text
stringlengths
128
2.05k
(b) Assign the coordinates in [MATH] using a [MATH] -biased distribution [MATH] Equivalently, we prove that the following distribution [MATH] -fools ROBPs of width- [MATH] and length- [MATH]
1. Pick [MATH] using a [MATH] -biased distribution with marginals [MATH] 2. Assign the coordinates in [MATH] uniformly at random.
3. (a) Pick [MATH] using a [MATH] -biased distribution with marginals [MATH] (b) Assign the coordinates in [MATH] uniformly at random.
(c) Assign the coordinates in [MATH] using a [MATH] -biased distribution [MATH] Let [MATH] . Let [MATH] be the event that [MATH] can be computed by a ROBP of the form [MATH] where [MATH] are defined over disjoint sets of at most [MATH] variables, and each [MATH] is a width- [MATH] ROBP with at most [MATH] vertices on t...
[MATH] Assuming that [MATH] happened, then by Theorem 4.2 [MATH] can be written as [MATH] where [MATH] are subprograms of [MATH] and [MATH] For each [MATH] , using Theorem we have that
[EQUATION] By linearity of expectation and the triangle inequality [EQUATION] [EQUATION] Overall, we get [EQUATION] [EQUATION] [EQUATION]
where the last equality is due to the fact for any [MATH] the distribution of [MATH] is the uniform distribution over [MATH] We bound Expression ( ) by
[EQUATION] where the second summand is bounded by [MATH] according to the above discussion using Theorem 4.2 and Theorem Pseudorandom restrictions for the XOR of short ROBPs
In this section, we prove Theorem Let [MATH] be pairwise disjoint subsets of [MATH] , each of size at most [MATH] . For [MATH] let [MATH] be a width [MATH] ROBP. We construct a pseudorandom generator that [MATH] -fools [MATH] We recall the statement of Theorem and the construction. See
Recall that the pseudorandom restriction assigns [MATH] fraction of the variables as follows: 1. Choose a set of coordinates [MATH] according to a [MATH] -biased distribution with marginals [MATH] , for [MATH]
2. Assign the variables in [MATH] according to a [MATH] -biased distribution, for [MATH] Analysis. We shall assume without loss of generality that for all [MATH] it holds that [MATH] We shall also assume without loss of generality that for all [MATH] it holds that [MATH] (i.e., that the functions are non-constant). Sin...
We partition the functions into [MATH] buckets according to their variance. Let [MATH] for every [MATH] , let [MATH] and [MATH] Let [MATH] be a sufficiently large constant. We consider two cases in our analysis:
Low-Variance Case: For every [MATH] we have [EQUATION] High-Variance Case: There exists a [MATH] with [EQUATION] Setting Up Parameters:
Let [MATH] be a sufficiently large constant. Set [EQUATION] 5.1 Low-Variance Case For [MATH] let [MATH] Thus, [MATH] Let [MATH] be any [MATH] -biased distribution with marginals [MATH] For [MATH] , we shall show that with probability at least [MATH] over the choice of [MATH] , it holds that
[EQUATION] for any [MATH] -biased distribution [MATH] over [MATH] Thus, by union bound Eq. ( ) holds for all [MATH] simultaneously with probability at least [MATH] over [MATH] Using the following XOR lemma for small-biased distributions from GMR 12 we get that any
[MATH] -biased distribution, fools [MATH] with error at most [MATH] (using [MATH] ). Lemma 5.1 GMR 12 , Thm. 4.1] , restated) Let [MATH] Let [MATH] be functions on disjoint input variables such that each [MATH] is [MATH] -fooled by any [MATH] -biased distribution. Let [MATH] be a multilinear function in its inputs. The...
In Appendix A.2 , we show how to derive Lemma 5.1 from GMR 12 , Thm. 4.1] In the remainder of this section, we focus on fooling a single [MATH] , that is, fooling the product (i.e., XOR) of functions [MATH] for which [MATH] We note that since we are in the “Low-Variance Case”, then
[EQUATION] We handle two cases depending on whether [MATH] is big or not. The case of [MATH] In this case there are at most [MATH] functions in [MATH] , each computed by a width- [MATH] ROBP on at most [MATH] bits. Thus, [MATH] can be computed by a ROBP of length at most [MATH] and width at most [MATH] Using Theorem 3....
The case of [MATH] In this case all variances in [MATH] are certainly smaller than [MATH] , and hence for all [MATH] , we have [MATH] Let
[EQUATION] Then, [EQUATION] We have [MATH] and [MATH] We will show that with high probability over [MATH] , any [MATH] -biased distribution fools [MATH]
For ease of notation, in this case we think of [MATH] as [MATH] and denote by [MATH] The proof strategy for this part follows the work of Gopalan and Yehudayoff GY14 We note that
[EQUATION] where [MATH] is the [MATH] -symmetric polynomial given by [MATH] We show that [MATH] and [MATH] fool the low-degree symmetric polynomials. Then, the following theorem by Gopalan and Yehudayoff GY14 bootstraps this to show that [MATH] and [MATH] also fool the sum of all high-degree symmetric polynomials.
Theorem 5.2 (Gopalan-Yehudayoff Tail Inequalities GY14 Let [MATH] Suppose [MATH] and [MATH] for some [MATH] and [MATH] Then, for every
[MATH] it holds that [MATH] Furthermore, if [MATH] , then [EQUATION] Analyzing the Symmetric Polynomials From Eq. ( ) and our assumption that
[MATH] we get that [MATH] Recall that [MATH] is a sufficiently large constant and recall the definition of [MATH] from Eqs. ( ), ( ) and ( ). We set
[EQUATION] In the following, we shall use the facts that [MATH] and [MATH] Claim 5.3 Let [MATH] Let [MATH] be a set of size at most [MATH] Then, with probability at least [MATH] over the choice of [MATH]
[MATH] has spectral-norm at most [MATH] Proof. Note that [MATH] can be computed by a ROBP with length [MATH] and width [MATH] (as in the case where [MATH] is big). Apply Theorem 3.10 to [MATH]
We say that [MATH] is a good set if for all sets [MATH] of size at most [MATH] the spectral-norm of [MATH] is at most [MATH] We observe that by Claim 5.3 , the probability that [MATH] is good is at least
[MATH] (using Eq. ( ) and ( )). Claim 5.4 If [MATH] is good, then for any [MATH] [MATH] has spectral-norm at most [MATH] Proof. We expand the [MATH] -symmetric polynomial:
[MATH] Since [MATH] is good, each summand has spectral-norm [EQUATION] (using [MATH] ). Summing over all [MATH] summands completes the proof.
We wish to show that with high probability the total variance under restrictions [MATH] is small. Towards this goal, we prove a bound on the [MATH] -th moment of the total variance.
Claim 5.5 [MATH] Proof. Fix [MATH] , not necessarily distinct indices. By Lemma 3.14 [EQUATION] from which we deduce [EQUATION] We are left to bound [MATH] By Fact 3.5 , for any [MATH] , the random variable [MATH] (whose value depends on the choice of [MATH] ) is bounded in [MATH] By Theorem 3.13 , its expected value i...
we have [MATH] In such a case [MATH] We get [MATH] , which gives [EQUATION] We say that a set [MATH] is excellent if [MATH] is good and [MATH]
Claim 5.6 [MATH] Proof. Note that [MATH] and apply Markov’s inequality on [MATH] using Claim 5.5 Claim 5.7 Let [MATH] be an excellent set. Let [MATH] be any [MATH] -biased distributions. Then, for [MATH] we have
[EQUATION] and [EQUATION] Proof. Recall that [MATH] and [MATH] . The first claim relies on the following: 1. [MATH] has small spectral-norm (using Claim 5.4 , since [MATH] is good) and hence is fooled by [MATH] . In details, its spectral-norm is at most [MATH] and [MATH] is [MATH] -biased. Thus
[EQUATION] 2. The expectation of [MATH] on a uniformly chosen [MATH] is at most [EQUATION] The second claim relies on the following:
1. [MATH] has small spectral-norm (using Claim 5.4 , since [MATH] is good) and hence is fooled by [MATH] . In details, its spectral-norm is at most [MATH] and [MATH] is [MATH] -biased. Thus
[EQUATION] 2. The expectation of [MATH] on a uniformly chosen [MATH] is [MATH] . ∎ The next lemma combined with Claim 5.7 concludes the low-variance case, since it shows that with high probability, [MATH] is excellent, and then [MATH] is an [MATH] -PRG for [MATH] (for a sufficiently large choice of [MATH] ).
Lemma 5.8 If [MATH] is excellent, then [MATH] Proof. Let [MATH] and let [MATH] be the event that [MATH] and [MATH] Picking [MATH] , and using Claim 5.7 the event [MATH] happens with probability at least [MATH] Assuming [MATH] occurs, Theorem 5.2 gives
[EQUATION] Furthermore, for sets of smaller cardinality, i.e., for [MATH] , Claim 5.7 gives [EQUATION] We would like to bound [MATH] for [MATH] . Towards this end, we consider the expectation of [MATH] by partitioning into the two cases depending on whether the event [MATH] occurred or not.
[EQUATION] Thus, [MATH] and we get [EQUATION] Since the [MATH] ’s and [MATH] ’s are bounded in [MATH] , we get [EQUATION] 5.2 High-Variance Case
In the high-variance case, there exists a [MATH] and an interval [MATH] (the constant [MATH] handles the case [MATH] ) satisfying:
[EQUATION] In this case, the expected value of [MATH] under the uniform distribution is rather small: [EQUATION] Recall that the pseudorandom restriction samples a set [MATH] according to some [MATH] -biased distribution [MATH] with marginals [MATH] , and a partial assignment to the bits in [MATH] according to some [MA...
[EQUATION] Thus, it suffices to show that for most [MATH] [MATH] we have [MATH] Theorem 5.9 (Theorem - High Variance Case) With probability [MATH] over [MATH] and [MATH] , it holds that [MATH]
Proof. Denote by [MATH] By our assumption, [MATH] Since all functions in [MATH] have variance at least [MATH] we have [EQUATION]
We remark that in this case, unlike the low-variance case, we do not know how to handle large [MATH] easily, so for the rest of the proof [MATH] can be anything between [MATH] and [MATH]
Fix [MATH] and [MATH] . We expand [MATH] [EQUATION] For any fixed [MATH] , using [MATH] gives [EQUATION] Claim 5.10 (Most [MATH] ’s preserve variance in expectation)
With probability at least [MATH] over the choice of [MATH] , it holds that [MATH] Proof. Since [MATH] , it suffices to show that with probability [MATH] over the choice of [MATH] we have [MATH] To show that
[MATH] is well-concentrated we analyze its [MATH] -th moment for [MATH] where [MATH] is a sufficiently large constant. [EQUATION]
Fix [MATH] , (not necessarily distinct), then by Lemma 3.14 [EQUATION] Overall, we get [EQUATION] To bound [MATH] we use the fact that by Theorem 3.13
[EQUATION] and then by Chernoff’s bound [MATH] with probability at least [MATH] Since [MATH] is always upper bounded by [MATH] , the [MATH] -moment of the sum is at most
[EQUATION] We get that [MATH] Since [MATH] this is at most [MATH] Thus, using Markov’s inequality, the probability that [MATH] is at most [MATH] which completes the proof.
Let [EQUATION] where [MATH] is a sufficiently large constant declared before Eq. ( ). Assume that [MATH] is an even integer. Recall that [MATH] We again define [MATH] to be a good set if [MATH] has spectral-norm at most [MATH] for all sets [MATH] of size at most [MATH] . As in Claim 5.3 the probability that [MATH] is g...
Claim 5.11 If [MATH] is a good set, then at most [MATH] of the [MATH] ’s have [MATH] Proof. If [MATH] have [MATH] , then their product has spectral-norm at least [MATH] , since [MATH] for functions defined on disjoint variables.
Fix an excellent set [MATH] . Let [MATH] be the of indices [MATH] with [MATH] . We show that with high probability over [MATH] [MATH] . We denote by
[EQUATION] Obviously [MATH] and [MATH] is bounded in [MATH] Furthermore, we have that [EQUATION] Thus, the error term have small spectral-norm since [MATH] . We use this fact to bound
[MATH] . (recall that [MATH] is an even integer.) Claim 5.12 [EQUATION] Proof. The spectral-norm of [MATH] is at most [MATH] . Thus, any [MATH] -biased distribution fools [MATH] with error at most [MATH] and we get
[EQUATION] To bound [MATH] we use Lemma 3.8 We observe that [MATH] are independent random variables, where each [MATH] is bounded in [MATH] with mean zero, and hence [MATH] (See Lemma 3.12 ). Applying Lemma 3.8 gives
[EQUATION] Since [MATH] , the upper bound on [MATH] is at most [MATH] Finally, the upper bound with respect to [MATH] is at most
[EQUATION] Using Markov’s Inequality and Claim 5.12 gives [EQUATION] using [MATH] and [MATH] in the last inequality. Furthermore, using Eqs. ( ) and ( 10 ): [MATH] In the complement event,
[EQUATION] Since [MATH] , we get that with probability at least [MATH] over [MATH] and [MATH] [MATH] . (End of Proof of Theorem 5.9
Assigning all the variables: a pseudorandom generator for the XOR of short ROBPs In Theorem , we proved that we can pseudorandomly assign [MATH] -fraction of the coordinates of [MATH] , while maintaining its acceptance probability up to an additive error of [MATH] , using [MATH] random bits. In this section, we will co...
[MATH] times, combined with Lovett’s Lov08 or Viola’s Vio08 pseudorandom generator for low-degree polynomials, and CHRT’s pseudorandom generator for constant-width ROBPs CHRT17 . Our main result is: See
Assigning [MATH] -fraction of the variables. The first step is rather standard. By making [MATH] recursive calls to Theorem we can assign all but [MATH] fraction of the coordinates while maintaining the acceptance probability. Note that we rely on the fact that under restrictions, the restricted function is still of th...
Claim 6.1 Let [MATH] , with block-length [MATH] Then, there is a pseudorandom restriction [MATH] using at most [MATH] random bits, and changing the acceptance probability by at most [MATH] Furthermore, [MATH] is [MATH] biased with marginals [MATH] , and
[MATH] is [MATH] biased. Proof Sketch. Apply Theorem with error [MATH] for [MATH] times recursively, with independent random bits per each iteration. Denoting by [MATH] , this generates [MATH] -pseudorandom restrictions
[MATH] where [MATH] and [MATH] for all [MATH] To be more precise, for each [MATH] [MATH] is the [MATH] -biased subset with marginals [MATH] in the description of the generator in Theorem and [MATH] is its assignment sampled from a [MATH] -biased distribution. We take [MATH] and [MATH] to be the concatenation of [MATH] ...
[EQUATION] The amount of random bits used to sample the restriction is [EQUATION] Next, we claim that [MATH] is a [MATH] -biased with marginals [MATH] Recall that [MATH] are [MATH] -biased with marginals [MATH] , and that [MATH] By Claim 3.7 , for any subset [MATH] of size at most [MATH] we have
[EQUATION] which implies by monotonicity that for larger subsets [MATH] , we have [MATH] Finally, we claim that conditioned on [MATH] and in fact for any choice of [MATH] [MATH] is [MATH] -biased. This is due to the fact that [MATH] is the concatenation of [MATH] where each [MATH] is [MATH] -biased.
We would like to claim that [MATH] simplifies after assigning [MATH] of the coordinates. For a particular function [MATH] , with high probability, at least [MATH] , the block length decreases under a random restriction by a factor of [MATH] This is due to the fact that on expectation at most [MATH] of the variables wil...
We run into trouble if at some iteration we have more than [MATH] functions of block-length [MATH] The first observation is that in this case the total variance of the functions is extremely high, exponential in [MATH] . Recall that the expected value of the product is exponentially small in the total variance. This me...
The nice thing about these “aggressive pseudorandom restrictions” is that they keep variables alive with such small probability that with high probability each function [MATH] will depend on at most [MATH] variables after the restriction, except for a small number of functions covering at most [MATH] “bad variables”. T...
6.1 PRG for the XOR of many functions with block-length [MATH] Let [MATH] be the class of functions of the form [MATH] where [MATH] are Boolean functions on disjoint sets of variables, [MATH] (the ‘junta’) depends on at most [MATH] variables, [MATH] are non-constant and depend on at most [MATH] variables and
[MATH] Lemma 6.2 There exists a constant [MATH] such that the following holds. For all [MATH] such that [MATH] , there exists a log-space explicit pseudorandom generator [MATH] that [MATH] -fools [MATH]
Algorithm 1 The Pseudorandom Generator [MATH] 1: A block-length [MATH] , the output length [MATH] , a junta-size [MATH] , an error parameter [MATH]
2: Set [MATH] 3: Pick [MATH] using a [MATH] -biased distribution with marginals [MATH] 4: Assign coordinates of [MATH] in [MATH] using a [MATH] -biased distribution.
5: Assign coordinates of [MATH] in [MATH] using Viola’s generator with error [MATH] and degree [MATH] 6: return [MATH] Lemma 6.3
Let [MATH] be a sufficiently large constant. Let [MATH] be some integer. Let [MATH] be non-constant Boolean functions that depend on disjoint sets of at most [MATH] variables each. Assume [MATH] Suppose [MATH] is [MATH] -biased distribution with marginals [MATH] Suppose [MATH] is sampled from a [MATH] -biased distribut...
Proof. Without loss of generality [MATH] Let [MATH] , and note that [MATH] since [MATH] for a sufficiently large constant [MATH]
Let [MATH] be the disjoint sets of variables on which [MATH] depend respectively. For any function [MATH] , there exists a sensitive pair of inputs [MATH] such that [MATH] and [MATH] differ in exactly one coordinate [MATH] and such that [MATH] We say that the sensitive pair “survives” the random restriction defined by ...
[EQUATION] We upper bound the first and second summands separately. By Lemma 3.8 and Lemma 3.12 , the first summand is upper bounded by [MATH] where
[EQUATION] Since [MATH] , the first summand is upper bounded by [MATH] Next, we upper bound the second summand. By Vazirani’s XOR lemma, since [MATH] is [MATH] -biased, we have that the marginal distribution of any set of at most [MATH] bits in [MATH] is [MATH] -close to uniform in statistical distance. Since [MATH] is...
[MATH] Combining the bounds on both summands we get [EQUATION] Using [MATH] we get [EQUATION] Using [MATH] and [MATH] we get [MATH] In the complement event, at least [MATH] of the functions [MATH] are non-constant.
Lemma 6.4 Let [MATH] Suppose [MATH] where [MATH] is a [MATH] -junta and [MATH] is a polynomial of degree- [MATH] over [MATH] If [MATH] fools degree- [MATH] polynomials over [MATH] with error [MATH] , then [MATH] fools [MATH] with error [MATH]
Proof. Let [MATH] be the set of variables on which [MATH] depends. Using the Fourier transform of [MATH] [MATH] we write [MATH] as
[MATH] Note that [MATH] is a polynomial of degree- [MATH] over [MATH] as well, thus we get [EQUATION] Proof of Lemma 6.2 First note that [MATH] has very small expectation under the uniform distribution
[EQUATION] using the assumption [MATH] Thus, we need to maintain low-expectancy under the pseudorandom assignment. By Lemma 6.3 , with probability at least [MATH] after the aggressive random restriction at least [MATH] of the functions [MATH] remain non-constant. Since [MATH] we maintained the low-expectancy under aggr...
Furthermore, we wish to show that with high probability, except for a set of at most [MATH] “bad variables” all functions have block-length at most [MATH] Recall that there are at most [MATH] functions. The probability that any particular [MATH] variables survive is at most [MATH] Pick [MATH] The probability that at le...
[EQUATION] If [MATH] , then this probability is at most [MATH] This means that, with high probability, there are less than [MATH] variables from all functions with more than [MATH] effective variables remaining. Otherwise, there would have been [MATH] functions accountable to a total number of more than [MATH] variable...
Overall, with probability at least [MATH] we are left with the XOR of a small-junta, on at most [MATH] variables, and an XOR of at least [MATH] non-constant functions on at most [MATH] variables (i.e., a degree [MATH] polynomial). Moreover, the restricted function has expected value at most [MATH] in absolute value und...
6.2 A thought experiment We are ready to describe the pseudo-random restriction process in full detail. We start by describing a process that iteratively “looks” at the restricted functions in order to decide which pseudorandom restriction to apply next: the one described in Lemma 6.1 or the one from Lemma 6.3 This ult...
We start with [MATH] blocks of length [MATH] We assume that [MATH] (if not set [MATH] ). Algorithm 2 an “adaptive pseudorandom generator”
1: for [MATH] do 2: Let [MATH] 3: if [MATH] then apply CHRT’s PRG on the remaining coordinates, and Halt! 4: if more than [MATH] of the restricted functions are non-constant and depend on at most [MATH] variables then
apply [MATH] from Lemma 6.2 on the remaining variables, and Halt! 5: else apply the pseudorandom restriction from Lemma 6.1 on the remaining variables.
Next, we show that the process yields a pseudorandom string fooling [MATH] First, note that the process either stops at Step 3 or at Step 4. In both cases we assign all the variables according to some pseudorandom generator, hence all the variables will be assigned by the end of the process.
For [MATH] Let [MATH] be the set of coordinates that remain alive at the beginning of the [MATH] -th iteration. Denote by [MATH] the [MATH] -th function under the restriction at the beginning of the [MATH] -th iteration. Define [MATH] to be the set of variables that affect the output of [MATH] For example if [MATH] is ...
Let [MATH] be the set of functions that depend on some but not more than [MATH] variables, [MATH] be the set of functions that depend on more than [MATH] variables and [MATH]
Claim 6.5 Let [MATH] Suppose [MATH] and [MATH] Then, with probability at least [MATH] we have [MATH] Proof. Under the assumptions we reach Step 5 in Algorithm . We show that:
1. With probability at least [MATH] , at most [MATH] of the variables in [MATH] remain alive in Step 5. 2. With probability at least [MATH] , at most [MATH] new variables are added to [MATH]
Both claims rely on the fact that any set of [MATH] variables remain alive under the pseudorandom restriction in Lemma 6.1 with probability at most [MATH]