[BCIKS20] Proximity Gaps 论文 soundness 解析#

论文 [BCIKS20] 对 [BBHR18] 中的 FRI 协议的 soundness 进行了改进,主要分析了 batched FRI 的情况。本文将详细解析 [BCIKS20] 论文中关于 batched FRI soundness 的内容。

Introduction#

在交互证明,分布式存储以及密码学等背景下,出现了各种协议,这些协议引出了关于一个线性编码 \(V \subset \mathbb{F}_q^n\) 的 proximity 问题,其中 \(\mathbb{F}_q\) 是有限域,\(V\) 的最小的相对距离为 \(\delta_V\) 。这些协议假设我们可以获得关于一批向量 \(\textbf{u} = \{u_0, \cdots, u_l \} \subset \mathbb{F}_q^n\) 的 oracle,它们的 soundness 要求每个向量 \(u_i\) 在相对 Hamming 距离上接近 \(V\) 。另外,soundness 是某些向量到 code \(V\) 之间的最大距离的一个函数,如果这个距离变大,那么 Verifier 拒绝的概率会降低。因此,我们想找到这样的协议,能够最小化对 \(\mathbf{u}\) 中元素的 query 的数量,同时最大化能识别某个向量 \(u_i\) 明显远离 \(V\) 的概率。

疑问

  • 如何理解下面这句话? Furthermore, soundness deteriorates as a function of the largest distance between some vector ui and the code V . soundeness 会随着某些向量 \(u_i\) 与 code \(V\) 之间的最大距离的增加而降低,也就是此时 Verifier 拒绝的概率就降低了?

  • 如何对 soundness 降低进行清楚地讲解呢?

由于 \(V\) 的线性性,一个自然的方法([RVW13])是:在 span(\(\textbf{u}\))(即 \(\textbf{u}\) 中各元素的线性组合) 中均匀地随机一个向量 \(u'\) ,记 \(u'\)\(V\) 之间的距离为 \(\Delta(u', V)\) ,将这个距离视为 \(\textbf{u}\) 中某些元素与 \(V\) 之间的最大距离的一个 proxy(代理)。为了证明 soundness ,我们想要即使只有一个 \(u_i\) 距离 \(V\) 中的所有元素有 \(\delta\)-远,那么随机选择的 \(u'\) 也距离 \(V\) 很远。

在下文中,用 \(\Delta\) 表示相对 Hamming 距离。当 \(\Delta(u, v) \le \delta\) 对某个 \(v \in V\) 成立时,称为 “ \(u\)\(V\) 的距离有 \(\delta\) -近”,记作 \(\Delta(u, V) \le \delta\) ;否则称为 “ \(u\)\(V\) 的距离有 \(\delta\) -远”,记作 \(\Delta(u, V) > \delta\)

关于这个问题,一些研究结果为:

  1. [AHIV17] 如果 \(\delta < \delta_V /4\) ,几乎所有的 \(u' \in \text{span}(\mathbf{u})\) 距离 \(V\)\(\delta\) -远。

  2. [RZ18] 将上述结果提高到 \(\delta < \delta_V /3\)

  3. [BKS18] 提高到 \(\delta < 1 - \sqrt[4]{1 - \delta_V}\)

  4. [BGKS20] 提高到 \(\delta < 1 - \sqrt[3]{1 - \delta_V}\) ,但这个界对 RS 编码是 tight 的,因为当 \(n = q\) 时可以达到这个界。

🤔 思考

  • 为什么研究的重心都想要提高这个 \(\delta\) 的上界呢?关于这个问题,目前我的想法是: \(\delta\) 的上界这里是和 \(\delta_V\) 有关的,对于 RS code ,\(\delta_V = 1 - \rho\) ,实质也就是和码率相关,那么提高上界也就意味着降低了码率,那么就意味着更多的冗余,如果以相同的安全性,或者同样的高概率来拒绝出错的情况,此时需要的query就更少了。或者这样说,如果对于同样一个协议,query 的数量固定,\(\delta\) 越大,拒绝的概率也就越大,也就提高了 soundness。

  • 上面分析的第 2 点与 “Furthermore, soundness deteriorates as a function of the largest distance between some vector \(u_i\) and the code V .” 似乎矛盾,这句话说的是 \(\delta\) 越大,soundness 越小?这一点该如何理解呢?

目前我们关心的一个问题是:对于哪些码以及什么范围的 \(\delta\) ,以下的陈述成立?

如果某个 \(u^* \in \text{span}(\mathbf{u})\)\(V\)\(\delta\) -远,那么对于几乎所有的 \(u' \in \text{span}(\mathbf{u})\)\(u'\)\(V\) 也有 \(\delta\) -远。

[BCIKS20] 论文的主要结论之一表明,当 \(V\) 是一个在足够大的域上的 RS 码(域的大小与码的块长度呈多项式关系)并且 \(\delta\) 小于 Johnson/Guruswami-Sudan list decoding 界时,上述陈述成立。接下来,我们称其为 proximity gap 。

Proximity Gaps#

先给出 Proximity Gaps 的定义。

Definition 1.1 [BCIKS20, Definition 1.1] (Proximity gap). Let \(P \subset \Sigma^n\) be a property and \(C \subset 2^{\Sigma^n}\) be a collection of sets. Let \(\Delta\) be a distance measure on \(\Sigma^n\) . We say that \(C\) displays a \((\delta, \epsilon)\) -proximity gap with respect to \(P\) under \(\Delta\) if every \(S \in C\) satisfies exactly one of the following:

  1. \(\Pr_{s \in S} [\Delta(s, P) \le \delta] = 1\) .

  2. \(\Pr_{s \in S} [\Delta(s, P) \le \delta] \le \epsilon\) .

We call \(\delta\) the proximity parameter and \(\epsilon\) is the error parameter. By default, \(\Delta\) denotes the relative Hamming distance measure.

对于 RS code ,如果 \(V \subset \mathbb{F}^n\) 是 RS 编码 ,对应上述定义中的 \(P\) ,并且 \(A \subset \mathbb{F}^n\) 是一个 affine space (仿射空间) ,对应于上述定义中的 \(S\) ,那么要么 \(A\) 中的所有元素离 \(V\)\(\delta\) -近,要么 \(A\) 中的几乎所有元素离 \(V\)\(\delta\) -远。换句话说,不会有这样的 affine space \(A\) ,其中大概一半的元素离 \(V\) 比较近,但同时另一半元素离 \(V\) 比较远。

如下图所示,\(A\) 是一个 affine space,这里用一条线表示,编码空间 \(V\) 中的元素用黑色的点表示,以这些点为圆心,以 \(\delta\) 为半径画圆。那么只有两种情况:

  1. 线 \(A\) 上的所有元素都落入了绿色的圆形区域内。

  2. 线上只有少量的元素落入了绿色的圆形区域内。 \(A\) 中的元素不可能一半在圆形区域内,一半在圆形区域外,这也是 gap 的含义,将 \(A\) 中的所有元素落入的情况分成了恰好分成了两种情况,而这两种情况之间根据相对 Hamming 距离形成了一个巨大的 gap 。

在下文中,用 \(\mathbb{F}_q\) 表示大小为 \(q\) 的有限域,\(\text{RS}[\mathbb{F}_q,\mathcal{D},k]\) 表示维数为 \(k+1\) ,块长度(blocklength) 为 \(n = |\mathcal{D}|\) 的 RS 编码,其码字是在 \(\mathcal{D}\) 上求值(evaluated),次数 \(\le k\) 的多项式。用 \(\rho\) 表示码率, 则 \(\rho = \frac{k+1}{n}\)\(\delta\) 表示相对于 RS code 的相对 Hamming 距离,\(\epsilon\) 表示 error 参数,也就是一个“坏事件(bad event)”发生的概率。

下面给出 RS code 的 Proximity gaps 定理。

Theorem 1.2 [BCIKS20, Theorem 1.2] (Proximity gap for RS codes). The collection \(C_{\text{Affine}}\) of affine spaces in \(\mathbb{F}_q^n\) displays a \((\delta, \epsilon)\) -proximity gap with respect to the RS code \(V := \text{RS}[\mathbb{F}_q, \mathcal{D}, k]\) of blocklength \(n\) and rate \(\rho = \frac{k+1}{n}\) , for any \(\delta \in (0, 1 - \sqrt{\rho})\) , and \(\epsilon = \epsilon(q, n, \rho, \delta)\) defined as the following piecewise function:

  • Unique decoding bound: For \(\delta \in [0,\frac{1 - \rho}{2})\) , the error parameter \(\epsilon\) is

\[ \epsilon = \epsilon_\text{U} = \epsilon_\text{U}(q, n) := \frac{n}{q} \tag{1.1} \]
  • List decoding bound: For \(\delta \in (\frac{1 - \rho}{2}, 1 - \sqrt{\rho})\) , setting \(\eta := 1 - \sqrt{\rho} - \delta\) , the error parameter \(\epsilon\) is

\[ \epsilon = \epsilon_\text{J} = \epsilon_\text{J}(q, n, \rho, \delta) := \frac{(k+1)^2}{\left(2 \min \left(\eta, \frac{\sqrt{\rho}}{20}\right)^7\right)q} = O \left(\frac{1}{(\eta \rho)^{O(1)}} \cdot \frac{n^2}{q} \right) \tag{1.2} \]

🤔 Question

  • \(\delta\) 越大,落入圆形区域的元素可能更多,也就是 \(\epsilon_{\text{J}}\) 相比 \(\epsilon_{\text{U}}\) 更大一些。是这个原因吗?

Correlated agreements#

论文中证明的主要定理是 correlated agreement ,对于在 \(\mathbb{F}^{\mathcal{D}}\) 中的两个向量 \(u_0, u_1 \in \mathbb{F}^{\mathcal{D}}\) ,在 \(\mathbb{F}\) 中选一个随机数 \(z\) ,我们关心用 \(z\) 进行线性组合后的 \(u_0 + zu_1\) 所形成的空间与 \(V\) 之间的距离,也就是一维的 affine space \(A = \{u_0 + z u_1 : z \in \mathbb{F}\}\) 。correlated agreement 结论说的是如果在 \(A\) 中有足够多的元素距离 RS code 空间 \(V\) 足够近( \(\delta\) - 近),那么一定存在一个非平凡的 subdomain \(\mathcal{D}' \subset \mathcal{D}\) ,其大小至少是 \(\mathcal{D}\) 大小的 \(1 - \delta\) 倍,使得限制 \(u_0, u_1\)\(\mathcal{D}'\) 上,有有效的 RS code \(v_0, v_1\) ,满足它们分别在 \(\mathcal{D}'\) 上与 \(u_0, u_1\) 一致。我们就说这样的 \(\mathcal{D}'\) 有 correlated agreement 性质,即 \(u_0, u_1\)\(A\) 中的元素不仅分别与 RS 码有很大的 agreement ,而且还共享一个共同的很大的 agreement 集合。这个结果有两个参数范围,一个是 unique decoding 范围内的 proximity 参数,另一个是 list decoding 范围内的 proximity 参数。

以下给出了三种情况的 correlated agreement。结合论文中其他关于 correlated agreement 的结论,如下表所示。

空间 \(U\)

\(\Delta_u(u,V)\)

\(\Delta_u(u,V)\) unique decoding

\(\Delta_u(u,V)\) list decoding

\(\text{agree}_{\mu}(u,V)\)

lines

\(\{u_0 + z u_1 : z \in \mathbb{F}\}\)

Theorem 1.4

Theorem 4.1

Theorem 5.1 & Theorem 5.2

low-degree parameterized curves

\(\text{curve}(\mathbf{u}) = \left\{u_z: = \sum_{i = 0}^{l}z^i \cdot u_i | z \in \mathbb{F}_q \right\}\)

Theorem 1.5

Theorem 6.1

Theorem 6.2

Theorem 7.1 & Theorem 7.2(Johnson bound 更精确版本)

affine spaces

\(u_0 + \text{span}\{u_1, \cdots, u_l\}\)

Theorem 1.6

Theorem 7.3 & Theorem 7.4(Johnson bound 更精确版本)

下面三个定理分别对应线、低次参数曲线以及 affine spaces 的 correlated agreement 定理。

Theorem 1.4 [BSCIK20, Theorem 1.4] (Main Theorem - Correlated agreement over lines). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. For \(u_0, u_1 \in \mathbb{F}_q^{\mathcal{D}}\) , if \(\delta \in (0, 1 - \sqrt{\rho})\) and

\[ \Pr_{z \in \mathbb{F}_q} [\Delta(u_0 + z \cdot u_1, V) \le \delta] > \epsilon, \]

where \(\epsilon\) is as defined in Theorem 1.2, then there exist \(D' \subset D\) and \(v_0, v_1 \in V\) satisfying

  • Density: \(|D'|/|D| \ge 1 - \delta\) , and

  • Agreement: \(v_0\) agrees with \(u_0\) and \(v_1\) agrees with \(u_1\) on all of \(D'\) .

\(\mathbf{u} = \{u_0, \ldots, u_l \} \subset \mathbb{F}_q^{\mathcal{D}}\) ,则次数为 \(l\) 的 parameterized curve 是由 \(\mathbf{u}\) 生成的如下的在 \(\mathbb{F}_q^{\mathcal{D}}\) 中的向量的集合,

\[ \text{curve}(\mathbf{u}) := \left\{u_z: = \sum_{i = 0}^{l}z^i \cdot u_i \Bigg| z \in \mathbb{F}_q \right\} \]

Theorem 1.5 [BSCIK20, Theorem 1.5] (Correlated agreement for low-degree parameterized curves). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) . If \(\delta \in (0, 1 - \sqrt{\rho})\) and

\[ \Pr_{u \in \text{curve}(u)} [\Delta(\mathbf{u}, V) \le \delta] > l \cdot \epsilon, \]

where \(\epsilon\) is as defined in Theorem 1.2, then there exist \(\mathcal{D}' \subset \mathcal{D}\) and \(v_0, \cdots, v_l \in V\) satisfying

  • Density: \(|\mathcal{D}'|/|\mathcal{D}| \ge 1 - \delta\) , and

  • Agreement: for all \(i \in \{0, \cdots, l\}\) , the functions \(u_i\) and \(v_i\) agree on all of \(\mathcal{D}'\) .

Theorem 1.6 [BSCIK20, Theorem 1.6] (Correlated agreement over affine spaces). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. For \(u_0, u_1, \cdots, u_l \in \mathbb{F}_q^{\mathcal{D}}\) let \(U = u_0 + \text{span}\{u_1, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) be an affine subspace. If \(\delta \in (0, 1 - \sqrt{\rho})\) and

\[ \Pr_{u \in U} [\Delta(u, V) \le \delta] > \epsilon, \]

where \(\epsilon\) is as defined in Theorem 1.2, then there exist \(\mathcal{D}' \subset \mathcal{D}\) and \(v_0, \cdots, v_l \in V\) satisfying

  • Density: \(|\mathcal{D}'|/|\mathcal{D}| \ge 1 - \delta\) , and

  • Agreement: for all \(i \in \{0, \cdots, l\}\) , the functions \(u_i\) and \(v_i\) agree on all of \(\mathcal{D}'\) .

Furthermore, in the unique decoding regime \(\delta \in \left(0, \frac{1 - \rho}{2}\right]\) , there exists a unique maximal \(\mathcal{D}'\) satisfying the above, with unique \(v_i\) .

Correlated Weighted Agreement#

如果要分析 FRI 协议的 soundness ,就需要 Theorem 1.5 的 weighted 版本。

对于一个给定的 weight 向量 \(\mu : \mathcal{D} \rightarrow [0,1]\)\(u,v\) 之间(相对的) \(\mu\) -agremment 定义为

\[ \text{agree}_{\mu}(u,v) := \frac{1}{|\mathcal{D}|} \sum_{x:u(x) = v(x)} \mu(x). \]

也就是在权重 \(\mu\) 之下看 \(u\)\(v\)\(\mathcal{D}\) 上一致的比例有多少。如果令 \(\mu \equiv 1\) ,那么

\[ \text{agree}_{\mu}(u,v) = \frac{1}{|\mathcal{D}|} \sum_{x:u(x) = v(x)} 1 = 1 - \frac{1}{|\mathcal{D}|} \sum_{x:u(x) \neq v(x)} 1 = 1 - \Delta(u,v). \]

一个 word \(u\) 与一个线性码 \(V\) 之间的 agreement 为 \(u\)\(V\) 之间的一个码字之间的最大的 agreement ,

\[ \text{agree}_{\mu}(u,V) := \max_{v \in V} \text{agree}_{\mu}(u,v). \]

定义 subdomain \(\mathcal{D}' \subset \mathcal{D}\) 的加权(weighted) 大小为

\[ \mu(\mathcal{D}') := \frac{1}{|\mathcal{D}|}\sum_{x \in \mathcal{D}'} \mu(x). \]

将上面定义中的 \(\mathcal{D}'\) 定义为 \(\{x \in \mathcal{D}: u(x) = v(x)\}\) ,则 agreement 满足 \(\text{agree}_{\mu}(u,v) = \mu(\{x \in \mathcal{D}: u(x) = v(x)\})\)

最后,对于 \(\mathbf{u} = \{u_0, \cdots, u_l\}\) ,其中 \(u_i \in \mathbb{F}_q^{\mathcal{D}}\) 是一组 word ,\(\mu\) -weighted correlated agreement 是 subdomain \(\mathcal{D}' \subset \mathcal{D}\) 的最大 \(\mu\) -weighted 大小,使得 \(\mathbf{u}\)\(\mathcal{D}'\) 上的限制属于 \(V|_{\mathcal{D}'}\) ,即对于每个 \(i = 0, \cdots, l\) ,存在 \(v_i \in V\) 使得 \(u_i|_{\mathcal{D}'} = v_i|_{\mathcal{D}'}\) 。当 \(\mu\) 未指定时,它被设置为常数权重函数 1 ,这恢复了在前面讨论的 correlated agreement 度量的概念。

接下来,我们假设权重函数 \(\mu\) 具有某些结构,具体来说,所有权重 \(\mu(x)\) 都是形式 \(\mu(x) = \frac{a_x}{M}\) ,其中 \(a_x\) 是变化的整数,且有共同的分母 \(M\) 。对于 FRI soundness 的特殊情况(其中 \(M\) 等于应用 FRI 协议的 RS 码的 blocklength),这个假设确实成立。以下是定理 1.5 的加权推广。

Theorem 7.1 [BSCIK20, Theorem 7.1] (Weighted correlated agreement over curves – Version I). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) . Let \(\alpha \in (\sqrt{\rho}, 1)\) and let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Suppose

\[ \Pr_{u \in \text{curve}(\mathbf{u})} [\text{agree}_{\mu}(u,V) \ge \alpha] > l \cdot \epsilon, \]

where \(\epsilon\) is as defined in Theorem 1.2 (with \(\eta = \min(\alpha - \sqrt{\rho}, \frac{\sqrt{\rho}}{20})\) ), and additionally suppose

\[ \Pr_{u \in \text{curve}(\mathbf{u})} [\text{agree}_{\mu}(u,V) \ge \alpha] \ge \frac{l(M|\mathcal{D}|+1)}{q} \left(\frac{1}{\eta} + \frac{3}{\sqrt{\rho}}\right) . \]

Then there exists \(\mathcal{D}' \subset \mathcal{D}\) and \(v_0, \cdots, v_l \in V\) satisfying

  • Density: \(\mu(\mathcal{D}') \ge \alpha\) , and

  • Agreement: for all \(i \in \{0, \cdots, l\}\) , the functions \(u_i\) and \(v_i\) agree on all of \(\mathcal{D}'\) .

仅适用于 Johnson 界限范围的一种更精确的形式如下。

Theorem 7.2 [BSCIK20, Theorem 7.2] (Weighted correlated agreement over curves – Version II). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) . Let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Let \(m \ge 3\) and let

\[ \alpha \ge \alpha_0(\rho, m) := \sqrt{\rho} + \frac{\rho}{2m}. \]

Let

\[ S = \{z \in \mathbb{F}_q : \text{agree}_{\mu}(u_0 + zu_1 + \cdots + z^lu_l, V) \ge \alpha\} \]

and suppose

\[ |S| > \max\left(\frac{(1 + \frac{1}{2m})^7 m^7}{3 \rho^{3/2}}n^2 l, \frac{2m + 1}{\sqrt{\rho}}(M \cdot n + 1)l \right) . \tag{7.1} \]

Then \(u_0, \ldots, u_l\) have at least \(\alpha\) correlated \(\mu\) -agreement with \(V\) , i.e. \(\exists v_0, \cdots, v_l \in V\) such that

\[ \mu(\{x \in \mathcal{D}: \forall 0 \le i \le l, u_i(x) = v_i(x)\}) \ge \alpha . \]

Theorem 7.3 [BSCIK20, Theorem 7.3] (Weighted correlated agreement over affine spaces). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) and let \(U = u_0 + \text{span}\{u_1, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) be an affine subspace. Let \(\alpha \in (\sqrt{\rho}, 1)\) and let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Suppose

\[ \Pr_{u \in U} [\text{agree}_{\mu}(u,V) \ge \alpha] > \epsilon, \]

where \(\epsilon\) is as defined in Theorem 1.2 (with \(\eta = \min(\alpha - \sqrt{\rho}, \frac{\sqrt{\rho}}{20})\) ), and additionally suppose

\[ \Pr_{u \in U} [\text{agree}_{\mu}(u,V) \ge \alpha] \ge \frac{M|\mathcal{D}| + 1}{q} \left(\frac{1}{\eta} + \frac{3}{\sqrt{\rho}}\right) . \]

Then there exist \(\mathcal{D}' \subset \mathcal{D}\) and \(v_0, \cdots, v_l \in V\) satisfying

  • \(\mu\)-Density: \(\mu(\mathcal{D}') \ge \alpha\) , and

  • Agreement: for all \(i \in \{0, \cdots, l\}\) , the functions \(u_i\) and \(v_i\) agree on all of \(\mathcal{D}'\) .

同样地,对于 Theorem 7.3 也有关于 Johnson 界限的更精确的形式。

Theorem 7.4 [BSCIK20, Theorem 7.4] (Weighted correlated agreement over affine spaces – Version II). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) and let \(U = u_0 + \text{span}\{u_1, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) be an affine subspace. Let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Let \(m \ge 3\) and let

\[ \alpha \ge \alpha_0(\rho, m) := \sqrt{\rho} + \frac{\sqrt{\rho}}{2m} . \]

Suppose

\[ \Pr_{u \in U} [\text{agree}_{\mu}(u,V) \ge \alpha] > \max \left( \frac{(1 + \frac{1}{2m})^7m^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{M \cdot n + 1}{q} \right) . \tag{7.2} \]

Then \(u_0, \ldots, u_l\) have at least \(\alpha\) correlated \(\mu\) -agreement with \(V\) , i.e. \(\exists v_0, \cdots, v_l \in V\) such that

\[ \mu \left(\{x \in \mathcal{D} : \forall 0 \le i \le l, u_i(x) = v_i(x)\}\right) \ge \alpha . \]

FRI 协议#

FRI 协议的目的是为了在 IOP 模型中,去解决 Reed-Solomon proximity testing 问题,即对于一个接收到的 word \(f^{(0)}: \mathcal{D}^{(0)} \rightarrow \mathbb{F}\) ,验证它到 \(V^{(0)} := \text{RS}[\mathbb{F}, \mathcal{D}^{(0)}, k^{(0)}]\) 之间的 proximity ,如果 \(f^{(0)}\) 属于 \(V^{(0)}\) ,就接受;如果距离 \(V^{(0)}\)\(\delta\) 远,就拒绝。FRI 协议适用于任何 evaluation domain \(\mathcal{D}^{(0)}\)2-smooth 群 的陪集的情况,即对于任何的 \(\mathcal{D}^{(0)}\) ,其是一个大小为 \(2^s\) 的(加法或乘法)群的陪集,其中 \(s\) 是一个整数。因此,为简化描述,假设群 \(\mathcal{D}^{(0)}\) 是乘法的。FRI 协议有两个阶段,分别是 COMMIT 阶段与 QUERY 阶段。

在 COMMIT 阶段,经过有限次 \(r\) 轮的交互,会生成一系列的函数 \(f^{(1)}: \mathcal{D}^{(1)} \rightarrow \mathbb{F}, f^{(2)}: \mathcal{D}^{(2)} \rightarrow \mathbb{F}, \cdots,f^{(r)}: \mathcal{D}^{(r)} \rightarrow \mathbb{F}\) 。每一次迭代, domain 的大小 \(|\mathcal{D}^{(i)}|\) 都会缩小。假设对于一个诚实的 prover, \(f^{(0)}\) 是 low-degree ,那么对于每一个 \(f^{(i)}\) ,它们也都会是 low-degree 的(见命题 1)。在第 \(i\)-轮开始时,prover 的消息 \(f^{(i)}: \mathcal{D}^{(i)} \rightarrow \mathbb{F}\) 已经生成,并且 verifier 可以访问该消息的 oracle 。Verifier 现在发送一个均匀随机的 \(z^{(i)} \in \mathbb{F}\) ,然后 prover 回复一个新的函数 \(f^{(i+1)}: \mathcal{D}^{(i+1)} \rightarrow \mathbb{F}\) ,其中 \(\mathcal{D}^{(i+1)}\)\(\mathcal{D}^{(i)}\) 的一个(2-smooth)strict 子群,意思是 \(\mathcal{D}^{(i)}\) 不仅是 \(\mathcal{D}^{(i+1)}\) 的子群,同时也是其真子集。

\(\mathcal{D}^{(i+1)}\)\(\mathcal{D}^{(i)}\) 划分成大小为 \(l^{(i)} := |\mathcal{D}^{(i)}|/|\mathcal{D}^{(i+1)}|\) 的陪集。令 \(C_g^{(i)}\) 表示对应于 \(g \in \mathcal{D}^{(i+1)}\) 的陪集,即

\[ \begin{equation} C_g^{(i)} := \{g' \in \mathcal{D}^{(i)} \mid (g')^{l^{(i)}} = g\}. \tag{8.1} \end{equation} \]

意思就是在 \(\mathcal{D}^{(i)}\) 中选取那些能够通过映射 \(q(x) = x^{l^{(i)}}\) 映射到 \(\mathcal{D}^{(i+1)}\) 中的 \(g\) 的元素,这些元素组成了集合 \(C_g^{(i)}\) ,其也是陪集。

对于每个陪集 \(C_g^{(i)}\)插值映射(interpolation map) \(M_g^{(i)}\) 是一个可逆的线性映射 \(M_g^{(i)}: \mathbb{F}^{C_g^{(i)}} \rightarrow \mathbb{F}^{l^{(i)}}\) ,它将 \(f^{(i)}|_{C_g^{(i)}}: C_g^{(i)} \rightarrow \mathbb{F}\) (即限制 \(f^{(i)}\) 在 domain \(C_g^{(i)} \subset \mathcal{D}^{(i)}\) 上)映射到多项式 \(P_{\mathbf{u}^{(i)}(g)}^{(i)}(Z) = \sum_{j<l^{(i)}} u_j^{(i)}(g) Z^j\) 的系数向量 \(\mathbf{u}^{(i)}(g) = (u_0^{(i)}(g), \ldots, u_{l^{(i)}-1}^{(i)}(g))^{\intercal}\) (这里与论文原文表示一致,都表示列向量,不过原文没有加上转置符号),其中 \(P_{\mathbf{u}^{(i)}(g)}^{(i)}(Z)\) 是插值 \(f^{(i)}|_{C_g^{(i)}}\) 的多项式。换句话说,\(M_g^{(i)}\) 是由 \(C_g^{(i)}\) 生成的 Vandermonde 矩阵的逆,这意味着 \(\left(M_g^{(i)}\right)^{-1} \cdot {\color{}(u_0, \ldots, u_{l^{(i)}-1})^{\intercal}}\) 是多项式 \(P_{\mathbf{u}}(X) = \sum_{i<l^{(i)}} u_i X^i\) 在陪集 \(C_g^{(i)}\) 上的 evaluation 。

👀 Notice 本文为保持前后一致,用 \((x_0, \ldots, x_n)\)表示行向量,而 \((x_0, \ldots, x_n)^{\intercal}\) 表示列向量,也可写为: $\( \begin{bmatrix} x_0\\ x_1\\ \vdots\\ x_n \end{bmatrix} \)$

下面详细解释下上面对插值映射的描述。根据 \(C_g^{(i)}\) 的定义,知道其中的元素有 \(l^{(i)}\) 个,设 \(C_g^{(i)} =\{g'_1, \cdots, g'_{l^{(i)}}\}\) ,我们可以将由 \(C_g^{(i)}\) 生成的 Vandermonde 矩阵写出来:

\[\begin{split} V_{C_g^{(i)}}= \begin{bmatrix} 1 & g'_1 & (g'_1)^2 & \cdots & (g'_1)^{l^{(i)}-1} \\ 1 & g'_2 & (g'_2)^2 & \cdots & (g'_2)^{l^{(i)}-1} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 1 & g'_{l^{(i)}} & (g'_{l^{(i)}})^2 & \cdots & (g'_{l^{(i)}})^{l^{(i)}-1} \end{bmatrix} \end{split}\]

\(M_g^{(i)} = V_{C_g^{(i)}}^{-1}\) ,是由 \(C_g^{(i)}\) 生成的 Vandermonde 矩阵的逆,因此

\[\begin{split} \begin{aligned} \left(M_g^{(i)}\right)^{-1} \cdot (u_0, \ldots, u_{l^{(i)}-1})^{\intercal} & = \left(V_{C_g^{(i)}}^{-1}\right)^{-1} \cdot (u_0, \ldots, u_{l^{(i)}-1})^{\intercal} \\ & = V_{C_g^{(i)}} \cdot (u_0, \ldots, u_{l^{(i)}-1})^{\intercal} \\ & = \begin{bmatrix} 1 & g'_1 & (g'_1)^2 & \cdots & (g'_1)^{l^{(i)}-1} \\ 1 & g'_2 & (g'_2)^2 & \cdots & (g'_2)^{l^{(i)}-1} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 1 & g'_{l^{(i)}} & (g'_{l^{(i)}})^2 & \cdots & (g'_{l^{(i)}})^{l^{(i)}-1} \end{bmatrix} \begin{bmatrix} u_0\\ u_1\\ \vdots\\ u_{l^{(i)}-1} \end{bmatrix} \\ & = \begin{bmatrix} u_0 + u_1 g'_1 + u_2 (g'_1)^2 + \cdots + u_{l^{(i)}-1} (g'_1)^{l^{(i)}-1}\\ u_0 + u_1 g'_2 + u_2 (g'_2)^2 + \cdots + u_{l^{(i)}-1} (g'_2)^{l^{(i)}-1}\\ \vdots \\ u_0 + u_1 g'_{l^{(i)}} + u_2 (g'_{l^{(i)}})^2 + \cdots + u_{l^{(i)}-1} (g'_{l^{(i)}})^{l^{(i)}-1} \end{bmatrix} \\ & = \begin{bmatrix} P_{\mathbf{u}}(g'_1)\\ P_{\mathbf{u}}(g'_2)\\ \vdots \\ P_{\mathbf{u}}(g'_{l^{(i)}}) \end{bmatrix} \end{aligned} \end{split}\]

通过上述推导看出 \(\left(P_{\mathbf{u}}(g'_1), P_{\mathbf{u}}(g'_2) \cdots, P_{\mathbf{u}}(g'_{l^{(i)}}) \right)^{\intercal}\) 是多项式 \(P_{\mathbf{u}}(X) = \sum_{i<l^{(i)}} u_i X^i\) 在陪集 \(C_g^{(i)}\) 上的 evaluation ,因此 \(\left(M_g^{(i)}\right)^{-1} \cdot (u_0, \ldots, u_{l^{(i)}-1})^{\intercal}\) 是多项式 \(P_{\mathbf{u}}(X) = \sum_{i<l^{(i)}} u_i X^i\) 在陪集 \(C_g^{(i)}\) 上的 evaluation 。

下面的命题使用了上述的符号,重新叙述了 [BBHR18, Section 4.1],与 [BBHR18, Section 4.1] 不同的是,是在乘法群而不是加法群上进行的。该命题描述的就是保持 low-degree 的性质。

Claim 1 [BCIKS20, Claim 8.1]. Suppose that \(f^{(i)} \in \text{RS}[\mathbb{F}, \mathcal{D}^{(i)}, k^{(i)}]\) where \(k^{(i)} + 1\) is an integral power of \(2\) . Then, for any \(z^{(i)} \in \mathbb{F}\), letting \(\mathbf{z}^{(i)} = \left(\left(z^{(i)}\right)^0, \left(z^{(i)}\right)^1, \ldots, \left(z^{(i)}\right)^{l^{(i)}-1}\right)^{\intercal}\), the function \(f_{f^{(i)},z^{(i)}}^{(i+1)}: \mathcal{D}^{(i+1)} \rightarrow \mathbb{F}\) defined on \(g \in \mathcal{D}^{(i+1)}\) by

\[ \begin{equation} f_{f^{(i)},z^{(i)}}^{(i+1)}(g) := {\color{}\left(\mathbf{z}^{(i)}\right)^{\intercal} \cdot \mathbf{u}^{(i)}(g) = \left(\mathbf{z}^{(i)}\right)^{\intercal} \cdot M_g^{(i)} \cdot f^{(i)}|_{C_g^{(i)}}} \tag{2} \end{equation} \]

is a valid codeword of \(V^{(i+1)} := \text{RS}[\mathbb{F}, \mathcal{D}^{(i+1)}, k^{(i+1)}]\) where \(k^{(i+1)} := \frac{k^{(i)}+1}{l^{(i)}} - 1\).

根据 [BBHR18] 以及上面的记号,在 FRI 协议的 COMMIT 阶段,固定一个 \(g \in \mathcal{D}^{(i+1)}\) ,下一步构造的 \(f_{f^{(i)},z^{(i)}}^{(i+1)}(g) := P_{\mathbf{u}^{(i)}(g)}^{(i)}(z^{(i)})\) ,下面理解下上述构造的式子。

\[\begin{split} \begin{aligned} f_{f^{(i)},z^{(i)}}^{(i+1)}(g) & = P_{\mathbf{u}^{(i)}(g)}^{(i)}(z^{(i)}) \\ & = \sum_{j<l^{(i)}} u_j^{(i)}(g) \cdot (z^{(i)})^j \\ & = \left(z^{(i)}\right)^0 \cdot u_0^{(i)}(g) + \left(z^{(i)}\right)^1 \cdot u_1^{(i)}(g) + \cdots + \left(z^{(i)}\right)^{l^{(i)}-1} \cdot u_{l^{(i)}-1}^{(i)}(g) \\ & = \begin{bmatrix} \left(z^{(i)}\right)^0 & \left(z^{(i)}\right)^1 & \cdots & \left(z^{(i)}\right)^{l^{(i)}-1} \end{bmatrix} \cdot \begin{bmatrix} u_0^{(i)}(g)\\ u_1^{(i)}(g)\\ \vdots\\ u_{l^{(i)}-1}^{(i)}(g) \end{bmatrix} \\ & = \left(\mathbf{z}^{(i)}\right)^{\intercal} \cdot \mathbf{u}^{(i)}(g) \end{aligned} \end{split}\]

下面说明下命题 1 给的第二个等式,即 \(\mathbf{u}^{(i)}(g) = M_g^{(i)} \cdot f^{(i)}|_{C_g^{(i)}}\) 。根据前面的分析,对于由 \(C_g^{(i)}\) 生成的 Vandermonde 矩阵有

\[\begin{split} V_{C_g^{(i)}} \cdot \begin{bmatrix} u_0^{(i)}\\ u_1^{(i)}\\ \vdots\\ u_{l^{(i)}-1}^{(i)} \end{bmatrix} = \begin{bmatrix} P_{\mathbf{u}}(g'_1)\\ P_{\mathbf{u}}(g'_2)\\ \vdots \\ P_{\mathbf{u}}(g'_{l^{(i)}}) \end{bmatrix} = \begin{bmatrix} f^{(i)}|_{C_g^{(i)}}(g'_1)\\ f^{(i)}|_{C_g^{(i)}}(g'_2)\\ \vdots \\ f^{(i)}|_{C_g^{(i)}}(g'_{l^{(i)}}) \end{bmatrix} \end{split}\]

因此

\[\begin{split} \begin{bmatrix} u_0^{(i)}\\ u_1^{(i)}\\ \vdots\\ u_{l^{(i)}-1}^{(i)} \end{bmatrix} = \left(V_{C_g^{(i)}}\right)^{-1} \cdot \begin{bmatrix} f^{(i)}|_{C_g^{(i)}}(g'_1)\\ f^{(i)}|_{C_g^{(i)}}(g'_2)\\ \vdots \\ f^{(i)}|_{C_g^{(i)}}(g'_{l^{(i)}}) \end{bmatrix} = M_g^{(i)} \cdot f^{(i)}|_{C_g^{(i)}} \end{split}\]

由此得到了 \(\left(\mathbf{u}^{(i)}(g)\right)^{\intercal} = M_g^{(i)} \cdot f^{(i)}|_{C_g^{(i)}}\)

Batching#

在某些情况下,第一个 prover 的 oracle \(f^{(0)}\) 是从一个仿射空间 \(F \subset \mathbb{F}^{\mathcal{D}^{(0)}}\) 的函数中采样的,这个仿射空间作为我们的输入,

\[ \begin{equation} F = \left\{ f_0^{(0)} + \sum_{i=1}^{t} x_i \cdot f_i^{(0)} \mid x_i \in \mathbb{F}, f_i : \mathcal{D}^{(0)} \to \mathbb{F} \right\} \tag{3} \end{equation} \]

当使用 FRI 协议来 “batch” 多个不同的 low degree testing 问题实例时,我们就通过随机线性组合将它们全部组合在一起,即上式中的 \(f_0^{(0)} + x_1 f_1^{(0)} + \cdots +x_t f_t^{(0)}\) 。在这种 batching 设置中,我们假设 prover 已经承诺了 \(f_1^{(0)}, \ldots, f_t^{(0)}\) (注意在这种情况下我们设 \(f_0^{(0)} = 0\)),并且 batched FRI 的 verifier 从 \(\mathbb{F}\) 中均匀随机地采样 \(x_1, \ldots, x_t \in \mathbb{F}\), prover 答复 \(f^{(0)}\) ,其应该等于 \(f_0^{(0)} + \sum_{i=1}^{t} x_i \cdot f_i^{(0)}\) ,现在 FRI 协议就应用于 \(f^{(0)}\) 了。相应地,batched FRI 的 QUERY 阶段也被扩展了,因此每次请求 \(f^{(0)}(g)\) 的查询时,验证者同时也查询了 \(f_0^{(0)}(g), \ldots, f^{(0)}_t(g)\) 并验证 \(f^{(0)}(g) = f_0^{(0)}(g) + \sum_{i=1}^{t} {\color{orange}x_i} \cdot f^{(0)}_i(g)\)

🐞 Fix

  • 论文这里的公式为 \(f^{(0)}(g) = f_0^{(0)}(g) + \sum_{i=1}^{t} f^{(0)}_i(g)\) ,我认为这里少了前面的系数 \(x_i\) ,应该为 \(f^{(0)}(g) = f_0^{(0)}(g) + \sum_{i=1}^{t} x_i \cdot f^{(0)}_i(g)\)

The (batched) FRI QUERY phase#

命题 1 表明对于诚实的 prover ,verifier 选取任意的值 \(z^{(i)}\) ,对于每一个 \(y \in D^{(i+1)}\) ,prover 都可以通过计算 \((2)\) 式,从一个码字 \(f^{(i)} \in V^{(i)}\) 构建一个新的码字 \(f^{(i+1)} \in V^{(i+1)}\) 。因此,我们将始终假设 \(f^{(r)} \in V^{(r)}\),例如,通过假设验证者总是查询 \(f^{(r)}\) 的前 \(k^{(r)}\) 个元素(按照某种规范顺序)并将 \(f^{(r)}\) 与该函数的插值多项式进行比较。

命题 1 给出了一种非常自然的测试方法,用来检查 \(f^{(i)}\)\(f^{(i+1)}\) 之间的一致性,并且 FRI 的查询阶段通过从 “顶部”(\(f^{(r)}\))到 “底部”(\(f^{(0)}\))迭代应用这种自然测试来遵循这一过程。

🤔 Question

  • 如何更好地去解释这里的自然的测试方法呢?

A single invocation of the FRI QUERY phase#

  1. \(\mathcal{D}^{(r)}\) 中均匀随机地选取 \(g^{(r)}\) 。对于 \(i = r, \cdots, 1\) ,从陪集 \(C_{g^{(i)}}^{(i-1)}\) 中均匀随机地选取 \(g^{(i-1)}\)

  2. 如果 \(f^{(0)}(g^{(0)}) \neq f_0^{(0)}(g^{(0)}) + \sum_{i = 1}^{t}x_i \cdot f_i^{(0)}(g^{(0)})\) ,则拒绝。

  3. 如果,对于任意地 \(i \in \{0, \cdots, r - 1\}\) ,有 \(f^{(i+1)}(g^{(i+1)}) \neq \left(\mathbf{z}^{(i)}\right)^{\intercal} \cdot M_g^{(i)} \cdot f^{(i)}|_{C_g^{(i)}}\) ,则拒绝。

  4. 否则 —— 如果上述条件中的所有等式都成立,则接受。

上述 QUERY 过程与 [BBHR18] 中 FRI 的 QUERY 过程有所不同,这里选取随机数是从最后一个 \(\mathcal{D}^{(r)}\) 中开始选取的,而不是从初始的 \(\mathcal{D}^{(0)}\) 中选取的。相比[BBHR18] 中的 QUERY 阶段,这里我们还想要验证在第 \(0\) 步时,batch 的是否正确,也就是 \(f^{(0)}(g^{(0)}) \neq f_0^{(0)}(g^{(0)}) + \sum_{i = 1}^{t}x_i \cdot f_i^{(0)}(g^{(0)})\)

Summary of the batched FRI protocol#

下面总结下目前提到的比较重要的性质,将会在下面的 soundness 分析中用到。

  1. 在协议的 COMMIT 阶段结束时,verifier 可以通过 oracle 访问一系列函数 \(f^{(0)}: \mathcal{D}^{(0)} \rightarrow \mathbb{F}, \ldots, f^{(r)}: \mathcal{D}^{(r)} \rightarrow \mathbb{F}\) ,其中 \(\mathcal{D}^{(0)} \supsetneq \ldots \supsetneq \mathcal{D}^{(r)}\) 是一系列 \(2\)-smooth 群,并且 \(f^{(i)}\) 任意依赖于 \(z^{(0)}, \ldots, z^{(i)}\) (以及 \(f^{(0)}, \ldots, f^{(i-1)}\))。我们假设 \(f^{(r)} \in V^{(r)}\)

  2. 存在一组 \(l^{(i)} \times l^{(i)}\) 的可逆矩阵 \(\{M_{g^{(i+1)}}^{(i)} : g^{(i+1)} \in D^{(i+1)}\}\) ,因此将 \(M_{g^{(i+1)}}^{(i)}\) 应用于 \(f^{(i)}|_{C_{g(i+1)}^{(i)}}\) 可以将 \(f^{(i)}\) 映射到一个向量序列 \(\mathbf{u} = \mathbf{u}^{(i)} = \{u_0^{(i)}, \ldots, u_{l(i)}^{(i)}\} \subset \mathbb{F}^{D^{(i+1)}}\) ,其中

\[ \mathbf{u}^{(i)}\left(g^{(i+1)}\right) = \left(u_0^{(i)}\left(g^{(i+1)}\right), \cdots, u_{l^{(i)}-1}^{(i)}\left(g^{(i+1)}\right) \right) = {\color{}M_{g^{(i+1)}}^{(i)} \cdot f^{(i)}|_{C_{g(i+1)}^{(i)}}}. \tag{4} \]

此外,如果 \(f^{(i)}\) 是在 \(D^{(i)}\) 上码率为 \(\rho\) 的有效 RS 码字,那么通过 \(\mathbf{u}^{(i)}\) 的参数化曲线上的每个向量也是在 \(D^{(i+1)}\) 上码率为 \(\rho\) 的有效 RS 码字。

  1. 在 QUERY 阶段的每次迭代会检查 \(f^{(i+1)}\) 是否是通过方程 \((2)\)\(f^{(i)}\) 构造的,并且(在 batched 的情况下)检查 \(f^{(0)}\) 是否是通过方程 \((3)\) 正确计算的。

Soundness#

Lemma 8.2 [BSCIK20, Lemma 8.2] (batched FRI error bound). Let \(V^{(0)} = \text{RS}[\mathbb{F}, \mathcal{D}^{(0)}, k^{(0)}]\) where \(\mathcal{D}^{(0)}\) is a coset of a \(2\)-smooth multiplicative group, and \(k^{(0)} + 1\) is a power of \(2\); set \(\rho = (k^{(0)} + 1)/|\mathcal{D}^{(0)}|\).

Let \(F \subseteq \mathbb{F}^{\mathcal{D}^{(0)}}\) be a space of functions as defined in Eq. \((3)\) whose correlated agreement density with \(V^{(0)}\) is \(\alpha\). For interger \(m \ge 3\) , let

\[ \alpha^{(0)}(\rho, m) = \max\left\{ \alpha, \sqrt{\rho}(1 + 1/2m) \right\}. \]

Assume the FRI protocol is used with \(r\) rounds, and let \(l^{(i)} = |\mathcal{D}^{(i)}|/|\mathcal{D}^{(i+1)}|\) denote the ratio between prover messages (oracles) \(i\) and \(i + 1\). Let \(\epsilon_Q\) denote the probability that the verifier accepts a single FRI QUERY invocation. Then,

\[ \Pr_{x_1, \ldots, x_t, z^{(0)}, \ldots, z^{(r-1)}}\left[ \epsilon_Q > \alpha^{(0)}(\rho, m) \right] \le \epsilon_{\text{C}}, \tag{8.5} \]

where

\[ \epsilon_{\text{C}} = \frac{\left(m + \frac{1}{2}\right)^7 \cdot \vert \mathcal{D}^{(0)}\vert^2}{2 \rho^{3/2} \vert \mathbb{F} \vert} + \frac{(2m+1) \cdot (\vert \mathcal{D}^{(0)}\vert + 1)}{\sqrt{\rho}} \cdot \frac{\sum_{i = 0}^{r - 1} l^{(i)}}{\vert \mathbb{F} \vert}. \]

In words: For any interactive FRI prover \(P^*\), the probability that the oracles \(f^{(0)}, \ldots, f^{(r)}\) sent by \(P^*\) will pass a single invocation of the batched FRI QUERY test with probability greater than \(\alpha^{(0)}(\rho, m)\), is smaller than \(\epsilon_{\text{C}}\). The probability is over the random variables \(x_1, \ldots, x_t\) used to sample \(f^{(0)}\) from \(F\) and over the random messages \(z^{(0)}, \ldots, z^{(r-1)}\) sent by the verifier during the COMMIT phase.

Theorem 8.3 [BSCIK20, Theorem 8.3] (Batched FRI Soundness). Let \(f_0^{(0)}, \ldots, f_t^{(r)}: \mathcal{D}^{(0)} \rightarrow \mathbb{F}\) be a sequence of functions and let \(V^{(0)} = \text{RS}[\mathbb{F}, \mathcal{D}^{(0)}, k^{(0)}]\) where \(\mathcal{D}^{(0)}\) is a coset of a \(2\)-smooth group of size \(n^{(0)} = |\mathcal{D}^{(0)}|\), and \(\rho = \frac{k^{(0)} + 1}{n^{(0)}}\) satisfies \(\rho = 2^{-\text{R}}\) for positive integer \(\text{R}\). Let \(\alpha = \sqrt{\rho}(1 + 1/2m)\) for integer \(m \ge 3\) and \(\epsilon_{\text{C}}\) be as defined in Lemma 8.2.

Assume the FRI protocol is used with \(r\) rounds. Let \(l^{(i)} = |\mathcal{D}^{(i)}|/|\mathcal{D}^{(i+1)}|\) denote the ratio between prover messages (oracles) \(i\) and \(i + 1\). Assume furthermore that \(s\) is the number of invocations of the FRI QUERY step.

Suppose there exists a batched FRI prover \(P^*\) that interacts with the batched FRI verifier and causes it to output “accept” with probability greater than

\[ \epsilon_{\text{FRI}} := \epsilon_{\text{C}} + \alpha^s = \frac{\left(m + \frac{1}{2}\right)^7 \cdot \vert \mathcal{D}^{(0)}\vert^2}{2 \rho^{3/2} \vert \mathbb{F} \vert} + \frac{(2m+1) \cdot (\vert \mathcal{D}^{(0)}\vert + 1)}{\sqrt{\rho}} \cdot \frac{\sum_{i = 0}^{r - 1} l^{(i)}}{\vert \mathbb{F} \vert} + \left(\sqrt{\rho} \cdot \left( 1 + \frac{1}{2m} \right) \right)^s . \]

Then \(f_0^{(0)}, \ldots, f_t^{(0)}\) have correlated agreement with \(V^{(0)}\) on a domain \(\mathcal{D}' \subset \mathcal{D}^{(0)}\) of density at least \(\alpha\).

定理 8.3 证明:反证法,然后直接通过引理 8.2 来证明。假设 \(f_0^{(0)}, \ldots, f_t^{(0)}\)\(V^{(0)}\) 的最大 correlated agreement 小于 \(\alpha^{(0)}(\rho, m) = \sqrt{\rho}(1 + 1/2m)\),但是同时接受的概率大于 \(\epsilon_{\text{C}} + (\alpha^{(0)}(\rho, m))^s\)

\(E\) 为在每次 FRI QUERY 阶段接受的概率大于 \(\alpha^{(0)}(\rho, m)\) 的事件。这个事件依赖于 \(x_1, \ldots, x_t, f^{(0)}, z^{(0)}, \ldots, z^{(r-1)}, f^{(r)}\) ,其中每个 \(f^{(i)}\)\(P^*\) 根据之前 Verifier 的消息生成的。通过引理 8.2,对于任意的 Prover \(P^*\) ,有事件 \(E\) 发生的概率不超过 \(\epsilon_{\text{C}}\) 。当事件 \(E\) 不成立时,那么 \(s\) 次独立调用 FRI QUERY 都返回 “accept” 的概率不超过 \((\alpha^{(0)}(\rho, m))^s\)

因此,FRI 的 Verifier 接受的概率不超过 \(\epsilon_{\text{C}} + (\alpha^{(0)}(\rho, m))^s\) ,这与假设矛盾。 \(\Box\)

Question

  • 这里每次 FRI QUERY 阶段接受的概率大于 \(\alpha^{(0)}(\rho, m)\) 的事件 \(E\) ,怎么联系调用 \(s\) 次的概率依然是不超过 \(\epsilon_{\text{C}}\) 呢?为什么不是 \((\epsilon_{\text{C}})^s\) 呢?

Proof of Lemma 8.2#

在证明 Lemma 8.2 之前,先介绍一种跟踪 verifier 检查 consistency 是否通过的方法。具体来说,Prover 会根据 Verifier 发送的随机数 \(z^{(i)}\) 来构造函数 \(f^{(i+1)}\) ,然后向 Verifier 回应函数 \(f^{(i+1)}\) 。在 FRI 的 QUERY 阶段, Verifier 会检查函数 \(f^{(i+1)}\) 与函数 \(f^{(i)}\) 的 consistency 。

定义一系列的加权(weight)函数,\(\mu^{(i)}:\mathcal{D}^{(i)} \rightarrow [0,1]\) 以及 \(\nu^{(i)}:\mathcal{D}^{(i)} \rightarrow [0,1]\) ,其中 \(i = 0, \ldots, r\) 。这些加权函数是通过归纳法来定义的。当 \(i = 0\) 时,用 \(\{0,1\}\) weights 来指示 \(f^{(0)}(g)\) 是否计算正确:

\[\begin{split} \mu^{(0)}(g) = \begin{cases} 1 & f^{(0)}(g) = f_0^{(0)}(g) + \sum_{i = 1}^t x_i f_i^{(0)}(g) \\ 0 & \text{otherwise} \end{cases} \end{split}\]

现在,可以通过归纳法得到的 \(\mu^{(i)}\) 可以来定义一个辅助 weight 函数 \(\nu^{(i+1)}: \mathcal{D}^{(i+1)} \rightarrow [0,1]\) 。在 \(\mathcal{D}^{(i+1)}\) 中取一个元素 \(g\) ,就可以得到 \(\mathcal{D}^{(i)}\) 中的陪集 \(C_g^{(i)} \subset \mathcal{D}^{(i)}\) ,这个集合是由那些在 \(\mathcal{D}^{(i)}\) 中能过通过映射 \(q^{(i)}(x) = x^{l^{(i)}}\) 得到 \(g\) 的所有元素组成的,即如 \((8.1)\) 所示,

\[ C_g^{(i)} := \{g' \in \mathcal{D}^{(i)} \mid (g')^{l^{(i)}} = g\}. \]

那么 \(\nu^{(i+1)}\) 的定义为

\[ \nu^{(i+1)}(g) = \mathbb{E}_{g' \in C_g^{(i)}}\left[ \mu^{(i)}(g') \right]. \tag{8.6} \]

换句话说,\(\nu^{(i+1)}(g)\) 是陪集 \(C_g^{(i)}\) 中的所有元素的 \(\mu^{(i)}\) weight 的期望值。最后,来定义函数 \(\mu^{(i+1)}\) ,对于每一个 \(g \in \mathcal{D}^{(i+1)}\)

\[\begin{split} \mu^{(i+1)}(g) = \begin{cases} \nu^{(i+1)}(g) & f^{(i+1)}(g) = f_{f^{(i)}, z^{(i)}}^{(i+1)}(g) \\ 0 & \text{otherwise} \end{cases} \end{split}\]

关于 \(\mu^{(i)}\) 的定义,一个很重要的性质就是,\(\mu^{(i)}(g)\) 是在 \(g\)\(f^{(i)}\) 中 query 的条件下,FRI QUERY 阶段成功的概率的一个度量,这也是下面的命题成立的一个重要原因。

Claim 8.5. The probability \(\epsilon_{\text{Q}}\) that a single invocation of the batched FRI QUERY accepts \(f^{(0)}, \ldots, f^{(r)}\), where \(f^{(r)} \in \text{RS}[\mathbb{F},\mathcal{D}^{(r)},k^{(r)}]\), satisfies

\[ \epsilon_{\text{Q}} = \mathbb{E}_{g^{(r)} \in \mathcal{D}^{(r)}}\left[ \mu^{(r)}(g^{(r)}) \right]. \]

证明:回顾下 FRI QUERY 的调用,会选择一系列随机的 \(g^{(r)}, \ldots, g^{(0)}\) ,其中 \(g^{(i-1)}\) 是从陪集 \(C_{g^{(i)}}^{(i-1)}\) 中均匀随机选取的。下面通过归纳法来证明,对于 \(i = 0, \ldots, r\)

\[ \mathbb{E}_{g^{(i)} \in \mathcal{D}^{(i)}}\left[ \mu^{(i)}(g^{(i)}) \right] \]

等于这样的概率,当均匀随机地选取 \(g^{(i)}\) ,并且其是从一个随机序列 \(g^{(i-1)} \in C_{g^{(i)}}^{(i-1)}, \ldots, g^{(0)} \in C_{g^{(1)}}^{(0)}\) 中生成的时,所有和 \(g^{(i)}\) 及其引发的测试都通过的概率。

归纳法证明的思路如下:

  1. 证明对于 \(i = 0\) 时,最基本的情况 \(\mu^{(0)}\) 成立

  2. 假设对于 \(i - 1\)\(\mu^{(i-1)}\) 成立,证明对于 \(i\)\(\mu^{(i)}\) 成立

从而就能证明该命题。

\(i = 0\) 时,由 \(\mu^{(0)}\) 的定义,

\[\begin{split} \mu^{(0)}(g^{(0)}) = \begin{cases} 1 & f^{(0)}(g^{(0)}) = f_0^{(0)}(g^{(0)}) + \sum_{i = 1}^t x_i f_i^{(0)}(g^{(0)}) \\ 0 & \text{otherwise} \end{cases} \end{split}\]

调用 FRI QUERY 通过的概率自然等于 \(\mathbb{E}_{g^{(0)} \in \mathcal{D}^{(0)}}\left[ \mu^{(0)}(g^{(0)}) \right]\)

假设对于 \(i - 1\)\(\mu^{(i-1)}\) 成立,现在分析 \(\mu^{(i)}(g^{(i)})\) ,如果 \(f^{(i)}(g^{(i)})\) 未按照式 \((8.2)\) 正确计算,那么 \(\mu^{(i)}(g^{(i)}) = 0\) ,否则的话,根据定义

\[ \mu^{(i)}(g^{(i)}) = \nu^{(i)}(g^{(i)}) = \mathbb{E}_{g^{(i-1)} \in C_{g^{(i)}}^{(i-1)}}\left[ \mu^{(i-1)}(g^{(i-1)}) \right]. \]

这说明 \(\mu^{(i)}(g^{(i)})\) 是在陪集 \(C_{g^{(i)}}^{(i-1)} \subseteq \mathcal{D}^{(i-1)}\)\(\mu^{(i-1)}\) 的值的平均值,由归纳法假设得,其是和 \(g^{(i-1)}, \ldots, g^{(0)}\) 相关的所有测试通过的概率,因此可得对于 \(i\)\(\mu^{(i)}\) 成立。 \(\Box\)

Lemma 8.2 需要估计的是在 FRI QUERY 阶段的概率,回顾上面的 batched FRI QUERY 阶段的协议,有两个地方涉及到随机数:

  1. 协议的第 2 步,使用 \(t\) 个随机数 \(x_1, \ldots , x_t\) 来 batch \(f_1^{(0)}, f_2^{(0)}, \ldots ,f_t^{(0)}\) ,这对应 affine space 的情况,要用到定理 7.4 对应的结论。

  2. 协议的第 3 步,用 \(\mathbf{z}^{(i)} = \left(\left(z^{(i)}\right)^0, \left(z^{(i)}\right)^1, \ldots, \left(z^{(i)}\right)^{l^{(i)}-1}\right)\) 来进行 batch ,对应 curves 的情况,会用到定理 7.2 的结论。

Lemma 8.2 证明:现在要证明引理 8.2 ,由命题 8.5,只需证明,在 verifier 的随机选择中,以大于概率 \(1 - \epsilon_C\)

\[ \mathbb{E}_{g \in \mathcal{D}^{(r)}}\left[ \mu^{(r)}(g) \right] \le \alpha^{(0)}(\rho, m). \tag{8.7} \]

如果证明了上述成立,那么也就是说当在 \(\mathbb{F}_q\) 中选取随机数时,如果有 \(\epsilon_Q > \alpha^{(0)}(\rho, m)\) ,那么其概率小于等于 \(\epsilon_C\) ,这也就证明了引理 8.2。

🤔 Question

  • 这里为什么不是 “以大于概率 \(1 - \epsilon_C\) 有”?

证明的思路是先定义一些列坏的事件 \(E^{(0)}, \ldots, E^{(r)}\) ,其中一些事件会发生的概率是各个事件发生的概率之和,证明这个概率小于等于 \(\epsilon_C\) 。接着再假设当没有坏的事件发生时,证明式子 \((8.7)\) 成立。

\(E^{(0)}\) 为事件

\[ \text{agree}_{\mu^{(0)}} \left(f^{(0)}, V^{(0)}\right) > \alpha^{(0)}(\rho, m). \]

\(\mu^{(0)}\) 的定义得事件 \(E^{(0)}\)

\[ \text{agree} \left(f_0^{(0)} + \sum_{i=1}^{t} x_i f_i^{(0)} , V^{(0)}\right) > \alpha^{(0)}(\rho, m) = \max\left\{ \alpha, \sqrt{\rho}(1 + 1/2m) \right\}. \]

🤔 Question

  • 这里的 \(\text{agree}\) 具体是什么含义?和 \(\text{agree}_{\mu^{(0)}}\) 的区别是什么?是表示常数 1 吗?

因此这个事件 \(E^{(0)}\) 主要取决于随机数 \(x_1, \ldots, x_t\) 。通过引理中的假设,有 \((f_0^{(0)}, \ldots, f_t^{(0)})\)\(V^{(0)}\) 的最大 correlated agreement density 不超过 \(\alpha\)

回顾下定理 7.4: Theorem 7.4 (Weighted correlated agreement over affine spaces – Version II). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) and let \(U = u_0 + \text{span}\{u_1, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) be an affine subspace. Let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Let \(m \ge 3\) and let

\[ > \alpha \ge \alpha_0(\rho, m) := \sqrt{\rho} + \frac{\sqrt{\rho}}{2m} . > \]

Suppose

\[ > \Pr_{u \in U} [\text{agree}_{\mu}(u,V) \ge \alpha] > \max \left( \frac{(1 + \frac{1}{2m})^7m^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{M \cdot n + 1}{q} \right) . \tag{7.2} > \]

Then \(u_0, \ldots, u_l\) have at least \(\alpha\) correlated \(\mu\) -agreement with \(V\) , i.e. \(\exists v_0, \cdots, v_l \in V\) such that

\[ > \mu \left(\{x \in \mathcal{D} : \forall 0 \le i \le l, u_i(x) = v_i(x)\}\right) \ge \alpha . > \]

则其逆否命题为:如果 \(u_0, \ldots, u_l\)\(V\) 有至多 \(\alpha\) 的 correlated \(\mu\) -agreement ,那么有

\[ > \Pr_{u \in U} [\text{agree}_{\mu}(u,V) \ge \alpha] \le \max \left( \frac{(1 + \frac{1}{2m})^7m^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{M \cdot n + 1}{q} \right) . > \]

由定理 7.4 的逆否命题,取 \(\alpha = \alpha^{(0)}(\rho, m)\)\(\mu \equiv 1\) 以及 \(M = 1\) ,有

\[\begin{split} \begin{aligned} \Pr_{x_1, \ldots, x_t} [E^{(0)}] & = \Pr_{u \in U} [\text{agree}_{\mu}(u,V) > \alpha^{(0)}(\rho, m)] \\ & {\color{orange}\text{(为什么这里括号里能直接改为严格的$>$?)}} \\ & \le \max \left( \frac{(1 + \frac{1}{2m})^7m^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{M \cdot n + 1}{q} \right) \\ & = \max \left( \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{n + 1}{q} \right) \end{aligned} \end{split}\]

注意,根据定理 7.4 以及定理 1.2,其中 \(V = \mathsf{RS}[\mathbb{F}_q, \mathcal{D}^{(0)}, k^{(0)}]\)\(n = |\mathcal{D}^{(0)}|\)\(\rho = \frac{k^{(0)} + 1}{n}\)

下面推导下

\[ \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q} > \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{n + 1}{q} \]

由于

\[\begin{split} \begin{aligned} & \frac{(m+\frac{1}{2})^7}{3} \ge 2m + 1 \\ \Rightarrow & \frac{(2m + 1)^7}{3 \times 2^7} \ge 2m + 1 \\ \Rightarrow & (2m + 1)^6 \ge 3 \times 2^7 \\ \end{aligned} \end{split}\]

由定理中的条件 \(m \ge 3\)\((2m + 1)^6\)\(m \ge 3\) 是增函数,因此 \((2m + 1)^6 \ge (2 \times 3 + 1)^6 = 7^6 = 117649\) ,同时上式的右边 \(3 \times 2^7 = 384\) ,满足 \((2m + 1)^6 \ge 117649 > 3 \times 2^7\) ,由此得到 \(\frac{(m+\frac{1}{2})^7}{3} > 2m + 1\) (取不到等号)成立。那么

\[\begin{split} \begin{aligned} \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q} & > \frac{2m + 1}{\rho^{3/2}} \cdot \frac{n^2}{q} \\ & = \frac{2m + 1}{\rho \cdot \rho^{1/2}} \cdot \frac{n^2}{q} \\ & {\color{blue}{\text{(由于 $\rho < 1$ )}}} \\ & > \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{n^2}{q} \\ & {\color{blue}{\text{(由于 $n \ge 2$ 时 $n^2 > n + 1$ )}}} \\ & > \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{n + 1}{q} \end{aligned} \end{split}\]

从而

\[\begin{split} \begin{aligned} \Pr_{x_1, \ldots, x_t} [E^{(0)}] & \le \max \left( \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q}, \frac{2m + 1}{\sqrt{\rho}} \cdot \frac{n + 1}{q} \right) \\ & = \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q} \end{aligned} \end{split}\]

\[ \epsilon = \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{n^2}{q} {\color{blue}\quad {\text{(注意其中 $n = |\mathcal{D}^{(0)}|$)}}} \]

\[ \Pr_{x_1, \ldots, x_t} [E^{(0)}] \le \epsilon \tag{8.8} \]

现在固定 \(i \in \{0, \ldots, r - 1\}\) 。定义事件 \(E^{(i+1)}\)

\[ \text{agree}_{\nu^{(i+1)}} \left(f_{f^{(i)},z^{(i)}}^{(i+1)}, V^{(i+1)}\right) > \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right) . \tag{8.9} \]

📖 Notes 理解下事件 \(E^{(i+1)}\) 。根据定义

\[\begin{split} > \begin{aligned} > \text{agree}_{\nu^{(i+1)}} \left(f_{f^{(i)},z^{(i)}}^{(i+1)}, V^{(i+1)}\right) & = \max_{g^{(i+1)} \in V^{(i+1)}} \text{agree}_{\nu^{(i+1)}} \left(f_{f^{(i)},z^{(i)}}^{(i+1)}, g^{(i+1)} \right) \\ > & = \max_{g^{(i+1)} \in V^{(i+1)}} \frac{1}{|\mathcal{D}^{(i+1)}|} \sum_{x:f_{f^{(i)},z^{(i)}}^{(i+1)}(x) = g^{(i+1)}(x)} \nu^{(i+1)}(x)\\ > & = \max_{g^{(i+1)} \in V^{(i+1)}} \frac{1}{|\mathcal{D}^{(i+1)}|} \sum_{x:f_{f^{(i)},z^{(i)}}^{(i+1)}(x) = g^{(i+1)}(x)} \mathbb{E}_{g' \in C_x^{(i)}} \left[ \mu^{(i)}(g') \right] \\ > \end{aligned} > \end{split}\]

衡量的是由 \(f^{(i)}\) 与随机数 \(z^{(i)}\) 构造得到 \(f_{f^{(i)},z^{(i)}}^{(i+1)}\) 后,在 \(\mathcal{D}^{(i+1)}\) 中找到使得 \(f_{f^{(i)},z^{(i)}}^{(i+1)}\) 能与 \(V^{(i+1)}\) 中的一个多项式 \(g^{(i+1)}\) 一致的 \(x\) ,再计算由这些 \(x\) 对应的在 \(\mathcal{D}^{(i)}\) 中的陪集中的元素的 \(\mu^{(i)}\) weight 的期望值之和。

\((8.9)\) 式右边中

\[ > \begin{aligned} > \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right) & = \max_{g^{(i)} \in V^{(i)}} \frac{1}{|\mathcal{D}^{(i)}|} \sum_{x:f^{(i)}(x) = g^{(i)}(x)} \mu^{(i)}(x) > \end{aligned} > \]

\(\mu^{(i)}(x)\) 衡量的是当从 \(f^{(i)}\) 中 qurey \(x\) 时,在 FRI QUERY 阶段能通过的概率。

\(E^{(i+1)}\) 事件想定义这样一些事件,通过 \(f^{(i)}\)\(z^{(i)}\) 构造得到的 \(f_{f^{(i)},z^{(i)}}^{(i+1)}\) ,对于 \(V^{(i+1)}\) 中的一个多项式 \(g^{(i+1)}\) ,取出使得它们值相同的点 \(x\) 的集合,去计算这些点对应陪集的 \(\mu^{(i)}\) weight 的期望之和与 \(\mathcal{D}^{(i+1)}\) 的大小的比值。

固定 \(f^{(i)}\)\(\mu^{(i)}\) ,则事件 \(E^{(i+1)}\) 是由随机数 \(z^{(i)}\) 决定的。根据 \(\mu^{(i+1)}\) 的定义,有

\[\begin{split} \mu^{(i+1)}(g) = \begin{cases} \nu^{(i+1)}(g) & f^{(i+1)}(g) = f_{f^{(i)}, z^{(i)}}^{(i+1)}(g) \\ 0 & \text{otherwise} \end{cases} \end{split}\]

当满足条件 \(f^{(i+1)}(g) = f_{f^{(i)}, z^{(i)}}^{(i+1)}(g)\) 时,\(\mu^{(i+1)}(g)\) 才会与 \(\nu^{(i+1)}(g)\) 相等。自然可以得到

\[ \text{agree}_{\mu^{(i+1)}} \left(f^{(i+1)}, V^{(i+1)}\right) \le \text{agree}_{\nu^{(i+1)}} \left(f_{f^{(i)},z^{(i)}}^{(i+1)}, V^{(i+1)}\right) \]

因此如果事件 \(E^{i+1}\) 不发生,那么再根据 \((8.9)\) 式可以得到

\[ \text{agree}_{\mu^{(i+1)}} \left(f^{(i+1)}, V^{(i+1)}\right) \le \text{agree}_{\nu^{(i+1)}} \left(f_{f^{(i)},z^{(i)}}^{(i+1)}, V^{(i+1)}\right) \le \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right) \]

\[ \text{agree}_{\mu^{(i+1)}} \left(f^{(i+1)}, V^{(i+1)}\right) \le \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right) \tag{8.10} \]

\(\alpha = \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right)\) 。根据定义,展开 \(f_{f^{(i)},z^{(i)}}^{(i+1)}\) ,得到事件 \(E^{(i+1)}\)

\[ \text{agree}_{\nu^{(i+1)}} \left(u_0 + z^{(i)}u_1 + \ldots + (z^{(i)})^{l^{(i)} -1} u_{l^{(i)} -1}, V^{(i+1)}\right) > \alpha, \]

其中 \(u_0, \ldots, u_{l^{(i)} - 1}: \mathcal{D}^{(i+1)} \rightarrow \mathbb{F}\) 为在 FRI 协议定义中由 \(f^{(i)}\) 得到的函数(见命题 8.1)。这正是定理 7.2 的所处理的情况。

📖 回顾定理 7.2

Theorem 7.2 (Weighted correlated agreement over curves – Version II). Let \(V, q, n, k\) and \(\rho\) be as defined in Theorem 1.2. Let \(\mathbf{u} = \{u_0, \cdots, u_l\} \subset \mathbb{F}_q^{\mathcal{D}}\) . Let \(\mu : \mathcal{D} \rightarrow [0,1]\) be a vector of weights, whose values all have denominator \(M\) . Let \(m \ge 3\) and let

\[ > \alpha \ge \alpha_0(\rho, m) := \sqrt{\rho} + \frac{\rho}{2m}. > \]

Let

\[ > S = \{z \in \mathbb{F}_q : \text{agree}_{\mu}(u_0 + zu_1 + \cdots + z^lu_l, V) \ge \alpha\} > \]

and suppose

\[ > |S| > \max\left(\frac{(1 + \frac{1}{2m})^7 m^7}{3 \rho^{3/2}}n^2 l, \frac{2m + 1}{\sqrt{\rho}}(M \cdot n + 1)l \right) . \tag{7.1} > \]

Then \(u_0, \ldots, u_l\) have at least \(\alpha\) correlated \(\mu\) -agreement with \(V\) , i.e. \(\exists v_0, \cdots, v_l \in V\) such that

\[ > \mu(\{x \in \mathcal{D}: \forall 0 \le i \le l, u_i(x) = v_i(x)\}) \ge \alpha . > \]

在定理 7.2 中取 \(M = |\mathcal{D}^{(0)}|/|\mathcal{D}^{(i+1)}|\) ,这时我们分析的是 \(i + 1\) 的情况,因此定理中 \(n = |\mathcal{D}^{i+1}|\) ,则 \(M \cdot n = |\mathcal{D}^{(0)}|\)。由于我们分析的 \(\mathbf{u} = \{u_0, \cdots, u_{l^{(i)} - 1}\}\) ,因此 式 \((7.1)\) 中的 \(l = l^{(i)} - 1\)。根据定理 7.2 ,如果

\[ \Pr_{z^{(i)}}\left[E^{(i+1)}\right] \ge (l^{(i)} - 1) \cdot \left(\epsilon^{(i)} + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \right) \]

其中,

\[ \epsilon^{(i)} = \frac{|\mathcal{D}^{(i+1)}|^2}{|\mathcal{D}^{(0)}|^2} \epsilon = \frac{\epsilon}{(l^{(0)}\cdots l^{(i)})^2} = \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{|\mathcal{D}^{(0)}|^2}{q} \cdot \frac{1}{(l^{(0)}\cdots l^{(i)})^2} \]

如果满足上述条件,对照定理 7.2 ,则有

\[\begin{split} \begin{aligned} |S| & \ge |\mathbb{F}| \cdot (l^{(i)} - 1) \cdot \left(\epsilon^{(i)} + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{\mathbb{F}} \right) \\ & = \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{|\mathcal{D}^{(0)}|^2}{q} \cdot \frac{|\mathcal{D}^{(i+1)}|^2}{|\mathcal{D}^{(0)}|^2} \cdot |\mathbb{F}| \cdot (l^{(i)} - 1) + \frac{2m + 1}{\sqrt{\rho}}\cdot (|\mathcal{D}^{(0)}| + 1) \cdot (l^{(i)} - 1) \\ & = \frac{(1 + \frac{1}{2m})^7 m^7}{3 \rho^{3/2}} \cdot |\mathcal{D}^{(i+1)}|^2 \cdot (l^{(i)} - 1) + \frac{2m + 1}{\sqrt{\rho}}\cdot (M \cdot n + 1) \cdot (l^{(i)} - 1) \\ & = \frac{(1 + \frac{1}{2m})^7 m^7}{3 \rho^{3/2}} \cdot n^2 \cdot (l^{(i)} - 1) + \frac{2m + 1}{\sqrt{\rho}}\cdot (M \cdot n + 1) \cdot (l^{(i)} - 1) \\ & > \max\left(\frac{(1 + \frac{1}{2m})^7 m^7}{3 \rho^{3/2}}n^2 l, \frac{2m + 1}{\sqrt{\rho}}(M \cdot n + 1)l \right) \end{aligned} \end{split}\]

满足式 \((7.1)\) ,因此由定理 7.2 可得存在一个集合 \(S \subseteq \mathcal{D}^{(i+1)}\) ,存在码字 \(v_0, \ldots, v_{l^{(i)} - 1} \in V\) ,满足 \(u_i\)\(v_i\)\(S\) 上一致,并且 \(\nu^{(i+1)}(S) > \alpha\) 。回顾式 \((8.4)\) ,知

\[ \mathbf{u}^{(i)}\left(g^{(i+1)}\right) = {\color{}M_{g^{(i+1)}}^{(i)} \cdot f^{(i)}|_{C_{g(i+1)}^{(i)}}} \]

可逆的插值映射 \(M_{g^{(i+1)}}^{(i)}\)\(f^{(i)}|_{C_{g(i+1)}^{(i)}}\) 映射到了 \(\mathbf{u}^{(i)}\left(g^{(i+1)}\right)\) 。使用其逆映射,即 evaluation 映射,对每一个 \(g^{(i+1)} \in \mathcal{D}^{(i+1)}\) ,将该逆映射作用到 \(v_0(g^{(i+1)}), \ldots, v_{l^{(i)}-1}(g^{(i+1)})\) 上,设 \(C_{g^{(i+1)}}^{(i)} =\{g'_0, \cdots, g'_{l^{(i)}-1}\}\) ,则作用后的结果为

\[\begin{split} \begin{aligned} \left(M_{g^{(i+1)}}^{(i)}\right)^{-1} \cdot \begin{bmatrix} v_0(g^{(i+1)}) \\ v_1(g^{(i+1)}) \\ \vdots \\ v_{l^{(i)}-1}(g^{(i+1)}) \end{bmatrix} & = V_{C_{g^{(i+1)}}^{(i)}} \cdot \begin{bmatrix} v_0(g^{(i+1)}) \\ v_1(g^{(i+1)}) \\ \vdots \\ v_{l^{(i)}-1}(g^{(i+1)}) \end{bmatrix} \\ & = \begin{bmatrix} 1 & g'_0 & (g'_0)^2 & \cdots & (g'_0)^{l^{(i)}-1} \\ 1 & g'_1 & (g'_1)^2 & \cdots & (g'_1)^{l^{(i)}-1} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 1 & g'_{l^{(i)-1}} & (g'_{l^{(i)}-1})^2 & \cdots & (g'_{l^{(i)}-1})^{l^{(i)}-1} \end{bmatrix} \begin{bmatrix} v_0(g^{(i+1)}) \\ v_1(g^{(i+1)}) \\ \vdots \\ v_{l^{(i)}-1}(g^{(i+1)}) \end{bmatrix} \\ & = \begin{bmatrix} v_0(g^{(i+1)}) + v_1(g^{(i+1)}) g'_0 + u_2 (g'_0)^2 + \cdots + v_{l^{(i)}-1}(g^{(i+1)}) (g'_0)^{l^{(i)}-1}\\ v_0(g^{(i+1)}) + v_1(g^{(i+1)}) g'_1 + u_2 (g'_1)^2 + \cdots + v_{l^{(i)}-1}(g^{(i+1)}) (g'_1)^{l^{(i)}-1}\\ \vdots \\ v_0(g^{(i+1)}) + v_1(g^{(i+1)}) g'_{l^{(i)-1}} + u_2 (g'_{l^{(i)}-1})^2 + \cdots + v_{l^{(i)}-1}(g^{(i+1)}) (g'_{l^{(i)}-1})^{l^{(i)}-1} \end{bmatrix} \\ & = \begin{bmatrix} h^{(i)}(g'_0)\\ h^{(i)}(g'_1)\\ \vdots \\ h^{(i)}(g'_{l^{(i)}-1}) \end{bmatrix} \end{aligned} \end{split}\]

可以得到函数 \(h^{(i)}: \mathcal{D}^{(i)} \rightarrow \mathbb{F}\) ,对于每一个 \(g^{(i)} \in C_{g^{(i+1)}}^{(i)}\)

\[ h^{(i)}(g^{(i)}) = \sum_{j = 0}^{l^{(i)}-1} \left(g^{(i)}\right)^j \cdot v_j\left({\color{blue}{g^{(i+1)}}}\right) = \sum_{j = 0}^{l^{(i)}-1} \left(g^{(i)}\right)^j \cdot v_j \left( {\color{blue}{\left(g^{(i)}\right)^{l^{(i)}}}} \right). \]

因此,由于 \(v_j \in V^{(i+1)}\) ,因此 \(h^{(i)} \in V^{(i)}\) 。另外,根据定义

\[\begin{split} \begin{aligned} \text{agree}_{\mu^{(i)}}\left( f^{(i)}, V^{(i)} \right) & = \max_{v \in V^{(i)}} \text{agree}_{\mu^{(i)}}\left( f^{(i)}, v \right)\\ & \ge \text{agree}_{\mu^{(i)}}\left( f^{(i)}, h^{(i)} \right) \\ & = \frac{1}{|\mathcal{D}^{(i)}|} \sum_{x: f^{(i)}(x) = h^{(i)}(x)} \mu^{(i)}(x) \\ & = \nu^{(i+1)}(S) \\ & > \alpha, \end{aligned} \end{split}\]

这与 \(\alpha\) 的定义 \(\alpha = \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right)\) 是矛盾的。那么说明我们在应用定理 7.2 时所给的假设是不成立的,也就是下式成立:

\[ \Pr_{z^{(i)}}\left[E^{(i+1)}\right] < (l^{(i)} - 1) \cdot \left(\epsilon^{(i)} + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \right) . \]

因此,如果没有事件 \(E^{(i+1)}\) 发生,根据 \((8.10)\) 式,对于所有的 \(i \in 0, 1, \ldots, r - 1\) 都有:

\[ \text{agree}_{\mu^{(i+1)}} \left(f^{(i+1)}, V^{(i+1)}\right) \le \max \left( \text{agree}_{\mu^{(i)}} \left(f^{(i)}, V^{(i)}\right), \sqrt{\rho}(1 + 1/2m) \right) \]

根据式 \((8.8)\) 得到

\[ \Pr_{x_1, \ldots, x_t} [E^{(0)}] \le \epsilon , \quad 其中 \epsilon = \frac{(m + \frac{1}{2})^7}{3 \rho^{3/2}} \cdot \frac{{|\mathcal{D}^{(0)}|}^2}{q} {\color{blue}}. \]

如果事件 \(E^{(0)}\) 或者一些 \(E^{(i+1)}\) 发生的概率估计为

\[\begin{split} \begin{aligned} \Pr_{x_1,\ldots, x_t} \left[E^{(0)}\right] + \sum_{i=0}^{r-1} \Pr_{z^{(i)}} \left[E^{(i+1)}\right] & \le \epsilon + \sum_{i=0}^{r-1} \left( (l^{(i)} - 1) \cdot \left(\epsilon^{(i)} + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \right) \right) \\ & = \epsilon + \sum_{i=0}^{r-1} (l^{(i)} - 1) \cdot \epsilon^{(i)} + \sum_{i=0}^{r-1} \left((l^{(i)} - 1) \cdot \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|}\right) \\ & = \left(1 + \sum_{i=0}^{r-1} \frac{l^{(i)} - 1}{(l^{(0)}\cdots l^{(i)})^2} \right) \epsilon + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \cdot \sum_{i=0}^{r-1} (l^{(i)} - 1) \end{aligned} \end{split}\]

下面估计下 \(\sum_{i=0}^{r-1} \frac{l^{(i)} - 1}{(l^{(0)}\cdots l^{(i)})^2}\) ,由于对于 \(i \in \{0, \ldots, r - 1\}\)\(l^{(i)} \ge 2\) ,因此

\[\begin{split} \begin{aligned} \sum_{i=0}^{r-1} \frac{l^{(i)} - 1}{(l^{(0)}\cdots l^{(i)})^2} & = \sum_{i=0}^{r-1} \left(\frac{l^{(i)}}{(l^{(0)}\cdots l^{(i)})^2} - \frac{1}{(l^{(0)}\cdots l^{(i)})^2}\right) \\ & = \sum_{i=0}^{r-1} \left(\frac{1}{(l^{(0)}\cdots l^{(i-1)})^2 l^{(i)}} - \frac{1}{(l^{(0)}\cdots l^{(i)})^2}\right) \\ & = \frac{1}{l^{(0)}} + \left(- \frac{1}{(l^{(0)})^2} + \frac{1}{(l^{(0)})^2l^{(1)}}\right) + \left(- \frac{1}{(l^{(0)}l^{(1)})^2} + \frac{1}{(l^{(0)}l^{(1)})^2l^{(2)}} \right)\\ & \quad + \ldots + \left(- \frac{1}{(l^{(0)}\cdots l^{(r-2)})^2} + \frac{1}{(l^{(0)}\cdots l^{(r-2)})^2l^{(r-1)}} \right) - \frac{1}{(l^{(0)}\cdots l^{(r-1)})^2} \\ & \le \frac{1}{l^{(0)}} + \left(- \frac{1}{(l^{(0)})^2} + \frac{1}{(l^{(0)})^2 \cdot 2}\right) + \left(- \frac{1}{(l^{(0)}l^{(1)})^2} + \frac{1}{(l^{(0)}l^{(1)})^2 \cdot 2} \right)\\ & \quad + \ldots + \left(- \frac{1}{(l^{(0)}\cdots l^{(r-2)})^2} + \frac{1}{(l^{(0)}\cdots l^{(r-2)})^2 \cdot 2} \right) - \frac{1}{(l^{(0)}\cdots l^{(r-1)})^2} \\ & = \frac{1}{l^{(0)}} - \frac{1}{(l^{(0)})^2 \cdot 2} - \frac{1}{(l^{(0)}l^{(1)})^2 \cdot 2} - \ldots - \frac{1}{(l^{(0)}\cdots l^{(r-2)})^2 \cdot 2}\\ & \quad - \frac{1}{(l^{(0)}\cdots l^{(r-1)})^2} \\ & < \frac{1}{l^{(0)}} \\ & < \frac{1}{2} \end{aligned} \end{split}\]

📖 另一种证明方法:利用等比数列求和 $\( \begin{aligned} \sum_{i=0}^{r-1} \frac{l^{(i)} - 1}{(l^{(0)}\cdots l^{(i)})^2} & = \sum_{i=0}^{r-1} \left(\frac{l^{(i)}}{(l^{(0)}\cdots l^{(i)})^2} - \frac{1}{(l^{(0)}\cdots l^{(i)})^2}\right) \\ & = \sum_{i=0}^{r-1} \left(\frac{1}{(l^{(0)}\cdots l^{(i-1)})^2 l^{(i)}} - \frac{1}{(l^{(0)}\cdots l^{(i)})^2}\right) \\ & < \sum_{i=0}^{r-1} \frac{1}{(l^{(0)}\cdots l^{(i-1)})^2 l^{(i)}} \\ & {\color{blue}(因为 \frac{1}{(l^{(0)}\cdots l^{(i)})^2} > 0)} \\ & < \sum_{i=0}^{r-1} \frac{1}{l^{(0)}\cdots l^{(i-1)} l^{(i)}} \\ & {\color{blue}(因为 l^{(i)} \ge 2, 因此 {l^{(i)}}^2 > l^{(i)} )} \\ & < \sum_{i=0}^{r-1} \left(\frac{1}{2}\right) ^{i+1} \\ & = \frac{1}{2}\sum_{i=0}^{r-1} \left(\frac{1}{2}\right) ^{i} \\ & < \frac{1}{2} \cdot \frac{1}{2} \\ & < \frac{1}{2} \end{aligned} \)$

🤔 Question

  • 关于上述证明还有更简洁的方式吗?

因此

\[\begin{split} \begin{aligned} \Pr_{x_1,\ldots, x_t} \left[E^{(0)}\right] + \sum_{i=0}^{r-1} \Pr_{z^{(i)}} \left[E^{(i+1)}\right] & \le \left(1 + \sum_{i=0}^{r-1} \frac{l^{(i)} - 1}{(l^{(0)}\cdots l^{(i)})^2} \right) \epsilon + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \cdot \sum_{i=0}^{r-1} (l^{(i)} - 1) \\ & < \left(1 + \frac{1}{2} \right) \epsilon + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \cdot \sum_{i=0}^{r-1} (l^{(i)} - 1) \\ & < \frac{3}{2} \epsilon + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \cdot \sum_{i=0}^{r-1} l^{(i)}. \end{aligned} \end{split}\]

综上所述,我们得到了当某些坏的事件 \(E^{(i)}\) 发生时,其概率严格小于

\[ \frac{3}{2} \epsilon + \frac{2m + 1}{\sqrt{\rho}}\cdot \frac{|\mathcal{D}^{(0)}| + 1}{|\mathbb{F}|} \cdot \sum_{i=0}^{r-1} l^{(i)} = \epsilon_C, \]

当没有坏的事件发生时,下面的式子成立

\[\begin{split} \begin{aligned} \text{agree}_{\mu^{(r)}} \left(f^{(r)}, V^{(r)}\right) & = \mathbb{E}_{g^{(r)} \in \mathcal{D}^{(r)}}\left[ \mu^{(r)}(g^{(r)}) \right] \\ & \color{blue}{(因为 f^{(r)} \in V^{(r)})} \\ & \le \max \left( \text{agree}_{\mu^{(r - 1)}} \left(f^{(r - 1)}, V^{(r - 1)}\right), \sqrt{\rho}(1 + 1/2m) \right) \\ & \color{blue}{(根据事件 E^{(i+1)} 的定义, 见式 (8.9) )} \\ & \le \max \left( \text{agree}_{\mu^{(r - 2)}} \left(f^{(r - 2)}, V^{(r - 2)}\right), \sqrt{\rho}(1 + 1/2m) \right) \\ & \le \ldots \\ & \le \max \left( \text{agree}_{\mu^{(0)}} \left(f^{(0)}, V^{(0)}\right), \sqrt{\rho}(1 + 1/2m) \right) \\ & = \max \left( \alpha, \sqrt{\rho}(1 + 1/2m) \right) \\ & = \alpha^{(0)}(\rho, m) \end{aligned}. \end{split}\]

至此,证得式 \((8.7)\) 成立,由此得证引理 8.2。 \(\Box\)

参考文献#

  • [BBHR18] Eli Ben-Sasson, Iddo Bentov, Yinon Horesh, and Michael Riabzev. “Fast Reed–Solomon Interactive Oracle Proofs of Proximity”. In: Proceedings of the 45th International Colloquium on Automata, Languages and Programming (ICALP), 2018.

  • [BCIKS20] Eli Ben-Sasson, Dan Carmon, Yuval Ishai, Swastik Kopparty, and Shubhangi Saraf. Proximity Gaps for Reed–Solomon Codes. In Proceedings of the 61st Annual IEEE Symposium on Foundations of Computer Science, pages 900–909, 2020.

  • [RVW13] Guy N. Rothblum, Salil Vadhan, and Avi Wigderson. Interactive proofs of proximity: delegating computation in sublinear time. In Proceedings of the forty-fifth annual ACM symposium on Theory of computing, pages 793–802. ACM, 2013.