New upper bound on block sensitivity and certificate complexity in terms of sensitivity

Abstract

Sensitivity  [4, 5] and block sensitivity  [9] are two important complexity measures of Boolean functions. A longstanding open problem in decision tree complexity, the “Sensitivity versus Block Sensitivity” question, proposed by Nisan and Szegedy [10] in 1992, is whether these two complexity measures are polynomially related, i.e., whether $bs(f)=O(s(f)^{O(1)})$.

We prove an new upper bound on block sensitivity in terms of sensitivity: $bs(f)\leq 2^{s(f)-1}s(f)$. Previously, the best upper bound on block sensitivity was $bs(f)\leq(\frac{e}{\sqrt{2\pi}})e^{s(f)}\sqrt{s(f)}$ by Kenyon and Kutin [8]. We also prove that if $\min\{s_{0}(f),s_{1}(f)\}$ is a constant, then sensitivity and block sensitivity are linearly related, i.e. $bs(f)=O(s(f))$.

1 Introduction

Sensitivity  [4, 5] and block sensitivity  [9] are two important complexity measures in decision tree complexity. It has been proven that block sensitivity is polynomially related to a number of other complexity measures: certificate complexity, polynomial degree, approximate degree, deterministic/randomized/quantum query complexity, etc. [3]. However, it is not known whether the simpler-to-define sensitivity has a similar property. The sensitivity-block sensitivity conjecture proposed by Nisan and Szegedy [10] suggests that for any total Boolean function $f$, the block sensitivity complexity and sensitivity complexity are polynomially related, i.e. $bs(f)=O(s(f)^{O(1)})$.

The best known lower bound on $bs(f)$ in terms of $s(f)$ is quadratic and upper bound is exponential. Rubinstein [11] gave the first quadratic separation for block sensitivity and sensitivity by constructing a Boolean function $f$ with $bs(f)={1\over 2}s(f)^{2}$. Virza [13] improved the separation to $bs(f)={1\over 2}s(f)^{2}+s(f)$. Ambainis and Sun [2] further improved the gap to $bs(f)={2\over 3}s(f)^{2}-{1\over 2}s(f)$. However, the best known upper bound on block sensitivity is

$$bs(f)\leq(\frac{e}{\sqrt{2\pi}})e^{s(f)}\sqrt{s(f)},$$ (1)

given by Kenyon and Kutin [8]. More information about sensitivity-block sensitivity conjecture can be found in a survey paper by Hatami et al. [7] and in Scott Aaronson’s blog post [1].

In this paper, we give a new upper bound on certificate complexity (which is related to block sensitivity).

Theorem 1.

For any Boolean function $f$,

$$C_{1}(f)\leq 2^{s_{0}(f)-1}s_{1}(f),\ \ \ C_{0}(f)\leq 2^{s_{1}(f)-1}s_{0}(f).$$ (2)

This implies

Corollary 1.

$$bs(f)\leq C(f)\leq 2^{s(f)-1}s(f).$$

We also introduce a new complexity measure called consistent block sensitivity, see section 4 for the definition. Using it we prove another upper bound for block sensitivity.

Theorem 2.

For any Boolean function $f$,

$$bs(f)\leq\min\{2^{s_{0}(f)},2^{s_{1}(f)}\}s_{1}(f)s_{0}(f).$$ (3)

One consequence of this result is that,

Corollary 2.

If $\min\{s_{0}(f),\ s_{1}(f)\}=O(1)$, then $bs(f)=O(s(f))$.

Remark 1.

This result suggests a stronger version of sensitivity-block sensitivity conjecture:

$$bs(f)=O(s_{1}(f)s_{0}(f))?$$ (4)

The rest part of the paper is organized as follows: in Section 2 we give some basic definitions. We prove Theorem 1 in Section 3 and Theorem 2 in section 4. We conclude the paper in Section 5.

2 Preliminaries

In this section, we give the basic definitions we use in this paper. See [3] for a survey on more complexity measures for decision trees and the relationships between them. Let $f:\{0,1\}^{n}\rightarrow\{0,1\}$ be a total Boolean function.

Definition 1.

The sensitivity complexity $s(f,x)$ of $f$ on input $x$ is defined as the number of bits on which the function is sensitive, i.e. $s(f,x)=\big|\{i|f(x)\neq f(x^{(i)})\}\big|$, where $x^{(i)}$ is obtained by flipping the $i$-th bit of $x$. Define the sensitivity of $f$ as $s(f)=\max\big\{s(f,x)|x\in\{0,1\}^{n}\big\}$ and the 0-sensitivity and 1-sensitivity of $f$ as $s_{0}(f)=\max\big\{s(f,x)|x\in\{0,1\}^{n},f(x)=0\big\}$, $s_{1}(f)=\max\big\{s(f,x)|x\in\{0,1\}^{n},f(x)=1\big\}$.

Definition 2.

The block sensitivity complexity $bs(f,x)$ of $f$ on input $x$ is defined as maximum number of pairwise disjoint subsets $B_{1},...,B_{k}$ of $[n]$ such that $f(x)\neq f(x^{(B_{i})})$ for all $i\in[k]$. Here $x^{(B_{i})}$ is obtained by flipping all the bits $\{x_{j}|j\in B_{i}\}$ of $x$. Define the block sensitivity of $f$ as $bs(f)=\max\big\{bs(f,x)|x\in\{0,1\}^{n}\big\}$ and the 0-block sensitivity and 1-block sensitivity of $f$ as $bs_{0}(f)=\max\big\{bs(f,x)|x\in\{0,1\}^{n},f(x)=0\big\}$, $bs_{1}(f)=\max\big\{bs(f,x)|x\in\{0,1\}^{n},f(x)=1\big\}$.

Definition 3.

The certificate complexity $C(f,x)$ of $f$ on input $x$ is defined as the minimum length of a partial assignment of $x$ such that $f$ is constant on this restriction. Define the certificate complexity of $f$ as $C(f)=\max\big\{C(f,x)|x\in\{0,1\}^{n}\big\}$ and the 0-certificate and 1-certificate of $f$ as $C_{0}(f)=\max\big\{C(f,x)|x\in\{0,1\}^{n},f(x)=0\big\}$, $C_{1}(f)=\max\big\{C(f,x)|x\in\{0,1\}^{n},f(x)=1\big\}$.

We introduce a new notation called consistent block sensitivity here.

Definition 4.

The consistent block sensitivity of $f$ on an input $x$ (denoted $cbs(f,x)$) is equal to the maximum number $k$ of blocks $B_{1},\ldots,B_{k}$ such that $f(x)\neq f(x^{(B_{i})})$ and $x^{(B_{1})},\ldots,x^{(B_{k})}$ have certificates $C_{1},\ldots,C_{k}$ such that there is an input $y$ that satisfies all of $C_{i}$ simultaneously. The consistent block sensitivity of $f$ (denoted $cbs(f)$) is equal to the maximum of $cbs(f,x)$ over all inputs $x$. The $c$-consistent block sensitivity of $f$ (denoted $cbs_{c}(f)$) is equal to the maximum of $cbs(f,x)$ over all inputs $x$ such that $f(x)=c$.

3 Upper bound for certificate complexity in terms of sensitivity

In this section we prove theorem 1. Actually, we prove a slightly stronger result.

Theorem 3.

Let $f:\{0,1\}^{n}\rightarrow\{0,1\}$ be a Boolean function, then

$$C_{1}(f)\leq 2^{s_{0}(f)-1}s_{1}(f)-(s_{0}(f)-1),\ C_{0}(f)\leq 2^{s_{1}(f)-1}s_{0}(f)-(s_{1}(f)-1).$$ (5)

Proof By symmetry we only need to prove $C_{1}(f)\leq 2^{s_{0}(f)-1}s_{1}(f)-(s_{0}(f)-1)$. Without the loss of generality, we assume $C_{1}(f)=C(f,0^{n})$, i.e. the 1-certificate complexity is achieved on the input $0^{n}$. We have $f(0^{n})=1$. We assume that the minimal certificate of $0^{n}$ consists of $x_{1}=0,x_{2}=0,\ldots,x_{m}=0$, where $m=C(f,0^{n})=C_{1}(f)$.

Let $Q_{0}$ be the set of inputs $x$ that satisfies $x_{1}=x_{2}=\ldots=x_{m}=0$. Since $x_{1}=0,x_{2}=0,\ldots,x_{m}=0$ is a 1-certificate, we have $\forall~x\in Q_{0}$, $f(x)=1$.

For each $i\in[m]$, let $Q_{i}$ be the set of inputs $x$ with $x_{1}=\ldots=x_{i-1}=x_{i+1}=\ldots=x_{m}=0$ and $x_{i}=1$. Let $S$ be the total sensitivity of all inputs $x\in\bigcup_{i=1}^{m}Q_{i}$. It consists of three parts: sensitivity between $Q_{i}$ and $Q_{0}$ (denoted by $S_{1}$), sensitivity inside $Q_{i}$ (denoted by $S_{2}$) and sensitivity between $Q_{i}$ and other input (denoted by $S_{3}$), i.e.

$$S=\sum_{i=1}^{m}\sum_{x\in Q_{i}}s(f,x)=S_{1}+S_{2}+S_{3}.$$ (6)

In the following we estimate $S_{1},S_{2}$ and $S_{3}$ separately. We use $A_{1},\ldots,A_{m}$ to denote the set of $0$-inputs in $Q_{1},\ldots,Q_{m}$, respectively, i.e. $A_{i}=\{x\in Q_{i}|f(x)=0\}$ ($i\in[m]$). Since $x_{1}=\ldots=x_{m}=0$ is the minimal certificate, i.e. $Q_{0}$ is maximal, thus $A_{1},\ldots,A_{m}$ are all nonempty.

We will use the following isoperimetric inequality for the Boolean hypercube.

Lemma 1.

[6] For any $A\subseteq\{0,1\}^{n}$, the edges between $A$ and $\bar{A}=\{0,1\}^{n}\setminus A$ is lower bounded by

$$|E(A,\bar{A})|\geq|A|(n-\log_{2}|A|).$$

We also need the following lemma which follows from Lemma 1 but can be also proven without using it [12]:

Lemma 2.

[12] For any $i\in[m]$,

$$|A_{i}|\geq 2^{n-m-s_{0}(f)+1}.$$

The sensitivity between $Q_{i}$ and $Q_{0}$ is $|A_{i}|$. By summing over all possible $i$ we get:

$$S_{1}=\sum_{i=1}^{m}|A_{i}|.$$ (7)

Sensitivity inside $Q_{1},\ldots,Q_{m}$: By Lemma 1, for each $i\in[m]$, the number of edges between $A_{i}$ and $Q_{i}\setminus A_{i}$ is bounded by:

$$|E(A_{i},Q_{i}\setminus A_{i})|\geq|A_{i}|(\log_{2}|Q_{i}|-\log_{2}|A_{i}|)=|A_{i}|(n-m-\log_{2}|A_{i}|).$$

Therefore,

$$S_{2} = 2\sum_{i=1}^{m}|E(A_{i},Q_{i}\setminus A_{i})|$$ (8)

$$\geq 2\sum_{i=1}^{m}|A_{i}|(n-m-\log_{2}|A_{i}|).$$

Sensitivity between $Q_{i}$ and other inputs: For each $1\leq i<j\leq m$, let $Q_{i,j}$ be the set of inputs that are adjacent to both $Q_{i}$ and $Q_{j}$, i.e. $Q_{i,j}$ is the set of inputs $x$ that satisfy $x_{1}=\ldots x_{i-1}=x_{i+1}=\ldots x_{j-1}=x_{j+1}=\ldots x_{m}=0$ and $x_{i}=x_{j}=1$. The sensitivity between $Q_{i},Q_{j}$ and $Q_{i,j}$ is lower bounded by

$$\sum_{x\in Q_{0}}|f(x+e_{i})-f(x+e_{j})|.$$

where $e_{i}$ is the unit vector with the $i$-th coordinate equal to 1 and all other coordinates equal to 0. Then, $x+e_{i}$, $x+e_{j}$ are the neighbors of $x$ in $Q_{i}$ and $Q_{j}$, respectively. Summing over all possible pairs of $(i,j)$ we get

$$S_{3} \geq \sum_{1\leq i<j\leq m}\sum_{x\in Q_{0}}|f(x+e_{i})-f(x+e_{j})|$$ (9)

$$= \sum_{x\in Q_{0}}\left(\sum_{i=1}^{m}f(x+e_{i})\right)\left(m-\sum_{i=1}^{m}f(x+e_{i})\right)$$

$$= \sum_{x\in Q_{0}}s(f,x)(m-s(f,x)).$$

If we combine inequalities (7)-(9), we get

$$S = \sum_{i=1}^{m}\sum_{x\in Q_{i}}s(f,x)$$ (10)

$$\geq \sum_{i=1}^{m}|A_{i}|+2\sum_{i=1}^{m}|A_{i}|(n-m-\log_{2}|A_{i}|)+\sum_{x\in Q_{0}}s(f,x)(m-s(f,x)).$$

Since $s(f,x)$ is upper bounded by $s_{0}(f)$ or $s_{1}(f)$ (depending on whether $f(x)=0$ or $f(x)=1$), we have

$$\sum_{x\in Q_{i}}s(f,x) \leq |A_{i}|s_{0}(f)+(|Q_{i}|-|A_{i}|)s_{1}(f)$$

$$= |A_{i}|s_{0}(f)+(2^{n-m}-|A_{i}|)s_{1}(f)$$

Thus,

$$S=\sum_{i=1}^{m}\sum_{x\in Q_{i}}s(f,x)\leq\sum_{i=1}^{m}\Big(|A_{i}|s_{0}(f)+(2^{n-m}-|A_{i}|)s_{1}(f)\Big).$$ (11)

We use $w$ to denote the total number of $0$-inputs in $Q_{1},\ldots,Q_{m}$. Then,

$$w=\sum_{i=1}^{m}|A_{i}|=\sum_{x\in Q_{0}}s(f,x).$$

The inequality (11) can be rewritten as

$$S\leq w\cdot s_{0}(f)+(m\cdot 2^{n-m}-w)s_{1}(f).$$ (12)

Also, $s(f,x)\leq s_{1}(f)$ for each $x\in Q_{0}$. Thus, the right-hand side of inequality (10) is

$$\sum_{i=1}^{m}|A_{i}|+2\sum_{i=1}^{m}|A_{i}|(n-m-\log_{2}|A_{i}|)+\sum_{x\in Q_{0}}s(f,x)(m-s(f,x))$$ (13)

$$\geq w+2\sum_{i=1}^{m}|A_{i}|(n-m-\log_{2}|A_{i}|)+(m-s_{1}(f))\sum_{x\in Q_{0}}s(f,x)$$

$$= w+2w(n-m)-2\sum_{i=1}^{m}|A_{i}|\log_{2}|A_{i}|+(m-s_{1}(f))w$$

$$= w(1+2n-m-s_{1}(f))-2\sum_{i=1}^{m}|A_{i}|\log_{2}|A_{i}|.$$

By combining inequalities (10)-(13) we get

$$w(1+2n-m-s_{1}(f))-2\sum_{i=1}^{m}|A_{i}|\log_{2}|A_{i}|\leq w\cdot s_{0}(f)+(m\cdot 2^{n-m}-w)s_{1}(f).$$

By rearranging the inequality we get

$$w(1+2n-m-s_{0}(f))\leq 2\sum_{i=1}^{m}|A_{i}|\log_{2}|A_{i}|+m\cdot 2^{n-m}s_{1}(f).$$ (14)

Since the function $g(x)=x\log_{2}x$ is convex, we know that $|A_{i}|\leq|Q_{i}|=2^{n-m}$. From Lemma 2 $|A_{i}|\geq 2^{n-m-s_{0}(f)+1}$. Therefore,

$$g(|A_{i}|) = g\left(\frac{|A_{i}|-2^{n-m-s_{0}(f)+1}}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot 2^{n-m}+\frac{2^{n-m}-|A_{i}|}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot 2^{n-m-s_{0}(f)+1}\right)$$

$$\leq \frac{|A_{i}|-2^{n-m-s_{0}(f)+1}}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot g(2^{n-m})+\frac{2^{n-m}-|A_{i}|}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot g(2^{n-m-s_{0}(f)+1})$$

$$= \frac{|A_{i}|-2^{n-m-s_{0}(f)+1}}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot 2^{n-m}(n-m)$$

$$+\frac{2^{n-m}-|A_{i}|}{2^{n-m}-2^{n-m-s_{0}(f)+1}}\cdot 2^{n-m-s_{0}(f)+1}(n-m-s_{0}(f)+1)$$

$$= \frac{|A_{i}|-2^{n-m-s_{0}(f)+1}}{2^{s_{0}(f)-1}-1}\cdot 2^{s_{0}(f)-1}(n-m)+\frac{2^{n-m}-|A_{i}|}{2^{s_{0}(f)-1}-1}(n-m-s_{0}(f)+1)$$

$$= \left(\frac{|A_{i}|-2^{n-m-s_{0}(f)+1}}{2^{s_{0}(f)-1}-1}2^{s_{0}(f)-1}+\frac{2^{n-m}-|A_{i}|}{2^{s_{0}(f)-1}-1}\right)(n-m)-\frac{2^{n-m}-|A_{i}|}{2^{s_{0}(f)-1}-1}(s_{0}(f)-1)$$

$$= |A_{i}|(n-m)-\frac{2^{n-m}-|A_{i}|}{2^{s_{0}(f)-1}-1}(s_{0}(f)-1).$$

Hence

$$\sum_{i=1}^{m}|A_{i}|\log_{2}|A_{i}| = \sum_{i=1}^{m}g(|A_{i}|)$$ (15)

$$\leq \sum_{i=1}^{m}\left(|A_{i}|(n-m)-\frac{2^{n-m}-|A_{i}|}{2^{s_{0}(f)-1}-1}(s_{0}(f)-1)\right)$$

$$= w(n-m+\frac{s_{0}(f)-1}{2^{s_{0}(f)-1}-1})-2^{n-m}\frac{s_{0}(f)-1}{2^{s_{0}(f)-1}-1}.$$

By combining inequalities (14) and (15), we get

$$w(1+2n-m-s_{0}(f))$$

$$\leq 2\left(w(n-m+\frac{s_{0}(f)-1}{2^{s_{0}(f)-1}-1})-2^{n-m}\frac{s_{0}(f)-1}{2^{s_{0}(f)-1}-1}\right)+m\cdot 2^{n-m}s_{1}(f).$$

It implies that

$$w\left(1+m-s_{0}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\right)\leq m\cdot 2^{n-m}\left(s_{1}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\right),$$

Substituting $w=\sum_{i=1}^{m}|A_{i}|\geq m\cdot 2^{n-m-s_{0}(f)+1}$, we get

$$m\cdot 2^{n-m-s_{0}(f)+1}\left(1+m-s_{0}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\right)\leq m\cdot 2^{n-m}\left(s_{1}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\right),$$

i.e.

$$1+m-s_{0}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\leq 2^{s_{0}(f)-1}\left(s_{1}(f)-\frac{2(s_{0}(f)-1)}{2^{s_{0}(f)-1}-1}\right),$$

which implies

$$m\leq 2^{s_{0}(f)-1}s_{1}(f)-s_{0}(f)+1.$$

∎

4 Upper bound for block sensitivity in terms of sensitivity

First we show that the consistent block sensitivity is no more than sensitivity.

Theorem 4.

$s_{b}(f)\geq cbs_{b}(f)$ for any $b\in\{0,1\}$.

Proof We prove $s_{0}(f)\geq cbs_{0}(f)$ here. Without the loss of generality assume that the maximum $cbs_{0}(f)=k$ is achieved on the input $0^{n}$. Let $|B_{i}|=b_{i}$. We denote the variables in the block $B_{i}$ as $x_{i,1},x_{i,2}$, $\ldots$, $x_{i,b_{i}}$. We further assume that blocks $B_{i}$ are minimal.

Since $C_{i}$ is a certificate for $x^{(B_{i})}$, it must include conditions $x_{i,1}=1$, $x_{i,2}=1$, $\ldots$, $x_{i,b_{i}}=1$ and no other conditions $x_{j}=1$. (If $C_{i}$ included some other $x_{j}=1$, it would not be satisfied by $x^{(B_{i})}$. If $C_{i}$ did not include $x_{i,j}=1$, it would also be satisfied by $x^{(B_{i}-(i,j))}$ - contradicting the assumption that $B_{i}$ is minimal.)

Moreover, $C_{i}$ must not include a condition $x_{i^{\prime},j}=0$ for $i^{\prime}\neq i$. (Otherwise, it would be impossible to satisfy $C_{i}$ and $C_{i^{\prime}}$ simultaneously.) Hence, $C_{i}$ consists of $x_{i,1}=1$, $x_{i,2}=1$, $\ldots$, $x_{i,b_{i}}=1$ and, possibly, $x_{j}=0$ for some variables $j$ that do not belong to any of $B_{i}$.

We say that an input $y$ has a type $(t_{1},\ldots,t_{k})$ if $y_{i}=0$ for all $i\notin B_{1}\cup\ldots\cup B_{k}$ and, for every $i\in\{1,\ldots,k\}$, the number of $y_{i,j}=1$ is equal to $t_{i}$.

Claim 1.

Assume that $s_{0}(f)<k$. Then, for every input $y$ of type $(t_{1},\ldots,t_{k})$, $t_{1}<b_{1}$, $\ldots$, $t_{k}<b_{k}$, $f(y)=1$.

Substituting $t_{1}=\ldots=t_{k}=0$ leads to the conclusion $f(0^{n})=1$ which is in a contradiction with our initial assumption $f(0^{n})=0$. Hence, the assumption $s_{0}(f)<k$ is wrong and we must have $s_{0}(f)\geq k=cbs_{0}(f)$.

Proof (of Claim 1) By induction on $B=(b_{1}+\ldots+b_{k})-(t_{1}+\ldots+t_{k})$.

The smallest possible value is $B=k$ (if $t_{i}=b_{i}-1$ for all $i$). This is our base case. In this case, there is one variable $y_{i,j}=0$ for each $i$. Then $y^{(i,j)}$ satisfies the certificate $C_{i}$. Hence, $f(y^{(i,j)})=1$. This means that, if $f(y)=0$, then $y$ is sensitive to $k$ variables $y_{i,j}$ - one for each $i\in\{1,\ldots,k\}$. Thus, $s_{0}(f)<k$ implies $f(y)=1$.

For the inductive case, we assume that Claim is true for $B=B_{0}-1$ and prove that this implies Claim being true for $B=B_{0}$. Let $y$ be an input for which $B=B_{0}$. We choose $y_{i,j}=0$ for each $i\in\{1,\ldots,k\}$. (Such $y_{i,j}$ always exists, since $t_{i}<b_{i}$.) We claim that $f(y^{(i,j)})=1$. (If $t_{i}=b_{i}-1$, then, with $y_{i,j}=1$ we have $y_{i,1}=\ldots=y_{i,b_{i}}=1$ and the certificate $C_{i}$ is satisfied. Hence, $f(y^{(i,j)})=1$. If $t_{i}<b_{i}-1$, then $y^{(i,j)}$ is an input of type $(t_{1},\ldots,t_{i}+1,\ldots,t_{k})$ and $f(y^{(i,j)})=1$ follows from the inductive assumption.) By the same argument as in the previous case, this implies $f(y)=1$. ∎

Next we upper bound block sensitivity by certificate complexity and consistent block sensitivity.

Theorem 5.

$bs_{0}(f)\leq 2(C_{1}(f)-\frac{1}{2})cbs_{0}(f),\ bs_{1}(f)\leq 2(C_{0}(f)-\frac{1}{2})cbs_{1}(f)$.

Together with Theorem 4, this implies

Corollary 3.

$bs_{0}(f)\leq 2(C_{1}(f)-\frac{1}{2})s_{0}(f),\ bs_{1}(f)\leq 2(C_{0}(f)-\frac{1}{2})s_{1}(f)$.

Proof (of Theorem 5) We use the same notation as in the proof of Theorem 4. Also, we again assume that the maximum $bs_{0}(f)$ is achieved on the input $x=0^{n}$ and the sensitive blocks $B_{i}$ are minimal.

We consider a graph $G$ with $m=bs_{0}(f)$ vertices $v_{1}$, $\ldots$, $v_{bs_{0}(f)}$, with an edge $(v_{i},v_{j})$ if the certificates $C_{i}$ and $C_{j}$ are consistent. The size of the largest clique in this graph is at most $cbs_{0}(f)$. By Turan’s theorem, a graph not containing a clique of size $k+1$, $k=cbs_{0}(f)$ has at most $\frac{k-1}{2k}m^{2}$ edges.

Therefore $\overline{G}$ (the complement of graph $G$) has at least

$${m\choose 2}-\frac{k-1}{2k}m^{2}=\frac{m^{2}}{2k}-\frac{m}{2}$$ (16)

edges. Each of those edges corresponds to a pair of certificates $C_{i}$, $C_{j}$ that are not consistent.

Since $C_{i}$ is a certificate for $x^{(B_{i})}$ and $B_{i}$ is minimal, $C_{i}$ must consist of conditions $x_{i,1}=1$, $\ldots$, $x_{i,b_{i}}=1$ and $x_{j}=0$ for some $j\notin B_{i}$. (This follows in the same way as the similar statement in the proof of Theorem 4.)

If $C_{i}$ and $C_{j}$ are inconsistent, then either $C_{i}$ contains a condition $x_{j,l}=0$ (which is not consistent with $x_{j,l}=1$ in $C_{j}$) or $C_{j}$ contains a condition $x_{i,l}=0$ (which is not consistent with $x_{i,l}=1$ in $C_{i}$). Since $|C_{i}|\leq C_{1}(f)$, $C_{i}$ may contain conditions $x_{j,l}=0$ for at most $C_{1}(f)-1$ different $j$. This means that the number of edges in $\overline{G}$ is at most $m(C_{1}(f)-1)$. Together with (16), this implies

$$\frac{m^{2}}{2k}-\frac{m}{2}\leq m(C_{1}(f)-1),$$

$$\frac{m}{2k}-\frac{1}{2}\leq(C_{1}(f)-1),$$

$$m\leq 2k\left(C_{1}(f)-\frac{1}{2}\right).$$

∎

Now we prove Theorem 2.

Proof (of Theorem 2) From Theorem 1 $C_{0}(f)\leq 2^{s_{1}(f)-1}s_{0}(f)$, hence

$$bs_{0}(f)\leq C_{0}(f)\leq 2^{s_{1}(f)-1}s_{0}(f).$$

From Corollary 3 we have $bs_{0}(f)\leq 2(C_{1}(f)-\frac{1}{2})s_{0}(f)$, together with Theorem 1 we get

$$bs_{0}(f)\leq 2(2^{s_{0}(f)-1}s_{1}(f)-\frac{1}{2})s_{0}(f).$$

Therefore,

$$bs_{0}(f)\leq\min\{2^{s_{1}(f)}s_{0}(f)s_{1}(f),\ 2^{s_{0}(f)}s_{1}(f)s_{0}(f)\}.$$

Similarly we can show that

$$bs_{1}(f)\leq\min\{2^{s_{1}(f)}s_{0}(f)s_{1}(f),\ 2^{s_{0}(f)}s_{1}(f)s_{0}(f)\}.$$

∎

5 Conclusions

In this paper we proved an upper bound for certificate complexity in terms of sensitivity, which also implies upper bound on block sensitivity in terms of sensitivity. We also introduce the concept of consistent block sensitivity. We then use it to prove that block sensitivity and sensitivity are linear related if $\min\{s_{0}(f),s_{1}(f)\}=O(1)$. This suggests a refined version of the sensitivity-block sensitivity question: is it true that

$$bs(f)=O(s_{0}(f)s_{1}(f))?$$