Expander Graphs: The Unbalanced Case Omer Reingold The Weizmann Institute What's in This Talk? • Expander Graphs – an array of definitions. • Focus on most established notions, and open problems on explicit constructions. Mainly in the unbalanced case since this is – What applications often require – Where constructions are very far from optimal • Will flash one construction (no details) Unbalanced expanders based on ParvareshVardy Codes [Guruswami,Umans,Vadhan 06] Bipartite Graphs • As a preparation for the unbalanced case we will talk of bipartite expanders. • Can also capture undirected expanders: G - Undirected N D Symmetric N N D Vertex Expansion N S, |S| K N D |(S)| A |S| (A > 1) Every (not too large) set expands. Vertex Expansion N S, |S| K N D |(S)| A |S| (A > 1) •Goal: minimize D (i.e. constant D) • Degree 3 random graphs are expanders [Pin73] Vertex Expansion N S, |S| K N D |(S)| A |S| (A > 1) Also: maximize A. • Trivial upper bound: A D – even A ≲ D-1 • Random graphs: AD-1 nd 2 Eigenvalue Expansion N N D • 2nd eigenvalue (in absolute value) of (normalized) adjacency matrix is bounded away from 1 • Can be interpreted in terms of Renyi (l2) entropy Expanders Add Entropy N Prob. dist. X x N D Induced dist. X’ x’ • Vertex expansion: |Support(X’)| A |Support(X)| • Some applications rely on “less naïve” measures of entropy. • Col(X) = Pr[X(1)=X(2)] = ||X||2 nd 2 Eigenvalue Expansion N X N D X’ • Col(X’) –1/N 2 (Col(X) –1/N) • Renyi entropy (log 1/Col(X)) increases as long as: < 1 and Col(X) is not too small nd 2 Eigenvalue Expansion N X N D X’ • Interestingly, vertex expansion and 2ndeigenvalue expansion are essentially equivalent for constant degree graphs [Tan84, AM84, Alo86] Explicit Constructions Applications need explicit constructions: • Weakly explicit: easy to build the entire graph (in time poly N). • Strongly explicit: – Given vertex name x and edge label i easy to find the ith neighbor of x (in time poly log N). Explicit constructions – 2nd Eigenvalue • Celebrated sequence of algebraic constructions [Mar73, GG80,JM85,LPS86,AGM87,Mar88,Mor94,...]. • Optimal 2nd eigenvalue (Ramanujan graphs) • “Combinatorial” constructions: [Ajt87, RVW00, BL04]. • Open: Combinatorial constructions of strongly explicit Ramanujan (or almost Ramanujan) graphs. • Getting “close”: [Ben-Aroya,Ta-Shma 08] Explicit constructions – Vertex Expansion • Optimal 2nd eigenvalue expansion does not imply optimal vertex expansion • Exist Ramanujan graphs with vertex expansion D/2 [Kah95]. • Lossless Expander – Expansion > (1-) D • Why should we care? – Limitation of previous techniques – Many applications Property 1: A Very Strong Unique Neighbor Property S, |S| K, |(S)| 0.9 D |S| S Unique neighbor of S Non Unique neighbor S has 0.8 D |S| unique neighbors ! • We call graphs where every such S has even a single unique neighbor – unique neighbor expanders Property 2: Incredibly Fault Tolerant S, |S| K, |(S)| 0.9 D |S| Remains a lossless expander even if adversary removes (0.7 D) edges from each vertex. Explicit constructions – Vertex Expansion • Open: lossless expanders for the undirected case. – Unique neighbor expanders are known [AC02] • For the directed case (expansion only from left side), lossless expanders are known [CRVW02]. Expansion D-O(D). • Open: expansion D-O(1) (even with nonconstant degree). Unbalanced Expanders • Many applications need N N D Unbalanced Expanders • Many applications need unbalanced expanders: N M D Array of Definitions N • Many flavors: X M D – How unbalanced. – Measure of entropy. – Lossless vs. lossy. – Is X’ close to full entropy? – Lower vs. upper bound on entropy of X. –… X’ Vertex Expansion Revisited N S, |S|= N 0.9 M D |(S)| 10 D • Even previously trivial tasks require D = (log N/log M) • M << N Farewell constant degree Slightly-Unbalanced ConstantDegree Lossless Expanders M= N N S, |S| K D |(S)| (1-) D |S| CRVW02: 0<, 1 constants D constant & K= (N) In case someone asks: K= ( M/D) & D= poly(1/ , log (1/ )) (fully explicit: D= quasipoly(1/ , log (1/ ))) Open: More Unbalanced N M D • E.g. M=N0.5 and sets of size at most K=N0.2 expand. While being greedy: • Unique neighbor expanders • Lossless expanders • Minimal Degree Super-Constant Degree N S, |S| K M D |(S)| (1-)D |S| • State of the art [GUV06]: D=Poly(LogN), M=Poly(KD) (w. some tradeoff). • Open: M=O(KD) (known w. D=QuasiPoly(LogN)) • Open: D= O(LogN) Dispersers [Sipser 88] N S, |S|≥ K M D |(S)| > (1-) M • Bounds: • D ≥ 1/ log(N/K) • DK/M ≥ log 1/ -- must be lossy • Explicit constructions are (comparably) good but still not optimal … Increasing Entropy? N Prob. dist. X x M D Induced dist. X’ x’ •Can Renyi entropy increase ? • |Col(X’)| < |Col(X)| (essentially) D> min{M0.5, N/M} Extractors [NZ 93] N X M≪N D X’ • (k,)-extractor if Min-entropy(X) k X’ -close to uniform • Min-entropy(X) k if x, Pr[x] 2-k • X and Y are -close if maxT | Pr[XT] - Pr[YT] | = ½ ||X-Y||1 Equivalently Extractors = Mixing N S, |S|= K M T, D | e(S,T)/DK - |T|/N | < • Vertex Expansion – Sets on the left have many neighbors. • Mixing Lemma – the neighborhood of S hits any T with roughly the right proportion. 2-Source Extractors source of biased correlated bits another independent weak source EXT almost uniform output random bits • Recently – lots of attention and results • Randomness Extractors are a special case, where the 2nd source is truly random. Explicit Constructs. of Extractors • Extractors are highly motivated in Interpretation: extracting an applications. As a general rule of thumb: arbitrary constant fraction Interpretation: extracting all of “Anythingentropy expanders can do, extractors the entropy can do better” … • Lots of progress. Still very far from optimal. Best in one direction [LRVW03, GUV06]: D=Poly(LogN / ), M=2k(1-) • Selected open problem: M=2k with D=Poly(LogN / ) A Word About Techniques • Research on randomness extractors was invigorated with the discovery of a beautiful and surprising connection to pseudorandom generators [Tre99]. • This further led to discoveries of connections between extractors and error correcting codes [Tre99, RRV99, TZ01, TZS01, SU01]. • In particular, [GUV06] relies on Parvaresh-Vardy list-decodable codes [GUV06] - Basic Construction • Left vertex f Fqn (poly. of degree· n-1 over Fq) • Edge Label y F • Right vertices = Fqm+1 y’th neighbor of f = 2 m-1 h h h (y, f(y), (f mod E)(y), (f mod E)(y), …, (f mod E)(y)) where E(Y) = irreducible poly of degree n h = a parameter Thm: This is a (K,A) expander with K=hm, A = q-hnm. Conclusions • Many interesting variants of expander graphs • Constructions in general – very far from optimal • Any clean and useful algebraic characterization?