问题:如何进行下图中的转换:
在 “TEBD算法计算一维格点模型基态”以及 “MPO与一维热力学计算”中我们都提出了以上的问题,现在让我们来看看这个问题是如何解决的。
通过研读论文“Linearized Tensor Renormalization Group Algorithm for Thermodynamics of Quantum LatticeModels”(量子点阵模型热力学的线性张量重整化群算法(LTRG))我们将对整个LTRG算法会更加的熟悉。
LTRG算法:采用了无限时间演化的分组抽取技术,使我们能直接处理转移矩阵,使其更具有可扩展性。
step1:得到张量网络
①首先,我们给出一维量子多体模型中哈密顿量的形式:
H
=
∑
i
=
1
N
h
i
,
i
+
1
=
H
1
+
H
2
H
1
=
∑
i
=
1
N
/
2
h
2
i
−
1
,
2
i
,
H
2
=
∑
i
=
1
N
/
2
h
2
i
,
2
i
+
1
\begin{aligned} H &=\sum_{i=1}^{N} h_{i, i+1}=H_{1}+H_{2} \\ H_{1} &=\sum_{i=1}^{N / 2} h_{2 i-1,2 i}, H_{2}=\sum_{i=1}^{N / 2} h_{2 i, 2 i+1} \end{aligned}
HH1=i=1∑Nhi,i+1=H1+H2=i=1∑N/2h2i−1,2i,H2=i=1∑N/2h2i,2i+1我们将
H
1
,
H
2
H_1,H_2
H1,H2展开可知:
H
1
=
h
1
,
2
+
h
3
,
4
+
h
5
,
6
+
.
.
.
+
h
N
−
1
,
N
H
2
=
h
2
,
3
+
h
4
,
5
+
h
6
,
7
+
.
.
.
+
h
N
,
N
+
1
H_1 = h_{1,2} +h_{3,4}+h_{5,6}+...+h_{N-1,N}\\H_2 = h_{2,3} +h_{4,5}+h_{6,7}+...+h_{N,N+1}
H1=h1,2+h3,4+h5,6+...+hN−1,NH2=h2,3+h4,5+h6,7+...+hN,N+1即:
H
1
H_1
H1为作用在奇偶上的阿密顿量的和,
H
2
H_2
H2为作用在偶奇上的阿密顿量的和(即将哈密顿量按格点编号分为两部分)
其中,
N
N
N为格点的个数且为偶数
我们添加
2
K
2K
2K(
K
K
K很大)个完备基:
{
∣
σ
i
j
⟩
}
(
σ
i
j
=
1
,
⋯
,
D
)
\left\{\left|\sigma_{i}^{j}\right\rangle\right\}\left(\sigma_{i}^{j}=1, \cdots, D\right)
{∣∣∣σij⟩}(σij=1,⋯,D)其中,
i
i
i为格点指标,
j
j
j为
T
r
o
t
t
e
r
Trotter
Trotter指标。
②于是我们可以将该模型的配分函数表示为:
Z
N
≃
Tr
[
e
−
β
H
1
/
K
e
−
β
H
2
/
K
]
K
=
∑
{
σ
i
j
}
∏
j
=
1
K
⟨
σ
1
2
j
−
1
…
σ
N
2
j
−
1
∣
e
−
β
H
1
/
K
∣
σ
1
2
j
…
σ
N
2
j
⟩
×
⟨
σ
1
2
j
…
σ
N
2
j
∣
e
−
β
H
2
/
K
∣
σ
1
2
j
+
1
…
σ
N
2
j
+
1
⟩
\begin{aligned} Z_{N} & \simeq \operatorname{Tr}\left[e^{-\beta H_{1} / K} e^{-\beta H_{2} / K}\right]^{K} \\ &=\sum_{\left\{\sigma_{i}^{j}\right\}} \prod_{j=1}^{K}\left\langle\sigma_{1}^{2 j-1} \ldots \sigma_{N}^{2 j-1}\left|e^{-\beta H_{1} / K}\right| \sigma_{1}^{2 j} \ldots \sigma_{N}^{2 j}\right\rangle \\ & \times\left\langle\sigma_{1}^{2 j} \ldots \sigma_{N}^{2 j}\left|e^{-\beta H_{2} / K}\right| \sigma_{1}^{2 j+1} \ldots \sigma_{N}^{2 j+1}\right\rangle \end{aligned}
ZN≃Tr[e−βH1/Ke−βH2/K]K={σij}∑j=1∏K⟨σ12j−1…σN2j−1∣∣∣e−βH1/K∣∣∣σ12j…σN2j⟩×⟨σ12j…σN2j∣∣∣e−βH2/K∣∣∣σ12j+1…σN2j+1⟩该式还可以继续展开为:
Z
N
≃
∑
σ
i
j
⟨
σ
1
1
⋯
σ
N
1
∣
e
−
β
H
1
/
K
∣
σ
1
2
⋯
σ
N
2
⟩
⟨
σ
1
2
⋯
σ
N
2
∣
e
−
β
H
2
/
K
∣
σ
1
3
⋯
σ
N
3
⟩
⟨
σ
1
3
⋯
σ
N
3
∣
e
−
β
H
1
/
K
∣
σ
1
4
⋯
σ
N
4
⟩
⟨
σ
1
4
⋯
σ
N
4
∣
e
−
β
H
2
/
K
∣
σ
1
5
⋯
σ
N
5
⟩
⋯
⟨
σ
1
2
K
−
1
⋯
σ
N
2
K
−
1
∣
e
−
β
H
1
/
K
∣
σ
1
2
K
⋯
σ
N
2
K
⟩
⟨
σ
1
2
K
⋯
σ
N
2
K
∣
e
−
β
H
2
/
K
∣
σ
1
1
⋯
σ
N
1
⟩
\begin{array}{l} Z_{N} \simeq \\ \sum_{\sigma_{i}^{j}}\left\langle\sigma_{1}^{1} \cdots \sigma_{N}^{1}\left|e^{-\beta H_{1} / K}\right| \sigma_{1}^{2} \cdots \sigma_{N}^{2}\right\rangle\left\langle\sigma_{1}^{2} \cdots \sigma_{N}^{2}\left|e^{-\beta H_{2} / K}\right| \sigma_{1}^{3} \cdots \sigma_{N}^{3}\right. \\ \rangle\left\langle\sigma_{1}^{3} \cdots \sigma_{N}^{3}\left|e^{-\beta H_{1} / K}\right| \sigma_{1}^{4} \cdots \sigma_{N}^{4}\right\rangle\left\langle\sigma_{1}^{4} \cdots \sigma_{N}^{4}\left|e^{-\beta H_{2} / K}\right| \sigma_{1}^{5} \cdots \sigma_{N}^{5}\right\rangle \\ \cdots \\ \left\langle\sigma_{1}^{2 K-1} \cdots \sigma_{N}^{2 K-1}\left|e^{-\beta H_{1} / K}\right| \sigma_{1}^{2 K} \cdots \sigma_{N}^{2 K}\right\rangle\left\langle\sigma_{1}^{2 K} \cdots \sigma_{N}^{2 K}\left|e^{-\beta H_{2} / K}\right|\right. \\ \left.\sigma_{1}^{1} \cdots \sigma_{N}^{1}\right\rangle \end{array}
ZN≃∑σij⟨σ11⋯σN1∣∣e−βH1/K∣∣σ12⋯σN2⟩⟨σ12⋯σN2∣∣e−βH2/K∣∣σ13⋯σN3⟩⟨σ13⋯σN3∣∣e−βH1/K∣∣σ14⋯σN4⟩⟨σ14⋯σN4∣∣e−βH2/K∣∣σ15⋯σN5⟩⋯⟨σ12K−1⋯σN2K−1∣∣e−βH1/K∣∣σ12K⋯σN2K⟩⟨σ12K⋯σN2K∣∣e−βH2/K∣∣σ11⋯σN1⟩③假设空间和时间上都满足周期性边界条件:
σ
i
1
=
σ
i
2
K
+
1
\sigma_{i}^{1} = \sigma_{i}^{2K+1}
σi1=σi2K+1和
σ
1
j
=
σ
N
+
1
j
\sigma_{1}^{j} = \sigma_{N+1}^{j}
σ1j=σN+1j
对其中的每一项再进行分解:其中:
⟨
σ
1
1
σ
2
1
⋯
σ
N
1
∣
e
−
β
H
1
/
K
∣
σ
1
2
σ
2
2
⋯
σ
N
2
⟩
=
⟨
σ
1
1
σ
2
1
⋯
σ
N
1
∣
e
−
β
∑
i
=
o
d
d
h
i
,
i
+
1
/
K
∣
σ
1
2
σ
2
2
⋯
σ
N
2
⟩
=
⟨
σ
1
1
σ
2
1
∣
e
−
β
h
1
,
2
/
K
∣
σ
1
2
σ
2
2
⟩
⟨
σ
3
1
σ
4
1
∣
e
−
β
h
3
,
4
/
K
∣
σ
3
2
σ
4
2
⟩
⋯
⟨
σ
N
−
1
1
σ
N
1
∣
e
−
β
h
N
−
1
,
N
/
K
∣
σ
N
−
1
2
σ
N
2
⟩
=
ν
σ
1
1
σ
2
1
,
σ
1
2
σ
2
2
ν
σ
3
1
σ
4
1
,
σ
3
2
σ
4
2
⋯
ν
σ
N
−
1
1
σ
N
1
,
σ
N
−
1
2
σ
N
2
≡
ν
1
,
2
1
,
2
ν
3
,
4
1
,
2
⋯
ν
N
−
1
,
N
1
,
2
\begin{aligned} &\left\langle\sigma_{1}^{1} \sigma_{2}^{1} \cdots \sigma_{N}^{1}\left|e^{-\beta H_{1} / K}\right| \sigma_{1}^{2} \sigma_{2}^{2} \cdots \sigma_{N}^{2}\right\rangle \\ =&\left\langle\sigma_{1}^{1} \sigma_{2}^{1} \cdots \sigma_{N}^{1}\left|e^{-\beta \sum_{i=o d d} h_{i, i+1} / K}\right| \sigma_{1}^{2} \sigma_{2}^{2} \cdots \sigma_{N}^{2}\right\rangle \\ =&\left\langle\sigma_{1}^{1} \sigma_{2}^{1}\left|e^{-\beta h_{1,2} / K}\right| \sigma_{1}^{2} \sigma_{2}^{2}\right\rangle\left\langle\sigma_{3}^{1} \sigma_{4}^{1}\left|e^{-\beta h_{3,4} / K}\right| \sigma_{3}^{2} \sigma_{4}^{2}\right\rangle \\ & \cdots\left\langle\sigma_{N-1}^{1} \sigma_{N}^{1}\left|e^{-\beta h_{N-1, N} / K}\right| \sigma_{N-1}^{2} \sigma_{N}^{2}\right\rangle \\ =& \nu_{\sigma_{1}^{1} \sigma_{2}^{1}, \sigma_{1}^{2} \sigma_{2}^{2}} \nu_{\sigma_{3}^{1} \sigma_{4}^{1}, \sigma_{3}^{2} \sigma_{4}^{2} \cdots \nu_{\sigma_{N-1}^{1} \sigma_{N}^{1}, \sigma_{N-1}^{2} \sigma_{N}^{2}}} \\ \equiv & \nu_{1,2}^{1,2} \nu_{3,4}^{1,2} \cdots \nu_{N-1, N}^{1,2} \end{aligned}
===≡⟨σ11σ21⋯σN1∣∣∣e−βH1/K∣∣∣σ12σ22⋯σN2⟩⟨σ11σ21⋯σN1∣∣∣e−β∑i=oddhi,i+1/K∣∣∣σ12σ22⋯σN2⟩⟨σ11σ21∣∣∣e−βh1,2/K∣∣∣σ12σ22⟩⟨σ31σ41∣∣∣e−βh3,4/K∣∣∣σ32σ42⟩⋯⟨σN−11σN1∣∣∣e−βhN−1,N/K∣∣∣σN−12σN2⟩νσ11σ21,σ12σ22νσ31σ41,σ32σ42⋯νσN−11σN1,σN−12σN2ν1,21,2ν3,41,2⋯νN−1,N1,2
同样可得:
⟨
σ
1
2
σ
2
2
⋯
σ
N
2
∣
e
−
β
H
2
/
K
∣
σ
1
3
σ
2
3
⋯
σ
N
3
⟩
≡
ν
2
,
3
2
,
3
ν
4
,
5
2
,
3
⋯
ν
N
,
1
2
,
3
\begin{aligned} &\left\langle\sigma_{1}^{2} \sigma_{2}^{2} \cdots \sigma_{N}^{2}\left|e^{-\beta H_{2} / K}\right| \sigma_{1}^{3} \sigma_{2}^{3} \cdots \sigma_{N}^{3}\right\rangle \\ \equiv & \nu_{2,3}^{2,3} \nu_{4,5}^{2,3} \cdots \nu_{N, 1}^{2,3} \end{aligned}
≡⟨σ12σ22⋯σN2∣∣∣e−βH2/K∣∣∣σ13σ23⋯σN3⟩ν2,32,3ν4,52,3⋯νN,12,3
所以配分函数可以写为:
Z
N
≃
∑
σ
i
j
(
ν
1
,
2
1
,
2
ν
3
,
4
1
,
2
⋯
ν
N
−
1
,
N
1
,
2
)
(
ν
2
,
3
2
,
3
ν
4
,
5
2
,
3
⋯
ν
N
,
1
2
,
3
)
(
ν
1
,
2
3
,
4
ν
3
,
4
3
,
4
⋯
ν
N
−
1
,
N
3
,
4
)
(
ν
2
,
3
4
,
5
ν
4
,
5
4
,
5
⋯
ν
N
,
1
4
,
5
)
\begin{array}{l} Z_{N} \simeq \\ \sum_{\sigma_{i}^{j}}\left(\nu_{1,2}^{1,2} \nu_{3,4}^{1,2} \cdots \nu_{N-1, N}^{1,2}\right)\left(\nu_{2,3}^{2,3} \nu_{4,5}^{2,3} \cdots \nu_{N, 1}^{2,3}\right) \\ \left(\nu_{1,2}^{3,4} \nu_{3,4}^{3,4} \cdots \nu_{N-1, N}^{3,4}\right)\left(\nu_{2,3}^{4,5} \nu_{4,5}^{4,5} \cdots \nu_{N, 1}^{4,5}\right) \end{array}
ZN≃∑σij(ν1,21,2ν3,41,2⋯νN−1,N1,2)(ν2,32,3ν4,52,3⋯νN,12,3)(ν1,23,4ν3,43,4⋯νN−1,N3,4)(ν2,34,5ν4,54,5⋯νN,14,5)
⋯
\cdots
⋯
(
ν
1
,
2
2
K
−
1
,
2
K
ν
3
,
4
2
K
−
1
,
2
K
⋯
ν
N
−
1
,
N
2
K
−
1
,
2
K
)
(
ν
2
,
3
2
K
,
1
ν
4
,
5
2
K
,
1
⋯
ν
N
,
1
2
K
,
1
)
\left(\nu_{1,2}^{2 K-1,2 K} \nu_{3,4}^{2 K-1,2 K} \cdots \nu_{N-1, N}^{2 K-1,2 K}\right)\left(\nu_{2,3}^{2 K, 1} \nu_{4,5}^{2 K, 1} \cdots \nu_{N, 1}^{2 K, 1}\right)
(ν1,22K−1,2Kν3,42K−1,2K⋯νN−1,N2K−1,2K)(ν2,32K,1ν4,52K,1⋯νN,12K,1)
即:
Z
N
≃
∑
{
σ
i
j
}
∏
i
=
1
N
/
2
∏
j
=
1
K
v
σ
2
i
−
1
2
j
−
1
σ
2
i
2
j
−
1
,
σ
2
i
−
1
2
j
σ
2
i
2
j
v
σ
2
i
2
j
σ
2
i
+
1
2
j
,
σ
2
i
2
j
+
1
σ
2
i
+
1
2
j
+
1
Z_{N} \simeq \sum_{\left\{\sigma_{i}^{j}\right\}} \prod_{i=1}^{N / 2} \prod_{j=1}^{K} v_{\sigma_{2 i-1}^{2 j-1} \sigma_{2 i}^{2 j-1}, \sigma_{2 i-1}^{2 j} \sigma_{2 i}^{2 j}} v_{\sigma_{2 i}^{2 j} \sigma_{2 i+1}^{2 j}, \sigma_{2 i}^{2 j+1} \sigma_{2 i+1}^{2 j+1}}
ZN≃{σij}∑i=1∏N/2j=1∏Kvσ2i−12j−1σ2i2j−1,σ2i−12jσ2i2jvσ2i2jσ2i+12j,σ2i2j+1σ2i+12j+1于是我们的转移矩阵是一个四阶张量:
v
σ
1
σ
4
,
σ
2
σ
3
≡
⟨
σ
1
σ
4
∣
exp
(
−
β
h
i
,
i
+
1
/
K
)
∣
σ
2
σ
3
⟩
v_{\sigma_{1} \sigma_{4}, \sigma_{2} \sigma_{3}} \equiv \left\langle\sigma_{1} \sigma_{4}\left|\exp \left(-\beta h_{i, i+1} / K\right)\right| \sigma_{2} \sigma_{3}\right\rangle
vσ1σ4,σ2σ3≡⟨σ1σ4∣exp(−βhi,i+1/K)∣σ2σ3⟩于是可以得到配分函数的张量网格图:
图为对配分函数近似得到的转移矩阵张量网络
对所有的中间态
∣
σ
i
j
⟩
|\sigma_{i}^{j}\rangle
∣σij⟩进行求和,即缩并张量网络中所有的
σ
\sigma
σ键,我们就可以得到配分函数了
step2 收缩张量网络
对
v
v
v张量进行SVD分解
v
σ
1
σ
2
,
σ
3
σ
4
=
∑
x
=
1
D
2
U
σ
1
σ
2
,
x
λ
x
V
x
,
σ
3
σ
4
⊤
≡
∑
x
=
1
D
2
(
T
a
)
x
,
σ
1
,
σ
2
(
T
b
)
x
,
σ
3
,
σ
4
\begin{aligned} v_{\sigma_{1} \sigma_{2}, \sigma_{3} \sigma_{4}} &=\sum_{x=1}^{D^{2}} U_{\sigma_{1} \sigma_{2}, x} \lambda_{x} V_{x, \sigma_{3} \sigma_{4}}^{\top} \\ & \equiv \sum_{x=1}^{D^{2}}\left(T_{a}\right)_{x, \sigma_{1}, \sigma_{2}}\left(T_{b}\right)_{x, \sigma_{3}, \sigma_{4}} \end{aligned}
vσ1σ2,σ3σ4=x=1∑D2Uσ1σ2,xλxVx,σ3σ4⊤≡x=1∑D2(Ta)x,σ1,σ2(Tb)x,σ3,σ4其中对角矩阵
λ
\lambda
λ中有
D
2
D^{2}
D2个奇异值,为了方便,我们引入两个辅助张量:
(
T
a
)
x
,
σ
1
,
σ
2
≡
U
σ
1
σ
2
,
x
λ
x
\left(T_{a}\right)_{x, \sigma_{1}, \sigma_{2}} \equiv U_{\sigma_{1} \sigma_{2}, x} \sqrt{\lambda_{x}}
(Ta)x,σ1,σ2≡Uσ1σ2,xλx
(
T
b
)
x
,
σ
3
,
σ
4
≡
V
σ
3
σ
4
,
x
λ
x
\left(T_{b}\right)_{x, \sigma_{3}, \sigma_{4}} \equiv V_{\sigma_{3} \sigma_{4}, x} \sqrt{\lambda_{x}}
(Tb)x,σ3,σ4≡Vσ3σ4,xλx
经过这种变换后,方块的张量网络变成了具有两个三阶张量
T
a
Ta
Ta和
T
b
Tb
Tb的六角张量网络,如图:
我们将所有的方块张量网络全分解后,原网络变为如图:
现将上图中最后两行中虚线圈包围的
σ
\sigma
σ键进行收缩,这就我们就得到两个四阶张量
(
M
a
)
α
,
t
1
,
β
,
b
1
=
∑
y
=
1
D
(
T
a
)
β
,
b
1
,
y
(
T
b
)
α
,
t
1
,
y
(
M
b
)
β
,
t
2
,
γ
,
b
2
=
∑
z
=
1
D
(
T
a
)
γ
,
z
,
t
2
(
T
b
)
β
,
z
,
b
2
\begin{aligned} \left(M_{a}\right)_{\alpha, t_{1}, \beta, b_{1}} &=\sum_{y=1}^{D}\left(T_{a}\right)_{\beta, b_{1}, y}\left(T_{b}\right)_{\alpha, t_{1}, y} \\ \left(M_{b}\right)_{\beta, t_{2}, \gamma, b_{2}} &=\sum_{z=1}^{D}\left(T_{a}\right)_{\gamma, z, t_{2}}\left(T_{b}\right)_{\beta, z, b_{2}} \end{aligned}
(Ma)α,t1,β,b1(Mb)β,t2,γ,b2=y=1∑D(Ta)β,b1,y(Tb)α,t1,y=z=1∑D(Ta)γ,z,t2(Tb)β,z,b2如图:
在
M
a
M_a
Ma和
M
b
M_b
Mb水平方向的键上都有一个对角矩阵
λ
1
,
2
\lambda_{1,2}
λ1,2