text
stringlengths
128
2.05k
This first item follows since the probability that more than [MATH] variables in [MATH] survive is at most [EQUATION] As for the second item, we start with the case where [MATH] Assume that more than [MATH] new variables were added to [MATH] This implies that there is a set of [MATH] good functions in step [MATH] that ...
[EQUATION] (where we used [MATH] which finishes the case [MATH] In the case where [MATH] , we show that with high probability all good functions remain good. For each individual function, using Markov’s inequality
[EQUATION] Thus, we can apply a union bound and show that all good functions remain good with probability at least [MATH] Say the process finished. We shall assume that [MATH] for every iteration [MATH] until the process stopped. By Claim 6.5 this happens with probability at least [MATH] by applying a union bound on th...
1. We stopped on Step 3 at some iteration [MATH] . If [MATH] then [MATH] and at most [MATH] variables remain that affect the functions [MATH] Otherwise, since [MATH] and [MATH] , at most [MATH] variables remain that affect the functions [MATH] , and thus at most [MATH] variables remain that affect the functions [MATH] ...
2. We stopped at Step 4 at some iteration [MATH] Certainly, [MATH] Thus, we are in the case that was handled in Section 6.1 , with [MATH] . Indeed, Lemma 6.2 guarantees that [MATH] fools the remaining function with error at most [MATH] using [MATH] random bits.
6.3 The actual generator Algorithm described the pseudo-random generator as if we knew whether or not the condition in step 3 holds. However, a pseudorandom generator cannot depend on the function it tries to fool. To overcome this issue, we use the following general observation regarding pseudorandom generators.
Claim 6.6 Say there are two families of functions [MATH] and [MATH] that are both closed under shifts (i.e., closed under XORing a constant string to the input). Say that [MATH] is an [MATH] -PRG for [MATH] and [MATH] is an [MATH] -PRG for [MATH] then [MATH] is an [MATH] -PRG for [MATH]
Proof. Let [MATH] , we show that [MATH] fools [MATH] By symmetry assume [MATH] [EQUATION] where [MATH] . Since [MATH] is closed under shifts, we have that [MATH] thus [MATH]
[MATH] -fools [MATH] and we get [MATH] The actual generator would proceed as follows. Algorithm 3 The Pseudorandom Generator [MATH]
1: a set [MATH] of the “live” coordinates, a width [MATH] , an integer [MATH] , a parameter [MATH] 2: if [MATH] then return [MATH] for [MATH]
3: Let [MATH] for [MATH] 4: Pick [MATH] [MATH] according to Claim 6.1 5: Let [MATH] 6: return [MATH] Claim 6.7 (Proof of Correctness)
Let [MATH] Suppose [MATH] are functions on disjoint sets of [MATH] Suppose each function depends on at most [MATH] variables except for a total of at most [MATH] variables, and the number of non-constant functions is at most [MATH] Then, [MATH] fools [MATH] with error [MATH]
Proof. We prove the claim by induction on [MATH] If [MATH] then Theorem 3.11 implies correctness. If [MATH] then we consider the following two cases:
1. If there are more than [MATH] good functions, then [MATH] fools [MATH] with error [MATH] 2. Otherwise, there are at most [MATH] good functions and we apply Step 3. According to Claim 6.1 , the average acceptance probability of [MATH] is [MATH] close to that of [MATH] Furthermore, with probability at least [MATH] all...
Since we have a pseudorandom generator fooling the function in each case, Claim 6.6 shows that [MATH] fools [MATH] with error [MATH]
Claim 6.8 (Seed Length) The amount of random bits used to calculate [MATH] is at most [MATH] Proof. Unwrapping the recursive calls in the evaluation of [MATH] we see that there are at most [MATH] recursive calls to the procedure and that the error parameters are at least [MATH] in all of them.
We apply the generator from Theorem 3.11 only once during these recursive calls, on a ROBP of width- [MATH] and length [MATH] . Thus, the application of Theorem 3.11 uses at most [MATH] random bits.
The partial assignment from Claim 6.1 uses at most [MATH] each time we invoke it, and we invoke it at most [MATH] times. The generator [MATH] uses [MATH] random bits each time we invoke it, and we invoke it at most [MATH] times.
Claims 6.7 and 6.8 complete the proof Theorem with [MATH] as the generator. 6.4 Pseudorandom generator for read-once polynomials
Next, we restate and prove Theorem See Proof. We show that [MATH] fools any read-once polynomial with error at most [MATH] . Its seed length is [MATH]
A read-once polynomial can be written as the XOR of AND functions on disjoint variables, i.e., as the XOR of width- [MATH] ROBPs on disjoint variables. It remains to show that these ROBPs are short. Rather, we show that any PRG that [MATH] -fools read-once polynomials of degree at most [MATH] also [MATH] -fools all rea...
[EQUATION] be a read-once polynomial over [MATH] , where [MATH] are disjoint subsets of [MATH] Without loss of generality let [MATH] be the blocks of length bigger than [MATH] Let
[EQUATION] be the sum over monomials of degree at most [MATH] of [MATH] Let [MATH] . By triangle inequality [EQUATION] For [MATH] , since [MATH] , we have
[MATH] As for the distribution [MATH] , by monotonicity [EQUATION] where [MATH] is any arbitrary subset of exactly [MATH] variables from [MATH] . Since [MATH] fools degree- [MATH] read-once polynomials with error at most [MATH] , and [MATH] is such a polynomial, we get that [MATH] is at most [MATH] Plugging both bounds...
Pseudorandom generators for width-3 ROBPs In this section, we construct pseudorandom generators fooling width-3 ROBPs ( 3ROBPs , in short) with seed-length [MATH] . For ordered width-3 ROBPs we can guarantee error [MATH] using seed-length [MATH] See
Note that in comparison, even for constant [MATH] , the best previous generators had seed-length [MATH] for ordered 3ROBPs. We also get similar improvements for unordered 3ROBPs but with worse dependence on the error [MATH] See
7.1 Proof overview We heavily rely on the pseudorandom restriction from Theorem that assigns [MATH] of the variables while changing the acceptance probability by at most [MATH] . As a first step we assign a constant fraction of the coordinates.
Assigning most of the coordinates. The first step is rather simple: we apply iteratively [MATH] times the pseudorandom restriction from Theorem to get the following analog result to Claim 6.1 The proof is the same as that of Claim 6.1 and is omitted.
Claim 7.1 Let [MATH] . For all constants [MATH] , there is a pseudorandom restriction [MATH] using [MATH] random bits, changing the acceptance probability of 3ROBPs by at most [MATH] Furthermore, [MATH] is [MATH] biased with marginals [MATH] and
[MATH] is [MATH] biased. Let [MATH] be a 3ROBP of length- [MATH] . First, we claim that after applying the pseudorandom restriction [MATH] in Claim 7.1 , with high probability (at least [MATH] ), [MATH] has a simpler structure in that between any two width- [MATH] layers the subprogram has at most [MATH]
colliding layers . Concretely, we use the following definitions. Definition 7.2 Given a ROBP [MATH] , we call a layer of edges colliding if either the edges marked by [MATH] and the edges marked by [MATH] collide.
Definition 7.3 We call a ROBP [MATH] [MATH] -ROBP if [MATH] can be written as [MATH] , with each [MATH] being a width [MATH] ROBP with the first and last layers having at most two vertices and each [MATH] having at most [MATH] colliding layers.
We show that after applying the pseudorandom restriction [MATH] in Claim 7.1 , with high probability the restricting ROBP [MATH] is a [MATH] -ROBPs. Now, similar to Section , we wish to iteratively apply Claim 7.1 , making the ROBP simpler in each step. We will have one progress measures on the restricted ROBP: the max...
7.2 Reducing the length of [MATH] -ROBPs Here, we show that [MATH] -ROBPs can be approximated by [MATH] -ROBPs for some constant [MATH] A crucial point in the analysis is that we need the approximation to hold not just under the uniform distribution but also under the pseudo-random distribution. Fortunately, we are abl...
Lemma 7.4 (Main Structural Result) For any [MATH] the following holds. Any [MATH] -ROBP [MATH] can be written as [MATH] where [MATH] is a [MATH] -ROBP and either [MATH] or for any [MATH] [MATH] where [MATH] are non-zero events that can be computed by [MATH] -ROBPs on disjoint variables.
We shall also show (in the next claim) that any non-zero event [MATH] that can be computed by [MATH] -ROBP, happens with probability at least [MATH] under the uniform distribution. Thus, [MATH] which is doubly-exponentially small in [MATH] provided that [MATH] is a large enough constant.
For any vertex [MATH] in a ROBP, we denote by [MATH] the probability to reach [MATH] under a uniform random assignment to the inputs.
Claim 7.5 In a ROBP with width [MATH] and at most [MATH] colliding layers, every vertex whose [MATH] has [MATH] We remark that this bound is sharp.
Proof. We prove by induction (on the length of the program) that any program with width at most [MATH] , exactly [MATH] colliding layers and exactly [MATH] reachable states in the last layer, has [MATH] for any reachable vertex [MATH] Without loss of generality all nodes in the program are reachable (otherwise, we remo...
Consider a program [MATH] of length [MATH] with parameters [MATH] Removing the last layer gives a program [MATH] of length [MATH] with parameters [MATH] . By the induction hypothesis for any [MATH] in the last layer of [MATH] we have [MATH] for [MATH]
We perform a case analysis. The following simple bound will be used in all cases. Let [MATH] be a vertex in the last layer of [MATH] . Assume that [MATH] edges enter [MATH] from vertices in the second to last layer. Then, [MATH] . In particular, since we assumed all vertices are reachable, any vertex in the last layer ...
If [MATH] and [MATH] , then the last layer of edges in [MATH] is regular, i.e., any node in the last layer in [MATH] has exactly two ingoing edges. In this case any vertex [MATH] in the last layer has [MATH]
If [MATH] , then [MATH] , since there are no collisions in the last layer of edges. Since we already handled the case [MATH] , we may assume [MATH] For any vertex [MATH] in the last layer we have
[MATH] If [MATH] , then we consider two sub-cases: if [MATH] then only one vertex is reachable in the last layer and its [MATH] equals [MATH] Otherwise, [MATH] and [MATH] thus [MATH] and for any vertex [MATH] in the last layer we have [MATH]
We say that two vertices [MATH] and [MATH] in a ROBP are locally-equivalent if the [MATH] -edges exiting [MATH] and [MATH] reach the same vertex and the [MATH] -edges exiting [MATH] and [MATH] reach the same vertex. We say that a ROBP has no-redundant vertices if any vertex in the program is reachable, and there are no...
Claim 7.6 (Colliding Layers [MATH] Colliding) Let [MATH] be a 3ROBP with width-2 at the start and finish, at least one colliding layer and no-redundant vertices. Let [MATH] and [MATH] be the two start nodes. Then, there exists a string on which the two paths from [MATH] and [MATH] collide.
Proof. First consider the case that [MATH] has width [MATH] . Then, there exists a layer [MATH] and a value [MATH] such that the two edges marked by [MATH] in the [MATH] -th layer collide. Any string whose [MATH] -th bit equals [MATH] results in colliding paths.
For the rest of the proof assume that [MATH] has a layer with width [MATH] . Let [MATH] be the layers of vertices in [MATH] Let [MATH] denote the index of the last layer in [MATH] with width [MATH] Since [MATH] has width-2 at the end, [MATH]
There are six edges between [MATH] and [MATH] : three edges marked with [MATH] and three edges marked with [MATH] Since [MATH] , by the Pigeon-hole principle, there are two edges marked with [MATH] going to some vertex [MATH] , and two edges marked with [MATH] going to some vertex [MATH] [MATH] is not necessarily diffe...
and the two edges from [MATH] marked with [MATH] go to [MATH] Since [MATH] is reachable, there is an input [MATH] that leads from [MATH] or [MATH] to [MATH] Without loss of generality, we assume that [MATH] is reachable from [MATH] Let [MATH] be the vertex reached by following the same input [MATH] starting from the ot...
Claim 7.7 (“First Collisions” can be detected by 3ROBPs) Let [MATH] be a 3ROBP with 2 vertices at the first layer, denoted [MATH] Suppose there are at most [MATH] colliding layers in [MATH] and that there exists a string on which the two paths from [MATH] and [MATH] collide. Let [MATH] be the first vertex on which a co...
Proof. To simulate whether the paths starting from [MATH] and [MATH] collide at [MATH] , we consider the 3ROBP that keeps the unordered pair corresponding to the states of the two paths during the computation. In each layer until [MATH] , we have only states corresponding to [MATH] or [MATH] When we reach the layer of ...
We are now ready to prove the main structural lemma – Lemma 7.4 . In the following, we consider branching programs with two initial nodes [MATH] . We interpret the value of the program on input [MATH] as its average value on the two paths starting from [MATH] and [MATH] . That is, the program can get value [MATH] or [M...
Throughout this section we think of the error terms as [MATH] -indicators (instead of the usual [MATH] -notation for other Boolean functions). We shall use [MATH] and [MATH] to denote the standard AND and negation of these Boolean values.
Lemma 7.8 Let [MATH] be a ROBP where each [MATH] is a width- [MATH] ROBP with at most [MATH] vertices on the first and last layers. Then, for any [MATH] we can write [MATH] as the sum of [MATH] and an error term [MATH] , that is bounded in absolute value by [MATH] where [MATH] denotes the event that the two paths in [M...
Proof. Assume without loss of generality that no layer of vertices has width- [MATH] except for maybe the first. For [MATH] , let [MATH] and [MATH] be the two nodes at the first layer of the subprogram [MATH] If [MATH] has two nodes at the first layer, then denote them by [MATH] and [MATH] , otherwise denote the single...
Proof of Lemma 7.4 Let [MATH] be a [MATH] -ROBP [MATH] If [MATH] has no colliding layers, then there is nothing to prove since [MATH] itself is a [MATH] -ROBP. If [MATH] has colliding layers, then without loss of generality each [MATH] has at least one colliding layer (since otherwise we can merge subprograms with no c...
7.3 PRGs for ROBPs with few colliding layers In this section we show that we can [MATH] -fool ordered ROBPs with at most [MATH] -colliding layers with [MATH] seed-length.
Theorem 7.9 For any [MATH] , there is a log-space explicit PRG that [MATH] -fools ordered width [MATH] -ROBPs with length [MATH] and at most [MATH] colliding layers using seed length
[EQUATION] The above relies on the PRGs for regular branching programs and generalizations of them due to Braverman, Rao, Raz, and Yehudayoff BRRY14 In the following, we say that a read-once branching program [MATH] is [MATH] -reachable if for all reachable vertices [MATH] in [MATH] we have [MATH] , where
[EQUATION] We start by quoting a result by Braverman, Rao, Raz, Yehudayoff BRRY14 Theorem 7.10 BRRY14 There is a log-space explicit PRG that [MATH] -fools all
[MATH] -reachable ROBPs of length- [MATH] and width- [MATH] using seed length [EQUATION] Next, we reduce the task of fooling ROBPs with at most [MATH] -colliding layers to the task of fooling [MATH] -reachable ROBPs. The reduction is similar to that in CHRT17 . The main difference is that we simulate a ROBP with width ...
Lemma 7.11 Let [MATH] Let [MATH] be a distribution on [MATH] that [MATH] -fools all [MATH] -reachable ROBPs of length [MATH] and width [MATH] Then, [MATH] also fools width- [MATH] ROBPs with at most [MATH] colliding layers with error at most [MATH]
Proof. Let [MATH] be a distribution on [MATH] that [MATH] -fools all [MATH] -reachable ROBPs of length- [MATH] and width- [MATH] The first observation is that [MATH] also fools prefixes of these programs. This reason is simple: to simulate the prefix of length- [MATH] of a [MATH] -reachable ROBP [MATH] , one can just r...
Let [MATH] be a length [MATH] width- [MATH] ROBP with at most [MATH] colliding layers. Next, we introduce [MATH] , a [MATH] -reachable ROBP of length- [MATH] and width- [MATH] , that would help bound the difference between
[EQUATION] where [MATH] is the uniform distribution over [MATH] Let [MATH] be the the following modified version of [MATH] To construct [MATH] we consider a sequence of [MATH] branching programs [MATH] where [MATH] and [MATH] Let [MATH] be the colliding layers in [MATH] For [MATH] we take [MATH] to be [MATH] except we ...
First, we claim that any reachable vertex [MATH] in [MATH] has [MATH] Let [MATH] for convenience. We apply induction and show that for [MATH] any vertex reachable by [MATH] in layers [MATH] has [MATH] The base case holds because up to layer [MATH] the branching program has no colliding layers and we may apply Claim 7.5...
Next, we bound [MATH] by using the triangle inequality [EQUATION] and bounding each of the three terms separately. 1. The first term is bounded by the probability of reaching one of the nodes in [MATH] in [MATH] when taking a uniform random walk. This follows since if the path defined by [MATH] didn’t pass through [MAT...
2. The second term is at most [MATH] since the program [MATH] is [MATH] -reachable. 3. Similarly to the first term, the third term is bounded by the probability of reaching one of the nodes in [MATH] in [MATH] when taking a walk sampled by [MATH]
[EQUATION] However since [MATH] is pseudorandom for prefixes of [MATH] , for each [MATH] the probability of reaching [MATH] when walking according to [MATH] is [MATH] -close to the probability of reaching [MATH] when walking according to [MATH]
[EQUATION] Summing the upper bound on the three terms in Eq. ( 12 ) gives: [EQUATION] Proof of Theorem 7.9 Take [MATH] and [MATH] Take the generator from Theorem 7.10 with parameters [MATH] and [MATH] Applying Lemma 7.11 , the error of this generator on the class of ROBPs with width [MATH] length [MATH] and at most [MA...
[MATH] By Theorem 7.10 , its seed length is [EQUATION] which is at most [MATH] 7.4 Proof of Theorem We are now ready to prove our main result on fooling 3ROBPs. Our generator is obtained by applying Claim 7.1 iteratively [MATH] times and then using a PRG fooling 3ROBPs with at most [MATH] colliding layers as in Theorem...
Let [MATH] be a 3ROBP and let [MATH] be a pseudorandom restriction as in Claim 7.1 . We first show that with probability at least [MATH] over [MATH] [MATH] is a [MATH] -ROBP for [MATH] . Let [MATH] where each [MATH] has at most [MATH] colliding layers and begins and ends with width two layers. Let [MATH] be an independ...
To carry the induction forward as outlined above, we need the following lemma that shows that the error terms simplify as well under the pseudorandom restrictions.
Lemma 7.12 For any constant [MATH] , there exists [MATH] such that the following holds. Let [MATH] be sufficiently large and [MATH] . Let [MATH] where [MATH] are non-zero events on disjoint variables computed by [MATH] -ROBPs. Let [MATH] be a pseudorandom restriction as in Claim 7.1 with parameter [MATH] and error para...
Proof. First, we show that with high probability, each [MATH] has at most [MATH] colliding layers under the pseudo-random restriction. To see it, note that any colliding layer that is restricted can be either:
Assigned to a value that reduces the width of the original program to [MATH] , and thus the width of [MATH] to 1, in which case any previous layer in [MATH] is not affecting its value.
Assigned to a value that applies a permutation on the states of the program, thus reducing the number of colliding layers. In either case, if [MATH] colliding layers are unassigned, then [MATH] can be a computed by a 3ROBP with at most [MATH] colliding layers. By Claim 7.1 the probability that less than [MATH] collidin...
We move to show that with high probability at least [MATH] of the functions [MATH] are non-zero. We apply the second moment method. Denote by [MATH] for [MATH] Let [MATH] be the events that [MATH] respectively, where [MATH] is the pseudo-random restriction from Claim 7.1 and [MATH] is uniformly distributed. By Claim 7....
[EQUATION] and by the next lemma, whose proof is deferred to Appendix A.3 , we get [EQUATION] Lemma 7.13 Let [MATH] be 3ROBPs on disjoint sets of variables of [MATH] Let [MATH] be any Boolean function. Then, [MATH] is [MATH] -fooled by the pseudorandom restriction in Claim 7.1
Thus, the covariance of the two events [MATH] and [MATH] is at most [MATH] . Denote by [MATH] By Claim 7.5 we have that [MATH] (since [MATH] [MATH] and [MATH] is sufficiently large). Let [MATH] Then, [MATH] and by Chebyshev’s inequality
[EQUATION] We bound [EQUATION] which gives [MATH] using [MATH] In the complement event, at least [MATH] of the events [MATH] occur, and in particular at least [MATH] of the restricted functions [MATH] are non-zero.
Suppose that at least [MATH] of the restricted functions [MATH] are non-zero, and that all restricted functions has at most [MATH] colliding layers. By the above analysis this happens with probability at least [MATH] . Under this assumption, we can reduce the number of functions to be exactly [MATH] , resulting in an u...
We are now ready to prove the main theorem, Theorem Proof of Theorem Let [MATH] Let [MATH] be a constant to be chosen later. Let [MATH] . Let [MATH] be a parameter to be chosen later and let [MATH] for [MATH]
Our generator is as follows. First choose [MATH] independent pseudo-random restrictions as in Claim 7.1 with parameter [MATH] and [MATH] . After iteratively applying the restrictions [MATH] , we set the remaining bits using the generator from Theorem 7.9 for a parameter [MATH] and error parameter [MATH] to be chosen la...
Let [MATH] . We first claim that [MATH] is a [MATH] -ROBP with high probability. In the following let [MATH] be uniformly random over [MATH]
Claim 7.14 With probability at least [MATH] [MATH] is a [MATH] -ROBP and [MATH] For [MATH] , let [MATH] . We will show the following claim by induction on [MATH]
Claim 7.15 For [MATH] , with probability at least [MATH] [MATH] can be written as [MATH] where [MATH] is a [MATH] -ROBP and the error terms [MATH] for [MATH] satisfy: Either [MATH] or [MATH] with [MATH] where [MATH] are non-zero events computed by [MATH] -ROBPs on disjoint sets of variables and [MATH]
Furthermore, [MATH] A crucial point in the above is that the functions [MATH] bounding the error terms are conjunctions of negations of [MATH] -ROBPs and there exactly [MATH] in each of them.
Proof. For [MATH] , the claim follows immediately by applying Lemma 7.4 to [MATH] . Now, suppose the claim is true for [MATH] . Suppose, we can write [MATH] , where [MATH] as in the claim. By the induction hypothesis, this happens with probability at least [MATH]
Clearly, [MATH] . Let [MATH] be a decomposition where each [MATH] has at most [MATH] colliding layers, starts and ends with width- [MATH] layers and [MATH]
Now, observe that as each [MATH] has at most [MATH] colliding layers, the probability that at least [MATH] of these colliding layers are unfixed under [MATH] is at most [MATH] by Claim 7.1 . Thus, by a union bound over [MATH] , with probability at least [MATH] (for a suitable choice of [MATH] ), over [MATH] [MATH] is a...
[EQUATION] where [MATH] , and [MATH] satisfy the conditions of the claim. We just need to argue that [MATH] can be written in the requisite form. To this end, note that for [MATH] [MATH] . By the induction hypothesis, we either have [MATH] or we can write [MATH] where [MATH] are [MATH] -ROBPs on disjoint sets of variab...
Adding up the failure probabilities over the choice of [MATH] , we get the desired decomposition for [MATH] with probability at least
[EQUATION] (since [MATH] ). The furthermore part follows immediately from Claim 7.1 . The claim now follows by induction. We are now ready to prove the theorem. By the above claim, we have that with probability at least [MATH] over the choice of [MATH] , we can write
[EQUATION] where [MATH] is a [MATH] -ROBP and [MATH] can be bounded by functions [MATH] that are conjunctions of negations of [MATH] non-zero events computed by [MATH] -ROBPs.
Note that each such [MATH] can be written as a width- [MATH] ROBP, say [MATH] , by adding an additional layer to compute the conjunction and that the number of collisions in the width [MATH] ROBP is at most [MATH] . Therefore, if we let [MATH] be the output distribution of the generator from Theorem 7.9 with [MATH] and...
[EQUATION] where we used Claim 7.5 to bound [MATH] . Since [MATH] [MATH] Combining the above inequalities we get that with probability at least [MATH] over the choice of [MATH]
[EQUATION] Finally, as we also have that [EQUATION] we get [EQUATION] To get [MATH] we set [MATH] . Furthermore, setting [MATH] and [MATH] , the above error bound becomes
[EQUATION] Finally, we estimate the seed-length of our generator. Choosing the random restrictions takes [MATH] random bits. Sampling [MATH] requires seed-length