1: \title{\vspace{-0.5cm}How many random edges make a dense hypergraph
2: non-$2$-colorable?}
3:
4: \author{
5: Benny Sudakov \thanks{Department of Mathematics,
6: Princeton University, Princeton, NJ 08544. E-mail:
7: bsudakov@math.princeton.edu.
8: Research supported in part by NSF
9: grant DMS-0355497, USA-Israeli BSF grant, and by an Alfred P. Sloan
10: fellowship.}
11: \and Jan Vondr\'ak \thanks{ Microsoft Research, Redmond, WA
12: 98502. E-mail: vondrak@microsoft.com.}
13: }
14:
15:
16: \documentclass [11pt]{article}
17: \usepackage{amsfonts}
18: \usepackage{amsmath}
19: \usepackage{latexsym}
20: \usepackage{epsfig}
21: \usepackage{graphicx}
22:
23: \oddsidemargin 0pt
24: \evensidemargin 0pt
25: \marginparwidth 40pt
26: \marginparsep 10pt
27: \topmargin 0pt
28: \headsep 10pt
29: \textheight 8.9in
30: \textwidth 6.6in
31:
32:
33: \newtheorem{theorem}{Theorem}[section]
34: \newtheorem{lemma}[theorem]{Lemma}
35: \newtheorem{proposition}[theorem]{Proposition}
36: \newtheorem{corollary}[theorem]{Corollary}
37: \newtheorem{conj}[theorem]{Conjecture}
38: \newtheorem{remark}[theorem]{Remark}
39: \newtheorem{propos}[theorem]{Proposition}
40: \newtheorem{claim}[theorem]{Claim}
41: \newtheorem{definition}[theorem]{Definition}
42:
43:
44: \def\D{\Delta}
45: \def\Ex{{\bf E}}
46: \def\HL{H} % notation for the lower-bound example
47:
48: \renewcommand{\baselinestretch}{1.2}
49: \date{}
50: \begin{document}
51: \maketitle
52: \begin{abstract}
53: We study a model of random uniform hypergraphs, where a random
54: instance is obtained by adding random edges to a large hypergraph
55: of a given density. The research on this model for graphs has been
56: started by Bohman et al. in \cite{BFM}, and continued in
57: \cite{BFKM} and \cite{KST}. Here we obtain a tight bound on the
58: number of random edges required to ensure non-$2$-colorability. We
59: prove that for any $k$-uniform hypergraph with
60: $\Omega(n^{k-\epsilon})$ edges, adding $\omega(n^{k\epsilon/2})$
61: random edges makes the hypergraph almost surely non-$2$-colorable.
62: This is essentially tight, since there is a $2$-colorable
63: hypergraph with $\Omega(n^{k-\epsilon})$ edges which almost surely
64: remains $2$-colorable even after adding $o(n^{k \epsilon / 2})$
65: random edges.
66: \end{abstract}
67:
68:
69: \section{Introduction}
70:
71: Research on random graphs and hypergraphs has a long history with
72: thousands of papers and two monographs by Bollob\'as \cite{B} and
73: by Janson et al. \cite{JLR} devoted to the subject and its diverse
74: applications. In the classical Erd\H{o}s-R\'{e}nyi model
75: \cite{ER}, a random graph is generated by starting from an empty
76: graph and then adding certain number of random edges. More
77: recently, Bohman, Frieze and Martin \cite{BFM} considered a
78: generalized model where one starts with a fixed graph $G=(V,E)$
79: and then inserts a collection $R$ of additional random edges. We
80: denote the resulting random graph by $G+R$. The initial graph $G$
81: can be regarded as given by an adversary, while the random
82: perturbation $R$ represents noise or uncertainty, independent of
83: the initial choice. This scenario is analogous to the {\em
84: smoothed analysis} of algorithms proposed by Spielman and Teng
85: \cite{ST}, where an algorithm is assumed to run on the worst-case
86: input, modified by a small random perturbation.
87:
88: Usually, one investigates {\em monotone properties} of random
89: graphs or hypergraphs; i.e., properties which cannot be destroyed by
90: adding more edges, like the property of containing a certain fixed
91: subgraph. Given a monotone property $\cal A$ of graphs on $n$ vertices,
92: we can ask what are the parameters for which a random
93: graph has property $\cal A$ almost surely, i.e. with probability
94: tending to $1$ as the number of vertices $n$ tends to infinity.
95: In our setting, we start with a fixed hypergraph $H$ and inquire
96: how many random edges $R$ we have to add so that $H+R$ has
97: property $\cal A$ almost surely. This question is too general
98: to get concrete and meaningful results, valid for all hypergraphs $H$.
99: Therefore, rather than considering a completely arbitrary
100: $H$, we start with a hypergraph from a certain natural class.
101: One such class of graphs was considered in
102: \cite{BFM}, where the authors analyze the question of how many
103: random edges need to be added to a graph $G$ of minimal degree at least $dn, 0<d<1$,
104: so that the resulting graph $G+R$ is almost
105: surely Hamiltonian. Further properties of random graphs in this model are
106: explored in \cite{BFKM}.
107:
108: In \cite{KST}, Krivelevich et al. considered a slightly more
109: general setting, in which one performs a small random perturbation
110: of a graph $G$ with at least $dn^2$ edges. Observe that since $G$
111: has at least $dn^2$ edges, removing a small set of random edges
112: would leave the total number of edges in $G$ essentially
113: unchanged. Therefore one only has to focus on the case of adding
114: random edges. In \cite{KST}, the authors obtained tight results
115: for the appearance of a fixed subgraph and for certain Ramsey
116: properties in this model. In the same paper, they also considered
117: random formulas obtained by adding random $k$-clauses
118: (disjunctions of $k$ literals) to a fixed $k$-SAT formula.
119: Krivelevich et al. proved that for any formula with at least $n^{k
120: - \epsilon}$ $k$-clauses, adding $\omega(n^{k \epsilon})$ random
121: clauses of size $k$ makes the formula almost surely unsatisfiable.
122: This is tight, since there is a $k$-SAT formula with
123: $n^{k-\epsilon}$ clauses which almost surely remains satisfiable
124: after adding $o(n^{k\epsilon})$ random clauses. A related
125: question, which was raised in \cite{KST}, is to find a threshold
126: for non-$2$-colorability of a random hypergraph obtained by adding
127: random edges to a large hypergraph of a given density.
128:
129: For an integer $k \geq 2$, a {\em $k$-uniform hypergraph}
130: is an ordered pair $H=(V,E)$, where $V$ is a finite non-empty set, called set of
131: {\em vertices} and $E$ is a family of distinct $k$-subsets of $V$, called
132: the {\em edges} of $H$. A {\em $2$-coloring} of a hypergraph $H$ is a partition of
133: its vertex set $V$ into two color classes so that no edge in $E$ is monochromatic.
134: A hypergraph which admits a $2$-coloring is called {\em $2$-colorable}.
135:
136: $2$-colorability is one of the fundamental properties of
137: hypergraphs, which was first introduced and studied by Bernstein
138: \cite{Ber} in 1908 for infinite hypergraphs. $2$-colorability in
139: the finite setting, also known as ``Property B" (a term coined by
140: Erd\H{o}s in reference to Bernstein), has been studied extensively
141: in the last forty years (see, e.g., \cite{Er, Er1, EL, Beck, RS}).
142: While $2$-colorability of graphs is well understood being
143: equivalent to non-existence of odd cycles, for $k$-uniform
144: hypergraphs with $k \geq 3$ it is already $NP$-complete to decide
145: whether a $2$-coloring exists \cite{L}. Consequently, there is no
146: efficient characterization of $2$-colorable hypergraphs. The
147: problem of $2$-colorability of random $k$-uniform hypergraphs for
148: $k \geq 3$ was first studied by Alon and Spencer \cite{AS}. They
149: proved that such hypergraphs with $m=(c2^k/k^2)n$ edges are almost
150: surely $2$-colorable. This bound was improved later by Achlioptas
151: et al. \cite{AKKT}. Recently, the threshold for $2$-colorability
152: has been determined very precisely. In \cite{AM} it was proved
153: that the number of edges for which a random $k$-uniform hypergraph
154: becomes almost surely non-$2$-colorable is $(2^{k-1} \ln 2 - O(1))
155: n$.
156:
157: Interestingly, the threshold for non-$2$-colorability is roughly
158: one half of the threshold for $k$-SAT. It has been shown in
159: \cite{AP} that a formula with $m$ random $k$-clauses becomes
160: almost surely unsatisfiable for $m = (2^k \ln 2 - O(k)) n$. The
161: two problems seem to be intimately related and it is natural to
162: ask what is their relationship in the case of a random
163: perturbation of a fixed instance. Recall that from \cite{KST} we
164: know that for any $k$-SAT formula with $n^{k-\epsilon}$ clauses,
165: adding $\omega(n^{k \epsilon})$ random clauses makes it almost
166: surely unsatisfiable. In fact, the same proof yields that for any
167: $k$-uniform hypergraph $H$ with $n^{k-\epsilon}$ edges, adding
168: $\omega(n^{k \epsilon})$ random edges destroys $2$-colorability
169: almost surely. Nonetheless, it turns out that this is not the
170: right answer. It is enough to use substantially fewer random edges
171: to destroy $2$-colorability: roughly a square root of the number
172: of random clauses necessary to destroy satisfiability. The
173: following is our main result.
174:
175: \begin{theorem}
176: \label{main} Let $k,\ell \geq 2$, $\epsilon \geq 0$ be fixed and
177: let $H$ be a $2$-colorable $k$-uniform hypergraph with
178: $\Omega(n^{k-\epsilon})$ edges. Then the hypergraph $H'$ obtained
179: by adding to $H$ a collection $R$ of $\omega\big(n^{\ell \epsilon
180: / 2}\big)$ random $\ell$-tuples is almost surely
181: non-$2$-colorable.
182: \end{theorem}
183:
184: Observe that for $\epsilon \geq 2 / \ell$, the result is easy.
185: Regardless of the hypergraph $H$, it is well known that a
186: collection of $\omega(n)$ random $\ell$-tuples on $n$ vertices is
187: almost surely non-$2$-colorable. So we will be only interested in
188: the case when $\epsilon < 2 / \ell$. For such $\epsilon$ we obtain
189: the following result, which shows that the assertion of Theorem
190: \ref{main} is essentially best possible.
191:
192: \begin{theorem}
193: \label{main-construction} For fixed $k,\ell \geq 2$ and $0 \leq
194: \epsilon < 2 / \ell$, there exists a $2$-colorable $k$-uniform
195: hypergraph $H$ with $\Omega(n^{k-\epsilon})$ edges such that its
196: union with a collection $R$ of $o\big(n^{\ell \epsilon / 2}\big)$
197: random $\ell$-tuples is almost surely $2$-colorable.
198: \end{theorem}
199:
200: The rest of this paper is organized as follows. In the next section we
201: present an example of the hypergraph which shows that our main result is
202: essentially best possible. In Section 3 we discuss some natural difficulties in
203: proving Theorem \ref{main} and describe how to deal with them in the case of
204: bipartite graphs. This result also serves as a basis for induction which we use
205: in Section 4 to prove the general case of $2$-colorable $k$-uniform hypergraphs.
206:
207: \begin{remark}
208: \label{equiv-model} We have two alternative ways of adding random
209: edges. Either we can sample a random $\ell$-tuple $|R|$ times,
210: each time uniformly and independently from the set of all ${n
211: \choose \ell}$ $\ell$-tuples. Or we can pick each $\ell$-tuple
212: randomly and independently with probability $p=|R|/{n \choose
213: \ell}$. Since $2$-colorability is a monotone property, it follows,
214: as in Bollob\'as \cite{B}, Theorem 2.2 and a similar remark in
215: \cite{KST}, that if the resulting hypergraph is almost surely
216: non-$2$-colorable ($2$-colorable) in one model then this is true
217: in the other model as well. This observation can sometimes
218: simplify our calculations.
219: \end{remark}
220:
221: \paragraph{\bf Notation.} Let $H = (V,E)$ be a $k$-uniform hypergraph.
222: In the following, we use the notions of {\em degree} and {\em
223: neighborhood}, generalizing their usual meaning in graph theory.
224: For a vertex $v \in V$, we define its degree $d(v)$ to be the
225: number of edges of $H$ that contain $v$. More generally, for a
226: subset of vertices $A \subset V, |A| < k$, we define its degree $
227: d(A) = |\{e \in E: A \subset e\}|.$ For a $(k-1)$-tuple of
228: vertices $A$, we define its {\em neighborhood} as $ N(A) = \{w \in
229: V \setminus A: A \cup \{w\} \in E\}.$ Also, for a $(k-2)$-tuple
230: of vertices $A$, we define its {\em link} as $ \Gamma(A) =
231: \{\{u,v\} \in V \setminus A: A \cup \{u,v\} \in E \}.$
232:
233: Throughout the paper we will systematically omit floor and ceiling
234: signs for the sake of clarity of presentation. Also, we use the
235: notations $a_n =\Theta(b_n)$, $a_n =O(b_n)$ or $a_n =\Omega(b_n)$
236: for $a_n, b_n >0$ and $n \to \infty$ if there are absolute
237: constants $C_1$ and $C_2$ such that $C_1\, b_n < a_n < C_2\, b_n$,
238: $a_n < C_2 b_n$ or $a_n > C_1 b_n$ respectively. The notation
239: $a_n=o(b_n)$ means that $a_n/b_n \rightarrow 0$ as $n \rightarrow
240: \infty$, and $a_n= \omega(b_n)$ means $a_n/b_n \rightarrow
241: \infty$. The parameters $k,\ell,\epsilon$ are considered constant.
242:
243: \section{The lower bound}
244: The following example proves Theorem \ref{main-construction} and
245: shows that our main result is essentially best possible.
246:
247: \vspace{-0.2cm}
248: \paragraph{\bf Construction.} Partition the set of vertices $[n]$ into
249: three disjoint subsets $X, Y, Z$ where $|X|=|Y|=n^{1-\epsilon/2}$.
250: Let $\HL$ be a $k$-uniform hypergraph whose edge set consists of
251: all $k$-tuples which have exactly one vertex in $X$, one vertex in
252: $Y$ and $k-2$ vertices in $Z$. By definition the number of edges
253: in $\HL$ is $\Theta(n^{k-\epsilon})$.
254:
255: \begin{figure}[!ht]
256: \begin{center}
257: \scalebox{0.6}{\includegraphics{example.eps}}
258: \caption{Construction of the hypergraph $\HL$.}
259: \end{center}
260: \end{figure}
261:
262: \vspace{-0.2cm}
263: \paragraph{\bf Claim.} Color all the vertices in $X$ by
264: color $1$ and vertices in $Y$ by color $2$. Note that no matter
265: how we assign colors to the remaining vertices, this gives a
266: proper $2$-coloring of $\HL$. Let $R$ be a set of $o\big(n^{\ell
267: \epsilon/2}\big)$ random $\ell$-tuples. Then almost
268: surely we can $2$-color $Z$ so that none of the $\ell$-tuples in
269: $R$ is monochromatic, i.e., there exists a proper $2$-coloring of
270: $\HL+R$.
271:
272: To prove this claim we transform $R$ into another random instance
273: $R'$ that contains only single vertices with a fixed {\em
274: prescribed color} and edges of size two which must not be
275: monochromatic. Following Remark~\ref{equiv-model} we can assume
276: that $R$ was obtained by choosing every $\ell$-tuple in $[n]$
277: randomly and independently with probability
278: $p=o\left(n^{\ell\epsilon/2-\ell}\right)$. First note that almost
279: surely there is no $\ell$-tuple in $R$ whose vertices are all in
280: $X$ or in $Y$. Indeed, since $|X|=|Y|=n^{1-\epsilon/2}$, the
281: probability that there is such an $\ell$-tuple is at most $2
282: {n^{1-\epsilon/2} \choose \ell}\,p =o(1)$. Also, every
283: $\ell$-tuple in $R$ which has vertices in both $X$ and $Y$ is
284: already $2$-colored so we discard it.
285:
286: For every $v \in Z$ we add it to $R'$ with prescribed color $1$ if
287: there is a subset $A$ of $Y$ of size $\ell-1$ such that $A \cup
288: \{v\} \in R$. Since $\epsilon< 2/\ell \leq 1$, the probability of
289: this event is
290: $$p_1={|Y| \choose \ell-1}\,p = {n^{1-\epsilon/2} \choose \ell-1}\, p \leq
291: n^{(\ell-1)(1-\epsilon/2)}
292: \,p=o\big(n^{-1+\epsilon/2}\big)=o\big(n^{-1/2}\big).$$ Similarly,
293: if there is a subset $B$ of $X$ of size $\ell-1$ such that $B \cup
294: \{v\} \in R$ then we add $v$ to $R'$ with prescribed color $2$.
295: The probability $p_2$ of this event is also $o\big(n^{-1/2}\big)$.
296:
297: Fix an ordering $v_1< v_2< \dots$ of all vertices in $Z$. For
298: every pair of vertices $u, w \in Z$ we add an edge $\{u,w\}$ to
299: $R'$ if there is an $\ell$-tuple $L \in R$ such that the two
300: smallest vertices in $L\cap Z$ are $u$ and $w$. Since the number of such
301: possible $\ell$-tuples is at most ${n \choose \ell-2}$, and
302: $\epsilon < 2/\ell$, the probability of this event is
303: $$p_3 \leq {n \choose \ell-2} p = O\left(n^{\ell-2}p\right) =
304: o\left(n^{\ell\epsilon/2-2}\right) = o\left(n^{-1}\right).$$
305: Also note that by definition all the above events are independent since they depend
306: on disjoint sets of $\ell$-tuples. By our construction, any
307: $2$-coloring of $Z$ in which singletons in $R'$ get prescribed
308: colors and no $2$-edge is monochromatic gives a proper
309: $2$-coloring of $R$. Therefore, to complete the proof of Theorem
310: \ref{main-construction}, it is enough to prove the following
311: simple statement.
312:
313: \begin{lemma}
314: \label{last} Let $R'$ be a random instance which is obtained as
315: follows. For $i=1,2$ we choose every vertex in $[n]$ with
316: probability $p_i= o\big(n^{-1/2}\big)$ (independently for $i=1,2$)
317: and prescribe to it color $i$. In addition we choose every pair of
318: vertices to be an edge in $R'$ with probability $p_3=o(n^{-1})$.
319: Then almost surely there exists a $2$-coloring of $[n]$ in which
320: all singletons in $R'$ get prescribed colors and no edge is
321: monochromatic.
322: \end{lemma}
323:
324: \noindent {\bf Proof.}\, Let $G$ be the graph formed by edges from
325: $R'$. The probability that there is a vertex with conflicting
326: prescribed colors is $n p_1 p_2 = o(1)$. The probability that $G$
327: contains a cycle is at most $\sum_{s=3}^n n^s p_3^s
328: = O(n^3 p_3^3) = o(1)$. Finally the probability that there exists a path
329: between two vertices with any prescribed color is also bounded by
330: $$ \sum_{s=1}^n {n \choose 2} (p_1+p_2)^2 n^{s-1} p_3^s
331: = o\big(n (p_1+p_2)^2\big) = o(1).$$
332:
333: Therefore almost surely no vertex gets prescribed conflicting
334: colors, every connected component of $G$ is a tree and contains at
335: most one vertex with prescribed color. This immediately implies
336: the assertion of the lemma, since every tree can be $2$-colored,
337: starting from the vertex with prescribed color (if any). \hfill
338: $\Box$
339:
340:
341: \section{Bipartite graphs}
342:
343: Now let's turn to Theorem~\ref{main}. First, consider the case of
344: $k=\ell=2$. Here, we claim that for any bipartite graph $G$ with
345: $\Omega\big(n^{2-\epsilon}\big)$ edges, adding
346: $\omega(n^{\epsilon})$ random edges makes the graph almost surely
347: non-bipartite. This will follow quite easily, since it turns out
348: that almost surely we will insert an edge inside one part of a
349: bipartite connected component of $G$, creating an odd cycle (see
350: the proof of Proposition \ref{bipartite}).
351:
352: However, with the more general hypergraph case in mind, we are
353: also interested in a scenario where random $\ell$-tuples are added
354: to a bipartite graph, and $\ell > 2$. Then we ask what is the
355: probability that the resulting hypergraph is $2$-colorable (i.e.,
356: no $2$-edge and no $\ell$-edge should be monochromatic). We prove
357: the following special case of Theorem~\ref{main}.
358:
359: \begin{proposition}
360: \label{bipartite} Let $\ell \geq 2$, $0 \leq \epsilon < 2/\ell$
361: and let $G$ be a bipartite graph with
362: $\Omega\big(n^{2-\epsilon}\big)$ edges. Then the hypergraph $H$
363: obtained by adding to $G$ a collection $R$ of $\omega\big(n^{\ell
364: \epsilon/2}\big)$ random $\ell$-tuples is almost surely
365: non-$2$-colorable.
366: \end{proposition}
367:
368: \begin{figure}[!ht]
369: \label{fig-bipartite}
370: \begin{center}
371: \scalebox{0.8}{\includegraphics{bipartite.eps}}
372: \caption{Components of the bipartite graph $G$.}
373: \end{center}
374: \end{figure}
375:
376:
377: \noindent {\bf Proof.}\, Consider the connected components of $G$
378: which are bipartite graphs on disjoint vertex sets $(A_1, B_1)$,
379: $(A_2, B_2), \ldots$ (see Figure 2). Denote $a_i = |A_i|$, $b_i =
380: |B_i|$ and assume $a_i \geq b_i$. The number of edges in each
381: component is at most $a_i b_i$. Since the total number of edges is
382: at least $c n^{2 - \epsilon}$ for some constant $c>0$, we have
383: $$ \sum{a_i^2} \geq \sum{a_i b_i} \geq c n^{2-\epsilon}.$$
384: Observe that for $\ell=2$, the number of pairs of vertices inside
385: the sets $\{A_i\}$ is $\sum{a_i \choose 2} \geq \frac12 (c
386: n^{2-\epsilon} - n) \geq c' n^{2-\epsilon}$, so a random edge
387: lands inside one of these sets with probability at least $c'
388: n^{-\epsilon}$. Consequently, the probability that none of the
389: $\omega(n^\epsilon)$ random edges ends up inside some $A_i$ is at
390: most $(1 - c' n^{-\epsilon}) ^ {\omega(n^{\epsilon})} = o(1)$.
391: Thus almost surely, $G + R$ contains an odd cycle.
392:
393: On the other hand, in the general case we are adding $\omega\big(n^{\ell
394: \epsilon/2}\big)$ random $\ell$-tuples, which might never end up
395: inside any vertex set $A_i$. The probability of hitting a specific
396: $A_i$ is ${a_i \choose \ell} / {n \choose \ell} = O\big(a_i ^ \ell /
397: n^\ell\big)$. For example, if $G$ has $n^\epsilon$ components with
398: $a_i = b_i = n^{1-\epsilon}$, then this probability is at most
399: $O\big(\sum{a_i^\ell} / n^\ell\big) = O\big(n^{-(\ell-1) \epsilon}\big)$. Hence we
400: need $\omega\big(n^{(\ell-1) \epsilon}\big)$ random $\ell$-tuples to hit
401: almost surely some $A_i$. This suggests a difficulty with the
402: attempt to place a random $\ell$-tuple in a set which is forced to
403: be monochromatic by the original graph. We have to allow ourselves
404: more freedom and consider sets which are monochromatic only under
405: certain colorings.
406:
407: More specifically, each of the sets $A_i$, $B_i$ must be
408: monochromatic under any coloring, and at least half of them must
409: share the same color. We do not know a priori which sets will
410: share the same color, yet we can estimate the probability that
411: {\em any} of these configurations allows a feasible coloring
412: together with the random $\ell$-tuples. First, it is convenient to
413: assume that the sets have roughly equal size, in which case we
414: have the following claim.
415:
416: \begin{lemma}
417: \label{clusters} Suppose we have $t$ disjoint subsets $A_1,
418: \ldots, A_t$ of $[n]$ of size $\Theta(n^{1-\alpha})$. Let $\alpha
419: \geq \epsilon/2$, $t = \Omega\big(n^{\frac{\ell}{\ell-1} (\alpha -
420: \epsilon/2)}\big)$ and let $R$ be a collection of $\omega\big(n^{\ell
421: \epsilon / 2}\big)$ random $\ell$-tuples on $[n]$. Then the
422: probability that $R$ can be $2$-colored in such a way that each
423: $A_i$ is monochromatic is at most $e^{-\omega(t)}$.
424: \end{lemma}
425:
426: \noindent {\bf Proof.} Consider the $2^t$ possible colorings in
427: which all $A_i$ are monochromatic. For each such coloring there is
428: a set of indices $I, |I|\geq t/2$ such that the sets $A_i, i \in
429: I$ share the same color. Since $A_i$ are disjoint we have
430: $|\cup_{i \in I}{A_i}| \geq c_1 t n^{1-\alpha}$ for some $c_1>0$.
431: The probability that one random $\ell$-tuple falls inside this set
432: is at least ${c_1 t n^{1-\alpha} \choose \ell} / {n \choose \ell}
433: \geq c_2 (t n^{-\alpha})^\ell$ for some $c_2 > 0$. Since
434: $t^{\ell-1} = \Omega\big(n^{\ell (\alpha - \epsilon/2)}\big)$, it
435: implies that
436: $$ \Pr\big[\mbox{ $\cup_{i\in I} A_i$ contains no $\ell$-tuple from
437: $R$}\big]
438: \leq \Big(1 - (c_2 t n^{-\alpha})^\ell \Big)^{\omega(n^{\ell \epsilon /
439: 2})} \leq e^{-\omega(t^\ell n^{-\ell (\alpha - \epsilon/2)})} =
440: e^{-\omega(t)}. $$
441: Therefore, by the union bound over all choices of $I$, we get
442: $$ \Pr\big[\exists\, I\, \mbox{such that $\cup_{i\in I} A_i$ contains no $\ell$-tuple from
443: $R$}\big] \leq 2^t e^{-\omega(t)} = e^{-\omega(t)}.$$
444: In particular, almost surely there is no $2$-coloring of $R$ in
445: which all $A_i$ are monochromatic.
446: \hfill $\Box$
447:
448: Now we can finish the proof of Proposition~\ref{bipartite} for
449: $\ell \geq 3$. Partition the components of $G$ according to their
450: size and let $G_s$ contain all the components with $|A_i| \in
451: [2^{s-1},2^s)$. If there is any $A_i$ of size at least $n^{1 -
452: \epsilon/2}$, we are done immediately because one of the
453: $\omega\big(n^{\ell \epsilon/2}\big)$ random $\ell$-tuples a.s. ends up in
454: $A_i$ and this destroys the $2$-colorability. So we can assume
455: that $s \leq \lfloor (1 - \epsilon/2) \log_2 n \rfloor$. Recall
456: that $\ell \geq 3$ and consider the following sum
457: $$ \sum_{s=1}^{\lfloor (1-\epsilon/2) \log_2 n \rfloor} 2^{\frac{\ell-2}{\ell-1} s}
458: n^{\frac{\ell}{\ell-1}(1 - \epsilon/2)}
459: \leq \frac{n^{\frac{\ell-2}{\ell-1}(1 - \epsilon/2)}}{1 -
460: 2^{-\frac{\ell-2}{\ell-1}}} \cdot n^{\frac{\ell}{\ell-1}(1 -
461: \epsilon/2)} \leq 4 n^{2 - \epsilon}. $$
462: Since $G$ has at least $c n^{2-\epsilon}$ edges, there is a
463: subgraph $G_s$ containing at least $\frac{c}{4}
464: 2^{\frac{\ell-2}{\ell-1} s} n^{\frac{\ell}{\ell-1}(1 -
465: \epsilon/2)}$ edges. As each component of $G_s$ has at most
466: $2^{2s}$ edges, the number of components of $G_s$ is $t =
467: \Omega\big(2^{-\frac{\ell}{\ell-1} s} n^{\frac{\ell}{\ell-1}(1 -
468: \epsilon/2)}\big)$. We set $2^s = n^{1 - \alpha}$, $\alpha \geq
469: \epsilon/2$ which means that $t =
470: \Omega\big(n^{\frac{\ell}{\ell-1}(\alpha - \epsilon/2)}\big)$. To
471: summarize, we have $t$ disjoint sets $A_i$ of size
472: $\Theta(n^{1-\alpha})$, each of which must be monochromatic under
473: any feasible coloring. Thus we can apply Lemma~\ref{clusters} to
474: conclude that for $H= G + R$, almost surely there is no feasible
475: $2$-coloring.
476: \hfill $\Box$
477:
478: \section{General hypergraphs}
479:
480: In this section we deal with the general case of a $2$-colorable
481: $k$-uniform hypergraph $H$, to which we add a collection of random
482: $\ell$-tuples $R$. Our goal is to prove Theorem~\ref{main} which
483: asserts that if $H$ has $\Omega\big(n^{k-\epsilon}\big)$ edges
484: then adding to it $\omega\big(n^{\ell \epsilon/ 2}\big)$ random
485: $\ell$-tuples makes it almost surely non-$2$-colorable. The proof
486: will proceed by induction on $k$. The base case when $k=2$ follows
487: from Proposition~\ref{bipartite}, so we can assume that $k
488: > 2$ and that the result holds for $k-1$.
489:
490: We start with a series of lemmas which allow us to make
491: simplifying assumptions. Depending on the hypergraph $H$, we
492: either reduce the problem to the $(k-1)$-uniform case or prove
493: directly that $H + R$ is not $2$-colorable.
494:
495: Since we have $\omega\big(n^{\ell \epsilon/2}\big)$ random $\ell$-tuples
496: available, we can divide them into a constant number of batches,
497: where each batch still has $\omega\big(n^{\ell \epsilon/2}\big)$
498: $\ell$-tuples. We will use a separate batch for each step of the
499: induction. We write $R = R_1 \cup R_2 \cup \ldots \cup R_k$ where
500: $|R_i| = \omega\big(n^{\ell \epsilon/2}\big)$ for each $i$.
501:
502:
503: \begin{lemma}
504: \label{k-1-degrees} Let $H_k$ be a $k$-uniform hypergraph on $n$
505: vertices with $c_1 n^{k-\epsilon}$ edges. Consider all
506: $(k-1)$-tuples $A \subset V(H_k)$ with degree greater than
507: $n^{1-\epsilon/2}$. If there are at least $\frac{c_1}{4}
508: n^{k-1-\epsilon}$ such $(k-1)$-tuples in $H_k$ then $H_k + R$ is
509: almost surely non-$2$-colorable.
510: \end{lemma}
511:
512: \noindent {\bf Proof.} For each $(k-1)$-tuple $A$ of degree
513: $>n^{1-\epsilon/2}$, the neighborhood $N(A)$ contains
514: $\Omega\big(n^{\ell - \ell \epsilon/2}\big)$ distinct
515: $\ell$-tuples. Therefore a random $\ell$-tuple lands inside $N(A)$
516: with probability $\Omega\big(n^{-\ell \epsilon/2}\big)$.
517: Consequently, the probability that none of $\omega\big(n^{\ell
518: \epsilon/2}\big)$ random $\ell$-tuples from $R_k$ ends up inside
519: $N(A)$ is at most $\big(1-\Omega(n^{-\ell
520: \epsilon/2})\big)^{\omega(n^{\ell \epsilon/2})}=o(1)$. If we have
521: $t \geq \frac{c_1}{4} n^{k-1-\epsilon}$ such $(k-1)$-tuples, then
522: the expected number of them, whose neighborhood does {\em not}
523: contain any $\ell$-tuple in $R_k$, is $o(t)$. Therefore, by
524: Markov's inequality, we get almost surely at least $\frac{t}{2}
525: \geq \frac{c_1}{8} n^{k-1-\epsilon}$ $(k-1)$-tuples with an
526: $\ell$-edge in their neighborhood. Denote by $H_{k-1}$ the
527: $(k-1)$-uniform hypergraph formed by these $(k-1)$-tuples.
528:
529: By induction, we know that $H_{k-1} + R_1 + \ldots + R_{k-1}$ is
530: almost surely non-$2$-colorable. Therefore for every $2$-coloring
531: respecting $R_1 \cup \ldots \cup R_{k-1}$, there is a
532: monochromatic $(k-1)$-tuple $A$ in $H_{k-1}$. Without loss of
533: generality assume that all vertices in $A$ are colored by $1$. By
534: definition, the neighborhood $N(A)$ contains an $\ell$-edge $L \in
535: R_k$. Either $L$ is monochromatic, or one of its vertices $x$ is
536: colored by $1$ as well. But then $A \cup \{x\}$ is a monochromatic
537: edge of $H_k$. This implies that there is no feasible $2$-coloring
538: for $H_k + R_1 + \ldots + R_k$. \hfill $\Box$
539:
540: \vspace{0.1cm}
541:
542: Thus we only need to treat the case where there are at most
543: $\frac{c_1}{4} n^{k-1-\epsilon}$ $(k-1)$-tuples with degree
544: greater than $n^{1-\epsilon/2}$, therefore at most $\frac{c_1}{4}
545: n^{k-\epsilon}$ edges through such $(k-1)$-tuples. We will get rid
546: of these high degrees by removing a constant fraction of edges and
547: making all degrees of $(k-1)$-tuples at most $n^{1-\epsilon/2}$.
548: This would also imply a bound of $n^{2-\epsilon/2}$ on the degrees
549: of $(k-2)$-tuples, etc. However, in the following we show that for
550: $(k-2)$-tuples we can assume an even stronger bound. More
551: specifically, we prove that if we have many edges through
552: $(k-2)$-tuples of degrees $n^{2-\delta}$ with $\delta \leq
553: \frac{\ell}{2(\ell-1)} \epsilon$, then we can proceed by
554: induction. For this purpose, we first show the following.
555:
556: \begin{lemma}
557: \label{families}
558: Let $\ell \geq 2$ and
559: let $G$ be a graph on $n$ vertices with $n^{2-\delta}$ edges.
560: Then $G$ contains $\frac12n^{1-\delta}$ disjoint
561: subsets of vertices $F_1, F_2, \ldots$ such that the vertices
562: in each $F_j$ have disjoint neighborhoods of sizes
563: $d_1, d_2,\ldots$, satisfying $d_i \geq \frac12n^{1-\delta}$ and
564: $$ \sum{d_i^\ell} \geq \frac{n^{\ell - (\ell-1) \delta}}{2^{\ell}}\,.$$
565: \end{lemma}
566:
567: \noindent{\bf Proof.} We iterate the following construction for
568: $j=1,2,\ldots,\frac12 n^{1-\delta}$.
569:
570: \begin{itemize}
571: \item Take the vertex $v_1$ of maximum degree $d_1$ and remove
572: all the edges incident to its neighbors. Note that by maximality
573: of $d_1$, at most $d_1^2$ edges are removed.
574: \item In step $i$, take the vertex $v_i$ of maximum degree $d_i$
575: in the remaining graph and remove the edges incident to its
576: neighbors (again, at most $d_i^2$ edges). Repeat these steps, as
577: long as $\sum{d_i^2} < \frac14 n^{2-\delta}$.
578: \item When the procedure terminates, define $F_j = \{v_1, v_2,
579: \ldots \}$. Then return to the original graph, but remove the
580: vertices in $F_j$ and all their edges permanently.
581: \end{itemize}
582:
583: \begin{figure}[!ht]
584: \begin{center}
585: \scalebox{0.8}{\includegraphics{families.eps}}
586: \caption{Construction of $F_j = \{v_1, v_2, \ldots \}$. The
587: neighborhood of $v_i$ is incident with at most $d_i^2$ edges.}
588: \end{center}
589: \end{figure}
590:
591: By construction, the neighborhoods of the vertices in every $F_j$
592: are disjoint and hence with each $F_j$, we remove $\sum{d_i} \leq
593: n$ edges from the graph. The sets $F_j$ are also disjoint
594: (although the neighborhoods of vertices from different $F_j$'s are
595: not necessarily disjoint). Since we constructed
596: $\frac12n^{1-\delta}$ sets $F_j$, there are at least
597: $n^{2-\delta}-\frac12n^{1-\delta} \cdot n=\frac12n^{2-\delta}$
598: edges available at the beginning of every construction.
599:
600: Inside the construction of $F_j$, we repeat as long as $\sum{d_i^2} < \frac14
601: n^{2-\delta}$ and therefore we remove at most $\frac14n^{2-\delta}$
602: edges from the graph we started with. Hence, at every step the remaining graph still
603: has at least $\frac14n^{2-\delta}$ edges and so its maximum degree is at least
604: $\frac12n^{1-\delta}$. When we terminate we have
605: $\sum{d_i^2} \geq \frac14 n^{2-\delta}$. This, together with the fact that
606: $d_i \geq \frac12n^{1-\delta}$, implies that for every $F_j$ we have
607: $$\hspace{4.8cm} \sum{d_i^\ell} \geq \left( \frac12 n^{1-\delta}\right)^{\ell-2}
608: \sum{d_i^2} \geq \frac{n^{\ell - (\ell-1) \delta}}{2^{\ell}}.
609: \hspace{4.8cm} \Box$$
610:
611:
612: \begin{lemma}
613: \label{k-2-degrees} Let $H_k$ be a $k$-uniform hypergraph on $n$
614: vertices with $c_1 n^{k-\epsilon}$ edges. Consider $(k-2)$-tuples
615: of degree $n^{2-\delta}$ where $\delta \leq \frac{\ell
616: }{2(\ell-1)} \epsilon$. If there are at least $\frac{c_1}{4}
617: n^{k-\epsilon}$ edges through such $(k-2)$-tuples then $H_k + R$
618: is almost surely non-$2$-colorable.
619: \end{lemma}
620:
621: \noindent {\bf Proof.} Consider a $(k-2)$-tuple $A$ of degree
622: $n^{2-\delta}$. The link of $A$ in $H_k$ is a graph $\Gamma(A)$
623: with $n^{2-\delta}$ edges. By Lemma~\ref{families}, we find
624: $\frac12 n^{1-\delta}$ subsets $F_j$ such that vertices in $F_j$
625: have disjoint neighborhoods in $\Gamma(A)$ with sizes satisfying
626: $\sum{d_i^\ell} \geq 2^{-\ell} n^{\ell - (\ell-1) \delta}$. We
627: repeat this construction for each $(k-2)$-tuple of degree
628: $n^{2-\delta}$ with $\delta \leq \frac{\ell}{2(\ell-1)} \epsilon$.
629: For each of them, we construct $\frac12 n^{1-\delta}$ sets as
630: above. Assuming that the total number of edges through such
631: $(k-2)$-tuples is at least $\frac{c_1}{4} n^{k-\epsilon}$, we get
632: $\frac{c_1}{8} n^{k-1-\epsilon}$ sets $F_j$ in total.
633:
634: Now fix a set $F_j$. Call it {\em good} if after adding random
635: $\ell$-tuples from $R_k$ there is at least one vertex in $F_j$
636: whose neighborhood in $\Gamma(A)$ contains a random $\ell$-tuple.
637: If this is not the case, call it {\em bad}. We estimate the
638: probability that $F_j$ is bad. By Lemma~\ref{families}, the total
639: number of $\ell$-tuples in the neighborhoods of vertices in $F_j$
640: is
641: $$ \sum{d_i \choose \ell}=\Omega \left(\sum\frac{d_i^\ell}{\ell!}\right) =
642: \Omega \left(
643: \frac{n^{\ell - (\ell-1) \delta}}{2^\ell \ell!}\right)=
644: \Omega \big( n^{\ell - \ell \epsilon/2}\big). $$ Thus the
645: probability that a random $\ell$-tuple falls inside some
646: neighborhood of $F_j$ is $\sum{d_i \choose \ell} / {n \choose
647: \ell} = \Omega \big( n^{-\ell \epsilon / 2}\big)$. After adding
648: the entire batch of random $\ell$-tuples $R_k$,
649: $$ \Pr\big[F_j \mbox{ is bad}\big]
650: = \left(1 - \Omega\big( n^{-\ell \epsilon / 2}\big)\right)^{-\omega(n^{\ell
651: \epsilon / 2})}= o(1).$$
652: Consequently, the expected fraction of bad $F_j$'s is $o(1)$. By
653: Markov's inequality, this fraction is almost surely at most one
654: half, which means that at least $\frac{c_1}{16} n^{k-1-\epsilon}$
655: sets $F_j$ have a vertex $v \in F_j$ whose neighborhood contains
656: some $\ell$-tuple from $R_k$. For each such $F_j$, we have a set
657: $A$ of size $k-2$ which together with $v$ forms a $(k-1)$-tuple
658: whose neighborhood in $H_k$ contains an $\ell$-tuple from $R_k$.
659: We could get the same $(k-1)$-tuple in $k-1$ different ways, but
660: in any case we have at least $\frac{c_1}{16k} n^{k-1-\epsilon}$
661: such $(k-1)$-tuples which form an edge set of a $(k-1)$-uniform
662: hypergraph $H_{k-1}$.
663:
664: By the induction hypothesis, $H_{k-1} + R_1 + \ldots + R_{k-1}$ is
665: almost surely non-$2$-colorable. Therefore, for any $2$-coloring
666: which respects the $\ell$-edges from $R_1 + \ldots + R_{k-1}$,
667: there must be a monochromatic $(k-1)$-edge $B$ in $H_{k-1}$.
668: However, since there is an $\ell$-edge from $R_k$ in the
669: neighborhood of $B$, one of its vertices should have the same
670: color as $B$. This would form a monochromatic edge in $H_k$ so
671: there is no feasible $2$-coloring for $H_k + R_1 + \ldots + R_k$.
672: \hfill $ \Box$
673:
674: Thus we can also assume that at most $\frac{c_1}{4}
675: n^{k-\epsilon}$ edges go through $(k-2)$-tuples of degree
676: $n^{2-\delta}, \delta \leq \frac{\ell}{2(\ell-1)} \epsilon$.
677: Before the last part of the proof, we make further restrictions on
678: the degree bounds and structure of our hypergraph, by finding a
679: subhypergraph $H_\alpha$ described in the following lemma.
680:
681: \begin{lemma}
682: \label{k-partite} Let $H_k = (V,E)$ be a $k$-uniform hypergraph
683: with $c_1 n^{k-\epsilon}$ edges, such that at most $\frac{c_1}{4}
684: n^{k-\epsilon}$ edges go through $(k-1)$-tuples of degree $\geq
685: n^{1-\epsilon/2}$ and at most $\frac{c_1}{4} n^{k-\epsilon}$ edges
686: go through $(k-2)$-tuples of degree $n^{2-\delta}, \delta \leq
687: \frac{\ell}{2(\ell-1)} \epsilon$. Then for some constant $\alpha
688: \geq \epsilon/2$, $H_k$ contains a subhypergraph $H_\alpha$ with
689: the following properties
690: \begin{enumerate}
691: \item $H_\alpha$ is $k$-partite, i.e. $V$ can be partitioned
692: into $V_1 \cup V_2 \cup \ldots \cup V_k$ so that every edge of
693: $H_\alpha$ intersects each $V_i$ in one vertex.
694: \item Every vertex has degree at most $n^{k-1 - \frac{\ell}
695: {2(\ell-1)} \epsilon}$.
696: \item The degree of every $(k-1)$-tuple in $V_1 \times V_2 \times \ldots \times V_{k-1}$
697: is either $0$ or between $n^{1-\alpha}$ and $2n^{1-\alpha}$.
698: \item The number of edges in $H_\alpha$ is at least
699: $$ c_5 \left(n^{k - \epsilon - \frac{\epsilon-\alpha}{\ell-1}} +
700: n^{k - \epsilon - \frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)}\right), $$
701: for some constant $c_5 =c_5(k,\ell,c_1)$.
702: \end{enumerate}
703: \end{lemma}
704:
705: \noindent {\bf Proof.} First, remove all edges through
706: $(k-1)$-tuples of degree $\geq n^{1-\epsilon/2}$ and through
707: $(k-2)$-tuples of degree $n^{2-\delta}, \delta \leq
708: \frac{\ell}{2(\ell-1)} \epsilon$. We get a hypergraph $H'$ such
709: that the degrees of all $(k-1)$-tuples are at most
710: $n^{1-\epsilon/2}$, the degrees of all $(k-2)$-tuples are at most
711: $n^{2 - \frac{\ell}{2(\ell-1)} \epsilon}$, and the number of edges
712: is at least $c_2 n^{k-\epsilon}$ edges, $c_2 = c_1/2$.
713: Consequently, the degree of every vertex in $H'$ is at most
714: $n^{k-3} \cdot n^{2 - \frac{\ell}{2(\ell-1)} \epsilon} = n^{k-1 -
715: \frac{\ell}{2(\ell-1)} \epsilon}$.
716:
717: Next, we use a well known fact, proved by Erd\H{o}s and Kleitman
718: \cite{EK} that every $k$-uniform hypergraph $H'$ with $c_2
719: n^{k-\epsilon}$ edges contains a $k$-partite subhypergraph with at
720: least $c_3 n^{k-\epsilon}$ edges where $c_3 = \frac{k!}{k^k}c_2$.
721: This can be achieved for example by taking a random partition of
722: the vertex set into $k$ parts and computing the expected number of
723: edges which intersect all of them. Let $(V_1, V_2, \ldots, V_k)$
724: be a partition, so that at least $c_3 n^{k-\epsilon}$ edges of
725: $H'$ have one vertex in every $V_i$. Discard all other edges and
726: denote this $k$-partite hypergraph by $H''$.
727:
728: Consider all $(k-1)$-tuples in $V_1 \times V_2 \times \ldots
729: \times V_{k-1}$ whose degree in $H''$ is less than $\frac{c_3}{2}
730: n^{1-\epsilon}$. Delete all their edges, which is at most ${n
731: \choose k-1} \frac{c_3}{2} n^{1-\epsilon} \leq \frac{c_3}{2}
732: n^{k-\epsilon}$ edges in total. We still have at least $c_4
733: n^{k-\epsilon}$ edges, where $c_4 = c_3/2$. Now the degree of
734: every $(k-1)$-tuple in $V_1 \times V_2 \times \ldots \times
735: V_{k-1}$ is either $0$ or between $c_4 n^{1-\epsilon}$ and
736: $n^{1-\epsilon/2}$. Finally, we are going to find a subhypergraph
737: in which all the non-zero degrees of $(k-1)$-tuples are
738: $\Theta(n^{1-\alpha})$ and the number of edges is at least
739: $$ c_5 \left(n^{k - \epsilon - \frac{\epsilon-\alpha}{\ell-1}} +
740: n^{k - \epsilon - \frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)}\right). $$
741: The existence of such a subhypergraph can be proved by an
742: elementary counting argument. Let $n^{1-\alpha} = 2^i$ and
743: partition $V_1 \times V_2 \times \ldots \times V_{k-1}$ into
744: groups of $(k-1)$-tuples with degrees in intervals $[2^i,
745: 2^{i+1})$, where $i$ ranging between $i_1 = \log_2 (c_4
746: n^{1-\epsilon})$ and $i_2 = \log_2 (n^{1 - \epsilon/2})$. Consider
747: the following two expressions:
748: $$\sum_{i=i_1}^{i_2} {2^{-i/(\ell-1)}}
749: \leq \frac{(c_4
750: n^{1-\epsilon})^{-\frac{1}{\ell-1}}}{1-2^{-\frac{1}{\ell-1}}}
751: \leq 2(\ell-1) c_4^{-1} n^{-\frac{1-\epsilon}{\ell-1}}$$
752: and
753: $$\sum_{i=i_1}^{i_2} {2^{\frac{\ell-2}{\ell-1} i}} \leq
754: \frac{n^{\frac{\ell-2}{\ell-1} (1 - \epsilon/2)}}{1 -
755: 2^{-\frac{\ell-2}{\ell-1}}}
756: \leq 4 n^{\frac{\ell-2}{\ell-1} (1 - \epsilon/2)}.$$
757: Normalizing by the right-hand side and taking the average, we get
758: $$ \sum_{i=i_1}^{i_2} \left( \frac{2^{-\frac{i}{\ell-1}}}
759: {4(\ell-1)c_4^{-1} n^{-\frac{1-\epsilon}{\ell-1}}} +
760: \frac{2^{\frac{\ell-2}{\ell-1}i}}{8
761: n^{\frac{\ell-2}{\ell-1}(1-\epsilon/2)}} \right) \leq 1$$ By the
762: pigeonhole principle, there is an $i$ such that the fraction of
763: edges through $(k-1)$-tuples with degree between
764: $2^i=n^{1-\alpha}$ and $2^{i+1}=2n^{1-\alpha}$ is at least
765: $$ \frac{2^{-\frac{i}{\ell-1}}}{4(\ell-1)c_4^{-1} n^{-\frac{1-\epsilon}{\ell-1}}} +
766: \frac{2^{\frac{\ell-2}{\ell-1}i}}{8 n^{\frac{\ell-2}{\ell-1}
767: (1-\epsilon/2)}} = \frac{c_4}{4 (\ell-1)}
768: n^{-\frac{\epsilon-\alpha}{\ell-1}} + \frac{1}{8}
769: n^{-\frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)} $$ so the lemma
770: holds with $c_5 = c_4 \cdot \min\big\{\frac{c_4}{4 (\ell-1)},
771: \frac{1}{8} \big\}$. \hfill $\Box$
772:
773: \vspace{0.1cm}
774:
775: Note that in this lemma, we lose more than a constant fraction of
776: the edges. However, from now on, we do not use induction anymore
777: and will prove directly that $H_\alpha + R$ is almost surely
778: non-$2$-colorable. We will proceed in $t = c_5 \ell^{-k}
779: n^{\frac{\ell}{\ell-1} (\alpha - \epsilon/2)}$ stages. For each
780: stage, we allocate a certain number of random $\ell$-tuples.
781: Namely, we set again $R = R_1 \cup R_2 \cup \ldots \cup R_k,\,
782: |R_i|=\omega\big(n^{\ell \epsilon / 2}\big)$. Furthermore, we
783: divide each $R_j$ for $j \leq k-1$ into $t$ parts $R_{1,j},
784: \ldots, R_{t,j}$ so that
785: $$ |R_{i,j}| = \omega\left(\frac{n^{\ell \epsilon / 2}}{t}\right)
786: = \omega\left(n^{\ell \epsilon/2 - \frac{\ell}{\ell-1} (\alpha -
787: \epsilon/2)}\right).$$
788: The random set $R_{i,j}$ will be used for the $j$-th ``level" of
789: the $i$-th stage. The following lemma describes one stage of the
790: construction. Finally, $R_k$ will be used in the last step of the proof.
791:
792: \begin{lemma}
793: \label{one-tree} Let $H_\alpha$ be a $k$-uniform $k$-partite
794: hypergraph where the degree of every $(k-1)$-tuple in $V_1 \times
795: V_2 \times \ldots \times V_{k-1}$ is either zero or
796: is in the interval $[n^{1-\alpha}, 2n^{1-\alpha}]$,
797: and the number of edges in $H_\alpha$ is at least
798: $$ c_5 n^{k - \epsilon - \frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)}.$$
799: Then almost surely, there exists a family of $q = \ell^{k-2}$ sets
800: $S_1, \ldots, S_q$, $n^{1-\alpha} \leq S_i \leq 2n^{1-\alpha}$,
801: such that for every feasible $2$-coloring of $H_\alpha + R_{i,1} +
802: \ldots + R_{i,k-1}$ at least one $S_i$ is monochromatic.
803: \end{lemma}
804:
805: \noindent {\bf Proof.} We are going to construct an $\ell$-ary
806: tree $T$ of depth $k-1$. We denote vertices on the $j$-th level by
807: $v_{a_1 a_2 \ldots a_{j-1}}$ where $a_i \in \{1,2,\ldots,\ell\}$.
808: $T$ is rooted at a vertex in $V_1$ and the $j$-th level is
809: contained in $V_j$. We construct $T$ in such a way that the
810: vertices along every path which starts at the root and has length
811: $k-1$ form a $(k-1)$-tuple with degree $\Theta(n^{1-\alpha})$ in
812: $H_\alpha$. The neighborhoods of all branches of length $k-1$ will
813: be our sets $S_i$ (not necessarily disjoint). In addition, the set
814: of $\ell$ children of every node on each level $j \leq k-2$, like
815: $\{v_{a_1 a_2 \ldots a_{j-1} 1}, v_{a_1 a_2 \ldots a_{j-1}
816: 2},\ldots, v_{a_1 a_2 \ldots a_{j-1} \ell}\}$, will form an edge
817: of $R_{i,j}$.
818:
819: \begin{figure}[!ht]
820: \begin{center}
821: \scalebox{0.8}{\includegraphics{tree.eps}} \caption{Construction
822: of the tree $T$, for $k=4$ and $\ell=3$. Branches of the tree form
823: active $(k-1)$-tuples, with neighborhoods $S_i$. Each set of
824: children on level $j+1$ forms an edge of $R_{i,j}$.}
825: \end{center}
826: \end{figure}
827:
828: Assuming the existence of such a tree, consider any $2$-coloring
829: of $H_\alpha + R_{i,1} + \ldots + R_{i,k-1}$. Since the children
830: of each vertex on level $j < k-1$ form an $\ell$-edge in
831: $R_{i,j}$, every vertex has children of both colors. In
832: particular, there is always one child with the same color as its
833: parent. Therefore, starting from the root, we can always find a
834: monochromatic branch $A$ of length $k-1$. Since all the extensions
835: of this branch to edges of $H_\alpha$ must be $2$-colored, all the
836: vertices in $S_i = N(A)$ must have the same color.
837:
838: We grow the tree level by level, maintaining the property that all
839: branches have sufficiently many extensions to edges of $H_\alpha$.
840: More precisely, we call an $r$-tuple in $V_1 \times \ldots \times
841: V_r$ {\em active} if its degree is at least
842: $$ \Delta_r = \frac{c_5}{2^r} n^{k-r-\epsilon - \frac{\ell-2}{\ell-1}(\alpha -
843: \epsilon/2)}.$$
844:
845: \noindent {\bf Claim.} Every active $r$-tuple $A$, $r \leq k-2$,
846: can be extended to at least
847: $$ d_r = \frac{\Delta_r}{4 n^{k-r-1-\alpha}} = \frac{c_5}{2^{r+2}}
848: n^{1 - \epsilon/2 + \frac{1}{\ell-1} (\alpha - \epsilon/2)} $$ active
849: $(r+1)$-tuples $A \cup \{x\}, x \in V_{r+1}$.
850:
851: {\em Proof.} Suppose that fewer than $d_r$ extensions of $A$ are
852: active. Since the degrees of $(k-1)$-tuples are at most $2
853: n^{1-\alpha}$, we get that any $(r+1)$-tuple has degree at most $2
854: n^{k-r-1-\alpha}$. Therefore the number of edges through all
855: active extensions of $A$ is smaller than $d_r \cdot 2
856: n^{k-r-1-\alpha}=\frac12\Delta_r$. We also have inactive
857: extensions of $A$ which have degrees less than $\Delta_{r+1}$. The
858: total number of edges through these extensions of $A$ is smaller
859: than $n \Delta_{r+1} = \frac12 \Delta_r$. But the total number of
860: edges through $A$ is at least $\Delta_r$. This contradiction
861: proves the claim. \hfill $\Box$
862:
863: We start our construction from an active vertex $v \in V_1$. Since
864: $H_\alpha$ has at least $n \Delta_1$ edges, such a vertex must
865: exist. By our claim, $v$ can be extended to at least $d_1$ active
866: pairs $\{v,x\}, x \in W_2 \subset V_2$. Consider this set of $d_1$
867: vertices $W_2$. The probability that a random $\ell$-tuple falls
868: inside $W_2$ is ${d_1 \choose \ell} / {n \choose \ell} =
869: \Omega(n^{-\ell \epsilon/2 + \frac{\ell}{\ell-1} (\alpha -
870: \epsilon/2)})$. Now we use $\omega(n^{\ell \epsilon/2 -
871: \frac{\ell}{\ell-1} (\alpha - \epsilon/2)})$ random $\ell$-tuples
872: from $R_{i,1}$ that we allocated for the first level of this
873: construction. This means that almost surely, we get an $\ell$-edge
874: $\{v_1, \ldots, v_{\ell}\} \in R_{i,1}$ such that $\{v, v_i\}$ is
875: an active pair for each $i=1,2,\ldots,\ell$.
876:
877: We continue growing the tree, using the random $\ell$-tuples of
878: $R_{i,j}$ on level $j$. Since we have ensured that each path from
879: the root to the level $j$ from an active $j$-tuple, it has at
880: least $d_j$ extensions to an active $(j+1)$-tuple. Again, the
881: probability that a random $\ell$-tuple hits the extension vertices
882: $W_{j+1} \subset V_{j+1}$ for a given path is ${d_j \choose \ell}
883: / {n \choose \ell} = \Omega\big(n^{-\ell \epsilon/2 +
884: \frac{\ell}{\ell-1} (\alpha - \epsilon/2)}\big)$. Almost surely,
885: one of the $\ell$-tuples in $R_{i,j}$ will hit these extension
886: vertices and we can extend this path to $\ell$ children on level
887: $j+1$. The number of paths from the root to level $j$ is bounded
888: by $\ell^{j-1}$ which is a constant, so in fact we will almost
889: surely succeed to build the entire level.
890:
891: In this way, we a.s. build the tree all the way to level $k-1$.
892: Every path from the root to one of the leaves forms an active
893: $(k-1)$-tuple and has degree $\in [n^{1-\alpha}, 2n^{1-\alpha}]$.
894: Define $S_1, S_2, \ldots, S_q$ to be the neighborhoods of all
895: these $q = \ell^{k-2}$ paths. By construction, for any feasible
896: $2$-coloring of $H_\alpha + R_{i,1} + \ldots + R_{i,k-1}$, one of
897: these paths is monochromatic which implies that the corresponding
898: set $S_i$ is monochromatic as well. \hfill $\Box$
899:
900: \begin{lemma}
901: \label{trees} Let $H_\alpha$ be a $k$-uniform $k$-partite
902: hypergraph where the degree of every vertex is at most
903: $n^{k-1-\frac{\ell}{2(\ell-1)} \epsilon}$, the degree of every
904: $(k-1)$-tuple in $V_1 \times V_2 \times \ldots \times V_{k-1}$ is
905: either zero or is in the interval $[n^{1-\alpha}, 2n^{1-\alpha}]$, and the number of
906: edges in $H_\alpha$ is at least
907: $$ c_5 n^{k - \epsilon - \frac{\epsilon-\alpha}{\ell-1}} +
908: c_5 n^{k - \epsilon - \frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)}.$$
909: Then almost surely, $H_\alpha + R$ is not $2$-colorable.
910: \end{lemma}
911:
912: \noindent {\bf Proof.} We apply Lemma~\ref{one-tree} repeatedly in
913: $t = c_5 \ell^{-k} n^{\frac{\ell}{\ell-1} (\alpha - \epsilon/2)}$
914: stages. In each stage $i$, we almost surely obtain $q=\ell^{k-2}$
915: sets $S_{i,1}, \ldots, S_{i,q}$, $n^{1-\alpha} \leq |S_{i,j}| \leq
916: 2n^{1-\alpha}$ such that for any $2$-coloring of the hypergraph
917: $H_\alpha+\sum R_{i,j}$, one of these sets must be monochromatic.
918: If this happens, we call such a stage ``successful". After each
919: successful stage, we remove all edges of $H_\alpha$ incident with
920: any of the sets $S_{i,1}, \ldots, S_{i,q}$. Since degrees are
921: bounded by $n^{k-1-\frac{\ell}{2(\ell-1)} \epsilon}$ and we repeat
922: $t = c_5 \ell^{-k} n^{\frac{\ell}{\ell-1} (\alpha - \epsilon/2)}$
923: times, the total number of edges we remove is at most
924: $$ \sum_{i=1}^{t} \sum_{j=1}^{q} |S_{i,j}| n^{k-1-\frac{\ell}{2(\ell-1)} \epsilon}
925: \leq t q \cdot 2 n^{1-\alpha} \cdot n^{k-1-\frac{\ell}{2(\ell-1)} \epsilon} =
926: 2 c_5 \ell^{-2}n^{k-\epsilon - \frac{\epsilon - \alpha}{\ell-1}}
927: \leq c_5 n^{k - \epsilon - \frac{\epsilon - \alpha}{\ell-1}}. $$
928: In particular, before every stage we still have at least $c_5 n^{k
929: - \epsilon - \frac{\ell-2}{\ell-1} (\alpha - \epsilon/2)}$ edges
930: available, so we can use Lemma~\ref{one-tree}. Since the expected
931: number of stages that are not successful is $o(t)$, by Markov's
932: inequality, we almost surely get at least $t/2$ successful stages.
933: Eventually, we obtain sets $S_{i,j}$ for $1 \leq i \leq t/2$ and
934: $1 \leq j \leq q$ such that
935: \begin{itemize}
936: \item For $i_1 \neq i_2$ and any $j_1, j_2$, $S_{i_1, j_1} \cap
937: S_{i_2, j_2} = \emptyset$.
938: \item For any $2$-coloring of $H_\alpha+\sum R_{i,j}$ and any $i$, there is $j_i$ such
939: that
940: $S_{i,j_i}$ is monochromatic.
941: \end{itemize}
942:
943: Finally, we add once again a collection $R_k$ of $\omega(n^{\ell
944: \epsilon/2})$ random $\ell$-tuples. We do not know a priori
945: which selection of sets $S_{i,j}$ will be monochromatic but there is only
946: exponential number of choices $q^{t/2} =
947: e^{O(t)}$. For any specific choice of sets to be monochromatic,
948: Lemma~\ref{clusters} says that the probability that after adding
949: $\omega(n^{\ell \epsilon/2})$ random $\ell$-tuples, there is a feasible
950: $2$-coloring keeping these sets monochromatic, is $e^{-\omega(t)}$.
951: By the union bound, the probability that there exist a proper $2$-coloring
952: of $H_\alpha+\sum R_{i,j}+R_k$ is at most $q^{t/2}e^{-\omega(t)}=o(1)$.
953: This completes the proof of this lemma together with the proof of Theorem~\ref{main}.
954: \hfill $\Box$
955:
956: \vspace{0.4cm}
957: \noindent
958: {\bf Acknowledgment.}\, The first author would like to thank Michael
959: Krivelevich for helpful and stimulating discussions.
960:
961:
962: \begin{thebibliography}{99}
963: \bibitem{AKKT}
964: D. Achlioptas, J.H. Kim, M. Krivelevich and P. Tetali,
965: Two-coloring random hypergraphs,
966: {\em Random Structures and Algorithms} 20 (2002), 249--259.
967:
968: \bibitem{AM}
969: D. Achlioptas and C. Moore,
970: On the $2$-colorability of random hypergraphs, in:
971: {\em Randomization and approximation techniques in computer science},
972: Lecture Notes in Comput. Sci. 2483, Springer, Berlin, 2002, 78--90.
973:
974: \bibitem{AP}
975: D. Achlioptas and Y.Peres,
976: The Threshold for Random $k$-SAT is $2^k \log 2 - O(k)$,
977: {\em Journal of the AMS} 17 (2004), 947--973.
978:
979: \bibitem{AS}
980: N. Alon and J. Spencer,
981: A note on coloring random $k$-sets, unpublished manuscript.
982:
983: \bibitem{Beck}
984: J. Beck, On $3$-chromatic hypergraphs,
985: {\em Discrete Math.} 24 (1978), 127--137.
986:
987: \bibitem{Ber}
988: F. Bernstein,
989: Zur Theorie der trigonometrischen Reihe,
990: {\em Leipz. Ber.} 60 (1908), 325--338.
991:
992: \bibitem{BFM} T. Bohman, A. Frieze and R. Martin, How many random
993: edges make a dense graph Hamiltonian?, {\em Random Structures and
994: Algorithms} 22 (2003), 33--42.
995:
996: \bibitem{BFKM} T. Bohman, A. Frieze, M. Krivelevich and R. Martin,
997: Adding random edges to dense graphs, {\em Random Structures and
998: Algorithms} 24 (2004), 105--117.
999:
1000: \bibitem{B} B. Bollob\'as, {\bf Random graphs}, $2^{nd}$ ed.
1001: Cambridge Studies in Advanced Mathematics, 73.
1002: Cambridge University Press, Cambridge, 2001.
1003:
1004: \bibitem{Er}
1005: P. Erd\H{o}s,
1006: On a combinatorial problem,
1007: {\em Nordisk Mat. Tidskr.} 11 (1963), 5--10.
1008:
1009: \bibitem{Er1}
1010: P. Erd\H{o}s,
1011: On a combinatorial problem. II,
1012: {\em Acta Math. Acad. Sci. Hungar} 15 (1964), 445--447.
1013:
1014: \bibitem{EK}
1015: P. Erd\H{o}s and D. Kleitman,
1016: On coloring graphs to maximize the proportion of multicolored $k$-edges,
1017: {\em J. Combinatorial Theory} 5 (1968), 164--169.
1018:
1019: \bibitem{EL}
1020: P. Erd\H{o}s and L. Lov\'asz,
1021: Problems and results on $3$-chromatic hypergraphs and some
1022: related questions, in: {\em Infinite and finite sets Vol. II},
1023: Colloq. Math. Soc. Janos Bolyai, Vol. 10,
1024: North-Holland, Amsterdam, 1975, 609--627.
1025:
1026: \bibitem{ER} P. Erd\H{o}s and A. R\'enyi, On the evolution of random
1027: graphs, {\em Publ. Math. Inst. Hungar. Acad. Sci.} 5 (1960),
1028: 17--61.
1029:
1030: \bibitem{JLR} S. Janson, T. \L uczak, and A. Ruci\'nski,
1031: {\bf Random Graphs}, Wiley, New York, 2000.
1032:
1033: \bibitem{KST}
1034: M. Krivelevich, B. Sudakov and P. Tetali,
1035: On smoothed analysis in dense graphs and formulas,
1036: {\em Random Structures and Algorithms}, to appear.
1037:
1038: \bibitem{L}
1039: L. Lov\'asz, Coverings and colorings of
1040: hypergraphs, in: {\em Proc. $4^{th}$ Southeastern Conf. on Combinatorics, Graph
1041: Theory and Computing}, 1973, Utilitas Math., 3--12.
1042:
1043: \bibitem{RS}
1044: J. Radhakrishnan and A. Srinivasan,
1045: Improved bounds and algorithms for hypergraph $2$-coloring,
1046: {\em Random Structures Algorithms} 16 (2000), 4--32.
1047:
1048: \bibitem{ST} D. A. Spielman and S.-H. Teng, Why the simplex algorithm
1049: usually takes polynomial time, {\em Proceedings of the 33rd Annual
1050: ACM Symposium on Theory of Computing} (2001), 296--305.
1051:
1052:
1053: \end{thebibliography}
1054:
1055: \end{document}
1056: