Scroll

Categories

Tags

Monthly Archives

Keyword

Back

2018.07.12

# 6. 变体

In Section 2 we described the PHANTOM’s greedy algorithm to mark blocks as blue or red (Algorithm 1). In fact, similar algorithms can provide similar guarantees. We describe several such variants below and explain the intuition behind them. All these variants can be thought of as greedy approximations to the Maximum $k$-cluster SubDAG problem described in Section 1.

A. Choosing the maximizing tip
In our original version of the colouring procedure, we chose the tip which has the highest score (Algorithm 1, line 6). Instead, Algorithm 3 below chooses the tip for which the score of the (virtual block of the) current DAG would be highest:

A.选择最大化末端 在我们的着色程序的原始版本中，我们选择了得分最高的末端（算法 1，行 6）。相反，下面的算法3选择当前DAG（的虚拟块）的得分最高的末端：

Algorithm 3 Selection of a blue set

Input: G – a block DAG, k – the propagation parameter

Output: $BLUE_k(G)$ – the dense-set of G

1. function CALC-BLUE(G, k )

2. if $G == {genesis}$ then

3. return {$genesis$}

4. for $B \in tips(G)$ do

5. $BLUE_k(B) \leftarrow CALC-BLUE(past (B) , k)$

6. $S_B \leftarrow BLUE_k(B) \cup {B}$

7. for $C \in anticone(B)$ in some arbitrary order do

8.  if $anticone(C) \cap S_B \leq k$ then
9. add C to $S_B$

10.  return arg $max{ S_B :B \in tips(G)}$ (and break ties arbitrarily)

B. Adding blocks to the greedily chosen blue set Another variant over the previous algorithms is the procedure to mark more blocks as blue, after the best tip and its blue set were selected. Previously, we required that the candidate block admit an anticone of size at most k in the current set. Instead, we can require that its anticone be counted only inside the originally chosen blue set. Thus, yet another variant is to replace lines 8-9 with

B.将块添加到贪婪选择出的蓝色集合 在选择最优末端和蓝色集合之后，先前算法的另一个变体是将更多块标记为蓝色的过程。在此之前，我们要求候选块在当前集合中承认一个至多为k的反锥体。相反，我们可以要求只在最初选择的蓝色集合内计算反锥体。因此，另一个变式将代替上面的8-9行

1.  if $anticone(C) \cap (BLUE_k(B) \cup {B}) \leq k$ then
2. add C to $S_B$

Observe that the resulting blue set may contain blocks which have an anticone larger than k within the set.

Finally, another viable alternative is to further relax the condition (in lines 8-9) by counting the anticone of the candidate B only against the resulting chain of blue blocks, $Chn(past(B))\cup {B}$, where $B$ is the selected tip:

1.  if $anticone(C) \cap Chn(past(B))\cup {B} \leq k$ then
2. add C to $S_B$

Compare this variant to Satoshi’s longest-chain rule, a rule that can be described as follows: each block in a chain $Chn$ increases its weight by 1, and we select the highest scoring chain. In contrast, in our last variant, each block whose gap from a chain $Chn$ is at most $k$ increases its weight by 1, and we select the highest scoring chain.

C. Iterative elimination of blocks We now introduce another variant based on an iterative method common in combinatorial optimization. The algorithm works as follows: Given a block DAG $G$, we iteratively eliminate from the blue set the block with largest blue anticone, and continue doing so until all blue blocks admit a blue anticone of size $k$ at most:

C.块的迭代消除 我们现在介绍另一种变体，基于一种常用于组合优化的迭代方法。算法描述如下：已知一个块DAG $G$，我们从蓝色集合中迭代消除有最大蓝色反锥体的块，并且一直重复直到所有蓝色块承认一个最大为$k$的蓝色反锥体：

Algorithm 4 Selection of a blue set

Input: G=(V,E) – a block DAG, k – the propagation parameter

Output: $BLUE_k(G)$ – the dense-set of G

1. function CALC-BLUE(G, k )

2. $BLUE_k(G) \leftarrow V$

3.  while $\exists B \in BLUE_k(G)$ with $anticone \cap BLUE_k(G) > k$ do
4.  $C \leftarrow ard_B\ max{ anticone(B)\cap BLUE_k(G) }$(with arbitrary tie-breaking)
5. remove $C$ from $BLUE_k(G)$

6. return $BLUE_k(G)$

We conjecture that Algorithm 1 can be replaced with each of the greedy algorithms described in this section. To this end, suffice it to repeat the proof of Theorem 5 with respect to these variants. We suspect that the same proof technique, namely, the use of Hourglass, would prove useful.

Top