0 η ξ η < ηN + 1 (3.2.13) ≤ − N1+1 1 1 1 50 0 (1) N1−1 0 because, noting that the length of ξ η is greater by M1 than N1 1, one has N1+1 − 0 0 0 N1−1 N1−1 N1−1 0 0 0 X X (1) N1−1 X X (1) N1−1 X N1−1 0 η ξ η N +1 = η ξ η N +1 + η N 0 −M ≤ − 1 − 1 1 1 1 1 1 N1 0 X X (1) X N1−1 (3.2.14) = η ξ + η N 0 −M − 1 1 1 < ηN1 + 1. Continuing with the induction, suppose that for some k N as in the previous k = 1 ∈ (k) (k) 0 case we are given a finite decreasing sequence ξ and Nk < Nk + Mk = `(ξ ) < Nk, (k) (k) with Mk k and with the last term of ξ being equal to ηN and ξ satisfying ≥ k Nk 0 X X (k) X Nk−1 0 η ξ < ηNk and η N 0 −M < k. (3.2.15) ≤ − k k 1 As in the previous k = 1 case, by an argument identical to that of (3.2.14), together these (3.2.15) inequalities imply 0 0 Nk−1 Nk−1 0 X X (k) Nk−1 0 η ξ η < ηN + k. (3.2.16) ≤ − Nk+1 k 1 1 (k+1) (k) 0 0 We then construct an extension ξ of ξ and Nk+1,Mk+1,Nk+1 with Nk < Nk+1 satisfying all of the properties in the preceding paragraph for k + 1 replacing k in all instances. The procedure mimics the base case as follows. Let Nk+1 be the smallest ηN0 0 k integer greater than N for which 0 < ηN . Along with (3.2.15) this inequality k k+1 ≤ 2 implies that Nk X (k) X ξ + 2ηN η + ηN 0 . k+1 ≤ k 1 51 Hence letting pk+1 be the largest positive integer for which Nk X (k) X ξ + pk+1ηN η + ηN 0 , (3.2.17) k+1 ≤ k 1 one has pk+1 2. And by the maximality of pk+1 one also has ≥ Nk X (k) X ξ + (p + 1)ηN > η + η 0 . (3.2.18) k+1 k+1 Nk 1 Adding to both sides of (3.2.17) and (3.2.18) the η-terms from Nk + 1 to Nk+1 excluding 0 (k+1) Nk, one defines ξ as denoted and obtains pk+1 times Nk+1 0 z }| { X (k) Nk−1 Nk+1 X ξ η N +1η N 0 +1 ηNk+1 , . . . , ηNk+1 η, (3.2.19) k k ≤ | {z } 1 ξ(k+1) and Nk+1 X (k+1) X ξ + ηNk+1 > η. (3.2.20) 1 So from (3.2.19)–(3.2.20) the difference of the (3.2.20) sums is nonnegative and less (k+1) than ηNk+1 . This shows that ξ , as defined in (3.2.19), satisfies the first inequalities of (3.2.15). Note further that ξ(k+1) is decreasing since ξ(k) and η are decreasing and (k) because the last term of ξ is ηNk . As for M1, and recalling that pk+1 2, set ≥ (k+1) (k) Mk+1:= `(ξ ) Nk+1 = (`(ξ ) + Nk+1 (Nk + 1) + pk+1) Nk+1 − − − = Mk 1 + pk+1 > Mk k, − ≥ ∗ 0 (k+1) and hence Mk+1 k + 1. Next, since η c we may choose some N > `(ξ ) ≥ ∈ 0 k+1 satisfying the last inequality of (3.2.15), for k + 1 replacing k. These facts again imply (3.2.16) for k + 1 replacing k by an argument identical to (3.2.14). 52 (k) 0 By induction, we have constructed ξ ,Nk,Mk,Nk with the desired properties (i.e., paragraph containing inequalities (3.2.15) and (3.2.16)). Furthermore, by construction each ξ(k) is an extension of the preceding ones. Thus, the infinite sequence ξ given by (k) (k) ξj := ξ when 1 j `(ξ ) is well-defined. Finally it suffices to show that ξ ∞ η j ≤ ≤ ≺ and ξ 4 η. In order to prove ξ ∞ η it suffices to observe the following two facts. Firstly, if ≺ 1 m N1 then ≤ ≤ m m m X X N1 X ξ = η ηN η. (3.2.21) 2 h 1 i ≤ 1 1 1 Secondly, for each k N, if Nk < m Nk+1 then ∈ ≤ m+M m X k X ξ η (3.2.22) ≤ 1 1 because m+Mk m+Mk 0 X X (k) Nk−1 Nk+1 ξ = ξ η η 0 ηNk+1 Nk+1 Nk+1 1 1 (k) `(ξ ) m+Mk 0 X (k) X (k) Nk−1 Nk+1 = ξ + ξ η η 0 ηNk+1 Nk+1 Nk+1 1 `(ξ(k))+1 (k) `(ξ ) m−Nk 0 X (k) X Nk−1 Nk+1 = ξ + η η 0 ηNk+1 Nk+1 Nk+1 1 1 Nk m−Nk m X X Nk+1 X η + η = η. ≤ Nk+1 1 1 1 Indeed, since Mk , these inequalities (3.2.21) and (3.2.22) imply ξ k η for infinitely ↑ ∞ ≺ many k N, i.e., ξ ∞ η. ∈ ≺ Finally, (3.2.16) implies ξ 4 η since 0 k N −1 X Xk lim inf (ηj ξj) lim inf (ηj ξj) lim inf(ηNk + k) = 0. k→∞ − ≤ k→∞ − ≤ k→∞ j=1 j=1 53 Operator consequences The following corollary of Theorem 3.2.4 and Proposition 3.2.8 gives a method of ensuring membership in E( (A)), for A +. The purpose of this U ∈ K corollary is to provide a more easily computable way to make this determination in special + cases. For if one is given sequences ξ, η c , establishing ξ p η or its negation seems ∈ 0 ≺ more difficult than verifying ξ η, which requires only ξ η and the strict positivity of 64 ≺ the associated lim inf condition. ⊥ Corollary 3.2.11 ([LW15, Corollary 2.10]). Suppose A, B +, B , Tr R ∈ K ∈ D B ≤ Tr R⊥, and s(B) s(A) but s(B) s(A), then B E( (A)). A ≺ 64 ∈ U Proof. By Proposition 3.2.8 one has s(B) ∞ s(A), and then using Theorem 3.2.4 one ≺ obtains B E( (A)). ∈ U 3.3. Approximate p-majorization (necessity) Theorem 3.2.4 of the last section shows that if p := Tr R⊥ Tr R⊥ 0 or when undefined A − B ≥ we set p = 0, then p-majorization (s(B) p s(A)) is a sufficient condition for membership ≺ in E( (A)), but Example 3.2.7 shows it is not necessary for 0 < p < . In our quest U ∞ to characterize E( (A)) in terms of sequence majorization, we introduce a new type of U majorization called approximate p-majorization[2], which is a necessary condition for membership in E( (A)). U Definition 3.3.1. Given ξ, η c+ and 0 p < , we say that ξ is approximately ∈ 0 ≤ ∞ p-majorized by η, denoted ξ p η, if ξ η and for every > 0, there exists an Np, N - ≺ ∈ such that for all n Np,, ≥ n+p n X X ξ∗ η∗ + η∗ . k ≤ k n+1 k=1 k=1 [2]The author, along with Gary Weiss, introduced the concept of approximate p-majorization in [LW15, Definition 3.1]. Cross-reference the footnote on page 37. The reader should note that the asterisks (indicating monotonization) which appear in the displayed inequality of Definition 3.2.2 are missing from [LW15, Definition 3.1]. This is a typo in [LW15]. 54 Furthermore, if ξ p η for infinitely many p N (equivalently obviously, for all p N), - ∈ ∈ this we call approximate -majorization and denote it by ξ ∞ η. ∞ - Remark 3.3.2 ([LW15, Remark 3.2]). Notice from the above definition that if ξ is p-majorized by η, then ξ is trivially approximately p-majorized by η. However, there is a partial converse with a small loss in that approximate p-majorization implies (p 1)- − majorization. That is, if p > 0 and ξ is approximately p-majorized by η, then by choosing < 1, from the above display ξ is (p 1)-majorized by η. Combining these two facts − yields that ξ ∞ η if and only if ξ ∞ η, which we will exploit later. Furthermore, by ≺ - Remark 3.2.5, if η has only finitely many nonzero terms, then any ξ which is majorized by η is -majorized by η and so also approximately -majorized by η. ∞ ∞ Example 3.3.3 ([LW15, Example 3.3]). It is important to note that p-majorization is distinct from approximate p-majorization. That is, for each 0 < p < we exhibit ∞ + sequences ξ, η c with ξ p η but ξ p η. When p = 1, it suffices to consider the ∈ 0 - 6≺ k+1 −k ∗ sequences ξ = (2 −3)/22k and η = 2 . Elementary calculations verify that ξ, η c ∈ 0 ∗ ∗ (that is, ξ = ξ and η = η ), and ξ 1 η but ξ 1 η. To produce analogous sequences for - 6≺ any p > 1, define p − 1 times (p) z }| { (p) ξ := 1,..., 1 , ξ1, ξ2,... and η := p 1, η1, η2,... . h i h − i (p) (p) (p) (p) Then ξ p η but ξ p η , which proves that p-majorization and approximate - 6≺ p-majorization are distinct. However, it should be noted that these examples were not nearly as easy for us to come by as those for p-majorization. In particular, the examples immediately preceding Remark 3.2.3 came naturally, but the single example above took some effort. For further discussion on pairs of sequences ξ p η but ξ p η see unifying - 6≺ Remark 3.3.9. 55 Our main theorem on necessity for membership in E( (A)) depends on approximate U p-majorization. Theorem 3.3.4 ([LW15, Theorem 3.4]). Suppose A + and B E( (A)). If ∈ K ∈ U ⊥ ⊥ p = min n N 0, Tr RA Tr RB + n , { ∈ ∪ { ∞} | ≤ } then s(B) -p s(A). Proof. Suppose A, B and p are as in the hypotheses of this theorem. We may assume that A has infinite rank for otherwise the conclusion holds because of Remark 3.3.2 and Theorem 3.2.4 proof of Case 1. We may also assume that N := Tr R⊥ < . Otherwise, Tr R⊥ = = Tr R⊥ by B ∞ B ∞ A Proposition 3.2.1 and therefore p = 0. Thus we would need to prove s(B) -0 s(A) which is equivalent to s(B) s(A), and this holds by Theorem 2.1.13. ≺ Note that since Tr R⊥ < , p satisfies Tr R⊥ = Tr R⊥ + p = N + p. It suffices to show B ∞ A B ⊥ s(B) r s(A) for all r N with r p. Because Tr R = N + p N + r, without loss - ∈ ≤ A ≥ of generality via unitary equivalence for A and permutations for B, and the fact that conjugation by a permutation commutes with the expectation E, we may assume that A = diag 0,..., 0, η˜ = diag η0 and B = diag 0,..., 0, s(B) = diag ξ, h| {z } i h| {z } i N+r N where η˜ is s(A) interspersed with p r zeros. (To aid intuition, in case p < we may − ∞ choose r = p and hence also η˜ = s(A).) Since B E( (A)) by Lemma 3.1.2(iv), there ∈ U 0 0 exists a unistochastic matrix Q = (qij) for which Qη = ξ. However, only the double stochasticity of Q is used here. For all m N one has ∈ m m m ∞ ∞ m X X 0 X X 0 0 X 0 X 0 ξi = (Qη )i = qijηj = ηj qij. (3.3.1) i=1 i=1 i=1 j=1 j=1 i=1 56 0 For a doubly stochastic matrix Q = (qij), denote the last quantity in equation (3.3.1) as 0 fm(Q, η ). Now fix any 0 < < 1, and choose N + r < Nr, N for which ∈ Nr, N+r X X N + r < q0 . (3.3.2) − ij i=1 j=1 The existence of Nr, follows from column-stochasticity, which yields ∞ N+r X X 0 qij = N + r. i=1 j=1 Certainly inequality (3.3.2) holds with Nr, replaced by any m Nr,, since Q is a doubly ≥ stochastic matrix and so its entries are nonnegative. Consider a permutation matrix Πm which fixes the first N + r coordinates and has the property that 0 Πmη = η := 0,..., 0, s1(A), s2(A), . . . , sm−N−r+1(A), ηm+2, ηm+3,... , h| {z } | {z } i N+r m−N−r+1 where ηm+2, ηm+3,... are the remaining η˜-terms. (To aid intuition, when p < and we ∞ choose r = p and η˜ = s(A), η0 is already in this form without need of the permutation Πm.) Then notice that Q is both doubly stochastic and satisfies inequality (3.3.2) if and −1 only if QΠm does the same. Notice also that inequality (3.3.2) depends only on the first N + r columns of Q. Then direct computations show 0 −1 0 −1 0 −1 fm(Q, η ) = fm(QΠm Πm, η ) = fm(QΠm , Πmη ) = fm(QΠm , η). (3.3.3) −1 Denote the entries of QΠm by qij. From the definition of η above it is clear that −1 ηm+1 ηk whenever k m + 1. This yields an upper bound for fm(QΠ , η) when ≥ ≥ m 57 m Nr, ≥ ∞ m −1 X X fm(QΠm , η) = ηj qij j=1 i=1 m m ∞ m X X X X = ηj qij + ηj qij j=N+r+1 i=1 j=m+1 i=1 m m ∞ m X X X X ηj qij + sup ηk qij ≤ j=N+r+1 i=1 k>m j=m+1 i=1 m m m m X X X X = ηj qij + ηm+1 m qij − j=N+r+1 i=1 j=1 i=1 m m m m ! X X X X = ηj qij + ηm+1 1 qij − j=N+r+1 i=1 j=1 i=1 m m m m ! X X X X = ηj qij + ηm+1 1 qij − j=N+r+1 i=1 j=N+r+1 i=1 N+r m ! X X + ηm+1 1 qij − j=1 i=1 m m m m ! X X X X ηj qij + ηj 1 qij ≤ − j=N+r+1 i=1 j=N+r+1 i=1 N+r m X X + ηm+1 N + r qij − j=1 i=1 m X < ηj + ηm+1 by (3.3.2). j=N+r+1 The inequalities above along with equation (3.3.3) yield m−N m X X 0 −1 si(B) = ξi = fm(Q, η ) = fm(QΠm , η) i=1 i=1 m X ηj + ηm+1 ≤ j=N+r+1 m−N−r X = si(A) + sm−N−r+1(A). i=1 58 Since is arbitrary, s(B) r s(A), and since the positive integer r p is arbitrary, - ≤ s(B) -p s(A). The next corollary is an analogue of Theorem 3.1.3. + Corollary 3.3.5. Let ξ, η c be sequences such that p := η−1(0) ξ−1(0) 0 (with ∈ 0 − ≥ the convention that p = 0 when both sets are infinite). Then = ξ p η ξ = Qη for some orthostochastic matrix Q ⇐ ≺ = ξ p η ⇒ - Proof. Apply Remark 3.2.6 for one implication. For the other implication, apply Lemma 3.1.2 and Theorem 3.3.4. One of our main results is Corollary 3.3.6 which, in the rather general setting where A has infinite rank and infinite dimensional kernel, we obtain a precise characterization of E( (A)) in terms of majorization and -majorization. U ∞ Corollary 3.3.6 ([LW15, Corollary 3.5]). Suppose A + has infinite rank and infinite ∈ K ⊥ dimensional kernel (Tr RA = = Tr R ). Then ∞ A E( (A)) = E( (A))fk E( (A))ik, U U t U the members of E( (A)) with finite dimensional kernel and infinite dimensional kernel, U respectively, are characterized by ⊥ E( (A))fk = B + s(B) ∞ s(A) and Tr R < U { ∈ D ∩ K | ≺ B ∞} and ⊥ E( (A))ik = B + s(B) s(A) and Tr R = . U { ∈ D ∩ K | ≺ B ∞} 59 Proof. If B E( (A)), then B E( (A)) and we know that s(B) s(A) by Theo- ∈ U ∈ V ≺ ⊥ rem 2.1.13. But when Tr R < , from Theorem 3.3.4 we know that s(B) ∞ s(A) B ∞ - which is equivalent to s(B) ∞ s(A). Thus the left-hand set in Corollary 3.3.6 is ≺ contained in the right-hand set. ⊥ Next suppose that B + lies in the right-hand set. If Tr R = , then ∈ D ∩ K B ∞ Theorem 3.2.4 with p = 0 shows that B E( (A)). Similarly, if Tr R⊥ < , then ∈ U B ∞ s(B) ∞ s(A) and again, by Theorem 3.2.4 for p = , we find that B E( (A)). ≺ ∞ ∈ U Corollary 3.3.6 can be expressed as [ n ⊥ o E( (A)) = B + Tr R + p = and s(B) p s(A) . U ∈ D ∩ K B ∞ - 0≤p≤∞ This motivates the following conjectured characterization for E( (A)), which remains an U open problem. And if this conjecture should prove false, is there a proper majorization characterization of E( (A))? U Conjecture 3.3.7 ([LW15, Conjecture 3.6]). Let A +. Then ∈ K [ n ⊥ ⊥ ⊥ o E( (A)) = B + s(B) p s(A) and Tr R Tr R Tr R + p . U ∈ D ∩ K - B ≤ A ≤ B 0≤p≤∞ In general, we have the following squeeze theorem for diagonals of positive compact operators. Corollary 3.3.8. Let A +. Then ∈ K [ n ⊥ ⊥ ⊥ o B + s(B) p s(A) and Tr R Tr R Tr R + p ∈ D ∩ K ≺ B ≤ A ≤ B 0≤p≤∞ E( (A)) ⊆ U ⊆ [ n ⊥ ⊥ ⊥ o B + s(B) p s(A) and Tr R Tr R Tr R + p . ∈ D ∩ K - B ≤ A ≤ B 0≤p≤∞ 60 Moreover, these are equal when Tr R⊥ 0, , otherwise the first inclusion is proper. A ∈ { ∞} Comparing the definitions of p-majorization (Definition 3.2.2) and approximate p- majorization (Definition 3.3.1), we see that they are somehow -close in a vague sense. Thus, although Corollary 3.3.8 does not provide a complete characterization of diagonals of positive compact operators with nonzero finite dimensional kernel, it does squeeze them between these two -close sets. The following remark provides a method for producing pairs of sequences ξ c+ and ∈ 0 ∗ η c for which ξ p η but ξ p η. ∈ 0 - 6≺ Remark 3.3.9 ([LW15, Remark 3.7]). Recall that Example 3.2.7 provided an orthos- tochastic matrix Q such that for every η c∗, setting ξ = Qη˜, where η˜ = 0, η , yields ∈ 0 h i ξ 1 η. Using Lemma 3.1.2 one has diag ξ E( (diag η)). By Theorem 3.3.4, 6≺ ∈ U ∗ ∗ ξ = s(diag ξ) -1 s(diagη ˜) =η ˜ = η, hence ξ -1 η. In general, given two finite sequences, say ϕ, ζ, of lengths n and n + p Pn Pn+p (where n is arbitrary) and having the same sum (i.e., j=1 ϕj = j=1 ζj), one can 0 0 0 0 prepend ϕ to η and ζ to ξ, clearly obtaining new sequences ξ , η with ξ -p η but 0 0 ξ p η . 6≺ Although we did not mention it earlier, in Example 3.2.7 there is a substantial amount of freedom in choosing Q. In particular, examination of [KW10, Example 6.11] ensures that each sum-1 strictly positive column vector followed by a Gram–Schmidt process produces a distinct orthostochastic matrix Q that can be used in Example 3.2.7. Perhaps these orthostochastic Q can be exploited or modified to prove Conjecture 3.3.7. 61 3.4. E(U(A)) convexity Historically, convexity played a central role and is ubiquitous in majorization theory. For example, Horn [Hor54, Theorem 1] integrates theorems of Hardy, Littlewood and P´olya [HLP88] and Birkhoff [Bir46] to prove n n ξ R ξ η = co η˜ R η˜k = ηπ(k), π Πn , { ∈ | ≺ } { ∈ | ∈ } where Πn is the set of n n permutation matrices. For operators, in [Hor54], using × ∗ [Sch23], Horn proved that E( (X)) is convex whenever X = X Mn(C) by establishing U ∈ the characterization n E( (X)) = diag d d R , d λ , U { | ∈ ≺ } where λ is the eigenvalue sequence of X (Theorem 1.2.5). However, the verification that E( (X)) is convex is immediate from its majorization characterization even without the U theorem of Horn integrating Hardy, Littlewood, P´olya and Birkhoff mentioned above. Likewise it is straightforward to verify that if η c+, then ∈ 0 ξ c+ ξ η { ∈ 0 | ≺ } is convex. In particular, this leads to the results of Kaftal and Weiss on the convexity of the expectation of the partial isometry orbit of a positive compact operator. Corollary 3.4.1 ([KW10, Corollary 6.7]). Let A +. Then ∈ K E( (A)) is convex. • V If RA = I or A has finite rank, then E( (A)) is convex. • U Since we have a characterization of E( (A)) when A + has both infinite rank and U ∈ K infinite dimensional kernel, it seems natural to ask if E( (A)) is convex in this case. U 62 The answer is positive, however the verification is much less obvious to us (see below Corollary 3.4.3). But first, a lemma. Lemma 3.4.2 ([LW15, Lemma 4.2]). Suppose that ξ, ζ, η c+, 0 < λ < 1, 0 p, q ∈ 0 ≤ ≤ −1 −1 −1 −1 such that ξ p η, ζ q η. If r = min p + ξ (0) ζ (0) , q + ζ (0) ξ (0) , ∞ - - { \ \ } then λξ + (1 λ)ζ r η. − - Proof. Set ϕ := λξ + (1 λ)ζ. There are two cases: either η has finite support, or − not. If the former, then since ξ η and ζ η, one easily has ϕ η by the comment ≺ ≺ ≺ immediately preceding Corollary 3.4.1. Then since η has finite support, we can improve this to ϕ ∞ η (see Remark 3.3.2 or proof of Theorem 3.2.4 Case 1), which is equivalent ≺ to ϕ -∞ η. Thus ϕ -r η. The second case: η has infinite support. For now, suppose both p, q are finite. Let −1 ∗ −1 π : N N ϕ (0) be a bijection monotonizing ϕ: ϕ = ϕπ(k). Since ϕ (0) = → \ k ξ−1(0) ζ−1(0), one has ∩ −1 −1 −1 −1 −1 −1 −1 N ϕ (0) = N (ξ (0) ζ (0)) ξ (0) ζ (0) ζ (0) ξ (0) , (3.4.1) \ \ ∪ t \ t \ which immediately yields the disjoint union −1 −1 π([1, m]) = π([1, m]) N (ξ (0) ζ (0)) (cardinality km) ∩ \ ∪ −1 −1 π([1, m]) ξ (0) ζ (0) (cardinality sm) t ∩ \ −1 −1 π([1, m]) ζ (0) ξ (0) (cardinality tm) t ∩ \ for each m N. Therefore m = km + sm + tm, and each term increases to the cardinality ∈ of its corresponding set from equation (3.4.1). From this cardinality equation, it is clear that m sm = km + tm and m tm = km + sm are both increasing sequences. However, − − 63 we may further conclude that they increase without bound. Indeed,