0% found this document useful (0 votes)
111 views3 pages

Diamond Graph

This document presents a proof that embedding the diamond graph Gk into Lp space for 1 < p ≤ 2 requires distortion of at least k(p-1)+1. As a corollary, this shows that there does not exist an analogue of the Johnson-Lindenstrauss dimension reduction lemma for L1 space. The proof relies on showing a inequality relating distances of points in Lp spaces and applying it to distances in the diamond graph embeddings. This establishes a lower bound on required distortion which then implies a lower bound on the dimension needed for embeddings into l1 space.

Uploaded by

dorax
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
111 views3 pages

Diamond Graph

This document presents a proof that embedding the diamond graph Gk into Lp space for 1 < p ≤ 2 requires distortion of at least k(p-1)+1. As a corollary, this shows that there does not exist an analogue of the Johnson-Lindenstrauss dimension reduction lemma for L1 space. The proof relies on showing a inequality relating distances of points in Lp spaces and applying it to distances in the diamond graph embeddings. This establishes a lower bound on required distortion which then implies a lower bound on the dimension needed for embeddings into l1 space.

Uploaded by

dorax
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Embedding the diamond graph in Lp and

dimension reduction in L1
James R. Lee∗ Assaf Naor
U.C. Berkeley Microsoft Research
arXiv:math.FA/0407520 v1 30 Jul 2004

Abstract
We show that any embedding of the level k diamond graph p of Newman and Rabi-
novich [6] into Lp , 1 < p ≤ 2, requires distortion at least k(p − 1) + 1. An imme-
diate corollary is that there exist arbitrarily large n-point sets X ⊆ L1 such that any
2
D-embedding of X into ℓd1 requires d ≥ nΩ(1/D ) . This gives a simple proof of a recent
result of Brinkman and Charikar [2] which settles the long standing question of whether
there is an L1 analogue of the Johnson-Lindenstrauss dimension reduction lemma [4].

1 The diamond graphs, distortion, and dimension


We recall the definition of the diamond graphs {Gk }∞k=0 whose shortest path metrics are known
to be uniformly bi-lipschitz equivalent to a subset of L1 (see [3] for the L1 embeddability of
general series-parallel graphs). The diamond graphs were used in [6] to obtain lower bounds
for the Euclidean distortion of planar graphs and similar arguments were previously used in a
different context by Laakso [5].
G0 consists of a single edge of length 1. Gi is obtained from Gi−1 as follows. Given an
edge (u, v) ∈ E(Gi−1 ), it is replaced by a quadrilateral u, a, v, b with edge lengths 2−i . In
what follows, (u, v) is called an edge of level i − 1, and (a, b) is called the level i anti-edge
corresponding to (u, v). Our main result is a lower bound on the distortion necessary to embed
Gk into Lp , for 1 < p ≤ 2.

Theorem 1.1. For every 1 < p ≤ 2, any embedding of Gk into Lp incurs distortion at least
p
1 + (p − 1)k.

The following corollary shows that the diamond graphs cannot be well-embedded into low-
dimensional ℓ1 spaces. In particular, an L1 analogue of the Johnson-Lindenstrauss dimension
reduction lemma does not exist. The same graphs were used in [2] as an example which shows
the impossibility of dimension reduction in L1 . Our proof is different and, unlike the linear
programming based argument appearing there, relies on geometric intuition. We proceed by
observing that a lower bound on the rate of decay of the distortion as p → 1 yields a lower
bound on the required dimension in ℓ1 .

Corollary 1.2. For every n ∈ N, there exists an n-point subset X ⊆ L1 such that for every
2
D > 1, if X D-embeds into ℓd1 , then d ≥ nΩ(1/D ) .

Work partially supported by NSF grant CCR-0121555 and an NSF Graduate Research Fellowship.

1
Proof. Since ℓd1 is O(1)-isomorphic to ℓdp when p = 1 + log1 d and Gk is O(1)-equivalent to a
q
subset X ⊆ L1 , it follows that 1 + logk d = O(D). Noting that k = Ω(log n) completes the
proof.

2 Proof
The proof is based on the following inequality. The case p = 2 is the well known “short
diagonals lemma” which was central to the argument in [5, 6].

Lemma 2.1. Fix 1 < p ≤ 2 and x, y, z, w ∈ Lp . Then,

ky − zk2p + (p − 1)kx − wk2p ≤ kx − yk2p + ky − wk2p + kw − zk2p + kz − xk2p .

Proof. For every a, b ∈ Lp , ||a + b||2p + (p − 1)||a − b||2p ≤ 2(||a||2p + ||b||2p ). A simple proof of
this classical fact can be found, for example, in [1]. Now,

||y − z||2p + (p − 1)||y − 2x + z||2p ≤ 2||y − x||2p + 2||x − z||2p

and
||y − z||2p + (p − 1)||y − 2w + z||2p ≤ 2||y − w||2p + 2||w − z||2p .
Averaging these two inequalities yields

||y − 2x + z||2p + ||y − 2w + z||2p


||y − z||2p + (p − 1) ≤ kx − yk2p + ky − wk2p + kw − zk2p + kz − xk2p .
2
The required inequality follows by convexity.

Lemma 2.2. Let Ai denote the set of anti-edges at level i and let {s, t} = V (G0 ), then for
1 < p ≤ 2 and any f : Gk → Lp ,

k
X X X
||f (s) − f (t)||2p + (p − 1) ||f (x) − f (y)||2p ≤ ||f (x) − f (y)||2p .
i=1 (x,y)∈Ai (x,y)∈E(Gk )

Proof. Let (a, b) be an edge of level i and (c, d) its corresponding anti-edge. By Lemma 2.1,
||f (a) − f (b)||2p + (p − 1)||f (c) − f (d)||2p ≤ ||f (a) − f (c)||2p + ||f (b) − f (c)||2p + ||f (d) − f (a)||2p +
||f (d) − f (b)||2p . Summing over all such edges and all i = 0, . . . , k − 1 yields the desired
result by noting that the terms ||f (x) − f (y)||2p corresponding to (x, y) ∈ E(Gi ) cancel for
i = 1, . . . , k − 1.

The main theorem now follows easily.

Proof of Theorem 1.1. Let f : Gk → Lp be a non-expansive D-embedding. Since |Ai | = 4i−1


and the length of a level i anti-edge is 21−i , applying Lemma 2.2 yields 1+(p−1)k
D2
≤ 1.

2
References
[1] K. Ball, E. A. Carlen and E. Lieb. Sharp uniform convexity and smoothness inequalities
for trace norms. Invent. Math. 115, 463-482 (1994).

[2] B. Brinkman and M. Charikar. On the Impossibility of Dimension Reduction in ℓ1 . In


Proceedings of the 44th Annual IEEE Conference on Foundations of Computer Science.
ACM, 2003.

[3] A. Gupta, I. Newman, Y. Rabinovich and A. Sinclair. Cuts, trees and ℓ1 embeddings. In
Proceedings of the 40th Annual Symposium on Foundations of Computer Science. ACM,
1999.

[4] W. B. Johnson and J. Lindenstrauss. Extensions of Lipschitz mappings into a Hilbert


space. In Conference in modern analysis and probability (New Haven, Conn., 1982), pages
189-206. Amer. Math. Soc., Providence, RI, 1984.

[5] T. J. Laakso. Ahlfors Q-regular spaces with arbitrary Q > 1 admitting weak Poincaré
inequality. Geom. Funct. Anal., 10(1):111-123, 2000.

[6] I. Newman and Y. Rabinovich. A Lower Bound on the Distortion of Embedding Planar
Metrics into Euclidean Space. Discrete Computational Geometry, 29 no. 1, 77-81 (2003).

You might also like