量子多体态的神经网络表示（一）

这是一大堆公式来袭！！！！

神经网络量子态

设$Q_1,Q_2,\dots ,Q_N$是具有尺寸为 $d_1,d_2,\dots ,d_N$ 的状态空间$H_1,H_2,\dots ,H_N$ 的 $N$ 个量子系统。考虑Q1，Q2，…，QN的复合系统 Q 的状态空间为$\mathcal{H}={\mathcal{H}}_1\otimes {\mathcal{H}}_2\otimes \dots \otimes {\mathcal{H}}_N$。设$S_1,S_2,\dots ,S_N$分别是系统$Q_1,Q_2,\dots ,Q_N$的非简并观测量，那么$\mathcal{S}={\mathcal{S}}_1\otimes {\mathcal{S}}_2\otimes \dots \otimes {\mathcal{S}}_N$是复合系统 $Q$ 的可观测值，于是

$$S_j|{\psi }_{k_j}⟩={\lambda }_{k_j}|{\psi }_{k_j}⟩\left(k_j=0,1,\dots ,d_j-1\right)$$

很容易知道$S=S_1\otimes S_2\otimes \dots \otimes S_N$对应的特征值${\lambda }_{k_1}{\lambda }_{k_2}\dots {\lambda }_{k_N}$和本征基$|{\psi }_{k_1}⟩\otimes |{\psi }_{k_2}⟩\otimes \dots \otimes |{\psi }_{k_N}⟩\left(k_j=0,1,\dots ,d_j-1\right)$

令输入空间为;
$$V(S)=\left\{{\Lambda }_{k_1{k}_2\dots k_N}\equiv {\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)}^{\mathrm{T}}:k_j=0,1,\dots ,d_j-1\right\}$$
其中参数$a={\left(a_1,a_2,\dots ,a_N\right)}^{\mathrm{T}}\in {\mathbb{C}}^N$,$b={\left(b_1,b_2,\dots ,b_M\right)}^{\mathrm{T}}\in {\mathbb{C}}^M$,$W=\left[W_{ij}\right]\in {\mathbb{C}}^{M\times N}$
令 $\Omega =(a,b,W)$，于是可以获得神经网络量子波函数neural network quantum wave function（NNQWF）：

$$\begin{array}{l}{\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)\\ ={\sum }_{h_i=\pm 1}\exp \left({\sum }_{j=1}^N{a}_j{\lambda }_{k_j}+{\sum }_{i=1}^M{b}_i{h}_i+{\sum }_{i=1}^M{\sum }_{j=1}^N{W}_{ij}{h}_i{\lambda }_{k_j}\right)\end{array}$$

对于一些${\Lambda }_{k_1}{k}_2\dots k_N$，假设${\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)\ne 0$，定义：
$$\begin{array}{rl}|{\Psi }_{S,\Omega }⟩=& \sum _{{\Lambda }_{k_1{k}_2\dots k_N}\in V(S)}{\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)\cdot |{\psi }_{k_1}⟩\otimes |{\psi }_{k_2}⟩\otimes \dots \otimes |{\psi }_{k_N}⟩\end{array}$$

它是希尔伯特空间H的非零向量（不一定归一化）。我们称其为由参数 $\Omega =(a,b,W)$ 和输入可观察到的 $S_1,S_2,\dots ,S_N$ 引起的神经网络量子态（NNQS）:

对NNQS进行编码的人工神经网络，具有一组N个可见的人工神经元（蓝色磁盘）和一组M个隐藏的神经元（黄色磁盘）， 对于输入的可观察到的S的每个值${\Lambda }_{k_1}{k}_2\dots k_N$，神经网络计算${\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)$的值。

使用内积符号，可以写成:
$$\begin{array}{l}{\sum }_{j=1}^N{a}_j{\lambda }_{k_j}+{\sum }_{i=1}^M{b}_i{h}_i+{\sum }_{i=1}^M{\sum }_{j=1}^N{W}_{ij}{h}_i{\lambda }_{k_j}\\ =⟨{\Lambda }_{k_1{k}_2\dots k_N},a⟩+⟨h,b+W{\Lambda }_{k_1{k}_2\dots k_N}⟩\end{array}$$
即${\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)={\mathrm{e}}^{⟨{\Lambda }_{k_1{k}_2,k_N},a⟩}\cdot {\sum }_{h_i=\pm 1}{\mathrm{e}}^{⟨h,b+W{\Lambda }_{k_1}{k}_2\dots k_N⟩}$

NNQWF的另外一种形式
已知${\mathrm{e}}^{⟨{\Lambda }_{k_1{k}_2\dots k_N},a⟩}={\mathrm{e}}^{{\sum }_{j=1}^N{a}_j{\lambda }_{k_j}}={\prod }_{j=1}^N{\mathrm{e}}^{a_j{\lambda }_{k_j}}$
令$x_i=b_i+{\sum }_{j=1}^N{W}_{ij}{\lambda }_{k_j}$

$$\begin{array}{l}{\sum }_{h_i=\pm 1}{\mathrm{e}}^{⟨h,b+W{\Lambda }_{k_1{k}_2\dots k_N}⟩}\\ ={\sum }_{h_i=\pm 1}{\mathrm{e}}^{{\sum }_{i=1}^M{h}_i\left[b_i+{\sum }_{j=1}^N{W}_{ij}{\lambda }_{k_j}\right]}\\ ={\sum }_{h_i=\pm 1}{\prod }_{i=1}^M{\mathrm{e}}^{x_i{h}_i}\\ ={\sum }_{h_1=\pm 1}\dots {\sum }_{h_M=\pm 1}{\mathrm{e}}^{x_1{h}_1}{\mathrm{e}}^{x_2{h}_2}\cdots {\mathrm{e}}^{x_M{h}_M}\\ =\left({\mathrm{e}}^{x_1}+{\mathrm{e}}^{-x_1}\right)\left({\mathrm{e}}^{x_2}+{\mathrm{e}}^{-x_2}\right)\cdots \left({\mathrm{e}}^{x_M}+{\mathrm{e}}^{-x_M}\right)\\ ={\prod }_{i=1}^{M}2\cosh \left(x_i\right)\end{array}$$

注释：$\cosh x=\frac{e^x+e^{-x}}{2}=\frac{e^{2x}+1}{2e^x}=\frac{1+e^{-2x}}{2e^{-x}}$

于是：
$$\begin{array}{l}{\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)\\ ={\prod }_{j=1}^N{\mathrm{e}}^{a_j{\lambda }_{k_j}}\cdot {\prod }_{i=1}^{M}2\cosh \left(b_i+{\sum }_{j=1}^N{W}_{ij}{\lambda }_{k_j}\right).\end{array}$$

我们将此网络称为量子人工神经网络，因为它的输入是量子可观察量的特征值，结果是NNQWF的值，而它的网络结构类似于通常的人工神经网络:

NNQS的张量积

当 $M>N$ ，$a\in {\mathbb{C}}^N,b\in {\mathbb{C}}^N,W\in {\mathbb{C}}^{N\times N}$
令
$$b^{\prime \prime }=\left(\begin{array}{c}\frac{\pi \mathrm{i}}{3}\\ \frac{\pi \mathrm{i}}{3}\\ ⋮\\ \frac{\pi \mathrm{i}}{3}\end{array}\right)\in {\mathbb{C}}^{M-N}$$

$$b^{\prime }=\left(\begin{array}{c}b\\ b^{\prime \prime }\end{array}\right)\in {\mathbb{C}}^M$$

$$W^{\prime }=\left(\begin{array}{c}W\\ 0\end{array}\right)=\left[W_{ij}^{\prime }\right]\in {\mathbb{C}}^{M\times N}$$

定义一个新的参数：${\Omega }^{\prime }=\left(a,b^{\prime },W^{\prime }\right)$

$$\begin{array}{l}{b}_i^{\prime }=b_i(i=1,2,\dots ,N),\\ b_i^{\prime }=\frac{\pi \mathrm{i}}{3}(i=N+1,N+2,\dots ,M)\\ W_{ij}^{\prime }=W_{ij}(1\le i\le N,1\le j\le N)\\ W_{ij}^{\prime }=0(i=N+1,N+2,\dots ,M,j=1,2,\dots ,N)\end{array}$$

当$i$大于$N$时
$$2\cosh \left(b_i^{\prime }+\sum _{j=1}^N{W}_{ij}^{\prime }{\lambda }_{k_j}\right)=2\cosh \frac{\pi \mathrm{i}}{3}=2\cos \frac{\pi }{3}=1$$

所以
$$\prod _{i=1}^{M}2\cosh \left(b_i^{\prime }+\sum _{j=1}^N{W}_{ij}^{\prime }{\lambda }_{k_j}\right)=\prod _{i=1}^{N}2\cosh \left(b_i+\sum _{j=1}^N{W}_{ij}{\lambda }_{k_j}\right)$$

$${\Psi }_{S,{\Omega }^{\prime }}\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right)={\Psi }_{S,\Omega }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_N}\right),\forall {\Lambda }_{k_1{k}_2\dots k_N}$$

同理
$$\begin{array}{l}{S}^{\prime }=S_1^{\prime }\otimes \dots \otimes S_{N^{\prime }}^{\prime },S^{\prime \prime }=S_1^{\prime \prime }\otimes \dots \otimes S_{N^{\prime \prime }}^{\prime \prime }\\ {\Omega }^{\prime }=\left(a^{\prime },b^{\prime },W^{\prime }\right),{\Omega }^{\prime \prime }=\left(a^{\prime \prime },b^{\prime \prime },W^{\prime \prime }\right)\end{array}$$

对于
$$\begin{array}{l}S=S^{\prime }\otimes S^{\prime \prime },N=N^{\prime }+N^{\prime \prime },M=M^{\prime }+M^{\prime \prime }\\ a=\left(\begin{array}{c}{a}^{\prime }\\ a^{\prime \prime }\end{array}\right),b=\left(\begin{array}{c}{b}^{\prime }\\ b^{\prime \prime }\end{array}\right),\\ W=\left[W_{ij}\right]=\left(\begin{array}{cc}{W}_{M^{\prime }\times N^{\prime }}^{\prime }& 0\\ 0& W_{M^{\prime \prime }\times N^{\prime \prime }}^{\prime \prime }\end{array}\right),\Omega =(a,b,W)\end{array}$$

$$\begin{array}{rl}|{\Psi }_{S^{\prime },{\Omega }^{\prime }}^{\prime }⟩=& \sum _{{\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_{N^{\prime }}}\right)}^{\mathrm{T}}\in V\left(S^{\prime }\right)}{\Psi }_{S^{\prime },{\Omega }^{\prime }}^{\prime }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_{N^{\prime }}}\right)\cdot |{\psi }_{k_1}⟩\otimes |{\psi }_{k_2}⟩\otimes \dots \otimes |{\psi }_{k_{N^{\prime }}}⟩\end{array}$$

$$\begin{array}{rl}|{\Psi }_{S^{\prime \prime },{\Omega }^{\prime \prime }}^{\prime \prime }⟩=& \sum _{{\left({\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)}^{\mathrm{T}}\in V\left(S^{\prime \prime }\right)}{\Psi }_{S^{\prime \prime },{\Omega }^{\prime \prime }}^{\prime \prime }\left({\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)\cdot |{\varphi }_{m_1}⟩\otimes |{\varphi }_{m_2}⟩\otimes \dots \otimes |{\varphi }_{m_{N^{\prime \prime }}}⟩\end{array}$$

$${\Psi }_{S^{\prime },{\Omega }^{\prime }}^{\prime }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_{N^{\prime }}}\right)\cdot {\Psi }_{S^{\prime \prime },{\Omega }^{\prime \prime }}^{\prime \prime }\left({\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)$$

$$\begin{array}{rl}=& \left(\prod _{j=1}^{N^{\prime }}{\mathrm{e}}^{a_j^{\prime }{\lambda }_{k_j}}\right)\cdot \left(\prod _{j=1}^{N^{\prime \prime }}{\mathrm{e}}^{a_j^{\prime \prime }{\mu }_{m_j}}\right)\cdot \left(\prod _{i=1}^{M^{\prime }}2\cosh \left(b_i^{\prime }+\sum _{j=1}^{N^{\prime }}{W}_{ij}^{\prime }{\lambda }_{k_j}\right)\right)\\ & \cdot \left(\prod _{i=1}^{M^{\prime \prime }}2\cosh \left(b_i^{\prime \prime }+\sum _{j=1}^{N^{\prime \prime }}{W}_{ij}^{\prime \prime }{\mu }_{m_j}\right)\right)\\ =& \left(\prod _{j=1}^N{\mathrm{e}}^{a_j{\xi }_{j_j}}\right)\cdot \left(\prod _{i=1}^{M}2\cosh \left(b_i+\sum _{j=1}^N{W}_{ij}{\xi }_j\right)\right)\\ =& {\Psi }_{S,\Omega }\left({\xi }_{l_1},{\xi }_{l_2},\dots ,{\xi }_{l_N}\right)\end{array}$$
其中$\left({\xi }_{l_1},{\xi }_{l_2},\dots ,{\xi }_{l_N}\right)=\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_{N^{\prime }}},{\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)$

$$\begin{array}{l}|{\Psi }_{S^{\prime },{\Omega }^{\prime }}^{\prime }⟩\otimes |{\Psi }_{S^{\prime \prime },{\Omega }^{\prime \prime }}^{\prime \prime }⟩\\ ={\sum }_{{\left({\lambda }_1,{\lambda }_{k_2},\dots ,{\lambda }_{N^{\prime }}\right)}^{\mathrm{T}}\in V\left(S^{\prime }\right){\left({\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)}^{\mathrm{T}}\in V\left(S^{\prime \prime }\right)}\\ {\Psi }_{S^{\prime },{\Omega }^{\prime }}^{\prime }\left({\lambda }_{k_1},{\lambda }_{k_2},\dots ,{\lambda }_{k_{N^{\prime }}}\right)\cdot {\Psi }_{S^{\prime \prime },{\Omega }^{\prime \prime }}^{\prime \prime }\left({\mu }_{m_1},{\mu }_{m_2},\dots ,{\mu }_{m_{N^{\prime \prime }}}\right)\\ \cdot |{\psi }_{k_1}⟩\otimes |{\psi }_{k_2}⟩\otimes \dots \otimes |{\psi }_{k_{N^{\prime }}}⟩\otimes |{\varphi }_{m_1}⟩\otimes |{\varphi }_{m_2}⟩\otimes \dots \otimes |{\varphi }_{m_{N^{\prime \prime }}}⟩\\ ={\sum }_{{\left({\xi }_l,{\xi }_{l_2},\dots ,{\xi }_{l_N}\right)}^{\mathrm{T}}\in V(S)}{\Psi }_{S,\Omega }\left({\xi }_{l_1},{\xi }_{l_2},\dots ,{\xi }_{l_N}\right)|{\phi }_{l_1}⟩\otimes \dots \otimes |{\phi }_{l_N}⟩\\ =|{\Phi }_{S,\Omega }⟩\end{array}$$

这表明两个NNQS的张量积也是NNQS

局部统一运算（LUO）对NNQS的影响

有前面的内容我们知道

$$|{\Psi }_{S,\Omega }⟩=\sum _{{\Lambda }_{k_1,u_N}\in V(S)}{\Psi }_{S,\Omega }\left({\lambda }_{k_1},\dots ,{\lambda }_{k_N}\right)|{\psi }_{k_1}⟩\otimes \dots \otimes |{\psi }_{k_N}⟩$$

令$U=U_1\otimes \dots \otimes U_N$,因为 $U_i{S}_i{U}_i^{†}|{\varphi }_{k_i}⟩=U_i{S}_i|{\psi }_{k_i}⟩=U_i\left({\lambda }_{k_i}|{\psi }_{k_i}⟩\right)={\lambda }_{k_i}|{\varphi }_{k_i}⟩$

于是
$$\begin{array}{l}U|{\Psi }_{S,\Omega }⟩\\ ={\sum }_{{\Lambda }_{k_1,k_N}\in V(S)}{\Psi }_{S,\Omega }\left({\lambda }_{k_1},\dots ,{\lambda }_{k_N}\right)|U_1{\psi }_{k_1}⟩\otimes \dots \otimes |U_N{\psi }_{k_N}⟩\\ ={\sum }_{{\Lambda }_{k_1\dots k_N}\in V\left(US{U}^{†}\right)}{\Psi }_{US{U}^{†},\Omega }\left({\lambda }_{k_1},\dots ,{\lambda }_{k_N}\right)|{\varphi }_{k_1}⟩\otimes \dots \otimes |{\varphi }_{k_N}⟩\\ =|{\Psi }_{US{U}^{†},\Omega }⟩\end{array}$$

这表明结果状态$U|{\Psi }_{S,\Omega }⟩$也是具有输入可观察到的$US{U}^{†}$和参数$\Omega$的NNQS，并且与$U|{\Psi }_{S,\Omega }⟩$具有相同的NNQWF。

当$S={\sigma }_1^z\otimes {\sigma }_2^z\otimes \dots \otimes {\sigma }_N^z$

∣ Ψ S , Ω ⟩ = ∑ Λ k 1 k 2 , k N ∈ { − 1 , 1 } N ( ∏ i = 1 M 2 Λ i ) ∣ ψ k 1 ⟩ ⊗ ∣ ψ k 2 ⟩ ⊗ … ⊗ ∣ ψ k N ⟩ \left|\Psi_{S, \Omega}\right\rangle=\sum_{\Lambda_{k_{1}} k_{2}, k_{N} \in\{-1,1\}^{N}}\left(\prod_{i=1}^{M} 2 \Lambda_{i}\right)\left|\psi_{k_{1}}\right\rangle \otimes\left|\psi_{k_{2}}\right\rangle \otimes \ldots \otimes\left|\psi_{k_{N}}\right\rangle ∣ΨS,Ω​⟩=Λk1​​k2​,kN​∈{−1,1}N∑​(i=1∏M​2Λi​)∣ψk1​​⟩⊗∣ψk2​​⟩⊗…⊗∣ψkN​​⟩

其中${\Lambda }_i=\cos \left(b+{\omega }_1{\lambda }_{k_{i-1}}+{\omega }_0{\lambda }_{k_i}+{\omega }_{-1}{\lambda }_{k_{i+1}}\right)$