Tropical Fan Sampler

φ³ · d = 3 · two-point function

Monte Carlo grows Feynman diagrams on the moduli space, weighs them, and anneals them toward the true integrand, live in your browser.

蒙特卡罗在模空间上长出费曼图、给它们加权,再退火到真实被积函数;一切都在浏览器里实时计算。

    Press Next step to watch one diagram being drawn, one edge at a time.

    按“下一步”,看一张图怎样一条边一条边地画出来。

    $w = {(\frac{U^{tr}}{U})}^{3/2}\,Π_e\,u_e^{δ}$

    $s_L = P_2(L)\,E_μ[w]$

    1 · Where the draw lands1 · 样本落在哪里

    Left: the moduli space as a tropical fan, one cone per diagram, glued along faces (5 cones drawn; each draw is placed in one of them). The needle is this draw’s top edge $t$. Right: the diagram this draw produced.左:画成热带扇的模空间,每张图一个锥,锥沿公共面粘合(只画出 5 个锥,每个样本画在其中一个里)。指针是这次样本的顶边 $t$。右:这次抽到的图。

    Left: the moduli space as a tropical fan (5 cones drawn). The teal trail is the annealing chain moving between cones; teal dots are where finished runs ended. Right: the chain’s current diagram.左:画成热带扇的模空间(只画出 5 个锥)。青色路径是退火链在锥之间的移动,青色点是已完成各轮的终点。右:链当前的图。

    Diagram tropical keeps one diagram fixed. Left: its cone, cut into Hepp sectors. Each wedge is one choice of the largest edge $σ_1$, with width equal to its exact probability; thin rays split it by the second largest $σ_2$. Right: the diagram, every edge numbered.单图热带固定一张图。左:它的锥,切成 Hepp 扇区。每个楔形对应最大边 $σ_1$ 的一种选择,宽度等于它的精确概率;细线再按第二大的边 $σ_2$ 细分。右:这张图,每条边都编了号。

    Tropical fan · moduli space热带扇 · 模空间One cone · Hepp sectors一个锥 · Hepp 扇区

    schematic · 5 of the many cones示意图 · 众多锥中的 5 个exact · wedge width = P($σ_1$)精确 · 楔形宽度 = P($σ_1$)

    Current diagram当前的图

    draw with large $w$$w$ 大的样本 draw with small $w$$w$ 小的样本 this draw’s $σ_1$, then $σ_2$这次的 $σ_1$,再到 $σ_2$ closes a loop (factor of $U^{tr}$)闭合一个圈($U^{tr}$ 的因子) chain state now (pulsing) and its trail链的当前状态(闪动)和它走过的路径 end points of finished runs已完成各轮的终点 needle = top edge $t$指针 = 顶边 $t$ external leg外腿 just drawn刚画出 just regrown (accepted)刚重长(被接受) rejected move被拒绝的移动

    2 · How the diagram is built2 · 这张图怎样造出来

    Left: the subset recursion. Each row is one step: its bar is split among the edges still left, in proportion to the probability that each one is the largest; the filled piece is the edge drawn. Right: the same diagram.左:子集递推。每一行是一步:横条按“剩下的每条边是最大边”的概率分给这些边,填色的那段就是抽中的边。右:同一张图。

    Left: the recursion tree. Right: the same diagram. Ring A is the loop through top edge A; each smaller ring is a piece drawn by the same rule. Hover either side to see the match.左:递推树。右:同一张图。环 A 是经过顶边 A 的圈;每个更小的环是按同样规则画出的一块。悬停任一边可看到对应关系。

    Recursion tree递推树Subset recursion子集递推

    one ring per loop · same letters as the diagram每个圈一个环 · 字母与图相同one row per step · 每步一行 · $P(e) = \frac{J(γ∖e)}{ω(γ∖e)\,J(γ)}$

    Current diagram当前的图

    top edge · A, B, C… name the rings顶边 · A、B、C… 是环的名字 bridge桥 closes a loop of what is left: factor of $U^{tr}$闭合剩余部分的一个圈:$U^{tr}$ 的因子 bridge of what is left: in the spanning tree剩余部分的桥:在生成树里 vertex顶点 external leg外腿 piece with loops → its own ring含圈的子图 → 自己的环 not drawn yet尚未画出 just drawn刚画出 planned, not drawn yet计划中,尚未画出 just drawn刚画出 just regrown (accepted)刚重长(被接受) rejected move被拒绝的移动 picked for this move这次移动选中的部分 grid at top right: column = round of $β$, row = move; teal kept, coral rejected, pale = edge redraw右上角格子:列 = $β$ 的回合,行 = 移动;青色保留,珊瑚色拒绝,浅色 = 重抽一条边
    1. A ring is one loop of the diagram. Its thick violet arc is the top edge, the largest $u$ in that piece; the same letter marks that edge in the diagram. Going round the rest of the ring you meet beads joined by gray bridges, and every bridge is shorter than the top edge.一个环就是图里的一个圈。粗紫弧是顶边,即这块里最大的 $u$;图里同一个字母标出这条边。沿环的其余部分是由灰色桥连起的珠子,每座桥都比顶边短。
    2. A bead is a vertex or a smaller piece. A dot is one vertex (an amber stub means it carries an external leg). A thick light segment is a piece that still has loops; it is drawn by the same rule as its own ring, one column to the right.珠子是一个顶点或一块更小的子图。点是单个顶点(琥珀短线表示带外腿);粗浅色段是仍含圈的子图,它按同样的规则画成右边一列里自己的环。
    3. Beads first. The tables first fix only the sizes: at every level, how many beads and how the loops and pendants are shared out (the dashed rings). Then each bead is strung completely before it is used, innermost first: its threads, then its clasp, and a smaller 1PI diagram is finished. Finished beads are strung on the next thread, and the next clasp closes them. Lengths are drawn in the units of the necklace being made (its clasp = 1) and scaled when it is hung in the next one, so this is the same exact draw, told from the inside out.先串珠子。先由表只定下大小:每一层有几颗珠子、圈和吊坠怎么分(虚线的环)。然后每颗珠子在用之前先完整地穿好,从最里面开始:先穿线,再扣上扣子,一个更小的 1PI 图就做好了。做好的珠子再穿到外一层的线上,由外一层的扣子扣起来。长度都以正在穿的那条项链为单位(它的扣子 = 1),挂到外一层时再整体缩放,所以这是同一个精确的抽样,只是从里往外讲。
    4. Each Next draws one edge. First a ring's top edge, whose dashed part is the plan for its chain; then, along the chain, its bridges and inner rings in order. The new edge is pink in the diagram, the tree and the bars. The click after the last edge is Weight: it gives $w$, drops the point on the fan and updates the estimate at once; % labels then show how far the true $U$ agrees with the tropical $U^{tr}$, and coral marks where $w$ loses weight.每按一次“下一步”画一条边。先画一个环的顶边(虚线部分是这条链的计划),再沿着链依次画桥和里面的环。新画的边在图、树和柱状图里都是粉色。最后一条边之后的那一下是“权重”:同时给出 $w$、把点落到扇上并更新估计值;百分比显示真实 $U$ 与热带 $U^{tr}$ 的一致程度,珊瑚色标出权重损失的地方。
    5. Each Next draws one edge, largest first. Among the edges still left, edge e is the largest with probability $\frac{J(γ∖e)}{ω(γ∖e)\,J(γ)}$, read off the table $J$ over all $2^E$ edge subsets of this one diagram. The first edge gets $α = 1$; each later one is smaller by a factor $ξ^{1/ω(γ)}$ with $ξ$ uniform on $(0,1)$. Violet means the edge closes a loop of what is left, so its $α$ is a factor of $U^{tr}$; gray means it is a bridge of what is left. The order of all $E$ edges is the Hepp sector.每按一次“下一步”画一条边,从大到小。在剩下的边里,边 e 是最大的概率为 $\frac{J(γ∖e)}{ω(γ∖e)\,J(γ)}$,查自这张图全部 $2^E$ 个边子集上的表 $J$。第一条边取 $α = 1$;之后每条边再乘上一个因子 $ξ^{1/ω(γ)}$($ξ$ 在 $(0,1)$ 上均匀分布)。紫色表示这条边闭合剩余部分的一个圈,它的 $α$ 是 $U^{tr}$ 的因子;灰色表示它是剩余部分的桥。全部 $E$ 条边的顺序就是 Hepp 扇区。
    6. The price is the table. $J$ has $2^E$ entries for every diagram: $2^{20} ≈ 10^6$ at $L = 7$, $2^{29} ≈ 5×10^8$ at $L = 10$, and the number of diagrams grows factorially. That is why this mode stops at $L = 7$, and why the paper moves on to global tropical sampling, which never lists diagrams.代价是这张表。每张图的 $J$ 有 $2^E$ 项:$L = 7$ 时 $2^{20} ≈ 10^6$,$L = 10$ 时 $2^{29} ≈ 5×10^8$,而图的数目按阶乘增长。所以这个模式只到 $L = 7$,论文也因此转向从不列举图的全局热带采样。
    7. Each Next is one step of the annealing chain. Raise $β$ a little and bank $Δβ\,\text{log}\,w$, or make one move: regrow a ring and everything inside it from the exact tropical rule (teal if accepted, coral if rejected), or redraw one edge. After $K$ steps of $β$ the run ends and its weight $W$ lands on the fan.每按一次“下一步”是退火链的一步。把 $β$ 升高一点并记下 $Δβ\,\text{log}\,w$,或者做一次移动:按精确的热带规则重长一个环和它里面的一切(被接受为青色,被拒绝为珊瑚色),或者重抽一条边。$β$ 走完 $K$ 步后这一轮结束,它的权重 $W$ 落到扇上。

    3 · What is actually sampled3 · 真正被抽的是什么

    Left: the recursion tree again. Right: the $u$ of every edge, largest first.左:还是递推树。右:每条边的 $u$,从大到小。Left: the subset recursion again. Right: the $α$ of every edge, largest first.左:还是子集递推。右:每条边的 $α$,从大到小。

    Recursion tree递推树Subset recursion子集递推

    one ring per loop · same letters as the diagram每个圈一个环 · 字母与图相同one row per step · 每步一行 · $P(e) = \frac{J(γ∖e)}{ω(γ∖e)\,J(γ)}$

    Its Schwinger parameters它的 Schwinger 参数

    largest first: this order is the Hepp sector从大到小:这个顺序就是 Hepp 扇区

    top edge of a ring (letter = the ring)环的顶边(字母 = 那个环) bridge桥 closes a loop: exactly the rings’ top edges, and $U^{tr}$ is their product闭合一个圈:恰好就是各环的顶边,$U^{tr}$ 是它们的乘积 closes a loop: $U^{tr}$ is the product of these $α$闭合一个圈:$U^{tr}$ 是这些 $α$ 的乘积 7 under a bar = the edge’s number in the diagram柱下的数 = 这条边在图里的编号 96 above a bar or on an edge = agreement with $U^{tr}$ (%)柱上或边上的数 = 与 $U^{tr}$ 的一致程度(%)

    This is all the sampler draws: a tree of rings and one $u$ per edge. The probability of a draw and the tropical weight $U^{tr} = Π\,(\text{top edge of each ring})$ follow from these alone, so no list of diagrams is ever made. The diagram is glued together from the tree only to compute the true $U$, and with it $w = {(\frac{U^{tr}}{U})}^{3/2}\,Π_e\,u_e^{δ}$.

    采样器抽的只有这些:一棵环的树,加上每条边一个 $u$。一个样本的概率和热带权重 $U^{tr} = Π$(每个环的顶边)只由它们决定,所以从不需要图的列表。只有在计算真实的 $U$、从而得到 $w = {(\frac{U^{tr}}{U})}^{3/2}\,Π_e\,u_e^{δ}$ 时,才把树拼成一张具体的图。

    This is all diagram tropical draws: an order of the $E$ edges (the Hepp sector) and $E−1$ ratios. In that sector $U^{tr}$ is the product of the violet $α$’s, so $w = {(\frac{U^{tr}}{U})}^{3/2}(Σ_e α_e)^{−ω} ≤ 1$, and $s_G = \frac{Γ(ω)\,J(G)}{|\text{Aut}\,G|}\,\text{mean}(w)$. It is exact and precise for one diagram, but it needs the whole table $J$ of this diagram first, and every diagram must be done separately.

    单图热带抽的只有这些:$E$ 条边的一个顺序(Hepp 扇区)和 $E−1$ 个比值。在这个扇区里 $U^{tr}$ 是紫色各 $α$ 的乘积,所以 $w = {(\frac{U^{tr}}{U})}^{3/2}(Σ_e α_e)^{−ω} ≤ 1$,$s_G = \frac{Γ(ω)\,J(G)}{|\text{Aut}\,G|}\,\text{mean}(w)$。对一张图它精确而且精度高,但先要算出这张图的整张表 $J$,而且每张图都得单独做一遍。

    Estimates估计值

    same $L$, two samplers, one reference同一个 $L$,两种采样,一个参考值one diagram, two samplers同一张图,两种采样

    method方法$ŝ_L$$ŝ_G$rel. error相对误差pull vs ref.与参考值的偏差samples样本数$\text{Var}\,\text{log}\,w$largest share最大份额
    Diagram tropical单图热带———0——
    Global tropical, this diagram全局热带,这张图———0——
    Global tropical全局热带———0——
    Annealed tropical退火热带———0——

    “Largest share” is the biggest single weight divided by the sum of all weights. Near 1 means one sample dominates and the error bar cannot be trusted.“最大份额”是最大的单个权重除以所有权重之和。接近 1 表示一个样本主导,误差不可信。

    Log-weight spread权重分布

    Amber: $\text{log}\,w$ of independent tropical draws. Teal: $\text{log}\,W$ of annealed tropical runs. A narrower teal curve is the whole point of annealing.琥珀色是独立热带样本的 $\text{log}\,w$,青色是退火热带各轮的 $\text{log}\,W$;青色更窄就是退火的意义。

    Amber: $\text{log}\,w$ of global tropical draws, all diagrams together. Violet: $\text{log}\,w$ of diagram tropical draws of the chosen diagram. These are different weights: $(U^{tr}/U)^{3/2}\,Π_e u_e^{δ}$ on the cube for the first, $(U^{tr}/U)^{3/2}\,(Σ_e α_e)^{−ω}$ on one projective diagram for the second.琥珀色是全局热带样本的 $\text{log}\,w$,所有图放在一起;紫色是所选图的单图热带样本的 $\text{log}\,w$。两者是不同的权重:前者是立方体上的 $(U^{tr}/U)^{3/2}\,Π_e u_e^{δ}$,后者是单张图射影坐标下的 $(U^{tr}/U)^{3/2}\,(Σ_e α_e)^{−ω}$。

    Each diagram’s share每张图的贡献

    from the global tropical draws at this $L$来自这个 $L$ 的全局热带样本from the annealed runs at this $L$来自这个 $L$ 的各轮退火

    #diagram图tropical measure picks it热带测度抽中的频率runs start here (violet) / end here (teal)从这里出发(紫)/ 在这里结束(青)true share of $s_L$在 $s_L$ 中的真实份额its contribution $s_G$这张图的贡献 $s_G$mean $w$平均 $w$mean $W$ of runs ending here在此结束的各轮平均 $W$

    Click a row to integrate that diagram on its own with diagram tropical sampling (the highlighted row is the one in use).点一行,就用单图热带单独积分那张图(高亮的一行是当前所用的图)。

    The tropical measure picks diagram $G$ with probability $p_G$ (violet). Its true share of $s_L$ is $p_G\,E[w|G]/E[w]$ (amber). The weight $w$ is exactly what turns the violet bar into the amber one.热带测度以概率 $p_G$ 抽中图 $G$(紫色);它在 $s_L$ 中的真实份额是 $p_G\,E[w|G]/E[w]$(琥珀色)。权重 $w$ 正是把紫色变成琥珀色的那一步。

    Each run starts at a tropical draw (violet), and annealing moves the chain toward the true measure, so where runs end (teal) drifts toward the true share (amber). The contribution uses the weight $W$ of each run at its end: $s_G = P_2\,\text{mean}(W\,[\text{run ends in }G])$, unbiased for any $K$.每一轮从一个热带样本出发(紫色),退火把链推向真实测度,所以各轮的终点(青色)向真实份额(琥珀色)靠拢。贡献用每轮终点的权重 $W$:$s_G = P_2\,\text{mean}(W\,[\text{该轮停在 }G])$,对任何 $K$ 都无偏。

    The space空间

    Every 1PI two-point cubic diagram $G$ with $E = 3L−1$ edges is one cone, the cube $(0,1)^E/\text{Aut}\,G$ in $u = 1 − e^{−α}$. Cones meet at the UV apex and are glued where an edge length shrinks to zero. Inside a cone, $U^{tr}$ is a single monomial in each Hepp sector.

    每张 $E = 3L−1$ 条边的三价 1PI 两点图是一个锥,即 $u = 1 − e^{−α}$ 坐标下的 $(0,1)^E/\text{Aut}\,G$。所有锥在 UV 顶点相交,并在某条边长缩为零的地方粘合。锥内每个 Hepp 扇区里 $U^{tr}$ 只是一个单项式。

    $μ_δ = \frac{Π_e\,u_e^{−δ}\,du_e}{(U^{tr})^{3/2}\,|\text{Aut}\,G|}, \quad ω_δ = (1−δ)E − \frac{3L}{2}$

    Diagram tropical单图热带

    Borinsky’s method integrates one diagram at a time. In each Hepp sector, an ordering of the $α$’s, $U$ is replaced by its largest monomial $U^{tr}$, which can be integrated exactly. A recursion over all $2^E$ edge subsets gives the total $J(G)$ and draws a sector edge by edge, largest first; the ratios follow from uniform numbers. The weight $w ≤ 1$ corrects for $U^{tr} ≤ U$.

    Borinsky 的方法一次积分一张图。在每个 Hepp 扇区($α$ 的一种排序)里,把 $U$ 换成它最大的单项式 $U^{tr}$,后者可以精确积分。对全部 $2^E$ 个边子集的递推给出总量 $J(G)$,并从大到小逐条边抽出一个扇区;比值由均匀随机数得到。权重 $w ≤ 1$ 修正 $U^{tr} ≤ U$ 的差别。

    $J(γ) = Σ_{e∈γ} \frac{J(γ∖e)}{ω(γ∖e)}, \quad ω(γ) = |γ| − \frac{3}{2}L(γ), \quad s_G = \frac{Γ(ω)\,J(G)}{|\text{Aut}\,G|}\,E[w]$

    Global tropical全局热带

    Draw the largest edge $t$ first. Cutting it leaves a chain of beads joined by bridges below $t$, and each bead is a smaller 1PI graph drawn the same way one scale down. The draw is exact from $μ_δ$, so no list of diagrams is ever needed. Then reweight by $w ≤ 1$.

    先抽最大的边 $t$。剪开它,剩下一条由桥连起的珠子链,桥都小于 $t$;每个珠子是更小的 1PI 图,在下一个尺度上用同样的规则抽。抽样精确服从 $μ_δ$,不需要图的列表。最后用 $w ≤ 1$ 加权。

    $ω_δ(n,L)\,P_n(L) = \frac{1}{2}\,\text{ch}(n, L−1), \quad ŝ_L = P_2(L)\,\text{mean}(w)$

    Annealed tropical退火热带

    Start from one exact tropical draw and raise $β$ from 0 to 1 in $K$ steps, banking $Δβ\,\text{log}\,w$ at each step. In between, Markov moves leave $μ_δ\,w^β$ invariant: regrow a whole subtree from the exact tropical rule (a jump to another wedge), or redraw one edge (same wedge). Accept with $(w′/w)^β$.

    从一个精确的热带样本出发,用 $K$ 步把 $β$ 从 0 升到 1,每步记下 $Δβ\,\text{log}\,w$。中间用保持 $μ_δ\,w^β$ 不变的马尔可夫移动:按精确热带规则重长一整棵子树(跳到另一个楔形),或重抽一条边(留在同一楔形)。接受率 $(w′/w)^β$。

    $E[W] = \frac{s_L}{P_2(L)}\ \text{for every}\ K, \quad \text{Var}\,\text{log}\,W ≈ V_f + \frac{c}{K}$