TOPICS IN CONVEX GEOMETRY AND PHENOMENA IN HIGH DIMENSION
by
DEPING YE
Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Dissertation Advisors: Dr. Stanislaw J. Szarek and Dr. Elisabeth M. Werner
Department of Mathematics
CASE WESTERN RESERVE UNIVERSITY
August, 2009 CASE WESTERN RESERVE UNIVERSITY
SCHOOL OF GRADUATE STUDIES
We hereby approve the thesis/dissertation of
______
candidate for the ______degree *.
(signed)______(chair of the committee)
______
______
______
______
______
(date) ______
*We also certify that written approval has been obtained for any proprietary material contained therein. Table of Contents
Table of Contents ...... iii Dedication ...... v Acknowledgement ...... vi Abstract ...... vii
1 Introduction and Background 1 1.1 Introduction and Overview of Results ...... 1 1.2 Background on convex geometry and asymptotic geometric analysis . 7 1.2.1 Preliminaries on Convex Bodies ...... 7 1.2.2 Lp affine surface area and mixed p-affine surface area . . . . . 12 1.2.3 Inequalities related to the volume radius ...... 14 1.2.4 Symmetrization and Rogers-Shephard inequality ...... 19 1.2.5 John ellipsoid and L¨ownerellipsoid ...... 20 1.2.6 Tensor products of convex bodies ...... 21 1.3 Background on Geometry of Quantum States ...... 23 1.3.1 Mathematical framework ...... 23 1.3.2 Peres-Horodecki’ positive partial transpose criterion ...... 24 1.3.3 Bures metric on D ...... 26 1.3.4 Hilbert-Schmidt and Bures measures on D ...... 32
2 New Lp Affine Isoperimetric Inequalities 40 2.1 L n affine surface area of the polar body ...... 40 − n+2 2.2 Lp affine surface areas ...... 47 2.3 Lp affine isoperimetric inequalities ...... 53
3 Inequalities for mixed p-affine surface area 61 3.1 Mixed p-affine surface area and related inequalities ...... 62 3.1.1 Inequalities for mixed p-affine surface area ...... 62 3.1.2 i-th mixed p-affine surface area and related inequalities . . . . 73 3.2 Illumination surface bodies ...... 80 3.3 Geometric interpretation of functionals on convex bodies ...... 86
iii 4 On the Bures Volume of Separable Quantum States 99 4.1 Hilbert-Schmidt volume of separable quantum states ...... 100 4.2 Bures volume of separable quantum states ...... 106 4.3 Optimality of the bounds ...... 116 4.3.1 Optimality of the lower bound ...... 116 4.3.2 Optimality of the upper bound ...... 117 4.4 Conclusion and Comments ...... 119 Bibliography ...... 121
iv Dedication
To my new born baby boy: Joseph Hengrui Ye. He is the best gift for my 30-years birthday and the completion of my Ph. D. degree in mathematics. I hope that he will have a bright future, a happy and healthy life.
To my lovely wife: Minglu Qian. My dissertation was started right after our marriage. Her strong support and encouragement were essential for the completion of my Ph.D. dissertation.
v ACKNOWLEDGEMENTS
I am greatly indebted to my advisors Dr. Stanislaw J. Szarek and Dr. Elisa- beth M. Werner, who are not only great advisors but also true mentors, for their patience, kindness and many beneficial discussions. Their guidance made me become more and more mature, both academically and as a person. Their achievements and professionalism as mathematician were and will always be my models.
My thanks also go to the Mathematics Department at Case Western Reserve University. I would like to thank the Institut Henri Poincar´eand the Mathematics Department at Texas A&M University for providing excellent research environments during my visits. Special thanks go to my defense committee members Dr. Harsh Mathur and Dr. Mark Meckes.
My research has been partially supported by grants from the National Science Foundation (U.S.A.).
Finally, I would like to give my deepest gratitude to my wife Minglu for her encouragement and understanding during the past few years. My appreciation also goes to my supportive family: my parents Jinhe and Baihua, my sister Zhongying, my grandmother-in-law Yu’e, and my parents-in-law Guangzhen and Jihong for their dedication and ongoing support.
vi Topics in Convex Geometry and Phenomena in High Dimension
Abstract
by
DEPING YE
This dissertation deals with topics in convex geometry and phenomena in high dimension. Convex geometry studies geometry of convex bodies of fixed dimension, while phenomena in high dimension aims to understand the structure of large dimen- sional objects.
Our objects of study in convex geometry are related to Lp affine surface areas
and mixed p-affine surface areas. In particular, we prove new Lp affine isoperimetric inequalities for all p ∈ [−∞, 1). For p ≥ 1, such inequalities had been established ear- lier. We generalize these inequalities to the analogous inequalities for mixed p-affine surface area. We also prove new Alexandrov-Fenchel type inequalities for mixed p- affine surface area. To generalize the Lp affine surface and mixed p-affine surface area to all p, we study the asymptotic behavior of the volume of the illumination surface bodies and the polar bodies of the surface bodies. Properties of the illumination sur- face bodies are established, for instance, we show that the illumination surface bodies are star-convex, but not necessarily convex. As applications of our asymptotic results, we give geometric interpretations of functionals associated with convex bodies, such as Lp affine surface area, surface area, and mixed p-affine surface area. Moreover, we establish, for all p 6= −n, a duality formula which shows that Lp affine surface area
vii ◦ of a convex body K equals L n2 affine surface area of the polar body K . p In phenomena in high dimension, we will study the relative sizes of various sets of quantum states. We obtain two sided estimates for the Bures volume of an arbitrary subset of the set of N × N density matrices in terms of the Hilbert-Schmidt volume of that subset. For general subsets, our results are essentially optimal (for large N). As applications, we derive, in particular, nontrivial lower and upper bounds for the Bures volume of sets of separable states and for sets of states with positive partial transpose.
Key words: affine isoperimetric inequality, Lp affine surface area, Lp Brunn- Minkowski theory, mixed p-affine surface area, Bures metric, Bures volume, sepa- rable states, positive partial transpose.
viii Chapter 1
Introduction and Background
1.1 Introduction and Overview of Results
Convex geometry studies the geometry of convex bodies in fixed dimension, while asymptotic geometric analysis strives to understand structures in high dimensions by using tools of convex geometry, functional analysis and probability. Asymptotic geometric analysis provides “isomorphic” rather than “isometric” solutions of many open problems in geometry and such a relaxation often makes these problems solvable.
One part of the dissertation will focus on Lp affine surface area and mixed p- affine surface area. In particular, we prove new isoperimetric inequalities and new Alexandrov-Fenchel type inequalities. Another part of the dissertation is about ob- taining nontrivial bounds for the Bures volume of sets of quantum separable states for large dimension.
Isoperimetric inequalities play an important role in many areas of mathematics. For instance, the classical isoperimetric inequality, which compares the area of a convex body K with its volume, is an extremely powerful tool in geometry and related areas. Many isoperimetric inequalities have an affine invariant flavor. As an example we mention the celebrated Blaschke-Santal´oinequality which can be obtained as a consequence of the classical affine isoperimetric inequality [78]. This classical affine
1 2 isoperimetric inequality provides an upper bound of as(K), the affine surface area of a convex body K, in terms of its volume. More generally, affine isoperimetric inequalities compare two functionals associated with convex bodies (or more general sets) where the ratio of the functionals is invariant under non-degenerate linear or affine transformations.
The affine surface area, as(K), of a convex body K was introduced by Blaschke in 1923 for dimension 2 and 3 in [13]. Originally a basic concept from the field of affine differential geometry, it has recently attracted increasing attention (e.g. [4, 5, 10, 48, 53, 55, 56, 60, 69, 84]). Affine surface area has many nice and useful properties. Aside from the fact that there is an affine isoperimetric inequality associate with it, it is invariant under affine maps and it is a valuation. It-in a sense-“measures” the boundary behavior of a convex body and therefore appears naturally and plays crucial roles in the theory of approximation of convex bodies by polytopes. We refer the readers to [33, 86, 57] for the details and more references. Also it is the subject of the affine Plateau problem solved in R3 by Trudinger and Wang [102], and Wang [108].
In the ground breaking paper [62], Lutwak introduced a generalization of the
classical affine surface area, Lp affine surface area for all p ≥ 1. When p = 1, L1 affine
surface area is just the classical affine surface area. The notion of Lp affine surface area was further extended to all p and to general convex bodies in, e.g., [70, 85, 86]. In fact, extensions were obtained by investigating the asymptotic behavior of the volume of certain families of (convex) bodies. These ideas were already successfully used in the study of the geometric interpretation of the classical affine surface area via the convex floating body by Sch¨uttand Werner [84] and the convex illumination body by Werner [109], (see also [51, 69].)
Lp affine surface area is now at the core of the rapidly developing Lp Brunn- Minkowski theory. Contributions here include the discovery of new ellipsoids [54, 64], the study of solutions of nontrivial ordinary and, respectively, partial differential 3 equations (see e.g. Chen [18], Chou and Wang [20], Stancu [96, 97]), and the study of the Lp Christoffel-Minkowski problem by Hu, Ma and Shen [44].
A generalization of Lp affine surface area is mixed p-affine surface area which, for p ≥ 1, was also introduced by Lutwak in [62]. Aside from Lp affine surface area, other special cases of mixed p-affine surface area are the dual mixed volume and the surface area. In this dissertation, we will extend the definition of the mixed p-affine surface area to all p ∈ [−∞, ∞]. To do so, we will provide new geometric interpretations for functionals on convex bodies. In particular, for Lp affine surface area, mixed p- affine surface area, and i-th mixed p-affine surface area (see below for the definitions). In Chapter 2, we investigate the asymptotic behavior of the volume of the polar of surface bodies, and give a new geometric interpretation of Lp affine surface area. As ◦ ◦ an application, we show the duality formula as n2 (K ) = asp(K) for p 6= −n where K p is the polar body of K. This duality formula was proved by Hug [47] for p > 0 using a different method. In Chapter 3, we construct a new class of bodies, the illumination surface bodies, and study the asymptotic behavior of their volumes. We show that the illumination surface bodies are not necessarily convex, thus introducing a novel idea in the theory of geometric characterizations of functionals on convex bodies, where to date only convex bodies were used (e.g. [70, 85, 86]).
The classical affine isoperimetric inequality is fundamental in many problems. Besides the ones already mentioned earlier, see also, e.g., [30, 31, 63, 82]. More recently, it was used, e.g., by Andrews [6, 7], Sapiro and Tannenbaum [79] to show the uniqueness of self-similar solutions of the affine curvature flow and to study its
asymptotic behavior. Lp affine isoperimetric inequalities for all p ≥ 1 were first proved
by Lutwak in [62]. He gave upper bounds of the Lp affine surface area of K for all p ≥ 1
in terms of the volume of K. We derive new Lp affine isoperimetric inequalities for all
p ∈ [−∞, 1) in Chapter 2. These Lp affine isoperimetric inequalities can be viewed as extensions of the Blaschke-Santal´oinequality and inverse Santal´oinequality (due to Bourgain and Milman [14] (see also Kuperberg [49])). We also provide examples 4 to show that these inequalities cannot be improved.
In Chapter 3, we prove isoperimetric inequalities for mixed p-affine surface area . For mixed p-affine surface area, Alexandrov-Fenchel type inequalities (for p = 1, ±∞) and affine isoperimetric inequalities (for 1 ≤ p ≤ n) were first established by Lutwak in [58, 59, 62]. We derive new Alexandrov-Fenchel type inequalities for mixed p-affine surface area for all p ∈ [−∞, ∞] and new mixed p-affine isoperimetric inequalities for all p ∈ [0, ∞]. Classification of the equality cases for all p in the Alexandrov-Fenchel type inequalities for mixed p-affine surface area is related to the uniqueness of solutions
of the Lp Minkowski problem (e.g., [18, 20, 61, 63, 65, 66, 96, 97]), which is unsolved for many cases. Thus the situation is similar to the classical Alexandrov-Fenchel inequalities for mixed volume, where the complete classification of the equality cases is also an unsolved problem.
Phenomena in high dimension studies geometric objects in dimension n, as n becomes large. The ultimate goal is to obtain uniform estimations independent of dimension n, or to obtain estimations with best dependence on dimension n. A typical result is the Bourgain-Milman’s version of inverse Santal´oinequality, which says that the volume radius product of K and K◦ is essentially a constant. This inverse Santal´o inequality provides an “isomorphic” solution of the Mahler conjecture (see Section 1.2.3 for details).
The techniques from phenomena in high dimension have been used to study the relative size of the set of separable quantum states within the set of all quantum states. In [8, 99], the authors obtained rather striking results: in large dimension, all but extremely few quantum states are entangled (not separable), in the sense of Hilbert-Schmidt measure. Moreover, they proved that the powerful positive partial transpose criterion is not precise in large dimension as a tool to detect the separability, in the sense of Hilbert-Schmidt measure. Their results rely on the Euclidean structure of the Hilbert-Schmidt metric. An arguably more important measure is the Bures measure, which is induced by the non-Euclidean Bures metric. In Chapter 4, we 5 deal with the relative size of the set of separable quantum states (and other sets of interest) within the set of all quantum states in terms of the Bures measure.
Quantum entanglement was discovered in 1930’s [25, 83] and is now at the heart of quantum computation and quantum information. The key ingredients in quantum algorithms such as Shor’s algorithm for integer factorization [87] or Deutsch-Jozsa algorithm (see e.g. [72]), are entangled quantum states, i.e., those states which can not be represented as a mixture of tensor products of states on subsystems. Fol- lowing [112], states that can be so represented are called separable states. Since determining whether a state is entangled or separable is in general a difficult prob- lem [34], sufficient and/or necessary conditions for separability are very important in quantum computation and quantum information theory, and have been studied extensively in the literature (see e.g. [38, 39, 40, 41, 42, 43, 73]). One well-known tool is the Peres-Horodecki’s positive partial transpose (PPT) criterion [38, 73], that is, if a state on H = CD1 ⊗ CD2 · · · ⊗ CDn is separable then its partial transpose must be positive. Equivalently, if a state on H does not have positive partial transpose, it must be entangled. This criterion works perfectly, namely, the set of separable states S = S(H) equals to the set of states with positive partial transpose PPT = PPT (H) for H = C2 ⊗ C2 (two-qubits), H = C2 ⊗ C3 (qubit-qutrit), and H = C3 ⊗ C2 (qutrit- qubit) [38, 79, 113]. However, entangled states with positive partial transpose appear in the composite Hilbert space H = C2 ⊗ C4 and H = C3 ⊗ C3 [39] (and of course in all “larger” composite spaces). The results in [8, 99] show that, in some measures, Qn the positive partial transpose criterion becomes less and less precise as N = i=1 Di grows to infinity. This is inferred by comparing the Hilbert-Schmidt volumes of S and PPT . The methods in [8, 99], which rely on the special geometric properties of the Hilbert-Schmidt metric and tools of phenomena in high dimension, can also be employed to derive tight estimates for the Hilbert-Schmidt volume of D = D(H) (the set of all states on H). However, a closed expression for the exact value of this volume is known; it was found in [115] via random matrix theory and calculating some nontrivial multivariate integrals. 6
Compared with the Hilbert-Schmidt metric, the Bures metric on D [16, 103] is, in some sense, more natural. It has attracted considerable attention (see e.g. [24, 23, 22, 46, 45, 106, 105, 104]). The Bures metric is Riemannian [74] but not flat. It is monotone, i.e., it does not increase under the action of any completely positive, trace preserving map. It induces the Bures measure [12, 35, 95], which has singularities on the boundary of D. The Bures volume of D has been calculated exactly in [95] and 2 1 happens to be equal to the volume of an (N − 1)-dimensional hemisphere of radius 2 [12, 95]. (This mysterious fact does not seem to have a satisfactory explanation.) On the other hand, the precise Bures (or Hilbert-Schmidt) volumes of S and PPT are rather difficult to calculate since the geometry of these sets is not very well understood and the relevant integrals seem quite intractable. These quantities can be used to measure the priori Bures probabilities of separability and of positive partial transpose within the set of all quantum states. (Here priori means that the state is selected randomly according to the Bures measure and no further information about it is available.) For small N, e.g., N = 2 × 2 and N = 2 × 3, the Bures volume of S (hence of PPT ) has been extensively studied by numerical methods in [93, 92, 91, 90, 89, 88]. For large N, the asymptotic behavior of the Hilbert-Schmidt volume of S and PPT was successfully studied in [8, 99]. Based on that work, we derive qualitatively similar “large N” results for the Bures volume. In summary, our results state that the relative size of S within D is extremely small for large N (see Corollaries 7 and 8 for detail). On the other hand, the corresponding relative size for PPT within D is, in the Bures volume radius sense (see Section 1.3.4 for a precise definition), bounded from below by a universal (independent of N) positive constant (see Corollary 9). In conclusion, the priori Bures probability of finding a separable state within PPT is exceedingly small is when N is large. In other words, we have shown that, as a tool to detect separability, the positive partial transpose criterion for large N is not precise in the priori Bures probability sense. Its effectiveness to detect entanglement is less clear (see comments following Corollary 9). 7 1.2 Background on convex geometry and asymp- totic geometric analysis
1.2.1 Preliminaries on Convex Bodies
A convex body K in Rn is a convex, compact subset of Rn with nonempty interior. Unless stated otherwise, here we will always assume that the centroid of a convex n n n body K in R is at the origin. We use B2 to denote the Euclidean ball in R and n−1 n n−1 S to denote the unit sphere, the boundary of B2 . For u ∈ S , the support function of a convex body K is defined as
hK (u) = maxhx, ui x∈K where h·, ·i is the standard inner product on Rn which induces the Euclidian norm k · k. The radial function of K at u ∈ Sn−1 is
ρK (u) = max{λ : λu ∈ K}.
n−1 + Both hK (·) and ρK (·) are functions from S to R and both determine a convex body uniquely. The volume of K can be calculated as Z 1 n |K| = ρK (u) dσ(u) n Sn−1 where σ is the usual surface area measure on Sn−1. More generally, for a set M, we use |M| to denote the Hausdorff content of its appropriate dimension.
For any convex body K with the origin in its interior, K◦ = {y ∈ Rn, hx, yi ≤ 1, ∀x ∈ K} is the polar body of K. K◦ is a convex body with the origin in its interior. The bipolar theorem (see [82]) states that (K◦)◦ = K. It is easy to check that for u ∈ Sn−1, 1 1 ρK (u) = , and ρK◦ (u) = . hK◦ (u) hK (u) Hence the volume of K◦ is Z Z ◦ 1 n 1 1 |K | = ρK◦ (u) dσ(u) = n dσ(u). n Sn−1 n Sn−1 hK (u) 8
n Let K be a convex body in R with the origin in its interior. The gauge k · kK of K is defined as
kxkK = inf{λ > 0 : x ∈ λK}.
Then one has for all u ∈ Sn−1,
1 kukK = = hK◦ (u). ρK (u)
If K is 0-symmetric, i.e., K = −K, k · kK is the norm which has K as its unit ball.
If K and L are two convex bodies in Rn with the origin in their interiors, we define the geometric distance between K and L as the product αβ where α and β are the smallest numbers such that β−1L ⊂ K ⊂ αL. This quantifies how the “shapes” of K and L differ.
2 We use ∂K to denote the boundary of convex body K. We write K ∈ C+, 2 if ∂K is C with everywhere strictly positive Gaussian curvature κK (x). We use
NK (x) to denote the unit outer normal vector of x ∈ ∂K. Thus, the hyperplane n H(x, NK (x)) = {y ∈ R : hNK (x), yi = hNK (x), xi} passing through x with normal 2 vector NK (x) is the tangent hyperplane of ∂K at x. In the case K ∈ C+, the Gauss n−1 1 map NK (·): ∂K → S is C . The hyperplane H(x, NK (x)) can be identified with n−1 R with (any) orthonormal basis, say {e˜1, ··· , e˜n−1}. Then {e˜1, ··· , e˜n−1,NK (x)} forms an orthonormal basis of Rn with the origin at x. In a neighborhood of x ∈ ∂K, ∂K can be represented by a convex function f : Rn−1 → R with f(0) = 0 as following: for any point z ∈ ∂K in a neighborhood of x,
z = x + (t1, ··· , tn−1, −f(t1, ··· , tn−1)).
Then κK (x), the Gaussian curvature of K at x ∈ ∂K, is defined as the determinant of 2 the Hessian matrix of f at 0. K ∈ C+ implies that κK (·) is a continuous function on ∂K. Moreover, the spherical measure σ is absolutely continuous with respect to the
pushforward of the surface area measure µK (or the (n − 1)-dimensional Hausdorff 9 measure) on ∂K along NK , and vice versa. Hence, by the Radon-Nikodym theorem, n−1 there is a function fK (·): S → R, such that
dµ (x) K = f (u) dσ(u) K and also dσ(u) = κK (x) dµK (x)
1 where u = NK (x). Consequently, κK (x) = . The function fK (u) is the curvature fK (u) 2 function of the convex body K ∈ C+ at the direction u and fK (·) is a continuous n−1 function on S . In fact, the curvature function fK (·) is closely related to the support function and can be determined as follows (see Corollary 2.5.3 in [82]).
Corollary 1 fK (u) equals to the product of all nonzero eigenvalues of the Hessian n n−1 matrix (with respect to any orthonormal basis of R ) of hK at u ∈ S .
For any convex body K, its volume can be calculated as follows:
1 Z |K| = hx, NK (x)i dµK (x). n ∂K
2 If K ∈ C+, the above formula is equivalent to
1 Z |K| = fK (u)hK (u) dσ(u). n Sn−1
For two convex bodies K and L with the origin in their interiors, and λ,η ≥ 0 (not both zero), the Minkowski linear combination λK + ηL is the convex body with n−1 support function hλK+ηL where for u ∈ S
hλK+ηL(u) = λhK (u) + ηhL(u).
A fundamental inequality in convex geometry is the Brunn-Minkowski inequality. This inequality also holds for more general compact sets instead of convex bodies. 10
Theorem 1 (Brunn-Minkowski inequality) Let K and L be two convex bodies in Rn. Then the volume measure is log-concave. That is, for all t ∈ [0, 1],
|tK + (1 − t)L| ≥ |K|t |L|1−t,
with equality for all t ∈ [0, 1] if and only if K = L + a for some a ∈ Rn. Equivalently,
1 1 1 |K + L| n ≥ |K| n + |L| n ,
with equality if and only if K = cL + a for some a ∈ Rn and c ∈ R.
The mixed volume of K and L, denoted by V1(K,L), is defined by
|K + L| − |K| V1(K,L) = lim . →0 n
n It is easy to verify that nV1(K,B2 ) coincides with the usual surface area of K, i.e., n nV1(K,B2 ) = µK (K). Moreover, V1(K,K) = |K|. Alexandrov [2] and Fenchel and Jessen [26] have shown that for each convex body K there is a positive Borel measure S(K, ·) on Sn−1, such that, for all convex bodies L,
1 Z V1(K,L) = hL(u) dS(K, u). n Sn−1
As noted in [60], the measure S(K, ·) has the following simple geometric description: n−1 for any Borel subset A of S , S(K,A) is just the µK measure of the set of points of ∂K which have an unit outer normal vector in A, namely
S(K,A) = µK ({x ∈ ∂K : ∃u ∈ A, s.t., H(x, u) is a support hyperplane of ∂K at x}).
2 Note that S(K, ·) is in general different from µK . For K ∈ C+, the measure pushfor-
ward of µK by NK coincides with S(K, ·), and hence S(K, ·) is absolutely continuous with respect to the spherical measure σ on Sn−1. The Radon-Nikodym theorem then implies that dS(K, u) = f (u). dσ(u) K 11
Good general references for this material are [17, 50].
The Minkowski linear combination and the mixed volume V1(·, ·) can be general- ized to all p ≥ 1. For convex bodies K and L in Rn with the origin in their interiors and λ, η ≥ 0 (not both zeros), the Firey p-sum λK +p ηL for p ≥ 1 [28] is the convex body with support function