9512.net

# On the embeddability of weighted graphs in Euclidean spaces

On the Embeddability of Weighted Graphs in Euclidean Spaces
Abdo Y. Alfakih Henry Wolkowicz y University of Waterloo Department of Combinatorics and Optimization Waterloo, Ontario N2L 3G1, Canada Research Report CORR 98{12 May 28, 1998

Key Words: weighted graphs, Euclidean distance matrices, semide nite programming, extreme points.
Given an incomplete edge-weighted graph, G = (V; E; !), G is said to be embeddable in < , or r-embeddable, if the vertices of G can be mapped to points in < such that every two adjacent vertices v , v of G are mapped to points x , x 2 < whose Euclidean distance is equal to the weight of the edge (v ; v ). Barvinok 3] proved that if G is r-embeddable p for some r, then it is r -embeddable where r = b( 8jE j + 1 ? 1)=2c. In this paper we provide a constructive proof of this result by presenting an algorithm to construct such an r -embedding.
r r i j r i j i j

Abstract

1 Introduction
Let G = (V; E; ! ) be an incomplete undirected edge-weighted graph with vertex set V = fv1 ; v2; : : :; v g, edge set E V V and a nonnegative weight
n

henry@orion.math.uwaterloo.ca.

yResearch supported by Natural Sciences Engineering Research Council Canada. E-mail

E-mail aalfakih@orion.math.uwaterloo.ca

1

! for each (v ; v ) 2 E . G is said to be r-embeddable if there exists a mapping : V ! < such that for every edge (v ; v ) 2 E , the Euclidean distance k (v ) ? (v )k = ! ; and, it is said to be embeddable if it is r-embeddable for some positive integer r. The Embeddability problem (r-embeddability problem) is the problem of determining whether G = (V; E; ! ) is embeddable (r-embeddable). In matrix theory literature, the r-embeddability problem is known as the Euclidean distance matrix completion problem (EDMCP ). An n n symmetric matrix D = (d ) with nonnegative elements and with zero diagonal is said to be a Euclidean distance matrix (EDM) if there exist points x1 ; x2; : : :; x 2 < for some r such that
ij i j r i j i j ij r ij n r

d = kx ? x k2; i; j = 1; 2; : : :; n: The smallest such r is called the embedding dimension of D. Note that r is always n ? 1. Let A be an n n symmetric partial matrix with some elements speci ed or xed and the rest are unspeci ed or free. The problem of determining whether A can be completed to a EDM with embedding dimension r, by assigning certain values to its free elements, is called EDMCP ; and, it is called EDMCP if the embedding dimension is unrestricted. The equivalence between the embeddability (r-embeddability) and EDMCP (EDMCP ) is immediate if, in the matrix A = A(! ), the elements a corresponding to (v ; v ) 62 E are free, while elements a corresponding to (v ; v ) 2 E are xed as a = ! 2 . Thus G = (V; E; ! ) is embeddable (r-embeddable) if and only if A = A(!) can be completed to a EDM (EDM with embedding dimension r). The r-embeddability problem or the (EDMCP ) has received much attention in recent years because of its many applications in elds such as chemistry, statistics, archaeology, genetics, and geography 1]. In some applications (e.g. the molecular conformation problems in chemistry 4, 15]), the embedding dimension r is required to be 2 or 3, while in others (e.g. multidimensional scaling in statistics 12, 13]), only a small r is desired. It was shown by Saxe 20] that for any given positive integer r, the r-embeddability problem of an integerweighted graph G = (V; E; ! ) is NP-Hard. It remains so even if the weights ! are restricted to 1 or 2 only. In 3], Barvinok proved that if G = (V; E; !) p is embeddable, then it is r -embeddable where r = b( 8jE j + 1 ? 1)=2c. His proof was based on the duality of semide nite programming and on the \corank formula" for the strata of singular quadratic forms. However, no method was presented to construct such an r -embedding if G is embeddable. Another proof of Barvinok's result is given in 6].
ij i j r r ij i j ij i j ij ij r ij

2

In this note we present a constructive proof of Barvinok's result. i.e., we present an algorithm to construct such an r -embedding if G is embeddable. This algorithm consists of two steps. In step 1, EDMCP is formulated as a semide nite programming problem. This problem is then solved, using an interior-point method, to determine whether G = (V; E; ! ) is embeddable or not. If it is embeddable, due to the nature of the interior-point methods, an r-embedding with large r is usually obtained. In step 2, a rounding procedure, exploiting the facial structure of the set of optimal solutions of the semide nite program, is used to construct an r -embedding starting from the r-embedding obtained in step 1.

2 Problem Formulation
Given an edge-weighted graph G(V; E; ! ), let H be its adjacency matrix and let ( 2 ! ;v ) 2 E A = A(!) = (a ) = 0 if (votherwise
ij ij i j

Note that wlog all free elements of A are originally set to zero. Consider the following problem :=
F

min f (D) = kH (D ? A)k2 subject to D 2 E;
F

F

(1)

where k k denote the Frobenius norm de ned as k A k = trace A A , ( ) denotes the Hadamard product or the element-wise product, and E denotes the cone of EDMs. Let D be the optimal solution of (1). Then, it immediately follows that G = (V; E; ! ) is embeddable if and only if f (D ) is zero. Furthermore, (1) can be posed as a semide nite programming problem be exploiting the relation between the cone E and P , the cone of positive semide nite matrices. This relation is established in the following subsection.
T

p

It is well known, e.g. 21, 7, 9, 22], that a symmetric matrix D with nonnegative elements and with zero diagonal is a EDM if and only if D is negative semide nite on n o M := x 2 < : x e = 0 ;
n T

2.1 Distance Geometry

3

the orthogonal complement of e, the vector of all ones. Let S denote the space of symmetric matrices of order n. De ne the centered and hollow subspaces of S as S := fB 2 S : Be = 0g; (2) S := fD 2 S : diag(D) = 0g;
n n C n H n

where diag(D) denotes the column vector formed from the diagonal of D. Following 5], de ne the two linear operators acting on S
n

K(B) := diag(B) e + e diag(B) ? 2B;
T T

(3) (4)

and
T

where J = I ? ee is the orthogonal projection matrix onto subspace M . n

1 T (D) := ? 2 JDJ;

Theorem 2.1 The linear operators satisfy K(S ) = S ; T (S ) = S ; and KjSC and TjSH are inverses of each other. Proof. See 8, 10].
C H H C

It can easily be veri ed that

2

formed from the vector De. Thus, a hollow matrix D is EDM if and only if B = T (D) is positive semide nite, to be denoted as B 0. Equivalently, D is EDM if and only if D = K(B ); for some B with Be = 0 and B 0. In this case, the embedding dimension r is given by the rank of B . If B = XX for some n r matrix X , then the rows of X are the coordinates of the points x1; x2; : : :; x that generate D. Furthermore, since Be = 0; it follows that the origin coincides with the centroid of these points. For these and other basic results on EDM see e.g. 8, 9, 10, 11, 21]. Let V be an n (n ? 1) matrix such that
T n

K (D) = 2 (Diag(De) ? D) (5) is the adjoint operator of K; where Diag(De) denotes the diagonal matrix

V e=0; V V =I ;
T T

(6)

4

where I 2 S ?1 is the identity matrix. Hence, the subspace M can be represented as the range of V . J , the orthogonal projection matrix onto M , is then given by J := V V = I ? ee . Now introduce the composite operators n
n T T

and

K (X ) := K(V XV );
V T T

(7)

T (D) := V T (D)V = ? 1 V DV; 2
V T

(8)

Lemma 2.1
K (S ?1) = S ; T (S ) = S ?1;
V n H V H n

and K and T are inverses of each other on these two spaces.
V V

It follows from (5) that

K (D) = V K (D)V (9) is the adjoint operator of K . The following corollary summarizes the relationships between E , the cone of Euclidean distance matrices of order n, and P ?1, the cone of positive semide nite matrices of order n ? 1. Corollary 2.1 Suppose that V is de ned as in (6). Then: K (P ?1) = E ; T (E ) = P ?1: Proof. We saw earlier that D is EDM if and only if D = K(B) with
t V V n V n V n t t t

Be = 0 and B 0. Let X = V BV , then since Be = 0 we have B = V XV . Therefore, V XV 0 if and only if X 0; and the result follows using (7) and Lemma 2.1. 2 Thus (1) is equivalent to the following semide nite programming problem 2 (EDMCP ) := min f (X ) = kH K0:(X ? B )k X where B = T (A). The optimal EDM D can be recovered from the optimal solution X of EDMCP using D = K (X ). Two comments are in order.
V F V V

5

First, the EDMCP is a convex problem which can be solved in polynomial time with an arbitrary precision. Second, the EDMCP is intractable since restricting the rank of X to r destroys the convexity of the problem. EDMCP is solved using a Primal-Dual Interior- Point algorithm introduced in 2]. Due to the nature of Interior-Point methods, optimal solutions of large ranks are often obtained. Rounding this optimal solution into another one with a lower rank necessitates the study of the set of optimal solution of EDMCP in the next section
r

3 Set of Optimal Solutions of EDMCP
Let X be the optimal solution of EDMCP and assume that = f (X ) = 0. i.e., graph G = (V; E; ! ) is embeddable. f (X 1) = f (X 2) = 0 if and only if Z = X 1 ? X 2 belong to the nullspace of A(X ), where A(X ) = H K (X ). Thus, , the set of optimal solutions of EDMCP, is given by = fX 2 S ?1 : X 0; X = X + Z; where A(Z ) = 0g; is a convex set; and in general it is non-polyhedral. If graph G is complete. i.e., if all o -diagonal elements of H are ones, it readily follows that f (X ) is a strictly convex function and is a single point. On the other hand, if G is disconnected then EDMCP can be solved more simply as two or more smaller subproblems; one for each connected component of G. Thus assume that G is connected. Since G is also assumed to be incomplete, X is not a singleton set. Next we study the facial structure of . (See also the theses 17, 18] and the papers 19, 16, 14] for characterizations for general sets). De nition 3.1 A matrix X 2 is said to be an extreme point if X can not be represented as a proper convex combination of two distinct points X 1 and X 2 in . Lemma 3.1 Let X , Z 2 S ?1 and let X be positive de nite and Z be indefinite. Then, there exist two positive numbers 1 and 2 such that X + 1Z and X ? 2 Z are both positive semide nite with rank n ? 2. Proof. Let X = Q Q be the spectral decomposition of X . Then, X + Z = Q 1 2(I + Z 0) 1 2Q where Z 0 = ?1 2Q ZQ ?1 2 . Let 1 2 : : : be the eigenvalues of 0 . The assertion follows by setting 1 = 1=j 1 j and 2 = 1= Z 2
V n n T = = T = T = n n

6

Lemma 3.2 Let the weighted graph G = (V; E; !) be connected and let Z be a nonzero symmetric matrix in the null space of A. Then Z is inde nite. Proof. Assume to the contrary that Z is de nite; and wlog assume that Z 0. Then D = (d ) = K (Z ) is a EDM. Since G is connected, there exists
a path between node v1 and node v . Hence there exist indices i1 ; i2; : : :; i ?2 such that H1 1 = H 1 2 = = H n?2 = 1. Since A(Z ) = H K (Z ) = H D = 0, it follows that d1 1 = d 1 2 = = d n?2 = 0. Thus D = 0 and consequently Z = 0. 2
n n ;i i ;i i ;n V ;i i ;i i ;n ij V

Corollary 3.1 Let the weighted graph G = (V; E; !) be connected. Then is
bounded.

Next we present a characterization of the extreme points of . Theorem 3.1 X 2 is not an extreme point if and only if there exists a nonzero symmetric matrix Z in the nullspace of A, such that the nullspace of Z contains the nullspace of X . Proof. Let Z 6= 0 be in the nullspace of A such that N (X ) N (Z ), where N (:) denotes the null space. Let rank(X ) = k and let X = Q Q be the spectral decomposition of X . Then, wlog we can assume that = diag( 1; 2; : : :; k; 0; : : :; 0) where are positive for all i = 1; 2; : : :; k. Hence, X = W diag( 1; : : :; k)WT where W = Q 1Q 2 Q ] (Q denotes the j th column of Q). Since N (X ) N (Z ), Z = WZ 0 W . From Lemmas 3.1 and 3.2 there exists a positive number = minf 1; 2g, such that X 1 = X ? Z 0 and X 2 = X + Z 0. Furthermore, X 1 and X 2 belong to since Z 2 N (A). Hence, X is not an extreme point of . To prove the converse, assume that X 2 is not an extreme point. Then 1 there exist two points X 1 6= X 2 in such that X = 2 X 1 + 1 X 2. Let Z = 2 1 (X 1 ? X 2). Then clearly Z 6= 0 and Z 2 N (A). Furthermore, X 1 = X + Z 2 and X 2 = X ? Z . Let rank(X )= k and let X = Q Q , = diag( ), be the spectral decomposition of X . As before, wlog assume that the rst k elements of are positive. Hence, the last (n ? k) rows and columns of the matrix Q XQ are zeros. Then since X + Z 0 and X ? Z 0, it follows that the last n ? k rows and columns of Q ZQ must also be zeros. Hence, N (X ) N (Z ). 2
T i : : :k :j T T T T

7

Corollary 3.2 X is an extreme point of if and only if there does not exist a nonzero symmetric matrix Y such that H K (WY W ) = 0, where the
columns of W form an orthonormal basis for the range of X .
V T

Corollary 3.3 Let the weighted graph G = (V;p !) be connected. Let X , of E; rank k, be an extreme point of . Then k b( 8jE j + 1 ? 1)=2c.
H = 1, i < j ; and k(k +1) variables since Y is symmetric. Thus if k(k +1) > 2jE j this system has a solution and X is not an extreme point of . 2 Next we present an algorithm which rounds anyp point of into an extreme point which by Corollary 3.3 has a rank k b( 8jE j + 1 ? 1)=2c. (An algorithm that does this (purifying step) for general semide nite programs is given in the thesis 17].)
ij

Proof. The system H K (WY W ) = 0 has jE j equations, one for each
1 2
V T

Algorithm 3.1 . Input: Tolerance ; H , the adjacency ( matrix of Graph G = (V; E; ! ); 2 p r = b( 8jE j + 1 ? 1)=2c ; A = (a ) = ! if (v ; v ) 2 E
ij

0

ij

otherwise

i

j

Step 1: solve the corresponding EDMCP to obtain
If If
T

nates of the vertices of G embedded in dimension r

endwhile Output: B = V X V , let B = PP then the rows of P give the coordiT T

and X > , stop. G is not embeddable in any dimension. , r= rank (X ), go to step 2. Step 2: while r > r Solve H K (WY W ) = 0 to get Y . Z = WY W , compute 1 from Lemma 3.1 X X + 1Z . r= rank (X ).
V T

r.

Note that 2 and X X ? 2Z can be used in Step 2 as well. Also, note that after each iteration in step 2, the rank of X decreases by at least one.

Example:

8

Let G = (V; E; ! ) be the 5-cycle with weights 1, 1, 2, 2, 2; and let = 10?9 . Then 00 1 0 0 41 00 1 0 0 11 C B B1 0 1 0 0C B B0 1 0 4 0C C B 0 1 0 1 0 C; A = B 1 0 1 0 0 C H=B B C B0 0 4 0 4C C B0 0 1 0 1C A @ A @ 4 0 0 4 0 1 0 0 1 0 Thus r = 2. Solving the EDMCP we get = 9:9 10?10 , X and D = K (X ) as follows
V

0 0:6557 B 0:3433 B X = B ?0:5383 @ ?0:6333

0:3433 1:0310 ?0:1666 ?0:5687

1 ?0:5383 ?0:6333 ?0:1666 ?0:5687 C C 2:6358 0:5408 C A
0:5408 2:4459

0 0 1:0000 1:8768 4:6143 4:0000 1 B 1:0000 0 1:0000 4:3681 4:3681 C C B D = B 1:8768 1:0000 0 4:0000 4:6143 C B C B 4:6143 4:3681 4:0000 0 4:0000 C A @
4:0000 4:3681 4:6143 4:0000 0 Here r = rank( X )=4. After the rst iteration of the algorithm we get:

4:0000 4:5483 4:9511 4:0000 0 where 1 = 0:9496. Here r= 3. After the second iteration we get: 1 0 0:8218 0:7857 ?0:9130 ?1:4906 C B 0: X = B ?07857 1::7497 0::8238 ?13203 C B :9130 0 8238 3 8979 1: :6228 C A @ ?1:4906 ?1:6228 1:3203 2:7428 9

1 0 0:5602 0:6484 ?0:6611 ?0:8080 C B 0: X = B ?06484 1::7367 0::2013 ?05190 C B :6611 0 2013 2 6660 0: :4211 C A @ ?0:8080 ?0:4211 0:5190 2:3721 0 0 1:0000 3:6757 4:9511 4:0000 1 B 1:0000 0 1:0000 4:5483 4:5483 C C B D = B 3:6757 1:0000 0 4:0000 4:9511 C B C B 4:9511 4:5483 4:0000 0 4:0000 C A @

4:0000 6:5457 7:7380 4:0000 0 where 1 = 2:7805. Here r= 2. Factorizing B = V XV = PP yields the following 2-dimensional embedding
T T

0 0 1:0000 3:4932 7:7380 4:0000 1 B 1:0000 0 1:0000 6:5457 6:5457 C C B D = B 3:4932 1:0000 0 4:0000 7:7380 C B C B 7:7380 6:5457 4:0000 0 4:0000 C A @

x1 = (?0:7284; ?0:9345) x2 = (?1:0844; ?0:0000) x3 = (?0:7284; 0:9345) x4 = ( 1:2706; 1:0000) x5 = ( 1:2706; ?1:0000) Thus we obtained a 2-dimensional embedding for G. However, for these weights, G has also a 1-dimensional embedding namely
T T T T T

x1 = x4 = 0; x2 = 1; x3 = 1; x5 = ?2
which we are not able to obtain. This should come as no surprise since it is NP-hard to nd the extreme points of with the smallest rank.

References
1] S. AL-HOMIDAN. Hybrid methods for optimization problems with positive semide nite amtrix constraints. PhD thesis, University of Dundee, 1993. 2] A. Y. ALFAKIH, A. KHANDANI, and H. WOLKOWICZ. Solving Euclidean distance matrix completion problems via semide nite programming. Technical report, CORR 97-9, Dept. of combinatorics and optimization, University of Waterloo, 1997. 3] A. I. BARVINOK. Problems of distance geometry and convex properties of quadratic maps. Discrete and Computational Geometry, 13:189{202, 1995. 4] G. M. CRIPPEN and T.F. HAVEL. Distance Geometry and Molecular Conformation. Wiley, New York, 1988. 10

5] F. CRITCHLEY. On certain linear mappings between inner-product and squared distance matrices. Linear Algebra Appl., 105:91{107, 1988. 6] M. DEZA and M. LAURENT. Geometry of Cuts and Metrics, Algorithms and Combinatorics. Vol 15. Springer, Berlin, 1997. 7] J. C. GOWER. Properties of Euclidean and non-Euclidean distance matrices. Linear Algebra Appl., 67:81{97, 1985. 8] J. C. GOWER. Properties of Euclidean and non-Euclidean distance matrices. Linear Algebra Appl., 67:81{97, 1985. 9] T. L. HAYDEN, J. WELLS, W-M. LIU, and P. TARAZAGA. The cone of distance matrices. Linear Algebra Appl., 144:153{169, 1991. 10] C.R. JOHNSON and P. TARAZAGA. Connections between the real positive semide nite and distance matrix completion problems. Linear Algebra Appl., 223/224:375{391, 1995. 11] M. LAURENT. A tour d'horizon on positive semide nite and Euclidean distance matrix completion problems. In Topics in Semide nite and Interior-Point Methods, volume 18 of The Fields Institute for Research in Mathematical Sciences, Communications Series, Providence, Rhode Island, 1998. American Mathematical Society. 12] J. DE LEEUW and W. HEISER. Theory of multidimensional scaling. In P. R. KRISHNAIAH and L. N. KANAL, editors, Handbook of Statistics, volume 2, pages 285{316. North-Holland, 1982. 13] SUBHASH LELE. Euclidean distance matrix analysis (EDMA): estimation of mean form and mean form di erence. Math. Geol., 25(5):573{602, 1993. 14] A. S. LEWIS. Eigenvalue-constrained faces. Linear Algebra Appl., (CORR 95-22), 1998. 15] P. M. PARDALOS, D. SHALLOWAY, and G. XUE, editors. Global minimization of nonconvex energy functions: molecular conformation and protein folding, volume 23 of DIMACS Series in Discrete Mathematics and Theoretical Computer Science. American Mathematical Society, Provi-

dence, RI, 1996. Papers from the DIMACS Workshop held as part of the DIMACS Special Year on Mathematical Support for Molecular Biology at Rutgers University, New Brunswick, New Jersey, March 20{21, 1995. 11

16] G. PATAKI. Cone-LP's and semide nite programs: geometry and a simplex-type method. In Integer programming and combinatorial optimization (Vancouver, BC, 1996), volume 1084 of Lecture Notes in Comput. Sci., pages 162{174. Springer, Berlin, 1996. 17] G. PATAKI. Cone programming and eigenvalue optimization: geometry and algorithms. PhD thesis, Carnegie Mellon University, Pittsburgh, PA, 1996. 18] M.V. RAMANA. An algorithmic analysis of multiquadratic and semidefinite programming problems. PhD thesis, Johns Hopkins University, Baltimore, Md, 1993. 19] M.V. RAMANA and A.J. GOLDMAN. Some geometric results in semide nite programming. Technical Report 1, 1995. 20] J. B. SAXE. Embeddability of weighted graphs in k-space is strongly NP-hard. Proc. 17th Allerton Conf. in Communications, Control, and Computing, pages 480{489, 1979. 21] I.J. SCHOENBERG. Remarks to Maurice Frechet's article: Sur la de nition axiomatique d'une classe d'espaces vectoriels distancies applicables vectoriellement sur l'espace de Hilbert. Ann. Math., 36:724{732, 1935. 22] WARREN S. TORGERSON. Multidimensional scaling. I. Theory and method. Psychometrika, 17:401{419, 1952.

12