LimTDD: A Compact Decision Diagram Integrating Tensor and Local Invertible Map Representations
Extracted source text · arXiv 2504.01168
% \documentclass{article}
% \documentclass[conference,10pt]{IEEEtran}
\documentclass[lettersize,journal]{IEEEtran}
% https://www.computer.org/digital-library/journals/tc/call-for-papers-general-submissions
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc} % use 8-bit T1 fonts
\usepackage{hyperref} % hyperlinks
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{nicefrac} % compact symbols for 1/2, etc.
\usepackage{microtype} % microtypography
\usepackage{lipsum}
\usepackage{graphicx}
\usepackage{subfigure}
\graphicspath{ {./images/} }
\usepackage{amsmath,amssymb,amsthm}
\usepackage{algorithm}
% \usepackage{algorithmic}
\usepackage{algorithmicx}
\usepackage{algpseudocode}
% \usepackage{qcircuit}
\usepackage{braket}
\usepackage{booktabs}
\usepackage{threeparttable}
\usepackage{xcolor}
\usepackage{supertabular}
\usepackage{longtable}
\usepackage{multirow}
\usepackage{array}
% \usepackage{subcaption}
% \usepackage[caption=false,font=normalsize,labelfont=sf,textfont=sf]{subfig}
\usepackage{tikz}
\definecolor{custompurple}{rgb}{0.5, 0.1, 0.5}
\usepackage{circuitikz}
\usetikzlibrary{quantikz}
\usepackage{setspace} % 使用setspace宏包
\usepackage{circledsteps}%require for circle
\pgfkeys{/csteps/inner ysep=10pt}
\usepackage{yquant}
\useyquantlanguage{groups}
% \newcommand*\circled[1]{\tikz[baseline=(char.base)]{
% \node[shape=circle,draw,inner sep=2pt] (char) {#1};}} % 定义 \circled 命令
% \newcommand*\circled[1]{\resizebox{2em}{!}{\textcircled{#1}}} % 定义 \circled 命令
\newtheorem{thm}{Theorem}
\newtheorem{lem}{Lemma}
\newtheorem{fact}{Fact}
\newtheorem{cor}{Corollary}
\newtheorem{defn}{Definition}
\newtheorem{exam}{Example}
\newtheorem{remark}{Remark}
\newcommand\vecc[1]{|#1\rangle\!\rangle}
\renewcommand{\algorithmicrequire}{ \textbf{Input:}}
\renewcommand{\algorithmicensure}{ \textbf{Output:}}
\newcommand{\bphi}{{\Phi}}
\newcommand {\TDD} {{\mathit{TDD}}}
\newcommand {\LimTDD} {{\mathit{LimTDD}}}
\newcommand {\cnot} {{\textsf{CX}}}
\newcommand {\F} {{\mathcal{F}}}
\newcommand {\G} {{\mathcal{H}}}
\newcommand {\cont} {{\texttt{cont}}}
\newcommand {\add} {{\texttt{add}}}
\newcommand {\low} {{\texttt{low}}}
\newcommand {\high} {{\texttt{high}}}
\newcommand {\locnorm} {{\texttt{loc\_norm}}}
\newcommand {\idx} {{\texttt{idx}}}
\newcommand {\w} {{\texttt{wt}}}
\newcommand {\val} {{\texttt{val}}}
\newcommand {\stab} {{\texttt{Stab}}}
\usepackage[normalem]{ulem}
\newcommand{\red}[1]{{\color{red} #1}}
\newcommand{\blue}[1]{{\color{black} #1}}
\newcommand{\addt}[1]{{\color{black} #1}}
\newcommand{\delete}[1]{{\color{cyan}\sout{#1}}}
\newcommand{\emp}[1]{{\color{red} #1}}
\newcommand{\znote}[1]{{\color{blue} \textbf{ZF: #1}}}
\newcommand{\magenta}[1]{{\color{magenta} \textbf{[SL: #1]}}}
% \usepackage{stfloats}
\hyphenation{op-tical net-works semi-conduc-tor IEEE-Xplore}
\begin{document}
\title{LimTDD: A Compact Decision Diagram Integrating Tensor and Local Invertible Map Representations}
% \author{}
% \date{}
\author{
Xin Hong$^1$, Aochu Dai$^2$, Dingchao Gao$^1$, Sanjiang Li$^3$, Zhengfeng Ji$^{2,4}$, Mingsheng Ying$^3$\\
~\IEEEmembership{$^1$Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, China}\\
~\IEEEmembership{$^2$Department of Computer Science and Technology, Tsinghua University, China}\\
~\IEEEmembership{$^3$Centre for Quantum Software and Information, University of Technology Sydney, Australia}\\
~\IEEEmembership{$^4$ Zhongguancun Laboratory, China}\\
% <-this % stops a space
\thanks{Emails: \{hongxin, gaodc\}@ios.ac.cn, dac22@mails.tsinghua.edu.cn, jizhengfeng@tsinghua.edu.cn, \{sanjiang.li, mingsheng.ying\}@uts.edu.au}
% , gaodc@ios.ac.cn,sanjiang.li@uts.edu.au,
\thanks{Work partially supported by the National Natural Science Foundation of China (Grant No. 12471437, 12347104), Beijing Nova Program (Grant No. 20220484128), National Key Research and Development Program of China (Grant No.\ 2023YFA1009403), and Beijing Natural Science Foundation (Grant No.\ Z220002).
}
% National Natural Science
% Foundation of China (Grant No.\ 12347104)
% \thanks{Manuscript received February
% 19, 2025; revised February
% 11, 2025.}
}
% The paper headers
%\markboth{Journal of \LaTeX\ Class Files,~Vol.~14, No.~8, February~2025}%
%{Shell \MakeLowercase{\textit{et al.}}: A Sample Article Using IEEEtran.cls for IEEE Journals}
% \IEEEpubid{0000--0000/00\$00.00~\copyright~2021 IEEE}
\maketitle
%相对于LIMDD，给了LIMDD一个张量的解释，并给出了对应的算法，推广了权重的取值，以及应用范围
%相对于TDD，提供了进一步压缩的可能，提高了运算的效率
%%SL:
%% 1. Only highlight major changes
%% 2. Is functionality computation a precise/good phrase? Tensor computation?
%% 3. US English vs. British English: optimisation -> optimization
\begin{abstract}
\addt{Tensor networks serve as a powerful tool for efficiently representing and manipulating high-dimensional data in applications such as quantum physics, machine learning, and data compression.} Tensor Decision Diagrams (TDDs) offer an efficient framework for tensor representation by leveraging decision diagram techniques. However, the current implementation of TDDs and other decision diagrams fail to exploit tensor isomorphisms, limiting their compression potential. This paper introduces Local Invertible Map Tensor Decision Diagrams (LimTDDs), an extension of TDDs that incorporates local invertible maps (LIMs) to achieve more compact representations. Unlike LIMDD, which uses Pauli operators for quantum states, LimTDD employs the $XP$-stabilizer group, enabling broader applicability across tensor-based tasks. We present efficient algorithms for normalization, slicing, addition, and contraction, critical for tensor network applications. Theoretical analysis demonstrates that LimTDDs achieve greater compactness than TDDs and, in best-case scenarios and for quantum state representations, offer exponential compression advantages over both TDDs and LIMDDs. Experimental results in quantum circuit tensor computation and simulation confirm LimTDD's superior efficiency. \addt{Open-source code is available at \url{https://github.com/Veriqc/LimTDD}.}
%\addt{As a useful tool, tensor network has been widely used in many aspects.} Tensor Decision Diagrams (TDDs) provide an efficient structure for representing tensors by combining techniques from decision diagrams. However, existing decision diagrams, including TDDs, fail to exploit isomorphisms within tensors, limiting their compression efficiency. \addt{Inspired by LIMDD,} This paper introduces Local Invertible Map Tensor Decision Diagrams (LimTDDs), an extension of TDD that integrates local invertible maps (LIMs) to achieve more compact representations. Unlike LIMDD, which applies Pauli operators to quantum states, LimTDD generalizes this approach using the $XP$-stabilizer group, enabling broader applicability. We develop efficient algorithms for normalization and key tensor operations, including slicing, addition, and contraction, essential for \addt{tensor network-based tasks}. Theoretical analysis shows that LimTDD surpasses TDD in compactness\addt{, and it even surpasses LIMDD in representing quantum states, where an exponential advantage over both TDD and LIMDD can be found} in the best-case scenarios. Experimental results \addt{in quantum circuit simulation and functionality computation} validate these improvements, demonstrating LimTDD's superior efficiency. \addt{\url{https://github.com/Veriqc/LimTDD/tree/C}}
% \addt{As a useful tool, tensor network has been widely used in many aspect. Tensor Decision Diagrams (TDDs) provide an efficient structure for representing tensors by combining techniques from decision diagrams. However, existing decision diagrams, including TDDs, fail to exploit isomorphisms within tensors, limiting their compression efficiency. Inspired by LIMDD, this paper introduces Local Invertible Map Tensor Decision Diagrams (LimTDDs), an extension of TDD that integrates local invertible maps (LIMs) to achieve more compact representations. Unlike LIMDD, which applies Pauli operators to quantum states, LimTDD generalizes this approach using the $XP$-stabilizer group, enabling broader applicability. We develop efficient algorithms for normalization and key tensor operations, including slicing, addition, and contraction, essential for tensor network-based tasks. Theoretical analysis shows that LimTDD surpasses TDD in compactness, and it even surpasses LIMDD in representing quantum states, where an exponential advantage over both TDD and LIMDD can be found in the best-case scenarios. Experimental results in quantum circuit simulation and functionality computation validate these improvements, demonstrating LimTDD's superior efficiency. \addt{\url{https://github.
% com/Veriqc/LimTDD/tree/C}}
% }
\end{abstract}
\begin{IEEEkeywords}
Tensor Decision Diagrams, Local Invertible Maps, Quantum Circuits, Tensor Networks
\end{IEEEkeywords}
\section{Introduction}
\IEEEPARstart{O}{ver} the past decade, tensor networks have evolved from a niche technique in condensed-matter physics into a universal toolkit for high-dimensional data. This transformation is driven by their ability to decompose exponentially large tensors into networks of low-rank cores, enabling high-efficiency algorithms for problems such as quantum many-body simulations, previously deemed intractable. Key milestones include DMRG-based PEPS simulators for quantum entanglement \cite{ran2020quantum}, MPS-driven discoveries of exotic quantum phases \cite{haller2023quantum}, and tensorized neural networks, which leverage tensor decomposition to match the performance of transformers on moderate-scale language-modeling benchmarks while using orders of magnitude fewer parameters \cite{ma2019tensorized}. Recent advances in automatic differentiation and GPU-accelerated contraction libraries (e.g., ITensor \cite{fishman2022itensor} and TensorNetwork-TensorFlow \cite{ganahl2019tensornetwork}) enable efficient optimization within deep-learning pipelines.
%\addt{\IEEEPARstart{O}{ver} the past decade, tensor networks have evolved from a niche technique in condensed-matter physics into a universal toolkit for high-dimensional data. By decomposing exponentially large tensors into networks of low-rank cores, they deliver polynomial-cost algorithms once deemed intractable. Key milestones include DMRG-based PEPS simulators of quantum entanglement \cite{ran2020quantum}, MPS-driven discoveries of exotic quantum phases \cite{haller2023quantum}, and tensorized neural networks that rival transformers on language tasks while using orders-of-magnitude fewer parameters \cite{ma2019tensorized}. Recent advances in automatic differentiation and GPU-accelerated contraction libraries—implemented in packages such as ITensor \cite{fishman2022itensor} and TensorNetwork-TensorFlow \cite{ganahl2019tensornetwork}—enable real-time optimisation within deep-learning pipelines.}
% \delete{\IEEEPARstart{Q}{uantum} computing has experienced rapid development in recent years, offering potential breakthroughs in
% cryptography~\cite{preskill2018quantum}, optimization~\cite{rebentrost2014quantum}, quantum chemistry~\cite{cao2019quantum}, and more. As quantum hardware scales, the need for efficient simulation and verification of quantum circuits becomes increasingly critical. Decision diagrams have emerged as a powerful approach for these tasks, enabling compact representations and efficient manipulations of quantum states and operations.}
\addt{In recent years, significant progress has been made in developing compact data structures for tensor representation and tensor network computations. One such structure, the Tensor Decision Diagram (TDD), combines tensor network techniques with decision diagram methodologies to enable efficient representation and manipulation of tensor structures. Besides quantum state simulation and tensor computation,} TDDs have been successfully applied to approximate equivalence checking of noisy quantum circuits~\cite{hong2021approximate} and equivalence checking of dynamic quantum circuits~\cite{10069104}. Notably, optimizing the contraction order of tensor networks can further enhance TDDs' computational efficiency~\cite{ftdd10528748}.
%\addt{Apart from the above progress, in recent years, there is also work that studies using more compact data structures to represent tensors and conduct tensor network computations. One of them is called the tensor decision diagram (TDD), which} leverages both tensor network techniques and decision diagram methodologies, enabling efficient compression and representation of tensor structures. Recently, TDD has been \addt{successfully} employed in the approximate equivalence checking of noisy quantum circuits \cite{hong2021approximate} and equivalence checking of dynamic quantum circuits \cite{10069104}. What is worth mentioning is that, by optimizing the contraction order of the tensor network, TDD has the potential to further enhance computational efficiency \cite{ftdd10528748}.
% Various decision diagram structures have been introduced, including Quantum Information Decision Diagrams (QuIDD)~\cite{viamontes2003improving}, Quantum Multiple-Valued Decision Diagrams (QMDD)~\cite{niemann2015qmdds}, X-Decomposition Quantum Decision Diagrams (XQDD)~\cite{wang2008xqdd}, and Tensor Decision Diagrams (TDD)~\cite{hong2020tensor}. These decision diagrams have been successfully applied in the simulation and verification of quantum circuits~\cite{viamontes2004high, burgholzer2020advanced, burgholzer2021qcec}, as well as quantum state preparation \cite{DBLP:conf/dac/MatoHW24} and quantum circuit design automation \cite{DBLP:conf/qsw/WilleBFKMPQRSSSSB24}.
% % {\color{blue}Ying: To be polite, we should say some advantages of other DDs FIRST. Then talk about TDD's advantage}
% Among these, QuIDD introduces the method of classical algebraic decision diagrams \cite{bahar1997algebric} to handle quantum information, QMDD takes advantage of the unique form of quantum circuits, XQDD fully exploits the commonly used symmetries in quantum gates, while TDD leverages both tensor network techniques and decision diagram methodologies, enabling efficient compression and representation of tensor structures in quantum circuits. What is worth mentioning is that, by optimizing the contraction order of the tensor network, TDD has the potential to further enhance computational efficiency \cite{ftdd10528748}.
% Recently, TDD has also been employed in the approximate equivalence checking of noisy quantum circuits \cite{hong2021approximate} and equivalence checking of dynamic quantum circuits \cite{10069104}.
% \begin{figure}[!t]
% \centering
% \begin{circuitikz}[transform shape]
% \node at (0,0) [anchor=north]{
% \begin{tikzcd}[row sep=0.3cm, column sep=0.3cm]
% & & &&\gate[1]{O_1}&\gate[3]{\psi}&\gate[1]{O_4}&\qw&& &&&\gate[3]{\psi}&\qw& \\
% &\lambda&\times &&\gate[1]{O_2}& &\gate[1]{O_5}&\qw&&\cong &&& &\qw& \\
% & & &&\gate[1]{O_3}& &\gate[1]{O_6}&\qw&& &&& &\qw&
% \end{tikzcd}
% };
% \end{circuitikz}
% \caption{Tensor isomorphism, where $\lambda\not=0$ is a complex number and each $O_i$ is an invertible operator.}
% \label{fig:tensoriso}
% \end{figure}
\begin{figure}[!t]
\centering
\input{draw_pic_dot/tensoriso}
\caption{Tensor isomorphism, where $\lambda\neq 0$ is a complex number and each $O_i$ is an invertible operator. The same edge color corresponds to the same index. \addt{After applying a local operator, we rename the corresponding index to keep it consistent with the original one.}}
\label{fig:tensoriso}
\end{figure}
While TDDs offer a compact and structured representation, they do not explicitly exploit isomorphic structures within tensors. In this paper, we define two tensors as \emph{isomorphic} if they encode the same essential information, up to a nonzero global coefficient and a local transformation applied to each index (see Fig.~\ref{fig:tensoriso} for an illustration). Exploiting these isomorphic relationships enables significant compression by merging structurally similar tensors into a single representation.
%While TDDs provide a structured and compact representation, they do not explicitly exploit isomorphic structures within tensors. In this paper, we consider two tensors as \emph{isomorphic} if they encode the same essential information up to a (nonzero) global coefficient and a local transformation applied to each index (see Fig.~\ref{fig:tensoriso} for an illustration). Recognizing and exploiting such isomorphic relationships may enable significant compression, as structurally similar tensors can be merged into a single representation. However, existing decision diagrams, including TDDs, do not leverage such tensor isomorphisms, leaving room for further compression.
\addt{The Local Invertible Map Decision Diagram (LIMDD), introduced in~\cite{vinkhuijzen2023limdd}, leverages a specific tensor isomorphism (i.e., quantum state isomorphism) to represent and manipulate quantum states. By identifying isomorphic relations and extracting local Pauli operators, LIMDD achieves superior compression efficiency, representing any stabilizer state with at most $n$ non-terminal nodes, where $n$ is the number of qubits. Although LIMDD has been extended to represent general tensors, it lacks mechanisms for tensor contraction and efficient tensor computations in its current implementation~\cite{vinkhuijzen2023limdd,vinkhuijzen2023efficient}.}
%A novel approach for representing quantum states has been implemented in Local Invertible Map Decision Diagrams (LIMDD)~\cite{vinkhuijzen2023limdd,vinkhuijzen2023efficient}. By establishing isomorphic relations and extracting local Pauli operators between quantum states, LIMDD achieves superior compression efficiency. Notably, it can represent any stabilizer state using at most $n$ nodes, where $n$ is the number of qubits. In the best-case scenario, LIMDD has been shown to be exponentially more compact than QMDD. However, LIMDD is specifically designed for quantum state representations and cannot be directly applied to general tensor representations \addt{and operations}.
% \red{, while QuIDD, XQDD and TDD have almost the same degree of compactness as QMDD}
% \addt{\delete{The main challenge for extending LIMDD to support general tensor representation and operations includes: (1) defining tensor isomorphisms, (2) giving the DD representation of a tensor, and defining the semantics, (3) defining the contraction operation between DD representations, (4) utlizing the tensor isomorphisms to improve the contraction operation.}}
This paper aims to incorporate isomorphic tensor compression into a decision diagram framework. To achieve this, we propose integrating Local Invertible Maps (LIMs)~\cite{vinkhuijzen2023limdd} into TDDs, creating a versatile decision diagram that applies to tensor network tasks while maintaining at least the same compression efficiency as TDD. However, integrating LIMs into TDD presents several key challenges:
\begin{enumerate}
\item \textbf{Efficiently identifying and utilizing tensor isomorphisms}: While the concept of tensor isomorphism is straightforward to define (Fig.~\ref{fig:tensoriso}), recognizing and applying these transformations within a decision diagram remains nontrivial. A structured approach is required to detect and exploit isomorphic structures while ensuring TDD’s operational efficiency is maintained.
\item \textbf{Extending beyond Pauli-based equivalences}: LIMDDs rely on Pauli operators to establish equivalences for quantum states, which limits their effectiveness to represent general tensors. A more expressive operator group is needed to enhance compression efficiency and extend applicability beyond simulating quantum states.
\item \textbf{Ensuring a canonical and compact representation}: Ensuring a canonical and compact representation demands a systematic approach to normalizing tensor structures, which involves refining weight normalization techniques.
%\delete{and leveraging efficient tensor transformation algorithms}.
\item \addt{\textbf{Efficiently implementing the contraction operation}: Efficient tensor network contraction operations require making full use of the isomorphism relationships among tensors. This requires us to handle the local operators on the edges of the decision diagram properly.
% This requires placing the isomorphic operators into the original tensor network.
}
\end{enumerate}
To address these challenges, we introduce the Local Invertible Map Tensor Decision Diagram (LimTDD), an extension of TDD that incorporates LIMs for enhanced compression. Our main contributions include:
% {\color{blue}Ying: Link each "contributions" to a "challenge"}
\begin{enumerate}
\item Addressing Challenge 1, we formalize tensor vectorization and present procedures to identify tensor isomorphisms within the LimTDD structure, enabling more effective compression while maintaining computational efficiency.
%\delete{establish a framework for identifying and utilizing tensor isomorphisms}
\item Addressing Challenge 2, we utilize the $XP$-stabilizer group \cite{webster_xp_2022} in LimTDD, extending its applicability beyond quantum states and improving compression efficiency compared to Pauli-based LIMDD.
\item Addressing Challenges 3, we develop optimized algorithms for normalization and provide theoretical proofs to ensure the canonicity and compactness of LimTDD.
%\delete{, slicing, addition, and contraction, enabling efficient computation of \red{quantum circuit functionality} and tensor network contraction}.
\item \addt{Addressing Challenges 4, we enhance the computational efficiency of LimTDD by removing the operators applied to open indices and transfer operators applied to contracted indices to one LimTDD, thus reducing the complexity of the LimTDDs that need to be contracted and improving the efficiency of reusing intermediate results.
% \red{restoring} the operators on the edges of a LimTDD to a tensor network and performing appropriate processing.
}
\item Additionally, we conduct extensive experiments demonstrating its advantages in compactness and computational efficiency for quantum circuit simulation and functionality computation.
%\delete{provide a theoretical comparison of LimTDD with existing decision diagrams and}
\end{enumerate}
\subsection*{Related Work}
\addt{
Decision diagrams, originally developed for analyzing Boolean functions, have been widely adopted across various fields through their diverse variants. By incorporating multi-terminal nodes~\cite{fujita1997multi} or valued edges~\cite{vrudhula1996edge}, decision diagrams have been extended to applications such as pseudo-Boolean functions~\cite{vrudhula1996edge,tafertshofer1997factored}, probabilistic distributions~\cite{kisa2014probabilistic}, algebraic systems~\cite{bahar1997algebric, sanner2005affine, wilson2005decision}, and matrix operations~\cite{vrudhula1996edge,tafertshofer1997factored}.
}
\addt{
To meet the demands of quantum system simulation and verification, several decision diagram models supporting complex domains and matrix manipulations have been introduced, including Quantum Information Decision Diagrams (QuIDD)~\cite{viamontes2003improving}, Quantum Multiple-Valued Decision Diagrams (QMDD)~\cite{niemann2015qmdds}, X-Decomposition Quantum Decision Diagrams (XQDD)~\cite{wang2008xqdd}, and Context-Free-Language Ordered Binary Decision Diagrams (CFLOBDD)~\cite{sistla2024cflobdds}. These decision diagrams have been successfully applied to quantum circuit simulation and verification~\cite{viamontes2004high,burgholzer2020advanced,burgholzer2021qcec}, quantum state preparation~\cite{DBLP:conf/dac/MatoHW24}, and quantum circuit design automation~\cite{DBLP:conf/qsw/WilleBFKMPQRSSSSB24}. Specifically, QuIDD adapts classical algebraic decision diagrams~\cite{bahar1997algebric} for quantum information, QMDD leverages the unique structure of quantum circuits, XQDD exploits symmetries in quantum gates, and CFLOBDD organizes diagram structures using a branching program and procedure-calling approach. Additionally, knowledge maps for decision diagrams have been explored to enhance their representational capabilities~\cite{fargier2014knowledge,vinkhuijzen2024knowledge}.
}
%\addt{Decision diagrams, originally developed as an analytical tool for Boolean functions, have been widely used in many fields with their subsequent rich variants. By introducing multi-terminal nodes~\cite{fujita1997multi} or valued edges~\cite{vrudhula1996edge}, decision diagrams can be applied to a variety of fields including pseudo-Boolean functions~\cite{vrudhula1996edge, tafertshofer1997factored}, probabilistic distributions~\cite{kisa2014probabilistic}, algebraic systems~\cite{bahar1997algebric, sanner2005affine, wilson2005decision}, and matrix operations~\cite{vrudhula1996edge, tafertshofer1997factored}.} \addt{With the demand for quantum system simulation and verification, many decision diagram models supporting complex domains and matrix manipulations have been introduced, }including Quantum Information Decision Diagrams (QuIDD)~\cite{viamontes2003improving}, Quantum Multiple-Valued Decision Diagrams (QMDD)~\cite{niemann2015qmdds}, X-Decomposition Quantum Decision Diagrams (XQDD)~\cite{wang2008xqdd}, \addt{and Context-Free-Language Ordered Binary Decision Diagrams (CFLOBDD~\cite{sistla2024cflobdds})}. These decision diagrams have been successfully applied in the simulation and verification of quantum circuits~\cite{viamontes2004high, burgholzer2020advanced, burgholzer2021qcec}, as well as quantum state preparation \cite{DBLP:conf/dac/MatoHW24} and quantum circuit design automation \cite{DBLP:conf/qsw/WilleBFKMPQRSSSSB24}.
% {\color{blue}Ying: To be polite, we should say some advantages of other DDs FIRST. Then talk about TDD's advantage}
%Among these, QuIDD introduces the method of classical algebraic decision diagrams \cite{bahar1997algebric} to handle quantum information, QMDD takes advantage of the unique form of quantum circuits, XQDD fully exploits the commonly used symmetries in quantum gates \addt{ and CFLOBDD organizes the diagram structure in a branching program and procedure calling manner}. \addt{The knowledge maps of decision diagram have also be studied \cite{fargier2014knowledge, vinkhuijzen2024knowledge}.}
The remainder of this paper is structured as follows. Section~\ref{sec:background} provides background on quantum computing, tensor networks, and decision diagrams (TDD and LIMDD). Section~\ref{sec:LimTDD} introduces the formal definition of LimTDD, along with its mathematical foundations. Section~\ref{sec:Algorithms} describes key tensor operations for LimTDD, including slicing, addition, and contraction. Section~\ref{sec:compare} presents a theoretical comparison between LimTDD, TDD, and LIMDD, analyzing their compression efficiency.
% Section \ref{sec:applications} introduces potential applications of LimTDD.
Section~\ref{sec:experiments} provides experimental results, demonstrating LimTDD’s advantages in quantum circuit simulation and tensor network computations. Finally, Section~\ref{sec:conclusion} concludes the paper. For clarity, proofs are provided in the appendix.
\input{background2.tex}
\section{LimTDD} \label{sec:LimTDD}
\addt{While TDD enables tensor compression, it suffers from relatively lower efficiency, especially when compared with the high compressing ratio of LIMDD in representing quantum states.} In this section, we introduce LimTDD, a novel approach \addt{inspired by LIMDD that enhances TDD efficiency.} By combining the respective advantages of TDD and LIMDD, LimTDD not only retains the ability to represent arbitrary tensors but also achieves greater compactness, offering enhanced efficiency and broader applicability.
%o this end, we first extend the notion of isomorphism from quantum states to tensors.
\subsection{Isomorphism Between Tensors}\label{sec:tensoriso}
To effectively represent tensors using LimTDD, we first characterize tensor isomorphism by introducing tensor vectorization. Let $\mathcal{G}$ be a subgroup of $\mathcal{O}$. As illustrated in Fig.~\ref{fig:tensoriso}, two tensors are considered $\mathcal{G}$-isomorphic if they differ only by a global coefficient and a local invertible transformation $O\in\mathcal{G}$. In order to express this isomorphic relation precisely,
we introduce the notion of tensor vectorization:
\begin{defn}[Tensor Vectorization \cite{biamonte2019lectures}]
Let $\phi_{x_n\cdots x_1}$ be a tensor with $n$ indices. Its \addt{vector representation with respect to the index order $x_n \prec \cdots \prec x_1$}, denoted by $\vecc{\phi_{x_n\cdots x_1}}$, is a $2^n$-dimensional vector whose $i$-th element is $\phi(i_n,\cdots, i_1)$, where $i_n\cdots i_1$ is the binary expansion of the integer $i$.
\end{defn}
\addt{For instance, consider the rank-2 tensor $\phi_{xy}$ represented by matrix $\left[\begin{array}{cccc}
1 & 2\\
0 & 3\\
\end{array}\right],$
where $x$ and $y$ denote the column and row indices, respectively. Its vector representation is $\vecc{\phi_{xy}} = [1,0,2,3]^\intercal$.
For the special case when a rank-$n$ tensor $\phi$ represents an $n$-qubit quantum state $\ket{\psi}$, then, under the same index/qubit order, $\vecc{\phi}$ coincides with the vector representation of the quantum state $\ket{\psi}$.
}
%\addt{Consider a rank-2 tensor $\phi_{xy}$ corresponding to the matrix $\left[\begin{array}{cccc}
% 1 & 2\\
% 0 & 3\\
%\end{array}\right]$. Here, we use $x$ and $y$ to represent the indices corresponding to columns and rows, respectively. Then the vectorization $\vecc{\phi_{xy}}$ of this tensor is the vector $[1,0,2,3]^\intercal$. Moreover,
%}
%\begin{exam}\label{exp:vecc}
%consider the tensor $\phi_{x_3x_2x_1}$ given in Example \ref{exp:tensor}, it is easy to see that $\vecc{\phi_{x_3x_2x_1}} = \frac{1}{2\sqrt{2}}[1,1,1,-1,-i,i,-1,-1]^\intercal$, \addt{consisting with the table representation given in Example \ref{exp:tensor} while providing us with mathematical convenience for performing calculations. Suppose a $P^2\otimes I \otimes I$ operator is applied to it, it will be changed to a tensor whose Vectorization is $P^2\otimes I \otimes I \vecc{\phi_{x_3x_2x_1}} = \frac{1}{2\sqrt{2}}[1,1,1,-1,i,-1,-i,-i]^\intercal$, corresponding to contraction an operator $P^2$ to the index $x_3$.}
%\end{exam}
\addt{Tensor vectorization provides a convenient way to define how an operator acts on a tensor. For instance, the action of an operator $O_3 \otimes O_2 \otimes O_1$ on tensor $\phi_{x_3x_2x_1}$ can be mathematically expressed as
\[
\vecc{\phi'_{x_3x_2x_1}} = (O_3 \otimes O_2 \otimes O_1)\,\vecc{\phi_{x_3x_2x_1}}.
\]
by using the matrix-vector multiplication.
%\delete{, which is equivalent to}
% \[
% \phi'_{x_3x_2x_1}
% =
% (\sum_{x_3,x_2,x_1}
% (O_3)_{y_3 x_3}\,
% (O_2)_{y_2 x_2}\,
% (O_1)_{y_1 x_1}\,
% \phi_{x_3x_2x_1})|_{y_3 \mapsto x_3,y_2\mapsto x_2,y_1\mapsto x_1}.
% \]
%\begin{align}
%\phi'_{y_3y_2y_1}
%&= \sum_{x_3,x_2,x_1}
% (O_3)_{y_3 x_3}\,
% (O_2)_{y_2 x_2}\,
% (O_1)_{y_1 x_1}\,
% \phi_{x_3x_2x_1}.\nonumber
%\\[-1ex]
%&\Bigl|\,_{y_3 \mapsto x_3,\; y_2 \mapsto x_2,\; y_1 \mapsto x_1}.
%\nonumber
%\end{align}
%Here, $y_3 \mapsto x_3,\; y_2 \mapsto x_2,\; y_1 \mapsto x_1$ represents the rename of the indices.
%\delete{After renaming $y_i$ as $x_i$, we have the desired tensor $\phi'$.}
\begin{exam}\label{exp:vecc}
Consider the tensor $\phi_{x_3x_2x_1}$ given in Example~\ref{exp:tensor}.
Its vector representation is
\[
\vecc{\phi_{x_3x_2x_1}} = \tfrac{1}{2\sqrt{2}}[1,\,1,\,1,\,-1,\,-i,\,i,\,-1,\,-1]^\intercal,
\]
which is consistent with the table representation in Example~\ref{exp:tensor} while offering greater mathematical convenience for computation.
Suppose we apply the operator $P^2 \otimes I \otimes I = \texttt{diag}(1,1,1,1,i,i,i,i)$ to this tensor.
The resulting tensor has the vector representation
\[
(P^2 \otimes I \otimes I)\,\vecc{\phi_{x_3x_2x_1}}
= \tfrac{1}{2\sqrt{2}}[1,\,1,\,1,\,-1,\,1,\,-1,\,-i,\,-i]^\intercal,
\]
which corresponds to contracting the operator $P^2 = \texttt{diag}(1,i)$ with the index~$x_3$.
% That is, when $x_3 =0$, the tensor is unchanged, when $x_3 = 1$, an phase $i$ is added.
\end{exam}
}
\addt{With vectorization, we could describe the isomorphism relation between tensors (as shown in Fig.~\ref{fig:tensoriso}) mathematically.}
%\delete{More specifically, the isomorphism between two tensors can be characterized as follows:}
\begin{fact}\label{fact}
Two rank-$n$ tensors $\phi_{x_n\cdots x_1}$ and $\psi_{x_n\cdots x_1}$ are $\mathcal{G}$-isomorphic if and only if there exist a nonzero complex number $\lambda$ and $O=\lambda \cdot O_n\otimes \cdots \otimes O_1 \in \mathcal{G}$ such that $\vecc{\phi_{x_n\cdots x_1}} = \lambda \cdot O_n\otimes \cdots \otimes O_1 \vecc{\psi_{x_n\cdots x_1}}$.
\end{fact}
This is because, apart from a global coefficient, applying $\lambda \cdot O_n\otimes \cdots \otimes O_1$ to $ \vecc{\psi_{x_n\cdots x_1}}$ is equivalent to contracting an operator $O_i$ over each index $x_i$ of $\psi$. The key point is to ensure that the indices of both tensors remain aligned.
% \begin{remark}
% \delete{
% We emphasize that tensor vectorization is used here solely to simplify notation. The expression $(O_n \otimes \cdots \otimes O_1) \vecc{\psi_{x_n \cdots x_1}}$ represents the contraction of an operator $O_i$ to each index $x_i$ of $\psi$, with the corresponding index retained as $x_i$. Even without vectorization, tensor isomorphisms can be defined, as illustrated in Fig.~\ref{fig:tensoriso}, and applied to LimTDD. Furthermore, LimTDD’s core operations, such as addition and contraction described below, rely on the representation of isomorphic operators rather than vectorization.}
% \end{remark}
%\magenta{The following discussion is out of place: We have not introduced LimTDD yet.}
%\delete{Suppose $\vecc{\phi_{x_n\cdots x_1}} = \lambda \cdot O_n\otimes \cdots \otimes O_1 \vecc{\psi_{x_n\cdots x_1}}$. \addt{Utilizing tensor isomorphism,} tensor $\phi$ can be represented by the decision diagram obtained by assigning a weight $\lambda \cdot O_n\otimes \cdots \otimes O_1 $ to the incoming edge of the decision diagram representing $\psi$. Note that this process can be recursively applied to the sub-tensors $\psi_{x_n=0}$ and $\psi_{x_n=1}$.}
%\delete{For a rank-$n$ tensor $\varphi$, an LIM $O\in\mathcal{O}_n$ is a stabilizer of $\varphi$ if $O\vecc{\varphi}=\vecc{\varphi}$. For a subgroup $\mathcal{G}$ of $\mathcal{O}$, the $\mathcal{G}$-stabilizer group of $\varphi$ is defined as $\stab(\vecc{\varphi})$.}
In the remainder of this paper, when there is no ambiguity, we will not distinguish between $\phi_{x_n\cdots x_1}$ and $\vecc{\phi_{x_n\cdots x_1}}$. Additionally, we will use the notations $\lambda \cdot \phi_{x_n\cdots x_1}$ and $O \cdot \phi_{x_n\cdots x_1}$ to represent multiplying a tensor by a scalar or contracting an operator on corresponding indices. Specially, $\vecc{O \cdot \phi_{x_n\cdots x_1}}=O \cdot \vecc{\phi_{x_n\cdots x_1}}$.
%\delete{In addition, we always use $\vecc{*}$ to represent the vectorization of a vector and use $\ket{*}$ to represent a quantum state.}
\subsection{LimTDD: Definition and Examples}
%当我们将所有的数据限制为量子态，并且权重只取X,Z的时候，LIMDD是LIMTDD的一个特殊情况；
%当我们将张量看做其Vector形式的时候，LIMTDD是LIMDD的一个推广。
%从电路的角度来论述，只比较TDD和LIMTDD
\begin{defn}[LimTDD]\label{def:limtdd}
Let $\mathcal{G}$ be a subgroup of $\mathcal{O}$. A $\mathcal{G}$-LimTDD $\mathcal{F}$ over a set of indices $S$ is a rooted and weighted DAG
$\mathcal{F} = (V, E, \idx, \low, \high, \w)$ defined as follows:
\begin{itemize}
% \item $V$ is a finite set of nodes which consists of non-terminal nodes $V_{NT}$ and a terminal node $v_T$ labeled with integer 1. Denote by $r_\mathcal{F}$ the unique root node of $\mathcal{F}$;
\item $V$, $V_{NT}$, $v_T$, $\low$, $\high$, and $E$ are interpreted analogously to their roles in LIMDDs, representing the nodes, non-terminal nodes, the terminal node, 0- and 1-successor mappings, and edges. In particular, we write $r_\F$ and $e_r$ for the root node and the incoming edge of $\F$.
\item $\idx: V_{NT} \rightarrow S$ assigns each non-terminal node an index in $S$. We call $\idx(r_\mathcal{F})$ the top index of $\F$, if $r_\mathcal{F}$ is not the terminal node;
% \item both $\low$ and $\high$ are mappings in $V_{NT} \rightarrow V$, which assign each non-terminal node with its 0- and 1-successors, respectively;
% \item $E = \{(v, \low(v)), (v, \high(v)) : v\in V_{NT}\}$ is the set of edges, where $(v, \low(v))$ and $(v, \high(v)) $ are called the low- and high-edges of $v$, respectively. For simplicity, we also assume the root node $r_\mathcal{F}$ has a unique incoming edge, denoted $e_r$, which has no source node;
\item $\w: E\rightarrow \mathcal{G}$ assigns each edge a weight in $\mathcal{G}$. $\w(e_r)$ is called the weight of $\mathcal{F}$, and denoted $w_\mathcal{F}$.
% \item $w: E\rightarrow \mathcal{O}(n)$ assigns each edge a weight in $\mathcal{O}(n)$, where $n$ is the level of the edge, $\mathcal{O}(n)=\{ c\cdot X^{b_n}D(\theta_n)\otimes \cdots \otimes X^{b_1}D(\theta_1)\ |\ c\in \mathbb{C},\ b_i \in \{0,1\},\ w_i \in [0,1) \}$. $\w(e_r)$ is called the weight of $\mathcal{F}$, and denoted $w_\mathcal{F}$.
\end{itemize}
\end{defn}
% \delete{For any edge $e$ pointing to a node $v'$, the tensor represented by $e$ is defined as $\Phi(e) = \w(e)\cdot \vecc{\Phi(v')}$; and for any internal node $v$ with index $x_v$,}
% \begin{equation}\label{eq:limtdd-intp}
% \delete{\Phi(v) = \overline{x}_v \cdot \Phi((v,\low(v)))
% +x_v \cdot \Phi((v,\high(v))),}
% \end{equation}
% \delete{and the terminal node represents the scalar 1. }
%\addt{For any edge $e$ pointing to a node $v'$ representing a rank-$k$ tensor, the tensor represented by $e$ is also a rank-$k$ tensor, whose vectorization is $\vecc{\Phi(e)} = \w(e)\cdot \vecc{\Phi(v')}$; and for any internal node $v$ with index $x_v$, suppose both its two outgoing edges represent a rank-$k$ tensor, $v$ represent a rank-$k+1$ tensor $\Phi(v)$, where $\Phi(v)_{x_v=0} = \Phi((v,\low(v)))$ and $\Phi(v)_{x_v=1} = \Phi((v,\high(v)))$. With a little abuse of the Boolean-Shannon decomposition,
%\begin{equation}\label{eq:limtdd-intp}
%\Phi(v) = \overline{x}_v \cdot \Phi((v,\low(v)))+x_v \cdot \Phi((v,\high(v))).
%\end{equation}
%The terminal node represents the scalar 1. }
\addt{As in LIMDD and TDD, the terminal node $v_T$ represents the scalar~1. For any edge $e$ pointing to a node $v'$ that represents a rank-$k$ tensor,
the tensor associated with $e$ is also a rank-$k$ tensor satisfies
% \magenta{Avoid using tensor vectorization.}
% \begin{align}\label{eq:vecc-edge}
% \vecc{\Phi(e)} = \w(e) \cdot \vecc{\Phi(v')}.
% \end{align}
\begin{align}\label{eq:vecc-edge}
\Phi(e) = \w(e) \cdot \Phi(v').
\end{align}
For any internal node $v$ with index $x_v$, suppose both of its outgoing edges
represent rank-$k$ tensors. Then $v$ itself represents a rank-$(k\!+\!1)$ tensor $\Phi(v)$,
whose two slices along the index $x_v$ are given by
\[
\Phi(v)_{x_v=0} = \Phi((v,\low(v))), \quad
\Phi(v)_{x_v=1} = \Phi((v,\high(v))).
\]
Using a slight extension of the Boolean–Shannon decomposition, we can write
\begin{equation}\label{eq:limtdd-intp}
\Phi(v)
= \overline{x}_v \cdot \Phi((v,\low(v)))
+ x_v \cdot \Phi((v,\high(v))).
\end{equation}
}
For simplicity, for a rank-$n$ tensor $\phi$ with indices $x_n, \ldots, x_1$, we always adopt the ordering $x_n\prec \cdots \prec x_1$, which implies that $x_n$ is above $x_1$ in the LimTDD representation of $\phi$.
% \znote{Have you defined $\prec$?}
% \begin{figure}
% \centering
% % \includegraphics[width=0.8\textwidth]{figures/fig1_2.png}
% % \useyquantlanguage{groups}
% \begin{tikzpicture}
% \begin{yquantgroup}
% \registers{
% qubit {} q[3];
% }
% \circuit{
% box {$O_1$} q[0];
% box {$O_2$} q[1];
% box {$O_3$} q[2];
% box {$U$} (q[0,1,2]);
% box {$O_4$} q[0];
% box {$O_5$} q[1];
% box {$O_6$} q[2];
% }
% \equals[$\cong $]
% \circuit{
% box {$U$} (q[0,1,2]);
% }
% \end{yquantgroup}
% \end{tikzpicture}
% \caption{The efficiency of LIMTDD vs TDD.}
% \label{fig:tdd_vs_limtdd}
% \end{figure}
% \begin{figure}
% \centering
% \subfigure[]{
% % \includegraphics[width=0.35\linewidth]{figures/run_exp_sub_tdd.png}
% \hspace{0.5em}
% \resizebox{0.16\textwidth}{!}{\input{figures/run_exp_sub_tdd}}
% }
% \subfigure[]{
% % \includegraphics[width=0.15\linewidth]{figures/run_exp_sub_limtdd.png}
% \hspace{2.2em}
% \resizebox{0.088\textwidth}{!}{\input{figures/run_exp_sub_limtdd}}
% }
% \caption{\delete{The TDD and LimTDD representations of the tensor shown in Example~\ref{exp:vecc}.}}
% \label{fig:run_exp_sub}
% \end{figure}
% \delete{We next provide two examples to illustrate the construction of LimTDD.}
% \begin{exam}
% \delete{Fig.~\ref{fig:run_exp_sub} illustrates the TDD and LimTDD representations of the tensor $\phi_{y_2x_2y_1x_1}$ from Example~\ref{exp:vecc}. The TDD representation consists of 8 nodes, while the LimTDD representation requires only 5 nodes. Notably, the high edge of the root node in the LimTDD representation carries the operator $P^4\otimes P^4 \otimes I$, which corresponds to the relationship $\vecc{\phi_{y_2=1}}=Z\otimes Z\otimes I \cdot \vecc{\phi_{y_2=0}}$ shown in Example~\ref{exp:vecc}. Similarly, the operator $P^4$ on the high edge of the $y_0$ node reflects the relationship $\vecc{\phi_{y_2x_2y_1=001}}=Z \cdot \vecc{\phi_{y_2x_2y_1=000}}$.}
% \end{exam}
% \delete{Additionally, LimTDD can represent quantum states, as quantum states can also be viewed as tensors.}
%\begin{exam}
%The LimTDD representation of the tensor $\phi_{x_3x_2x_1}$ given in Example~\ref{exp:tensor} is shown in Fig.~\ref{fig:run_exp} (c). \addt{The terminal node represents a scalar 1. The tensor represented by the node labeled with index $x_1$ is a rank-1 tensor $\psi_{x_1}$ where $\psi_{x_1}(0)=1$ and $\psi_{x_1}(1)=1$. Or, $\vecc{\psi_{x_1}}=[1,1]^\intercal$. The node labeled with index $x_2$ represents a rank-2 tensor $\psi^{\prime}_{x_2x_1}$, where $\vecc{\psi^{\prime}_{x_2=0}}=[1,1]^\intercal$ and $\vecc{\psi^{\prime}_{x_2=1}}=P^4[1,1]^\intercal=[1,-1]^\intercal$. Then, $\vecc{\psi^{\prime}_{x_2x_1}} = [1,1,1,-1]^\intercal$. The node labeled with index $x_3$ represent a rank-3 tensor $\psi^{''}_{x_3x_2x_1}$, where $\vecc{\psi^{''}_{x_3=0}} = [1,1,1,-1]^\intercal$ and $\vecc{\psi^{''}_{x_3=1}} = P^6\otimes P^4 [1,1,1,-1]^\intercal=[1,-1,-i,-i]^\intercal$. In other words, $\vecc{\psi^{''}_{x_3x_2x_1}} = [1,1,1,-1,1,-1,-i,-i]^\intercal$. Finally, the tensor represented by the whole LimTDD is a rank-3 tensor $\psi^{''}_{x_3x_2x_1}$, whose vectorization is $\vecc{\psi^{'''}_{x_3x_2x_1}} = \frac{1}{2\sqrt{2}}P^6\otimes I \otimes I\vecc{\psi^{''}_{x_3x_2x_1}}=\frac{1}{2\sqrt{2}}[1,1,1,-1,-i,i,-1,-1]^\intercal$.}
%\end{exam}
\addt{
\begin{exam}
Fig.~\ref{fig:run_exp}\,(c) shows the LimTDD $\F$ of the tensor
$\phi_{x_3x_2x_1}$ introduced in Example~\ref{exp:tensor}. For convenience, we write $v_i$ ($i=1,2,3$) for the node labeled with $x_i$ in $\F$ and write $\Phi(v_i)$ for the tensor represented by $v_i$.
Recall that the tensor corresponds to the terminal node $v_T$ is 1. As $\low(v_1)=v_T$ and the weight on the low edge of $v_1$ is 1, by Eq.~\ref{eq:vecc-edge}, we have
\[\Phi((v_1,\low(v_1)))=\text{wt}((v_1,\low(v_1)))\cdot\Phi(v_T) = 1\cdot 1 = 1.\]
Similarly, $\Phi((v_1,\high(v_1)))=1$. We have by Eq.~\ref{eq:limtdd-intp}
\[ \Phi(v_1)=\overline{x_1}\cdot \Phi((v_1,\low(v_1))+x_1\cdot\Phi(v_1,\high(v_1)) =\overline{x_1} + x_1.\]
This implies that $\Phi(v_1)(0)=1$ and $\Phi(v_1)(1)=1$. Thus tensor $\Phi(v_1)$ has vector representation $\vecc{\Phi(v_1)} = [1,\,1]^\intercal$.
For node $v_2$, we have $\Phi((v_2,\low(v_2))=\Phi(v_1)$, $\Phi((v_2,\high(v_2))=P^4\cdot\Phi(v_1)=Z\cdot \Phi(v_1)$, and
\[\Phi(v_2)=\overline{x_2}\cdot\Phi(v_1)+x_2\cdot Z\cdot \Phi(v_1).\]
As $\vecc{\Phi(v_1)} = [1,\,1]^\intercal$ and
$\vecc{Z\cdot \Phi(v_1)} = Z [1,\,1]^\intercal = [1,\,-1]^\intercal$, we have
$\vecc{\Phi(v_2)} = [1,\,1,\,1,\,-1]^\intercal$.
For node $v_3$, we have
$\Phi((v_3,\low(v_3))=\Phi(v_2)$, $\Phi((v_3,\high(v_3))=(P^6\otimes P^4)\cdot\Phi(v_2)=(ZS\otimes Z)\cdot \Phi(v_2)$, and
\[\Phi(v_3)=\overline{x_3}\cdot\Phi(v_2)+x_3\cdot (ZS\otimes Z)\cdot \Phi(v_2).\]
Because
$\vecc{(ZS\otimes Z)\cdot \Phi(v_2)} = (ZS\otimes Z) [1,\,1,\,1,\,-1]^\intercal = [1,\,-1,\,-i,\,-i]^\intercal$, we have
$\vecc{\Phi(v_3)} = [1,\,1,\,1,\,-1,\,1,\,-1,\,-i,\,-i]^\intercal$.
Finally, the tensor represented by $\mathcal{F}$ is the tensor $\Phi(\F)=\texttt{wt}(\F)\cdot\Phi(v_3)$, whose vector representation is
\begin{align*}
\vecc{\Phi(\F)} = & \texttt{wt}(\F)\cdot \vecc{\Phi(v_3)} \\
=& \tfrac{1}{2\sqrt{2}}(P^6 \!\otimes\! I \!\otimes\! I)[1,\,1,\,1,\,-1,\,1,\,-1,\,-i,\,-i]^\intercal \\
=& \tfrac{1}{2\sqrt{2}}[1,\,1,\,1,\,-1,\,-i,\,i,\,-1,\,-1]^\intercal.
\end{align*}
This is exactly the vector representation of $\phi_{x_3x_2x_1}$.
\end{exam}
}
%\begin{comment}
%\begin{exam}
%Fig.~\ref{fig:run_exp}\,(c) shows the LimTDD of the tensor $\phi_{x_3x_2x_1}$ introduced in Example~\ref{exp:tensor}.
%The terminal node represents the scalar~1. The node labeled with $x_1$ corresponds to a rank-1 tensor $\psi_{x_1}$ with $\psi_{x_1}(0)=1$ and $\psi_{x_1}(1)=1$, that is, $\vecc{\psi_{x_1}} = [1,\,1]^\intercal$. The node labeled with $x_2$ represents a rank-2 tensor $\psi'_{x_2x_1}$, where$\vecc{\psi'_{x_2=0}} = [1,\,1]^\intercal$, $\vecc{\psi'_{x_2=1}} = P^4 [1,\,1]^\intercal = [1,\,-1]^\intercal$. Hence, $\vecc{\psi'_{x_2x_1}} = [1,\,1,\,1,\,-1]^\intercal$. The node labeled with $x_3$ represents a rank-3 tensor $\psi''_{x_3x_2x_1}$, where $\vecc{\psi''_{x_3=0}} = [1,\,1,\,1,\,-1]^\intercal$, $\vecc{\psi''_{x_3=1}} = (P^6 \!\otimes\! P^4)[1,\,1,\,1,\,-1]^\intercal = [1,\,-1,\,-i,\,-i]^\intercal$. Therefore, $\vecc{\psi''_{x_3x_2x_1}} = [1,\,1,\,1,\,-1,\,1,\,-1,\,-i,\,-i]^\intercal$. Finally, the tensor represented by the entire LimTDD is the normalized rank-3 tensor $\psi'''_{x_3x_2x_1}$, whose \delete{vectorization} \red{vector representation} is
%\begin{align*}
%\vecc{\psi'''_{x_3x_2x_1}} =& \tfrac{1}{2\sqrt{2}}(P^6 \!\otimes\! I \!\otimes\! I)\,\vecc{\psi''_{x_3x_2x_1}} \\ =& \tfrac{1}{2\sqrt{2}}[1,\,1,\,1,\,-1,\,-i,\,i,\,-1,\,-1]^\intercal.
%\end{align*}
%\magenta{Should we replace $\psi'''_{x_3x_2x_1}$ with $\phi_{x_3x_2x_1}$ here?}
%\end{exam}
%\end{comment}
% \addt{What we want to emphasise here is that, although this example comes from a quantum state, and in this paper we mainly focus on tasks such as quantum circuit simulation and functionality representation, the ability of LimTDD to represent tensors gives us the advantage to cope with more complex tensor network based tasks.}
Typically, a node $v$ in a LimTDD is uniquely determined by its index, its two successors, and the weights on the two edges connecting to those successors. Following \cite{vinkhuijzen2023limdd}, the LimTDD that represents the tensor of $v$ is denoted as
\[\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}.\]
Similarly, a LimTDD is uniquely determined by its root node and the weight on the incoming edge, represented as:
\[\Big(w_\F,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big),\] or simply as \[\xrightarrow[]{w_\F}\Circled{v}.\]
%\red{It is straightforward to observe that two identical LimTDD structures, differing only in the weights on their incoming edges.}
% {\color{blue}Ying: Discuss relationship between TDD, LIMDD and LimTDD here. Any embedding/isomorphism/homomorhpism, with respect to, say, $XP$-operators?
% Thus, strengthen Theorem 2}
When the group $\mathcal{G}$ is restricted to the Pauli group,
LimTDD and LIMDD become equivalent in representing quantum states.
Furthermore, when $\mathcal{G}$ is restricted to scalar multiples of the identity
(i.e., $\mathcal{G} \subset \mathbb{C} \cdot I$),
LimTDD reduces exactly to the standard TDD.
In our practical implementations, we choose $\mathcal{G}$ to be the $XP$-stabilizer group introduced in Sec.~\ref{sec:xplim}.
%When the group $\mathcal{G}$ is restricted to Pauli group, LimTDD and LIMDD are the same when representing quantum states. Moreover, when $\mathcal{G}$ is restricted to the scalar multiples of the identity (i.e., $\mathcal{G}\subset \mathbb{C}\cdot I$), LimTDD reduces exactly to standard TDD. \blue{In practical implementations, we choose $\mathcal{G}$ to be the $XP$-stabilizer group introduced in Sec.~\ref{sec:xplim}.}
%\subsection{$XP$-operators}
%\magenta{This subsection will be merged into the previous one.}
% \znote{Allowing XP-stab weights seems to be the key to improvement over limdd and to achieve linear size for qft. I think it helps to talk about the design choice here. Why XP-stab, and why not more general?}
%While the definition of tensor isomorphism allows for general $2\times 2$ invertible maps $O_i$, in practice, the operators are restricted to a smaller group for computational efficiency. For LIMDD, this group is restricted to Pauli operators, as their algebraic properties simplify representation and computation.
%In this subsection, we extend these operators using $XP$-operators~\cite{webster_xp_2022}. This extension allows for more compact decision diagrams while leveraging the algebraic properties of $XP$-operators to ensure canonicity. \addt{At the same time, it ensures the efficiency of operations such as Slicing.} The basic notations and results of this subsection come from \cite{webster_xp_2022}.
%An $n$-qubit $XP$-operator of precision $N$ (where $\omega=e^{\pi i/N}$) is an operator of the form:
%\begin{equation}
% XP_N(p|\bold{x}|\bold{z}) := \omega^p \bigotimes_{0\leq i<n}X^{\bold{x}[i]}P^{\bold{z}[i]},
%\end{equation}
% \znote{Define $\omega$ and $P$?}
%where $p$, $\bold{x}[i]$, and $\bold{z}[i]$ are integers satisfying:
%\begin{itemize}
% \item $0\leq p<2N$,
% \item $0\leq \bold{x}[i]<2$,
% \item $0\leq \bold{z}[i]<N$.
%\end{itemize}
%We call $p$, $\bold{x}$, and $\bold{z}$ the phase component, $X$-component, and $Z$-component of the $XP$-operator, respectively. It is easy to see that two $XP$-operators $XP_N(p|\bold{x}|\bold{z})$ and $XP_N(p'|\bold{x'}|\bold{z'})$ are identical if and only if $p=p'$, $\bold{x}=\bold{x'}$, and $\bold{z}=\bold{z'}$.
%An $XP$-operator can be uniquely represented by the vector $(p|\bold{x}|\bold{z}) \in \mathbb{Z}_{2N}\times \mathbb{Z}_{2}^{n} \times \mathbb{Z}_{N}^n$. We denote this operator as $XP_N(p|\bold{x}|\bold{z})$,
%\[XP_N(p|\bold{x}|\bold{z}) = XP_N(p\mod{2N}|\bold{x} \mod{2}|\bold{z}\mod{N}),\]
%A natural order of $XP$-operators can be defined as follows: $XP_N(p_1|\bold{x_1}|\bold{z_1}) < XP_N(p_2|\bold{x_2}|\bold{z_2})$ if and only if $(p_1|\bold{x_1}|\bold{z_1}) < (p_2|\bold{x_2}|\bold{z_2})$ in the lexicographical order.
%Understanding the group generated by $XP$-operators requires characterizing the unit operator, multiplication, and inverses of $XP$-operators. We will need the following notation: for any integer vector $\bold{z} \in \mathbb{Z}^n$, we write $D_N(\bold{z}) := XP_N(\sum_i{\bold{z}[i]}|\bold{0}|{-}\bold{z})$, denoting the antisymmetric operator for $\bold{z}$.
%First, we note that
%$$I = XP_N(0|\bold{0}|\bold{0}),\quad -I=XP_N(N|\bold{0}|\bold{0}).$$
%Second, suppose $\bold{u}_1 = (p_1|\bold{x_1}|\bold{z_1})$ and $\bold{u_2} = (p_2|\bold{x_2}|\bold{z_2})$. The multiplication of two $XP$-operators $XP_N(\bold{u_1})$ and $XP_N(\bold{u_2})$ can then be computed as:
%\begin{equation}\label{eq:XPD1}
% XP_N(\bold{u_1})XP_N(\bold{u_2}) = XP_N(\bold{u}_1+\bold{u}_2)D_N(2\bold{x_2}\bold{z_1}),
%\end{equation}
%where addition and multiplication of vectors are performed component-wise and with proper modulos. That is:
%\begin{eqnarray*}
% (\bold{a}+\bold{b})[j] &=& \bold{a}[j]+\bold{b}[j] \mod{\bold{m}[j]},\\
% (\bold{a}\bold{b})[j] &=& \bold{a}[j]\bold{b}[j] \mod{\bold{m}[j]},
%\end{eqnarray*}
%where $\bold{m}$ is a $(2n+1)$-dimensional vector such that
%\[
%\bold{m}[j]=\begin{cases}
% 2N, & \text{if $j=0$}\\
% 2, & \text{if $1\leq j\leq n$}\\
% N, & \text{if $n+1\leq j\leq 2n$}\\
% \end{cases}
%\]
%Lastly, the inverse of the operator $XP_N(p|\bold{x}|\bold{z})$ is given by:
%\begin{equation}
%\label{eq:XPD2}
%XP_N(-p|\bold{x}|-\bold{z})D_N(-2\bold{x}\bold{z}).
%\end{equation}
%Furthermore, since the $X$-components of $D_N(2\bold{x_2}\bold{z_1})$ and $D_N(-2\bold{x}\bold{z})$ are $\bold{0}$, right multiplication of these operators in \eqref{eq:XPD1} and \eqref{eq:XPD2} only modifies the phase and $Z$-components of the original $XP$-operator.
%As established in \cite[Proposition 4.1]{webster_xp_2022}, $XP$-operators possess useful group properties. Examples include elegant characterizations of the unit operator, multiplication, and inverses of $XP$-operators% (cf.~\ref{sec:appxp}). Of particular importance is the following result: Let $G = \{g_1,\cdots, g_m\}$ be a set of $XP$-operators, and denote the group generated by $G$ as $\langle G\rangle$. It was shown in \cite{webster_xp_2022} that there exists a unique set of diagonal operators $S_Z:=\{ B_j:0 \leq j <s\}$ and non-diagonal operators $ S_X:=\{ A_i:0\leq i<r\}$, called the \emph{canonical generators} of $\langle G\rangle$, such that all group elements $g \in \langle G\rangle$ can be expressed as $g=S_X^{\bold{a}} S_Z^{\bold{b}}$, where $\bold{a} \in \mathbb{Z}_2^{|S_X|}$ and $\bold{b} \in \mathbb{Z}_2^{|S_Z|}$. Furthermore, two sets of $XP$-operators of precision $N$ generate the same group if and only if they have the same canonical generators. \addt{In the following part of this paper, we set $\mathcal{G}$ to be the group formed by $XP$-operators.}
%The following properties ensure the effective representation and operation of $XP$-stabilizer groups.
%\begin{thm}[Proposition 4.1 of \cite{webster_xp_2022}]
% \label{thm:cangen}
%Let $G = \{g_1,\cdots, g_m\}$ be a set of $XP$-operators. There exists a unique set of diagonal operators $S_Z:=\{ B_j:0 \leq j <s\}$ and non-diagonal operators $ S_X:=\{ A_i:0\leq i<r\}$, called canonical generators, with the following form:
%(1) The binary matrix formed from the $X$-components of the $S_X$ is in Reduced Row Echelon Form (RREF),
%(2) Let $B_j$ has the form $\omega^p\cdot \bigotimes_{0\leq i <n} P^{\bold{z}[i]}$, the $s\times (n+1)$ matrix with rows formed by $(2\bold{z}|p)$, denoted as $B$, is in Howell Matrix Form,
% Let $S_{Z_p}$ be matrix whose rows are the image of $S_Z$ under the $Z_p$ map of $Z_p(XP_N (p|0|z)) = (2z|p))$. The matrix $S_{Z_p}$ is in Howell Matrix Form,
%(3) For $A_i = \omega^p\cdot \bigotimes_{0\leq i <n} X^{\bold{x}[i]}P^{\bold{z}[i]}$, the matrix $\begin{bmatrix}1 & (2\bold{z}|p) \\ 0 &B\end{bmatrix}$ is in Howell Matrix Form.
%With this set of canonical generators, all group elements $g \in \langle G\rangle$ can be expressed as $g=S_X^{\bold{a}} S_Z^{\bold{b}}$ where $\bold{a} \in \mathbb{Z}_2^{|S_X|}$ and $\bold{b} \in \mathbb{Z}_2^{|S_Z|}$, and two sets of $XP$-operators of precision N generate the same group if and only if they have the same canonical generators.
%\end{thm}
% Equipped with the above tools, we now are ready to provide a characterization for $\langle F\rangle$, the group generated by a finite set $F$ of $XP$-operators.
% %Any set of $XP$-operators $\bold{G}$ generates a group. For this group, we can find a set of canonical generators in the following sense.
% We have the following result:
% \begin{thm}
% \label{thm:cangen}
% Let $G$ be a group generated by a finite set of $XP$-operators. There exists a unique set of diagonal $XP$-operators $S_Z\subset G$ and a unique set of non-diagonal $XP$-operators $S_X\subset G$ such that every element $g$ in $G$ can be uniquely expressed as $g=S_X^{\vec{a}} S_Z^{\vec{b}}$, where $\vec{a} \in \mathbb{Z}_2^{|S_X|}$ and $\vec{b} \in \mathbb{Z}_2^{|S_Z|}$.
% \end{thm}
% In the following, we call $S_X$ and $S_Z$ the canonical generators of the group $G$. The proof of the above theorem is deferred to the appendix.
% \magenta{This part is not clear. The proof is not given. Complete the proof and put it in the appendix. The meaning of $g=S_X^{\vec{a}} S_Z^{\vec{b}}$ is not clear. Note $S_X$ and $S_Z$ are subsets not individual operators.}\xh{These parts are the existing conclusions in \cite{webster_xp_2022}, not the conclusions we derive.}
% We first assert that there exists a unique set of diagonal operators $S_Z:=\{ B_j:0 \leq j <s\}$ and a unique set of non-diagonal operators $S_X:=\{ A_i:0\leq i<r\}$ with the following properties:
% \begin{enumerate}
% \item The binary matrix formed from the $X$-components of the $S_X$ operators is in Reduced Row Echelon Form (RREF).
% \item For each diagonal operator $B_j=\omega^p\cdot \bigotimes_{0\leq i <n} P^{\vec{z}[i]}$, the matrix formed by $(2\vec{z}|p)$, denoted as $B$, is in Howell Matrix Form.
% \item For each non-diagonal operator $A_i = \omega^p\cdot \bigotimes_{0\leq i <n} X^{\vec{x}[i]}P^{\vec{z}[i]}$, the matrix $\begin{bmatrix}1 & (2\vec{z}|p) \\ 0 &B\end{bmatrix}$ is in Howell Matrix Form, where $B$ is the matrix from property 2).
% \end{enumerate}
% Then, every element $g$ in the group generated by $G$, denoted $\langle G\rangle$, can be uniquely expressed as $g=S_X^{\vec{a}} S_Z^{\vec{b}}$, where $\vec{a} \in \mathbb{Z}_2^{|S_X|}$ and $\vec{b} \in \mathbb{Z}_2^{|S_Z|}$. Furthermore, two sets of $XP$-operators of precision $N$ generate the same group if and only if they have the same canonical generators.
\begin{remark}
LIMDDs exploit the algebraic properties of Pauli operators to achieve compact representations and efficient computations. In LimTDDs, we extend these to $XP$-operators~\cite{webster_xp_2022}, enabling more compact decision diagrams while preserving canonicity through the algebraic structure of $\mathcal{XP}$. This extension also ensures efficient operations, such as slicing, which extracts sub-tensors from the diagram.
We explain the rationale for selecting $\mathcal{XP}$ over more general groups for LimTDD. Several factors must be considered to ensure canonical representations and efficient computations in the decision diagram:
\begin{enumerate}
\item The operators must have a canonical form, with multiplication and inversion operations being computationally efficient.
\item These operators should be efficiently extractable from the decision diagram structure. For example, an \( X \) operator can be extracted by swapping a node’s outgoing edges, while a \( P \) operator can be applied by adjusting the phase of the high edge.
\item Computing the stabilizer group for a node (cf. Eq.~\ref{eq:stab_state}) within this group must be efficient to maintain the diagram’s canonical form; otherwise, adding stabilizers to edges without altering functionality could disrupt canonicity.
\item The stabilizer group should have a compact representation.
\item Computing sub-diagrams representing sub-tensors must be straightforward, as LimTDDs’ core algorithms, such as addition and contraction, rely on this operation.
\end{enumerate}
Fortunately, $XP$-operators satisfy all these requirements \cite{webster_xp_2022}.
\end{remark}
% \vspace{-1.5em}
\subsection{LimTDD Normalization}
%For weighted decision diagrams, the functionality is determined by multiplying the weights along a path. Thus, it is possible to construct different decision diagrams that represent the same functionality by modifying the weights. For example, $\Big(w,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$, $\Big(w\cdot w_0,\ \Circled{v_{0}}\overset{1}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1/w_0}\Circled{v_{1}}\Big)$. normalization is a step in reducing this ambiguity and making the representation canonical. The core idea is to add some restrictions to the weights on different edges and make the representation unique. In this example, we ask the first non-zero weight on the outgoing edges of every internal node to be 1 and make this one the only one acceptable.
In decision diagrams with weighted paths, the function value associated with each path is obtained by multiplying the weights along that path. As a result, multiple decision diagrams with different weight assignments can represent the same function. \addt{The following lemma provides examples of LimTDDs, in general forms, that represent the same tensor.}
%Normalization addresses this ambiguity by establishing a canonical representation, ensuring each function corresponds to a unique diagram. For LimTDD, where edge weights are Local Invertible Maps (LIMs), the normalization process is more complex than for TDDs.
%In decision diagrams that use weighted paths, the \blue{function value corresponding to each path is obtained by multiplying the weights along the path.} This can lead to a situation where different decision diagrams, with different weight assignments, represent the same underlying functionality. Normalization is a process that reduces this ambiguity, aiming to create a canonical representation where each functionality has a unique diagram representation. %Normalization achieves this by imposing specific constraints on the weights of edges within the decision diagram.
%For LimTDD, as the weight on an edge can be a LIM, the normalization process is more complicated than TDD.
\begin{lem}
Suppose $v$ is a node in a LimTDD with form
\[\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}.\]
Let $g_i$ be a $XP$-stabilizer of $\vecc{v_i}$ $(i=0,1)$, and $0\leq k <N$.
Then the following LimTDDs represent the same tensor $\Phi(v)$:
\[\Big( 1 ,\ \Circled{v_{0}}\overset{w_0\cdot g_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1\cdot g_1}\Circled{v_{1}}\Big),\]
\[\Big(P^ k \otimes (w_0\cdot g_0),\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{ \omega^{-2k}\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1}\Circled{v_{1}}\Big),\]
\[\Big( X \otimes (w_1\cdot g_1),\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{g_1^\dagger\cdot w_1^\dagger\cdot w_0\cdot g_0}\Circled{v_{0}}\Big).\]
%Suppose $v$ is an internal node in a LimTDD, $v_0=\low(v)$ and $v_1=\high(v)$ its child nodes, and $w_0, w_1$ the weights on the low- and high-edges of $v$, respectively. \emp{Let $g_i$ be any stabilizer of $\vecc{v_i}$ $(i=0,1)$, and $0\leq k <N$.}
%The following LimTDDs
%\[\Big( 1 ,\ \Circled{v_{0}}\overset{w_0\cdot g_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1\cdot g_1}\Circled{v_{1}}\Big),\]
%\[\Big(P^ k \otimes (w_0\cdot g_0),\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{ \omega^{-2k}\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1}\Circled{v_{1}}\Big),\]
%\[\Big( X \otimes (w_1\cdot g_1),\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{g_1^\dagger\cdot w_1^\dagger\cdot w_0\cdot g_0}\Circled{v_{0}}\Big)\]
%represent the same tensor as
%\[\Big( 1 ,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big).\]
\end{lem}
\addt{In the above result, the successors $v_0,v_1$ can represent arbitrary tensors. We next give a concrete example, where $v_0,v_1$ represent rank 1 tensors.}
%\addt{While we focus on $XP$-operators, the stabilizer $g_i$ in the above lemma could be any operators $O$ such that $O\vecc{v_i}=\vecc{v_i}$.}
\begin{exam}\label{exp:id_dds}
\addt{Consider a node $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$. Suppose $\vecc{v_0}=[1,0]^\intercal, \vecc{v_1} = [1,1]^\intercal$. The vectorization of $\Phi(v)$ is $[1,0,1,1]^\intercal$, where $\Phi(v)$ is the tensor corresponds to node $v$. Because $P^k[1,0]^\intercal=[1,0]^\intercal$ for any $k \in \mathbb{N}$ and $X[1,1]^\intercal=[1,1]^\intercal$, we know $\Big( 1 ,\ \Circled{v_{0}}\overset{P^k}{\dashleftarrow}\Circled{v}\xrightarrow[]{X}\Circled{v_{1}}\Big)$ represents $\Phi(v)$.
%the same tensor as $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$, whose vectorization is $[1,0,1,1]^\intercal$.
As $X\otimes I [1,1,1,0]^\intercal = [1,0,1,1]^\intercal$, $\Big( X\otimes I ,\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{0}}\Big)$ also
represents $\Phi(v)$.}
%\addt{Consider a node $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$. Suppose $\vecc{v_0}=[1,0]^\intercal, \vecc{v_1} = [1,1]^\intercal$. Since $P^k[1,0]^\intercal=[1,0]^\intercal$ for any $k \in \mathbb{N}$ and $X[1,1]^\intercal=[1,1]^\intercal$. Then, $\Big( 1 ,\ \Circled{v_{0}}\overset{P^k}{\dashleftarrow}\Circled{v}\xrightarrow[]{X}\Circled{v_{1}}\Big)$ represent the same tensor as $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$, whose vectorization is $[1,0,1,1]^\intercal$. Also, $\Big( X\otimes I ,\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{0}}\Big)$ represent the same tensor since $X\otimes I [1,1,1,0]^\intercal = [1,0,1,1]^\intercal$.}
\end{exam}
The above lemma and example show that the same tensor can be represented by several different LimTDDs. Thus, a normalization process should be conducted.
For any node $v$ in a LimTDD, we write
% \begin{equation}\label{eq:stab_v}
% \stab(v) = \{O\in \mathcal{O}\ |\ O\vecc{\Phi(v)} =\vecc{\Phi(v)} \}.
% \end{equation}
\begin{equation}\label{eq:stab_v}
\stab(v) = \{O\in \mathcal{XP}\ |\ O\vecc{\Phi(v)} =\vecc{\Phi(v)}\}
\end{equation}
for the set of stabilizers of $v$ in $\mathcal{XP}$. The procedure for calculating $\stab(v)$ is described in \ref{sec:stab}. The complexity for calculating $\stab(v)$ is polynomial to the rank of the tensor represented by $v$.
% \magenta{Explain how important and difficult it is to calculate $\stab(v)$.}
\addt{Given a node $\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$, to ensure a unique representation, two sets of weights need to be considered:}
\begin{eqnarray*}\label{eq:W0}
W_0 := \{g_0^{\dag} \cdot w_{0}^\dag \cdot w_{1} \cdot g_1 \ |\ g_0 \in \stab(v_0), g_1 \in \stab(v_1)\}, \\ \label{eq:W1}
W_1 :=
\{g_1^{\dag} \cdot w_{1}^\dag \cdot w_{0} \cdot g_0\ |\ g_0 \in \stab(v_0), g_1 \in \stab(v_1)\}.
\end{eqnarray*}
For LimTDD, the normalization process incorporates three principles to ensure a unique representation for each internal node $v$:
\begin{enumerate}
\item $\low(v) \leq \high(v)$.
\item $\w((v,\low(v))) =I$, unless it is 0.
\item If $\low(v) < \high(v)$, then $\w((v,\high(v)))$ is the smallest one in $W_1$, up to a complex coefficient; otherwise, $\low(v) = \high(v)$ and $\w((v,\high(v)))$ is the smallest in $W_0\cup W_1$, up to a complex coefficient.
\end{enumerate}
% To select a unique representation, we adopt the following assumptions:
Here, for two nodes $v_1, v_2$, $v_1 < v_2$ if $v_1$ is created before $v_2$ (typically with a smaller id). For two weights $w_1 = r_1 e^{2\pi i \theta_1} XP_N(p_1|\bold{x_1}|\bold{z_1})$ and $w_2 = r_2e^{2\pi i \theta_2}XP_N(p_2|\bold{x_2}|\bold{z_2})$,
\[w_1<w_2\quad \text{if}\quad (\bold{x_1}|\bold{z_1}|r_1|\theta_1|p_1) < (\bold{x_2}|\bold{z_2}|r_2|\theta_2|p_2)\]
in lexicographical order, where $r_1, r_2, \theta_1, \theta_2 \in \mathbb{R}$, and $r_1,r_2\geq 0$. $r_1 e^{2\pi i \theta_1}$ and $r_2 e^{2\pi i \theta_2}$ are the polar forms of complex coefficients \addt{(cf. the notations after Definition~\ref{def:lim})}.
\begin{algorithm}
\caption{$\locnorm(x,\F_0,\F_1)$}
\begin{algorithmic}[1]
\Require{Two normalized LimTDDs $\xrightarrow[]{w_i}\Circled{v_i}$ ($i=0,1)$; and an index $x$. We require $v_0\leq v_1$}.
%\Require{Two normalized LimTDDs $\F_i$ with root $v_i$ and weight $w_i$ ($i=0,1)$; and an index $x$. We require $v_0\leq v_1$. }
\Ensure{A normalized LimTDD $\Big(\tilde{w},\Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{\hat{w}}\Circled{v_{1}}\Big)$}.
\State $b_n = 0$\hfill \# \texttt{Are $v_0,v_1$ exchanged?}
%\If{$v_1 < v_0$ }
%\State $\F_0, \F_1 = \F_1, \F_0$
%\State $v_0, v_1 = v_1, v_0$
% \znote{what is $r_{\F_1}$?}
%\State $b_n \leftarrow 1$
%\EndIf
\If{$w_0=w_1=0$ }
\State \Return $\texttt{make\_dd}(0,x,0,v_0,0,v_1)$
%\hfill \# \texttt{a DD representing the 0 tensor}
\EndIf
\State $w_\text{temp} \leftarrow \min_{g_0 \in \stab(v_0),g_1 \in \stab(v_1)} {g_0^{\dag} \cdot w_0^\dag \cdot w_1 \cdot g_1}$
\State $w_\text{temp}' \leftarrow \min_{g_0 \in \stab(v_0),g_1 \in \stab(v_1)} {g_1^{\dag} \cdot w_1^\dag \cdot w_0 \cdot g_0}$
\If{$v_1 = v_0$ and $w_\text{temp}' < w_\text{temp}$}
\State $b_n \leftarrow 1$
\State $g_1^\star \leftarrow \text{the stabilizer of } v_1 \text{ yielding } w_\text{temp}'$
\State $w \leftarrow w_1\cdot g_1^\star$ \hfill %\# \texttt{$g_1^\star$ is the stabilizer of $v_1$ which yields $w_\text{temp}'$}
\State $w'_1 \leftarrow w_\text{temp}'$
\Else
\State $g_0^\star \leftarrow \text{the stabilizer of } v_0 \text{ yielding } w_\text{temp}$
\State $w \leftarrow w_0 \cdot g_0^\star$ \hfill
%\# \texttt{$g_0^\star$ is the stabilizer of $v_0$ which yields $w_\text{temp}$}
%\State $w'_0 \leftarrow I$
\State $w'_1 \leftarrow w_\text{temp}$
% \znote{so $w_0$ is set to $I$ all the time?}
\EndIf
%\State $w'_0 \leftarrow I$
% \State $\theta \leftarrow \texttt{ang}(w_1)$
\State $k \leftarrow \lfloor \texttt{ang}(w'_1)/2\omega \rfloor$
%\State $w'_1 \leftarrow w'_1/\omega^{2k}$
%\State $w \leftarrow X^{b_n}P^k \otimes w$
%\State $dd \leftarrow \texttt{make\_dd}(w,x,w'_0,v_0,w'_1,v_1)$
\State $dd \leftarrow \texttt{make\_dd}(X^{b_n}P^k \otimes w,x,I,v_0,w'_1/\omega^{2k},v_1)$
\State \Return $\texttt{check\_unique\_table}(dd)$
\end{algorithmic}
\label{alg:LimTDD_norm}
\end{algorithm}
The normalization process adopts a bottom-up approach, enforcing the three conditions on every node through the $\locnorm$ procedure (cf. Algorithm~\ref{alg:LimTDD_norm}). This procedure takes two sub-LimTDDs, $\F_0 = \xrightarrow[]{w_0}\Circled{v_0}$ and $\F_1 = \xrightarrow[]{w_1}\Circled{v_1}$, and the index of their intended parent node $x$ as inputs. Assume $v_0\leq v_1$; that is, $v_0<v_1$ or they are identical. Otherwise, we swap the two nodes and add a $X$ operator. If $v_0<v_1$, i.e., $v_0$ is created before $v_1$, we set $\low(v)=v_0$ and $\high(v)=v_1$, and select the smallest value in $W_1$ as the weight on the high edge; otherwise, $v_0=v_1$ and we select the smallest value $w'_1$ in $W_0\cup W_1$ as the weight on the high edge. After that, the weight of the low-edge will be set to $I$ (unless it is 0), and the weight of the resulting LimTDD will be adjusted correspondingly.
% Suppose $v_0$ and $v_1$ are the root nodes of the two sub-LimTDDs, $\F_0$ and $\F_1$, that are subjected to the normalization procedure and $v$ is their intended parent node with index $x$.
% Let $v_i=r_{\F_i}$ be the root node of $\F_i$ for $i=0,1$. We need to decide which of $v_0$ and $v_1$ should be the 0-successor of $v$, and determine the normalized weights on the low and high edges of $v$.
The subroutine $\texttt{make\_dd}(w,x,w_0,v_0,w_1,v_1)$ in this algorithm is used to construct a LimTDD $$\Big( w ,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big).$$
In the last step of the procedure, we check if the node of the resulted LimTDD already exists in the \texttt{unique\_table}; if so, reuse the node; otherwise, we generate a new LimTDD node and store it in the hash table with the key $(x,w_0,r_{\F_0},w_1,r_{\F_1})$. With these techniques, we make the LimTDD unique.
%The $g_0$ and $g_1$ in line 13 and 17 are the stabilizers making $g_0^{\dag} \cdot w_{\F_0}^\dag \cdot w_{\F_1} \cdot g_1$ and $g_1^{\dag} \cdot w_{\F_1}^\dag \cdot w_{\F_0} \cdot g_0$ minimal, respectively.
\begin{exam}
\addt{
Suppose $\xrightarrow[]{P^k}\Circled{v_0}$ and $\xrightarrow[]{X}\Circled{v_1}$ are two LimTDDs, where $\vecc{v_0} = [1,0]^\intercal$ and $\vecc{v_1}=[1,1]^\intercal$. We combine them to form a larger tensor using the $\locnorm$ operation. The resulting node is $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$ if $v_0\leq v_1$, or $\Big( X \otimes I,\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{0}}\Big)$ if $v_0> v_1$. Both representations are equivalent to $\Big( 1 ,\ \Circled{v_{0}}\overset{P^k}{\dashleftarrow}\Circled{v}\xrightarrow[]{X}\Circled{v_{1}}\Big)$, describing the same tensor.
% With the $\locnorm$ operation, when we encounter two LimTDDs $\xrightarrow[]{P^k}\Circled{v_0}$ and $\xrightarrow[]{X}\Circled{v_1}$, where $\vecc{v_0} = [1,0]^\intercal$ and $\vecc{v_1}=[1,1]^\intercal$, and want to combine them to represent a larger tensor, the output node will exactly be $\Big( 1 ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{1}}\Big)$, if $v_0\leq v_1$, and be $\Big( X \otimes I,\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{I}\Circled{v_{0}}\Big)$, if $v_0> v_1$. Both represent the same tensor as $\Big( 1 ,\ \Circled{v_{0}}\overset{P^k}{\dashleftarrow}\Circled{v}\xrightarrow[]{X}\Circled{v_{1}}\Big)$.
}
\end{exam}
New local normalizations can also be obtained from established ones.
\begin{lem}\label{lem:norm}
Suppose $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = \xrightarrow[]{w}\Circled{v}$. Then we have the following new local normalizations:
\begin{itemize}
\item[(i)] $\locnorm\Big(x,\xrightarrow[]{O\cdot w_0}\Circled{v_0},\xrightarrow[]{O\cdot w_1}\Circled{v_1}\Big) = \xrightarrow[]{I\otimes O\cdot w}\Circled{v}$, for any $O \in \mathcal{XP}$;
\item[(ii)] $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{\omega^{2k}w_1}\Circled{v_1}\Big) = \xrightarrow[]{(P^k\otimes I)\cdot w}\Circled{v}$;
\item[(iii)] $\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) =\xrightarrow[]{(X\otimes I) \cdot w}\Circled{v}$, if $v_0\neq v_1$ or $w_0 \neq w_1$;
\item[(iv)] $\locnorm\Big(x,\xrightarrow[]{w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{w_1 \cdot g_1}\Circled{v_1}\Big) = \xrightarrow[]{w}\Circled{v}$, for any $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$.
\end{itemize}
\end{lem}
\addt{This lemma provides further opportunities for merging nodes. Suppose, for example, we are given two sub-LimTDDs, $\xrightarrow[]{O\cdot w_0}\Circled{v_0}$ and $\xrightarrow[]{O\cdot w_1}\Circled{v_1}$. The normalization result share the same node $v$ obtained from normalizing $\xrightarrow[]{w_0}\Circled{v_0}$ and $\xrightarrow[]{w_1}\Circled{v_1}$.
}
\begin{remark}
\addt{
The $\locnorm$ operation corresponds to the $MakeEdge$ operation described in~\cite{vinkhuijzen2023limdd}. However, computing $\stab(v)$ for $XP$-operators, as detailed in Appendix~\ref{sec:stab}, is more complex than for XZ-stabilizers. Additionally, for any $k \in \mathbb{N}$, we can extract the phase $\omega^{2k}$ from the weight of a node’s high edge and replace it with a $P^k$ operator on the incoming edge, whereas only the phase $-1$ can be extracted for XZ-stabilizers. This provides more opportunity for merging two nodes (cf. Lemma~\ref{lem:norm} (ii)).
}
% Note that the $\locnorm$ operation corresponds to the $MakeEdge$ operation given in \cite{vinkhuijzen2023limdd}. The main difference is that it is more complex to calculate $\stab(v)$ for $XP$-operators as shown in Appendix \ref{sec:stab} than XZ-stabilizers. In addition, for any $k \in \mathbb{N}$, we could extract the phase $\omega^{2k}$ in the weight of the high-edge of a node and replace it with a $P^k$ operator on the incoming edge of the node, while only phase -1 can be extracted for XZ-stabilizers.
\end{remark}
% \begin{remark}
% \addt{
% The primary computational overhead of the $\locnorm$ operation lies in calculating $w_{\text{temp}}$ and $w_{\text{temp}}'$, which requires computing $\stab(v_0)$ and $\stab(v_1)$ and identifying minimal operators. This process involves a polynomial-time complexity proportional to the operator length, ensuring LimTDD's canonical representation but significantly reducing operational efficiency, as core LimTDD operations rely heavily on $\locnorm$ and invoke it frequently. To mitigate this, our implementation assumes $\stab(v) = {I}$ for all nodes $v$, allowing direct computation of $w_{\text{temp}} = w_0^\dagger \cdot w_1$ and $w_{\text{temp}}' = w_1^\dagger \cdot w_0$.
% \addt{This breaks the canonicity to some extent, but will not change the tensor represented by the created node.} Consequently, the complexity of the $\locnorm$ operation is significantly reduced, enhancing overall efficiency.
% % \magenta{Explain why this does not affect the validity of the result.}
% %The main time-consuming of the $\locnorm$ process is in calculating $w_\text{temp}$ and $w_\text{temp}'$, where $stav(v_0)$ and $stab(v_1)$ need to be calculated and the minimal operators should be found, which needs a time-consuming polynomial to the length of the operators. This ensures the canonical representation of LimTDD, but significantly reduces the efficiency of the operations of LimTDD, as they all rely on $\locnorm$ operation and will call this procedure numerously. So, in our implementation, we just assume that $stab(v) = {I}$ for all $v$, thus, we directly have $w_\text{temp} = w_0^\dagger\cdot w_1$ and $w_\text{temp}' = w_1^\dagger\cdot w_0$. As a result, the complexity of the $\locnorm$ itself changes to be linear in the length of the operator in $w_0$ and $w_1$.
% }
% \end{remark}
% \begin{lem}\label{lem:norm}
% If $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = (w, node)$, then
% \begin{itemize}
% \item[(i)] $\locnorm\Big(x,\xrightarrow[]{O\cdot w_0}\Circled{v_0},\xrightarrow[]{O\cdot w_1}\Circled{v_1}\Big) = (O\cdot w, node)$;
% \item[(ii)] $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{\omega^{2k}w_1}\Circled{v_1}\Big) = \big((P^k\otimes I)\cdot w, node\big)$.
% \item[(iii)] $\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) = \big((X\otimes I) \cdot w, node\big)$, if $v_0\neq v_1$ or $w_0 \neq w_1$;
% \item[(iv)] $\locnorm\Big(x,\xrightarrow[]{w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{w_1 \cdot g_1}\Circled{v_1}\Big) = (w, node)$, for any $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$.
% \end{itemize}
% \end{lem}
\section{Algorithms}\label{sec:Algorithms}
This section presents algorithms for constructing LimTDDs from tensors and performing key operations such as addition and contraction. All algorithms are implemented recursively. As with TDDs, we employ local normalization whenever a node is created. Additionally, we use \texttt{computed\_table} (a hash table technique) to prevent redundant calculations.
\addt{These algorithms have standard counterparts in traditional decision diagrams: slicing corresponds to restriction, addition corresponds to apply, and contraction corresponds to matrix multiplication (cf. \cite[Sec.~3.2]{bahar1997algebric}). However, unlike matrix multiplication, contraction allows us to handle operations on data with incompatible dimensions more easily. Here, we provide a concise restatement of these algorithms from a tensor network perspective. This approach introduces operator actions on edges, while also allowing users familiar with tensor networks to recognize and leverage these operations effectively.}
\subsection{LimTDD Generation}
To generate the LimTDD representation of a tensor, we recursively construct the LimTDD representations of its two sub-tensors and then apply the normalization procedure. For a scalar, the representation is a LimTDD with a single terminal node weighted by that scalar. This process aligns with the TDD generation method described in~\cite{hong2020tensor} (see Algorithm \ref{alg:LimTDD_generate}). The construction has a time complexity linear in the tensor's size (i.e., the number of values in the tensor).
\begin{algorithm}
\caption{$\LimTDD\_generate(\phi)$}
\begin{algorithmic}[1]
\Require{A tensor $\phi$ over a linearly ordered index set $S$.}
\Ensure{The normalized LimTDD of $\phi$.}
%\If{the rank of $\phi$ is 0}
\If{$\phi \equiv c$ is a constant}
% \State \Return the trivial LimTDD with weight $c$
\State \Return $\xrightarrow[]{\ c\ }\Circled{v_T}$
\EndIf
\State $x\leftarrow$ the smallest index of $\phi$
% \State $dd\leftarrow$ an empty LimTDD
% \State $dd.root \leftarrow$ a new node $v$ with index $x$
\State $r_0 \leftarrow \LimTDD\_generate(\phi_{x=0})$
\State $r_1 \leftarrow \LimTDD\_generate(\phi_{x=1})$
% \If{$l>h$}
% \State $l, h = h,l$
% \State $b_n = 1$
% \EndIf
% \State $dd = normlise(l,h)$
% \State $\w(dd) = X^{b_n}\w(dd)$
% \State \Return $\texttt{check\_unique\_table}(dd)$
\State \Return $\locnorm(x,r_0,r_1)$
\end{algorithmic}
\label{alg:LimTDD_generate}
\end{algorithm}
%To generate the LimTDD representation of a tensor, we recursively generate the LimTDD representations of its two sub-tensors and then apply the normalization procedure. For a scalar, the representation is a LimTDD with a single node (the terminal node) weighted by that scalar. \addt{This process aligns with the TDD generation method described in \cite{hong2020tensor} (see also Algorithm~\ref{LimTDD_generate} in the appendix).} The construction has a time complexity that is linear in the size of the tensor (i.e., the number of values in the tensor).
\subsection{LimTDD Slicing}
%权重在DD上的作用
\addt{LimTDD operations, such as addition and contraction, rely on a slicing step, which involves extracting the sub-LimTDD of a sub-tensor from a LimTDD.} Let $\mathcal{F}$ be the LimTDD representation of a tensor $\phi_{x_n \cdots x_1}$. In this section, we describe how to extract from $\mathcal{F}$ the LimTDDs $\mathcal{F}{x_i = c}$ that represent the tensors $\phi_{x_i=c}$ for $c \in \{0,1\}$. Unlike TDDs, this process must account for edge operators, ensuring their effects are propagated to sub-LimTDDs.
%\addt{The operation of LimTDD, such as addition and contraction, all rely on a step slicing, that is to obtain the sub-LimTDD of a sub-tensor from a LimTDD.} Let $\F$ be the LimTDD representation of tensor $\phi_{x_n\cdots x_1}$. In this section, we describe how to extract from $\F$ the LimTDDs $\F_{x_i = c}$ that represent the tensors $\phi_{x_i=c}$ for $c\in \{0,1\}$. \addt{Compared to TDD, we must cope with the operators on the edges, such that the effect of the operators can be passed into sub-LimTDDs.}
First, consider the case when $x_i=x_n$ is the top index. Suppose $\F = \Big(w_\F,\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$, with $w_\F= X^{b_n}P^{z_n}\otimes w$. Here, $P^{z_n}$ adds a phase $\omega^{2z_n}$ to the high edge's weight, and $X^{b_n}$ swaps the 0-successor and 1-successor if $b_n=1$. Thus,
\begin{itemize}
\item if $b_n=0$, then $\F_{x_i = c} = \xrightarrow[]{w \cdot w_{c}}\Circled{v_{c}}$, and
\item if $b_n=1$, then $\F_{x_i = c} = \xrightarrow[]{w \cdot w_{1-c}}\Circled{v_{1-c}}$.
\end{itemize}
Note that $w_1 = \omega^{2z_n}w_1$ in the high edge's weight.
If $x_i\neq x_n$ (i.e., $x_i$ is not the top index), we first calculate $\F_{x_n = b}$ for $b \in \{0,1\}$, then recursively calculate $(\F_{x_n = b})_{x_i=c}$, and combine the results. If $x_i$ is not an index of $\F$, $\F_{x_n = c}=\F$ for both $c=0$ and $1$.
\begin{algorithm}
\caption{$\texttt{Slicing}(\F,x,c)$}
\begin{algorithmic}[1]
\Require{A LimTDD $\F$ with weight $w_\F = \lambda X^{b_n}P^{z_n}\otimes w$ representing the tensor $\phi$; an index $x$ and a value $c\in\set{0,1}$.}
\Ensure{The sub-LimTDD of $\F$ that represents $\phi_{x=c}$.}
\If{$\F$ is a trivial LimTDD}
\State \Return $\F$
\EndIf
\State $x^{\prime}\leftarrow \idx(r_{\F})$
\If{$x < x^{\prime}$}
\State \Return $\F$
\EndIf
\If{$x = x^{\prime}$}
\State $dd\leftarrow$ an empty LimTDD
\State $v_0 \leftarrow \low(r_\F)$
\State $v_1 \leftarrow \high(r_\F)$
\State $w_0 \leftarrow \w(r_\F,\low(r_\F))$
\State $w_1 \leftarrow \w(r_\F,\high(r_\F))$
\If{$c=0$}
\State $dd.node = v_{b_n}$
\State $dd.weight = \lambda \cdot \omega^{2z b_n} \cdot w \cdot w_{b_n}$
\Else
\State $dd.node = v_{1-b_n}$
\State $dd.weight = \lambda \cdot \omega^{2z (1-b_n)} \cdot w \cdot w_{1-b_n}$
\EndIf
\State \Return $dd$
\EndIf
\If{$x > x^{\prime}$}
\State $r_0 \leftarrow \texttt{Slicing}(\texttt{Slicing}(\F,x^{\prime},0),x,c)$
\State $r_1 \leftarrow \texttt{Slicing}(\texttt{Slicing}(\F,x^{\prime},1),x,c)$
\State \Return $\locnorm(x^{\prime},r_0,r_1)$
\EndIf
\end{algorithmic}
\label{alg:LimTDD_Slicing}
\end{algorithm}
% In this procedure, if $x_i=x_n$ is the top index, we could obtain the result by choosing the corresponding successor, since $\xrightarrow[]{\lambda w_\F[n-1)w_{b_n}e^{2\pi i\theta b_n}}\Circled{v_{b_n}}$ represent the sub-tensor $\phi_{x_n=0}$, and $\xrightarrow[]{\lambda w_\F[n-1)w_{1-b_n}e^{2\pi i\theta (1-b_n)}}\Circled{v_{1-b_n}}$ represent the sub-tensor $\phi_{x_n=1}$, suppose $\F = (w_\F,\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}})$, $w_\F=\lambda X^{b_n}P(\theta_n)\otimes \cdots \otimes X^{b_1}P(\theta_1)$. If $x_i\neq x_n$ is not the top index, we should first calculate $\F_{x_n = b}$ for $b \in \{0,1\}$ and then calculate $(\F_{x_n = b})_{x_i=c}$ recursively, and then combine the two parts.
The general slicing procedure for LimTDD is similar to that for TDD described in \cite{hong2020tensor} (see Algorithm \ref{alg:LimTDD_Slicing}). This procedure extracts the portion of data in $\Phi(\F)$ with $x_i=c$ and leads to the following result:
\begin{lem}\label{lem:sub-limtdd}
Let $\F$ be the LimTDD of tensor $\phi_{x_n\cdots x_1}$. For $c\in \{0,1\}$, define $\F_{x_i=c} = \texttt{Slicing}(\F,x_i,c)$. Then, $\F_{x_i=c}$ is the LimTDD representation of the tensor $\phi_{x_i=c}$.
\end{lem}
For the root node of a LimTDD, the \texttt{Slicing} operation is the inverse operation of $\locnorm$: applying the $\locnorm$ operation to the two sliced sub-LimTDDs with respect to $x_n$ retrieves $\F$.
\begin{lem}\label{lem:sub-limtdd_com}
Let $\F$ be the LimTDD representation of tensor $\phi_{x_n\cdots x_1}$. Then
$\locnorm(x_n,\F_{x_n=0},\F_{x_n=1}) = \F$.
\end{lem}
\addt{With the slicing operation defined, we now introduce the core LimTDD operations, namely addition and contraction. These operations enable LimTDD to address most tensor network-based tasks.}
\subsection{LimTDD Addition}
\begin{figure*}[!tbh]
\centering
\input{draw_pic_dot/limtdd_add}
\caption{The addition of two tensors using LimTDD.}
\label{fig:limtdd_add}
\end{figure*}
Let $\F$ and $\G$ be two LimTDDs over an index set $S$. The sum of $\F$ and $\G$, denoted $\F+\G$, is a LimTDD representing the tensor $\bphi(\F)+\bphi(\G)$. Similar to TDD addition, the addition of two LimTDDs can be calculated recursively by adding their sub-LimTDDs. For any $x\in S$ where $x\preceq \idx(r_\F)$ and $x\preceq \idx(r_\G)$, the LimTDD version of the Boole-Shannon expansion gives us
\begin{align*}
\bphi(\F)+ \bphi(\G) = &\ \overline{x} \cdot (\bphi(\F_{x=0})+ \bphi(\G_{x=0}))\\
&\quad\quad + x \cdot (\bphi(\F_{x=1}) + \bphi(\G_{x=1})).
\end{align*}
Here $\F_{x=c}$ and $\G_{x=c}$ are the sub-LimTDDs as defined in Lemma~\ref{lem:sub-limtdd} for $c\in \{0,1\}$.
\begin{algorithm}
\caption{$\add(\F, \G)$}
\begin{algorithmic}[1]
\Require{
Two LimTDDs $\F$ and $\G$}.
\Ensure {The LimTDD representation of $\bphi(\F)+\bphi(\G)$}.
\If{$r_\F= r_\G$ and $\w(\F) = c \cdot \w(\G)$}
\State $dd\leftarrow \F$
\State $dd.weight\leftarrow w_\F\ + \ w_\G$
\State \Return $dd$
\EndIf
\State $l \leftarrow$ the smaller one of $\F$ and $\G$
\State $h \leftarrow$ the bigger one of $\F$ and $\G$
\State $temp\_w \leftarrow \w(h)$
\State $\w(l) \leftarrow \w(h)^\dag\cdot \w(l)$
\State $\w(h) \leftarrow I$
\State $dd\leftarrow \texttt{find\_computed\_table}(l,h,+)$
\If {$dd$ is empty}
\State $x\leftarrow$ the smaller index of $r_\F$ and $r_\G$
\State $r_0 \leftarrow \add(l_{x=0},h_{x=0})$
\State $r_1 \leftarrow \add(l_{x=1},h_{x=1})$
\State $dd\leftarrow \locnorm(x,r_0,r_1)$
\EndIf
\State $dd.weight\leftarrow temp\_w \cdot dd.weight$
\State \Return $dd$
\end{algorithmic}
\label{alg:LimTDD_Add}
\end{algorithm}
The addition procedure for LimTDD is similar to that for TDD \cite{hong2020tensor} (see Algorithm \ref{alg:LimTDD_Add}).
%of the $\add$ operation for LimTDDs in a node-wise manner. In the algorithm, $\F < \G$ means that $r_\F < r_\G$ or $r_\F=r_\G$ and $w_\F<w_\G$.
To enhance the efficiency of reusing intermediate results, we first transfer the weights of one LimTDD to the other before performing the recursive step. As illustrated in Fig.~\ref{fig:limtdd_add}, to calculate the sum $O_3\otimes O_2\otimes O_1 \vecc{\phi}+O_{6}\otimes O_5 \otimes O_4\vecc{\psi}$, we first compute $(O_{6}\otimes O_5 \otimes O_4)^\dag \cdot O_3\otimes O_2 \otimes O_1 \vecc{\phi}+\vecc{\psi}$ and then multiply the result's weight by $O_{6}\otimes O_5 \otimes O_4$. Second, this approach enables reuse of the intermediate result when computing two tensors that differ only by a multiplicative factor. For example, when calculating $O\cdot O_3\otimes O_2 \otimes O_1 \vecc{\phi}+ O\cdot O_{6}\otimes O_5 \otimes O_4\vecc{\psi}$, the calculated sum $O_3\otimes O_2\otimes O_1 \vecc{\phi}+O_{6}\otimes O_5 \otimes O_4\vecc{\psi}$ can be reused, improving computational efficiency.
\addt{LimTDD performs well in practice, but, like TDD, its worst-case complexity depends on the number of non-zero paths in the two diagrams involved in the addition. Specifically, if $\mathcal{P}(\F)$ and $\mathcal{P}(\G)$ denote the number of non-zero paths in $\F$ and $\G$, the worst-case complexity is $\mathcal{O}(\mathcal{P}(\F) \cdot \mathcal{P}(\G))$. This contrasts with standard BDDs, where complexity typically depends only on the number of nodes. The reason is that, when adding two nodes $v_0$ and $v_1$ with weights $w_0$ and $w_1$, the operation cannot be reduced to simply computing $v_0 + v_1$ and then adding $w_0 + w_1$ to the edge weights.}
% \begin{figure}
% \centering
% \includegraphics[width=\textwidth]{figures/LimTDD_Add.png}
% \caption{The addition of two tensors using LimTDD.}
% \label{fig:limtdd_add}
% \end{figure}
% \begin{figure*}[!tbh]
% \centering
% \begin{circuitikz}[scale=0.75, transform shape]
% \node at (0,0) [anchor=north]{
% \begin{tikzcd}[row sep=0.3cm, column sep=0.3cm]
% &\gate[1]{O_1}&\gate[3]{\phi}&\gate[1]{O_4}&\qw& &&\gate[1]{O_7}&\gate[3]{\psi}&\gate[1]{O_{10}}&\qw& &&\gate[1]{O_7^\dag}\gategroup[3,steps=2,style={dashed,rounded
% corners,fill=blue!20, inner
% xsep=2pt},background,label style={label
% position=below,anchor=north,yshift=-0.2cm}]{}&\gate[1]{O_1}&\gate[3]{\phi}&\gate[1]{O_4}\gategroup[3,steps=2,style={dashed,rounded
% corners,fill=blue!20, inner
% xsep=2pt},background,label style={label
% position=below,anchor=north,yshift=-0.2cm}]{}&\gate[1]{O_{10}^\dag}&\qw& &&\gate[3]{\psi}&\qw \\
% &\gate[1]{O_2}& &\gate[1]{O_5}&\qw&+ &&\gate[1]{O_8}& &\gate[1]{O_{11}}&\qw&\cong &&\gate[1]{O_8^\dag}&\gate[1]{O_2}& &\gate[1]{O_5}&\gate[1]{O_{11}^\dag}&\qw&+ && &\qw \\
% &\gate[1]{O_3}& &\gate[1]{O_6}&\qw& &&\gate[1]{O_9}& &\gate[1]{O_{12}}&\qw& &&\gate[1]{O_9^\dag}&\gate[1]{O_3}& &\gate[1]{O_6}&\gate[1]{O_{12}^\dag}&\qw& && &\qw
% \end{tikzcd}
% };\end{circuitikz}
% \caption{The addition of two tensors using LimTDD.}
% \label{fig:limtdd_add}
% \end{figure*}
% \vspace{-0.3em}
\subsection{LimTDD Contraction}
% \begin{figure}
% \centering
% \includegraphics[width=\textwidth]{figures/LimTDD_Cont2.png}
% \caption{The contraction of two tensors using LimTDD.}
% \label{fig:limtdd_cont}
% \end{figure}
% \begin{figure*}[tbh]
% \centering
% \begin{circuitikz}[scale=0.8, transform shape]
% \node at (0,0) [anchor=north]{
% \begin{tikzcd}[row sep=0.3cm, column sep=0.3cm]
% &\gate[1]{O_1}&\gate[3]{\phi}&\gate[1]{O_4}&\qw& &&\gate[1]{O_7}&\gate[3]{\psi}&\gate[1]{O_{10}}&\qw& &&\qw&\gate[3]{\phi}&\gate[1]{O_4}\gategroup[3,steps=2,style={dashed,rounded
% corners,fill=blue!20, inner
% xsep=2pt},background,label style={label
% position=below,anchor=north,yshift=-0.2cm}]{}&\gate[1]{O_{7}}&\qw& &&\gate[3]{\psi}&\qw \\
% &\gate[1]{O_2}& &\gate[1]{O_5}&\qw&\times &&\gate[1]{O_8}& &\gate[1]{O_{11}}&\qw&\cong &&\qw& &\gate[1]{O_5}&\gate[1]{O_{8}}&\qw&\times && &\qw \\
% &\gate[1]{O_3}& &\gate[1]{O_6}&\qw& &&\gate[1]{O_9}& &\gate[1]{O_{12}}&\qw& &&\qw& &\gate[1]{O_6}&\gate[1]{O_{9}}&\qw& && &\qw
% \end{tikzcd}
% };
% \end{circuitikz}
% \caption{The contraction of two tensors using LimTDD, where $\times$ denotes tensor contraction.}
% \label{fig:limtdd_cont}
% \end{figure*}
\begin{figure*}[tbh]
\centering
\input{draw_pic_dot/limtdd_cont}
\caption{The contraction of two tensors using LimTDD, where $\times$ denotes tensor contraction.}
\label{fig:limtdd_cont}
\end{figure*}
Contraction is a fundamental operation in tensor networks. This subsection details how to efficiently implement contraction using LimTDDs. The procedure is also similar to that presented in \cite{hong2020tensor} for TDD.
Let $\F$ and $\G$ be two LimTDDs, $S$ the set of indices appearing in $\F$ or $\G$, and $var$ a subset of $S$ representing the indices to be contacted. We use $\cont(\cdot)$ for both tensor and LimTDD contractions. For any $x\in S$ with $x\preceq \idx(r_\F)$ and $x\preceq \idx(r_\G)$, Eq.~\ref{eq:contdef} implies that if $x\in var$, then $\cont(\left(\bphi(\F),\bphi(\G), var\right)$ equals
$$\sum_{c=0}^{1} \cont(\bphi(\F_{x=c}),\bphi(\G_{x=c}), var\backslash\{x\});$$
otherwise, it equals
\begin{align*}
\hfill \overline{x} \cdot \cont(\bphi(\F_{x=0}),\bphi(\G_{x=0}), var) \quad\quad\quad\quad \\
+ \ x \cdot \cont(\bphi(\F_{x=1}),\bphi(\G_{x=1}), var).
\end{align*}
This means that calculating the contraction of $\F$ and $\G$ involves calculating the contraction of their sub-LimTDDs and then adding or combining the results.
\begin{algorithm}
\caption{$\cont(\F, \G, var)$}
\begin{algorithmic}[1]
\Require {Two LimTDDs $\F$ and $\G$; the set $var$ of indices to be contracted.}
\Ensure {The contracted LimTDD over $var$.}
\If{both $\F$ and $\G$ are trivial}
\State $dd\leftarrow \F$
\State $dd.weight \leftarrow w_\F \cdot w_\G \cdot 2^{len(var)}$
\State \Return $dd$
\EndIf
\State $temp\_w_0 =$ operators in $\w(\F)$ applied to open indices
\State $temp\_w_1 =$ operators in $\w(\G)$ applied to open indices
\State $\w(\G) = w(\G)/temp\_w_1$
\State $\w(\F) = w(\G)^\dag\cdot \w(\F)/temp\_w_0$
\State $\w(\G) = I$
\State $dd\leftarrow \texttt{find\_computed\_table}(\F,\G,\times )$
\If {$dd$ is empty}
\If{$\G$ is identity}
\State $dd\leftarrow \F$
\Else
\State $x\leftarrow$ the smaller index of $r_\F$ and $r_\G$
\If{$x\in var$}
\State $r_0 \leftarrow \cont(\F_{x=0},\G_{x=0},var\backslash\{x\})$
\State $r_1 \leftarrow \cont(\F_{x=1},\G_{x=1},var\backslash\{x\})$
\State $dd\leftarrow \add(r_0,r_1)$
\Else
\State $r_0 \leftarrow \cont(\F_{x=0},\G_{x=0},var)$
\State $r_1 \leftarrow \cont(\F_{x=1},\G_{x=1},var)$
\State $dd \leftarrow \locnorm(x,r_0,r_1)$
\EndIf
\EndIf
\EndIf
\State $dd.weight = temp\_w_0\cdot temp\_w_1\cdot dd.weight$
\State \Return $dd$
\end{algorithmic}
\label{alg:LimTDD_Contraction}
\end{algorithm}
\addt{ The tensor network representation of operators in LimTDD improves efficiency by enabling the reuse of intermediate results, thereby reducing the computational complexity of LimTDD operations. For example, consider the contraction of two LimTDDs representing $(O_3 \otimes O_2 \otimes O_1) \vecc{\phi}$ and $(O_6 \otimes O_5 \otimes O_4) \vecc{\psi}$. As illustrated in Fig.~\ref{fig:limtdd_cont}, we first remove operators applied to open indices ($O_1, O_2; O_5, O_6$) and then transfer operators applied to contracted indices ($O_3; O_4$) to a single LimTDD. This strategy allows intermediate results to be reused when computing contractions of LimTDDs that differ only by operators on open indices (e.g., replacing $O_1$ with $O \cdot O_1$) or by operators on contracted indices where the differences cancel out (e.g., replacing $O_3$ and $O_4$ with $O_3 \cdot O^\dagger$ and $O \cdot O_4$, respectively).}
Algorithm~\ref{alg:LimTDD_Contraction} provides the detailed procedure for LimTDD contraction. While LimTDD is generally more compact and efficient, like TDD, its worst-case time complexity for contraction depends on the number of non-zero paths in the diagrams involved. Let $\mathcal{P}(\F)$ and $\mathcal{P}(\G)$ denote the number of non-zero paths in $\F$ and $\G$, respectively. Then the worst-case complexity is $\mathcal{O}(\mathcal{P}(\F)^2 \cdot \mathcal{P}(\G)^2)$.
% It's important to note that while LimTDD is generally more compact and efficient than TDD, its worst-case complexity, like TDD, is related to the number of non-zero paths in the two diagrams involved in the operation.
% \begin{remark}
% \addt{
% The tensor network representation of operators in LimTDD enables improved efficiency in reusing intermediate results, reducing the computational complexity of LimTDD operations. For example, consider the contraction of two LimTDDs representing $(O_3 \otimes O_2 \otimes O_1) \vecc{\phi}$ and $(O_6 \otimes O_5 \otimes O_4) \vecc{\psi}$. As illustrated in Fig.~\ref{fig:limtdd_cont}, we first remove operators applied to open indices ($O_1, O_2; O_5, O_6$) and then transfer operators applied to contracted indices ($O_3; O_4$) to one LimTDD. This approach allows the intermediate result to be reused when computing contractions of LimTDDs that differ only by operators on open indices (e.g., replacing $O_1$ with $O \cdot O_1$) or by operators on contracted indices where the differences cancel out (e.g., replacing $O_3$ and $O_4$ with $O_3 \cdot O^\dagger$ and $O \cdot O_4$, respectively).
% }
%\addt{What should be noted here is that with the advantage of the tensor network explanation of operators in a LimTDD, we could improve the efficiency of reusing intermediate results and reduce the complexity of the LimTDD involved in the calculation. For example, suppose we want to calculate the contraction of two LimTDDs representing $O_3\otimes \cdots \otimes O_1\vecc{\phi}$ and $O_{6}\otimes \cdots \otimes O_4\vecc{\psi}$. As shown in Fig. \ref{fig:limtdd_cont}, we could first remove the operators that applied to open indices ($O_1, O_2; O_5, O_6$) and then push operators that applied to contracted indices ($O_3; O_4$) to one of them. With this operation, the result can be reused when we calculate the contraction of two LimTDDs that differ only by operators on the open indices (for example, $O_1$ is changed to $O\cdot O_1$), or differ by operators on the contracted indices, but the difference can be canceled (for example $O_3$ and $O_4$ are changed to $O_3\cdot O^\dag$ and $O\cdot O_4$, respectively).
% \end{remark}
\section{Compactness}\label{sec:compare}
% In this section, we denote the LimTDD with $\omega$ set to be $e^{2\pi i/N}$ to be $N$-LimTDD, and the set of all $XP$-operators with precision $N$ to be $\mathcal{XP_N}$
%In this section, we prove that LimTDD is \addt{at least as compact as} LIMDD and TDD, and we demonstrate that the gap can be exponential in some cases.
\blue{To analyze the compactness of LimTDD, we first characterize isomorphic tensors in terms of their LimTDD representations.}
Recall that $\mathcal{XP_N}$ denote the group of $XP$-stabilizers with precision $N$, where $\omega=e^{\pi i/N}$. We use $N$-LimTDD as a shorthand for $\mathcal{XP_N}$-LimTDD. Suppose $\F$ and $\G$ are two $N$-LimTDDs on the same index set and with the same index order. We write $\F \cong \G$ if $r_\F =r_\G$.
\begin{thm}\label{thm:iso_nodes}
%Let $\F$ and $\G$ be the $N$-LimTDD representation of two tensors $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ with the same index order. Then $r_\F =r_\G$ if and only if there exists some $O \in \mathcal{XP_N}$ such that
%\[\vecc{\phi_{x_n\cdots x_1}} = O \vecc{\gamma_{x_n\cdots x_1}}.\] %Moreover, if $r_\F =r_\G$, then $w_\F = O \cdot w_\G \cdot g$ for some $g \in \stab(\Phi(r_\G))$.
Suppose $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ are two tensors with the same index set. Let $\F$ and $\G$ be the $N$-LimTDD representations of $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ with the same index order. The following two conditions are equivalent:
\begin{enumerate}
\item $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ are $\mathcal{XP_N}$-isomorphic.
\item $r_\F =r_\G$ and $w_\F = O \cdot w_\G \cdot g$ for some $O\in \mathcal{XP_N}$ and some $g \in \stab(r_\G)$.
\end{enumerate}
\end{thm}
% \begin{proof}
% $2\Rightarrow 1$ follows directly from the definition. We prove $1\Rightarrow 2$ by induction on the tensor rank.
% %That is, given two tensors $\vecc{\phi} = O \vecc{\gamma}$, let $\F$ and $\G$ be the LimTDD representation of $\phi$ and $\gamma$, we prove $r_\F = r_{\G}$.
% %We prove this by using induction on the rank of the tensor.
% For rank 0 tensors, the claim holds trivially. Now assume it holds for rank $n-1$, and consider the rank-$n$ case. Given that $\vecc{\phi_{x_n\cdots x_1}} = O \vecc{\gamma_{x_n\cdots x_1}}$ with $O= X^{b_n}P^{z_n}\otimes w$, we show $r_\F=r_\G$, and $w_\F = O \cdot w_\G \cdot g$ for some $g \in \stab(\Phi(r_\G))$.
% Suppose $\G_{x_n=0} = \xrightarrow[]{w_0}\Circled{v_0}$ and $\G_{x_n=1} = \xrightarrow[]{w_1}\Circled{v_1}$. Then $\G = \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = (w_\G, r_\G)$, because $\G$ is normalized.
% According to the value of $b_n$, there are two cases to examine:
% \textbf{Case 1}: If $b_n = 0$, then the components satisfy:
% $$\Phi(\F_{x_n=0}) = \phi_{x_n=0} = w \gamma_{x_n=0} = w \Phi(\G_{x_n=0}),$$
% $$\Phi(\F_{x_n=1}) = \phi_{x_n=1} = \omega^{2z_n}\cdot w \gamma_{x_n=1} = \omega^{2z_n}\cdot w \Phi(\G_{x_n=1}).$$
% By the induction hypothesis, there exist $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$ such that:
% \[\F_{x_n=0} = \xrightarrow[]{w\cdot w_0 \cdot g_0}\Circled{v_0}\ \text{and}\ \F_{x_n=1} = \xrightarrow[]{\omega^{2z_n} w\cdot w_1\cdot g_1}\Circled{v_1}.\]
% Because
% \[\F = \locnorm\Big(x_n,\xrightarrow[]{w\cdot w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{ \omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1}\Big),\]
% applying Lemma~\ref{lem:norm} in the order (i), (ii), and (iv), we conclude $\F \cong \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = \G$.
% \textbf{Case 2}: If $b_n=1$, then the components satisfy:
% $$\Phi(\F_{x_n=0}) = \phi_{x_n=0} = \omega^{2z_n}\cdot w \gamma_{x_n=1} = \omega^{2z_n}\cdot w \Phi(\G_{x_n=1}),$$
% $$\Phi(\F_{x_n=1}) = \phi_{x_n=1} = w \gamma_{x_n=0} = w \Phi(\G_{x_n=0}).$$
% There exist $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$ such that:
% \[\F_{x_n=0} = \xrightarrow[]{\omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1}\ \text{and}\ \F_{x_n=1} = \xrightarrow[]{w \cdot w_0\cdot g_0}\Circled{v_0}.\]
% Because
% \[\F = \locnorm\Big(x_n,\xrightarrow[]{\omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1},\xrightarrow[]{w \cdot w_0\cdot g_0}\Circled{v_0}\Big),\]
% applying Lemma~\ref{lem:norm} in the order (i), (ii), (iii), and (iv), we conclude
% \[\F \cong \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big)=G.\]
% This means that, for both cases, $r_\F =r_\G$. Since $w_\F \cdot \Phi(r_\F) =\vecc{\phi_{x_n\cdots x_1}} = O\cdot \vecc{\gamma_{x_n\cdots x_1}} =O\cdot w_\G \cdot \Phi(r_\G)$, we have $w_\G^\dag \cdot O^\dag \cdot w_\F \cdot \Phi(r_\G) = \Phi(r_\G)$. Let $g=w_\G^\dag \cdot O^\dag \cdot w_\F$. Then $g \in \stab(\Phi(r_\G))$ and $w_\F = O \cdot w_\G \cdot g$.
% \end{proof}
This theorem establishes that two $\mathcal{XP_N}$-isomorphic tensors have isomorphic $N$-LimTDD representations. Consequently, the canonicity of LimTDDs follows from this theorem.
%This theorem reveals that an $N$-LimTDD is able to merge two tensors that differ only by an operator in $\mathcal{XP_N}$.
As the number of operators grows with increasing $N$, more isomorphism operators can be extracted between two tensors. This leads to the following result:
\begin{cor}\label{cor:size}
For any tensor $\phi$ and any $N\geq 0$, let $\F$ be its $(N+1)$-LimTDD representation and $\G$ its $N$-LimTDD representation. Then, $size(\F)\leq size(\G)$, where $size(\F)$ and $size(\G)$ denote the number of nodes in $\F$ and $\G$, respectively.
\end{cor}
% We could obtain similar results for TDD and LIMDD as in theorem~\ref{thm:iso_nodes}. What is different is that for TDD, two tensors that want to share the same root node can differ only by a complex multiplier, and for LIMTDD, these local operators can only be Pauli operators. This limits their ability to share nodes. Therefore, in theory, LimTDD will be more compact than TDD and LIMDD.
For quantum states, LIMDDs are equivalent to $2$-LimTDDs. For general tensors, $1$-LimTDDs provide a more compact representation than TDDs by enabling further compression of tensors differing only by a series of $X$ operators. With a slight abuse of notation, we denote TDDs as $0$-LimTDDs. \blue{The following result asserts that LimTDD is at least as compact as TDD (LIMDD) when representing tensors (quantum states).}
\begin{thm}[LimTDD vs. TDD \& LIMDD]
\begin{enumerate}
\item Let $\F$ be the LimTDD representation and $\G$ the TDD representation of the same tensor $\phi_{x_n\cdots x_1}$ with the same index order. Then, $size(\F)\leq size(\G)$.
\item Let $\F$ be the N-LimTDD ($N\geq 2$) representation and $\G$ the LIMDD representation of the same quantum state $\ket{\phi_{x_n\cdots x_1}}$ with the same index (qubit) order. Then, $size(\F)\leq size(\G)$.
\end{enumerate}
\end{thm}
\blue{We demonstrate below that the size gap can be exponential in certain cases.}
\subsection{Tower LimTDD}\label{sec:tower}
A LimTDD is in \emph{tower form} (illustrated in Fig.~\ref{fig:run_exp}) if for every internal $v$, the 0- and 1-successors are identical, i.e., $\low(v)=\high(v)$. In such cases, the tensors represented by the low and high edges of any internal node $v$ differ only by a local operator. Conversely, if two tensors differ only by a local operator, they could be represented as two LimTDDs sharing the same root node. This leads to the following corollary:
\begin{cor}[Tower LimTDD]
A tensor $\phi_{x_n\cdots x_{1}}$ can be represented as a tower LimTDD if and only if for every $i \in \{ n, \cdots, 1\}$, there exists an operator $O_i \in \mathcal{O}(i-1)$ such that: $\vecc{\phi_{x_n\cdots x_{i}=0\cdots0 0}} = O_i \vecc{\phi_{x_n\cdots x_{i}=0\cdots0 1}}$.
\end{cor}
In \cite{vinkhuijzen2023limdd}, it is shown that all stabilizer states admit tower-form representations using LIMDDs. Below, we identify additional classes of quantum states that can be represented as tower-form LimTDDs:
%and demonstrate cases \red{where LimTDDs achieve exponentially greater compactness than LIMDDs}:
\begin{enumerate}
\item States prepared from $\ket{0}$ using the gate set \{$H$, $P_{N'}$\} admit tower $N$-LimTDD representations for $N \geq 0$ and $N'>0$.
\item States prepared from $\ket{0}$ using the gate set \{$X$, $CX$, $P_{N'}$, $CP_{N'}$\} admit tower $N$-LimTDD representations for $N \geq 0$ and $N'>0$.
\item States prepared from $\ket{0}$ using the gate set \{$H$, $CX$, $S$\} admit tower $N$-LimTDD representations for $N \geq 2$.
\item States prepared from $\ket{0}$ using the gate set \{$H$, $P_{N'}$, $CP_{N'}$\} with at most one $H$ gate per qubit, admit tower $N$-LimTDD representations for $N \geq N'>0$.
\end{enumerate}
The first type of circuit generates no entanglement, while the second type produces no superposition. The third type consists of Clifford circuits. For the fourth type, applying a single $H$ gate per qubit preserves the tower form representation for computational basis states, and phase gates ($P_{N'}$) or controlled phase gates ($CP_{N'}$) maintain this structure without disruption. Note that, before applying an $H$ gate, a qubit remains in $\ket{0}$, as $P_{N'}$ and $CP_{N'}$ can not alter its state.
%Noting that there is no entanglement in the first type of circuit, no superposition in the second type of circuit, and the third type of circuit is the Clifford circuit. For the fourth type of circuit, a single $H$ gate applied on each qubit will keep the tower form for a computational basis state, and the phase gate and the controlled phase gate do not change a tower form into a non-tower form.
\begin{remark}
\begin{enumerate}
\item Since LIMDDs are equivalent to 2-LimTDDs for quantum state representations and TDDs correspond to 0-LimTDDs, the first two circuit types admit tower-form representations across all three decision diagrams. The third type (Clifford circuits) requires $N\geq 2$ for tower-form representation in LimTDD.
\item \addt{Notably, the fourth type provide examples demonstrating an exponential advantage of LimTDDs over LIMDDs and TDDs. For instance, consider a circuit with $n$ qubits: starting from $\ket{0}^{\otimes n}$, apply an $H$ gate to each qubit, followed by a $CP_{2^{k-1}}$ gate between qubits $k$ and $1$ for each $k \in \{2,\cdots,n\}$. The output state can be represented with only $n+1$ nodes using a $2^n$-LimTDD, but requires $3\cdot 2^{n-2}$ nodes using an LIMDD and $2^n$ nodes using a TDD.} \addt{Here we continue using the index order $x_n \prec \cdots \prec x_1$. In this order, the $CP_{2^{k-1}}$ gate adds a different phase to the two branches of every $x_k$ node, causing node doubling for TDD when $k>1$. For LIMDD, node doubling starts when $k>2$.}
\item The theoretical compactness assumes strong canonicity, requiring full stabilizer checks in Algorithm~\ref{alg:LimTDD_norm} (lines 5 and 6). In practice, our implementation omits these checks, which may lead to slightly larger node counts than theoretically predicted. However, experimental data indicate this rarely exceeds TDD/LIMDD sizes (cf. Fig. \ref{expe:rand_clifford_t}, Fig. \ref{expe:rand_clifford_t_big} and Table \ref{tab:functionality}).
% \magenta{Link to the experimental data.}
\end{enumerate}
\end{remark}
% \section{Applications}\label{sec:applications}
% % {\color{blue}
% % Ying: Here I'd like to see a new section about applications.
% % The purpose of a more compact representation in LimTDD is to do tasks more efficiently, e.g. equivalence checking, ect. Otherwise, no point for introducing a compact representation.
% % }
% LimTDD aims to provide a more compact and efficient representation for tensors and quantum states, which can significantly enhance the efficiency of various tasks in quantum computing and tensor network computations. In this section, we explore several key applications where LimTDD's compactness translates into practical advantages.
% \subsection{Quantum Circuit Simulation}
% Quantum circuit simulation is a fundamental task in quantum computing, essential for understanding the behavior of quantum circuits, debugging, and optimizing circuit designs. However, simulating quantum circuits on classical computers is challenging due to the exponential growth of the state space with the number of qubits.
% LimTDD's compact representation allows it to handle larger quantum circuits with fewer resources. By exploiting tensor isomorphisms and using the $XP$-stabilizer group, LimTDD can represent quantum states and operations more efficiently, reducing the memory and computational requirements for simulation.
% \subsection{Equivalence Checking of Quantum Circuits}
% Equivalence checking is a critical task in quantum circuit design and verification, where two circuits are deemed equivalent if they produce the same output state for any given input state. LimTDD's compact representation significantly enhances the efficiency of this task.
% By leveraging the $XP$-stabilizer group and local invertible maps, LimTDD can merge structurally similar tensors, reducing the complexity of the decision diagram. This allows for checking the equivalence of two circuits with less memory space, and enables checking larger circuits.
% \subsection{Model Checking of Quantum Circuits}
% Model checking is another important application in quantum circuit verification, where the correctness of a quantum circuit is verified against a formal specification. LimTDD's compact representation allows for more efficient model checking by reducing the computational complexity of the verification process.
% \subsection{Further Applications}
% Beyond the specific applications mentioned above, LimTDD's compact representation can be applied to a wide range of tasks in quantum computing and tensor network computations. For instance, LimTDD can be used for:
% \begin{enumerate}
% \item Quantum State Preparation: Efficiently representing and manipulating quantum states is crucial for tasks such as state preparation. LimTDD's compactness allows for more efficient handling of these tasks.
% \item Quantum Algorithm Design: LimTDD can aid in the design and optimization of quantum algorithms by providing a more efficient way to represent and manipulate quantum operations.
% \item Tensor Network Computations: LimTDD's integration of tensor representations and local invertible maps makes it a powerful tool for general tensor network computations, which are widely used in fields such as quantum many-body physics and machine learning.
% \end{enumerate}
% In the next section, we will compare the efficiency of LimTDD and TDD through experiments, focusing on quantum circuit simulation and functionality construction.
\section{Experiments and Applications}\label{sec:experiments}
%随机Clifford;- 说明虽然不是强canonical,但是对整体效果影响不大 - max node
%Clifford+T三者比较 - 两个不同的t比例 - 在一些比较通用的电路上 LimTDD是有优势的；探索一下几个DD支持的最大规模
%随着qubit和T门比例变化 LimTDD表现的色温图-3个不同的p -- 进一步探索LimTDD在实际电路中的一些性质 --- 最主要要看到的就是p值跟pi/8;16/pi
%An Astonishing Example-I\otimes QFT
%常用算法的functionality:TDD和LimTDD比较-探索一下最大可以做到多少个比特
We conducted a comprehensive evaluation of LimTDD, TDD, and LIMDD, measuring their efficiency in quantum circuit simulation and functionality construction. Our benchmark suite includes:
\begin{itemize}
\item Standard quantum algorithms: QFT, GHZ, Grover, etc.
\item Random Clifford circuits
\item Random Clifford+T circuits with varying T-gate densities
\end{itemize}
For each method, we recorded total execution time and maximum node count during computation.
%Specifically, we measured the time and memory usage required for quantum circuit simulation and functionality construction for each method. We select commonly used algorithms such as QFT, GHZ, and Grover, as well as random Clifford and random Clifford+T circuits as our benchmarks. We collected data on the total time and maximum node count for each method when performing respective tasks.
All three tools were implemented in C++ and executed on the same platform (Intel(R) Core(TM) i5-13600KF CPU, 32GB RAM, GCC compiler). The LIMDD implementation is from \addt{\cite{vinkhuijzen2023efficient}} (\url{https://github.com/cda-tum/mqt-limdd}) and the TDD implementation is from (\url{https://github.com/Veriqc/TDD_C}). The C++ version of LimTDD can be found in \url{https://github.com/Veriqc/LimTDD/tree/C}, where a Python version is also provided in the main branch. All three tools use the same index (qubit) order, and all tensors (quantum gates) were processed from left to right according to the circuit, without any additional optimizations. \addt{The current LIMDD implementation~\cite{vinkhuijzen2023limdd,vinkhuijzen2023efficient} lacks support for matrix-matrix product operations, preventing its use in constructing quantum circuit functionality. Thus, we compare only TDDs and LimTDDs for this task. }
\subsection{Simulation with Random Clifford + T Circuits}
% Three sets of 1000-Random-400-Gates-10-Qubits experiments: T proportion 0.02, 1/7, 0.3. Figures are in dac_expr/xxxCliffT-yy.pdf
We first compare the three decision diagrams using random Clifford + T circuits, a universal gate set commonly used in quantum computation. The probability of T-gate occurrence is set to 0.02, and the circuits are simulated with the input state $\ket{0\cdots 0}$. These circuits, 1000 in total, are randomly generated from the gate set $\{X, Y, Z, S, H, CX, T\}$, each consisting of 10 qubits and 400 gates. The corresponding experiment data is shown in Fig.~\ref{expe:rand_clifford_t}. Experimental results show that LimTDDs achieve greater compactness than LIMDDs for these circuits, with TDDs performs the least efficiently, requiring the highest node counts and computation time.
We also conducted experiments on 1000 randomly generated circuits, each with 20 qubits and 600 gates, where the performance trend is even more pronounced. The average runtime for LimTDDs was less than 0.15 seconds, over 20 times faster than TDDs and 200 times faster than LIMDDs. Furthermore, LimTDDs consistently required fewer nodes, with an average maximum of approximately 484 nodes, compared to approximately 26484 nodes for TDDs and 135603 nodes for LIMDDs. Notably, LimTDDs outperformed both methods in efficiency and compactness in 992 out of 1000 experiments. These results underscore LimTDD's superior performance for Clifford + $T$ circuits, highlighting its practical advantages in quantum computation. Besides, LIMDDs performed worse than TDDs in these experiments, possibly due to implementation limitations in the current version of LIMDDs or incomplete adherence to the methodologies described in~\cite{vinkhuijzen2023limdd}.
\begin{figure*}[htbp]
\centering
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/TDDCliff-T-0.02-0720.pdf}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/LIMDDCliff-T-0.02-0720.pdf}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/LimTDDCliff-T-0.02-0720.pdf}
}
\caption{Comparison of time and space efficiency for TDD, LIMDD, and LimTDD over 1000 random Clifford+T circuits (10 qubits, 400 gates). Average node counts: TDD (\(\approx 391\)), LIMDD (\(\approx 182\)), LimTDD (\(\approx 92\)). The $y$-axes for node counts are scaled using scientific notation (units: $10^3$).
}
\label{expe:rand_clifford_t}
\end{figure*}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/tddCliff0.02big-0721.pdf}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/limddCliff0.02big-0721.pdf}
}
\subfigure[]{
\includegraphics[width=0.3\textwidth]{dac_expr/limtddCliff0.02big-0721.pdf}
}
\caption{Comparison of time and space efficiency for TDD, LimTDD, and LIMTDD over 1000 random Clifford+T circuits (20 qubits, 600 gates). Average node counts: TDD (\(\approx 26,488\)), LIMDD (\(\approx 135,603\)), LimTDD (\(\approx 484\)).
The $y$-axes for TDD and LimTDD node counts are scaled using scientific notation (units: $10^5$ for TDD, $10^6$ for LIMDD, and $10^4$ for LimTDD).}
\label{expe:rand_clifford_t_big}
\end{figure*}
\subsection{Quantum Circuit Functionality Construction}
We then compare the two decision diagrams, TDDs and LimTDDs, in their ability to construct quantum circuit functionality. For this comparison, we use several standard quantum algorithms, including QFT, GHZ state preparation, and Grover’s algorithm. Functionality representation, which precisely describes quantum circuit behavior, is critical for tasks such as equivalence checking. All circuits are sourced from the MQT benchmark suite~\cite{quetschlich2023mqtbench}.
%We then compare the \addt{two} decision diagrams—TDD, \addt{and} LimTDD—in their ability to construct quantum circuit functionality. For this comparison, we employ several common quantum algorithms, including Quantum Fourier Transform (QFT), GHZ state preparation, and Grover's algorithm. Functionality representation, which provides a precise description of quantum circuit behaviour, is essential for tasks such as equivalence checking. All circuits are taken from the MQT benchmark suite \addt{\cite{quetschlich2023mqtbench}}.
%\delete{\addt{Since the implementation of LIMDD \cite{vinkhuijzen2023efficient} does not support the functionality calculating of quantum circuits}, we only present results for TDD and LimTDD. }
The comparison results are summarized in Table~\ref{tab:functionality}. In the worst cases, LimTDDs match the node counts of TDDs and require slightly longer computation time due to their more complex internal operations. However, LimTDDs significantly outperform TDDs in node count for most scenarios. For example, for the `qpeexact\_10' circuit, LimTDDs require only 68,010 nodes, whereas TDDs need 466,988 nodes--approximately seven times more. Similar trends are observed for circuits like `ae' and `dj'. In the best cases, such as the `qft' circuit, LimTDDs exhibit exponential improvements in compactness and efficiency over TDDs. For the `qft\_10' circuit, LimTDDs require only 21 nodes, while TDDs require 525,311 nodes. When scaled to 12 qubits, LimTDDs complete the task with just 25 nodes, compared to 8,392,703 nodes for TDDs.
% times out after 3600 seconds due to its exponential growth in computation time.
% \begin{table}[htbp]
% \centering
% \begin{tabular}{lcc}
% \toprule
% Algorithms & LimTDD & TDD \\
% \midrule
% ae\_10 & 131111 & 174802 \\
% dj\_10 & 21 & 56 \\
% ghz\_10 & 21 & 55 \\
% graphstate\_10 & 21 & 101 \\
% qft\_10 & 21 & 525311 \\
% qftentangled\_10 & 153158 & 567597 \\
% qpeexact\_10 & 3928 & 26978 \\
% qpeinexact\_10 & 15424 & 87334 \\
% qnn\_8 & 65536 & 65536 \\
% portfoliovqe\_8 & 65536 & 65536 \\
% grover-noancilla\_7 & 16384 & 16384 \\
% \bottomrule
% \end{tabular}
% \caption{Comparison of functionality representation of common quantum algorithms.}
% \label{tab:functionality}
% \end{table}
% \begin{table}[htbp]
% \centering
% \caption{The experiment data for functionality of some common algorithms.}
% \label{tab:functionality}
% \setlength{\tabcolsep}{0.75mm}{
% \begin{tabular}{lcccccc}
% \hline
% \multicolumn{1}{c}{\multirow{2}{*}{Benchmarks}} & & \multicolumn{2}{c}{LimTDD} & & \multicolumn{2}{c}{TDD} \\ \cline{3-4} \cline{6-7}
% & & \textbf{Time (s)} & \textbf{Nodes} & & \textbf{Time (s)} & \textbf{Nodes} \\ \cline{1-1} \cline{3-4} \cline{6-7}
% ae\_10 & & 4.651 & 131111 & & 1.819 & 174802 \\
% dj\_10 & & 0.002 & 21 & & 0.004 & 56 \\
% ghz\_10 & & 0.001 & 21 & & 0.002 & 55 \\
% graphstate\_10 & & 0.001 & 21 & & 0.003 & 101 \\
% qft\_10 & & 0.019 & 21 & & 2.416 & 525311 \\
% qftentangled\_10 & & 5.762 & 153158 & & 5.840 & 567597 \\
% qpeexact\_10 & & 0.138 & 3928 & & 0.575 & 26978 \\
% qpeinexact\_10 & & 0.345 & 15424 & & 1.309 & 87334 \\
% qnn\_8 & & 8.483 & 65536 & & 1.938 & 65536 \\
% portfoliovqe\_8 & & 5.174 & 65536 & & 3.425 & 65536 \\
% qwalk-noancilla\_7 & & 5.606 & 11282 & & 3.013 & 15379 \\
% grover-noancilla\_7 & & 81.636 & 16384 & & 23.187 & 16384 \\ \hline
% \end{tabular}}
% \end{table}
\begin{table}[h]
\centering
\caption{The experiment data for functionality of some common algorithms.}
\label{tab:functionality}
\setlength{\tabcolsep}{0.75mm}{
\scalebox{0.8}{
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{Benchmarks}} & & \multicolumn{2}{c}{TDD} & & \multicolumn{2}{c}{LimTDD} \\ \cline{3-4} \cline{6-7}
& & \textbf{Time (s)} & \textbf{Nodes} & & \textbf{Time (s)} & \textbf{Nodes} \\ \cline{1-1} \cline{3-4} \cline{6-7}
ae\_10 & & 1.819 & 174802 & & 4.651 & 131111 \\
% dj\_10 & & 0.004 & 56 & & 0.002 & 21 \\
dj\_60 & & 0.079 & 356 & & 0.061 & 121 \\
dj\_120 & & 0.291 & 716 & & 0.285 & 241 \\
% ghz\_10 & & 0.002 & 55 & & 0.001 & 21 \\
ghz\_60 & & 0.012 & 355 & & 0.004 & 121 \\
ghz\_120 & & 0.033 & 715 & & 0.007 & 241 \\
% graphstate\_10 & & 0.003 & 101 & & 0.001 & 21 \\
graphstate\_30 & & 0.129 & 34301 & & 0.004 & 61 \\
graphstate\_60 & & 75.455 & 3798765 & & 0.011 & 121 \\
grover-noancilla\_7 & & 23.187 & 16384 & & 81.636 & 16384 \\
portfoliovqe\_8 & & 3.425 & 65536 & & 5.174 & 65536 \\
qft\_10 & & 2.416 & 525311 & & 0.019 & 21 \\
qft\_12 & & 137.745 & 8392703 & & 0.074 & 25 \\
% qft\_20 & & - & - & & 60.118 & 41 \\
% qft\_10 & & 0.033 & 3060 & & 0.015 & 21 \\
% qft\_20 & & 225.934 & 3145706 & & 56.674 & 41 \\
qftentangled\_10 & & 5.840 & 567597 & & 5.762 & 153158 \\
qnn\_8 & & 1.938 & 65536 & & 8.483 & 65536 \\
qpeexact\_10 & & 0.575 & 26978 & & 0.138 & 3928 \\
qpeexact\_12 & & 14.793 & 466988 & & 2.261 & 68010 \\
qpeinexact\_10 & & 1.309 & 87334 & & 0.345 & 15424 \\
qpeinexact\_12 & & 60.244 & 1379333 & & 11.398 & 237193 \\
qwalk-noancilla\_7 & & 3.013 & 15379 & & 5.606 & 11282 \\
\hline
\end{tabular}
}
}
\end{table}
% Results for QMDD:
% ae10 dj60 dj120 ghz60 ghz120 gs30 gs60 grov7 port8 qft10 qft12
% Time 1.195 0.019 0.041 0.012 0.025 0.016 4.541 0.029 55.800 1.075 235.254
% Nodes 58358 120 1 120 240 19326 2196638 33 21846 349526 5592406
% qften10 qnn8 qpe10 qpe12 qpein10 qpein12 qwalk7
% Time 6.740 9.218 0.045 9.292 0.171 258.640 0.015
% Nodes (3492700) 21846 17228 307004 57157 902817 28
\subsection{The Influence of Operator Precision}
% HeatMap in dac_expr/CompExpr-717.pdf and dac_expr/CompExprCliff.pdf
In this subsection, we investigate the impact of precision \(N\) on the compactness of LimTDDs in quantum circuit simulation. We consider four circuit types: Clifford, Clifford+\( T (\text{diag}(1, e^{2\pi i / 8})) \), Clifford+\( \sqrt{T} ( \text{diag}(1, e^{2\pi i / 16})) \), and Clifford+\( \sqrt[4]{T} ( \text{diag}(1, e^{2\pi i / 32}) )\). These circuits are selected due to their progressively decreasing minimum rotation angles, which may require higher precision for accurate simulation. Each circuit consists of 400 gates and 15 qubits, with rotation gate proportions of 0.1 and 0.3. We simulate 200 random circuits for each configuration and record the average maximum number of nodes \((\# \text{Nodes})\) required, along with the compression factor, defined as \(\# \text{Nodes} / 2^{15}\). The results are summarized in Fig.~\ref{tab:rot_p}.
%In this subsection, we investigate the impact of precision ($N$) on the compactness of LimTDDs in quantum circuit simulation. We consider four circuit types: Clifford, Clifford+ $T$ ($diag(1,e^{2\pi i/8})$), Clifford+$\sqrt{T}$ ($diag(1,e^{2\pi i/16})$), and Clifford+$\sqrt[4]{T}$ ($diag(1,e^{2\pi i/32})$). These circuits are chosen because their minimum rotation angles decrease progressively, potentially necessitating higher precision for accurate simulation. Each circuit comprises 400 gates and 15 qubits, with rotation gate proportions set at 0.1 and 0.3. We simulate 200 random circuits for each setting and record the average maximum number of nodes \addt{(\#Nodes)} required \addt{, and calculate $\frac{\#Nodes}{2^{15}}$, representing the compression factor of LimTDD}. The results are summarized in Fig.~\ref{tab:rot_p}.
As shown in Fig.~\ref{tab:rot_p}, the compression factor decreases with increasing precision $N$ for all circuit types. However, this trend stops for Clifford+$T$ circuits when $N>4$ and for Clifford+$\sqrt{T}$ circuits when $N>8$. For Clifford+$\sqrt[4]{T}$ circuits, the trend persists up to $N=32$. This shows that higher $N$ enhances LimTDD compactness, underscoring the need for LimTDDs in efficient quantum circuit simulation. However, it also highlights a limitation: beyond a certain $N$, further increases do not improve compactness due to a fundamental limit imposed by the minimum argument of the circuit's matrix elements, necessitating alternative techniques to overcome this constraint.
% \znote{Any intuition or explanation for why it cannot further decrease the size?}
\begin{figure}[htbp]
% \vspace{-0.5em}
\centering
\includegraphics[width=\linewidth]{dac_expr/CompExpr-717.pdf}
\vspace{-1em}
\caption{Results of simulating random circuits using LimTDD with different precision levels ($N$th root of unity). Rows represent different circuit types (Clifford, and Clifford with various rotation gates and proportions; for example, $0.1\_PI/8$ represents a gate $diag(1,e^{2\pi i/8})$ with proportions 0.1). Data shows the compression factor, defined as the average maximum number of nodes divided by $2^{15}$. We use the gradation from light to dark colors to represent the continuous increase in compression factor.
% All data are normalized by the maximum number of nodes that can be generated with the number of qubits, that is, $\frac{\text{\#Nodes}}{2^{\text{qubit}}}$.
}
\label{tab:rot_p}
% \vspace{-1em}
\end{figure}
% \vspace{-1em}
\subsection{An Example of LimTDD's Exponential Advantage}
The impact of precision $N$ on LimTDD compactness for various circuit types, combined with the theoretical comparisons in Section~\ref{sec:compare}, motivates us to identify circuits exhibiting an exponential advantage over other decision diagrams, such as TDDs and LIMDDs. This subsection presents such an example, featuring QFT as its core component. A QFT circuit includes Controlled-$\sqrt[k]{T}$ gates for $k$ in $\{ 2^0,2^1,\cdots, 2^{n-3}\}$. Fig.~\ref{expe:qft} presents the experimental results for simulating the circuit depicted in Fig.~\ref{fig:qc_qft}. The circuit applies an $I\otimes QFT$ operation on the maximally entangled state $\ket{\Psi} = \frac{1}{\sqrt{2^n}}\sum_{i=0}^{2^n-1}\ket{ii}$. The results demonstrate that LimTDDs are exponentially more efficient and compact than TDDs and LIMDDs. For $n=15$ qubits, LimTDDs complete the simulation in 0.5 seconds with a maximum node count of 31. In contrast, TDDs requires 8.6 seconds with 98,302 nodes, and LIMDDs take 37 seconds with 32,797 nodes. These results highlight LimTDD's significant advantages in computational efficiency and memory usage.
\begin{figure}
\centerline{
\begin{tikzcd}[column sep=0.2cm,row sep=0.12cm]
&\qw &\gate{H} &\ctrl{2} &\qw &\qw &\qw &\qw&\qw&\qw\\
&\qw &\gate{H} &\qw &\ctrl{2} &\qw &\qw &\qw&\qw&\qw\\
&\qw &\qw &\gate{X} &\qw &\qw &\gate[2]{\ \ QFT\ \ } &\qw&\qw&\qw\\
&\qw &\qw &\qw &\gate{X} &\qw &\qw &\qw&\qw&\qw\\
\end{tikzcd}
}
\caption{A quantum circuit for preparing $(I\otimes QFT)\ket{\Psi}$, where $\ket{\Psi} = \frac{1}{\sqrt{2^n}}\sum_{i=0}^{2^n-1}\ket{ii}$.}
\label{fig:qc_qft}
\end{figure}
\begin{figure}[h]
\centering
\subfigure[]{
\includegraphics[height=0.3\textwidth]{dac_expr/spaceQFTthin.pdf}
}
\subfigure[]{
\includegraphics[height=0.3\textwidth]{dac_expr/timeQFTthin.pdf}
}
\caption{Time and space comparison of TDD, LimDD, and LimTDD in simulating the circuit shown in Fig.~\ref{fig:qc_qft}. LimTDD used a precision ($N$) of $2^{15}=32,768$.
}
\label{expe:qft}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
This paper presented LimTDD, a novel decision diagram that integrates tensor representations with local invertible maps (LIMs) to achieve compact and efficient representations of quantum states and general tensors. By generalizing LIMDD's use of Pauli operators to the more flexible $XP$-stabilizer group, LimTDD extends applicability beyond quantum states to arbitrary tensor networks while maintaining superior compression. Theoretical analysis demonstrates that LimTDDs are at least as compact as TDDs and LIMDDs, with exponential advantages in best-case scenarios, such as circuits involving QFT and controlled-phase gates. Experimental results confirm these improvements, showing significant reductions in node counts and computation times for quantum circuit simulation and functionality computation tasks.
\addt{Based on LimTDD, Hong et al. propose a novel approach for quantum state preparation in \cite{xin2025limtddqsp}. Examples indicate that, in the best-case scenario, the LimTDD-based method can achieve exponential efficiency gains over state-of-the-art algorithms.} Future work will focus on extending LimTDD's operator support to include multi-qubit gates and arbitrary diagonal operations, further enhancing its compression capabilities. Additionally, we plan to investigate applications in broader tensor network computations beyond quantum computing. These advancements will establish LimTDDs as a versatile and powerful tool for efficient representation and manipulation of high-dimensional data in quantum and classical settings.
% % \appendix
% \section*{Appendix}\label{sec:appendix}
% \subsection{Calculate $\stab(v)$}\label{sec:stab}
% We show how to calculate a generating set for $\stab(v)$ for an internal node $v$, and how to calculate $\min_{g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})} {g_0^{\dag} \cdot w \cdot g_1}$.
% Suppose $v$ is a node with two successors $v_0$, $v_1$, and the label on the high edge is $h$; then there are two cases. First, if $v_0 \neq v_1$, then only operators of the form $P^z \otimes g$ can be in $\stab(v)$, and $g$ must stabilize $v_0$ and $\omega^{2z} h^{\dagger} g h$ must stabilize $v_1$ for some $z$. Otherwise, operators of the form $P^z \otimes g$ and $XP^z \otimes g$ all can be in $\stab(v)$. More specifically:
% (1) If $v_0 \neq v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\}$;
% (2) If $v_0 = v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\} \cup \{ XP^z \otimes g | hg\in \stab(v_0)\ and\ \omega^{2z} g h \in \stab(v_1) \}$.
% For the rank 1 cases, the tensor represented by a normalized node $v$ can only be $[1,r]^\intercal$, where $r \in \mathbb{R}$, and $0 \leq r \leq 1$. Then, there are three cases:
% (1) If $r=0$, then $\stab(v) = \langle P \rangle$;
% (2) If $r=1$, then $\stab(v) = \langle X \rangle$;
% (3) $\stab(v) = \emptyset$, otherwise.
% Suppose $G_0$ and $G_1$ are the generating sets of $\stab(v_0)$ and $\stab(v_1)$ ignoring their phase components. To calculate a generating set for $\stab(v)$, we only need to calculate $\langle G_0 \rangle \cap \langle h G_1 h^\dag \rangle$ and then add a suitable phase for each item. Note that the intersection of two groups of XP Stabilizers can be calculated by solving a system of linear equations $S_{X0}^{\vec{a_0}} S_{Z0}^{\vec{b_0}} = S_{X1}^{\vec{a_1}} S_{Z1}^{\vec{b_1}}$ in the canonical generating set form, and the time complexity is polynomial with the length of the operators. The minimum value of $ {g_0^{\dag} \cdot w \cdot g_1}$ for $g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})$ can also be calculated using their canonical generating set form.
% Note that in lines 9 and 10 of Algorithm~\ref{alg:LimTDD_norm}, we are supposed to find the minimal element in the following sets
% \begin{align*}
% \{g_0^{\dag} \cdot w_{\F_0}^\dag \cdot w_{\F_1} \cdot g_1 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\} \\
% \{g_1^{\dag} \cdot w_{\F_1}^\dag \cdot w_{\F_0} \cdot g_0 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\},
% \end{align*}
% which is useful for making the representation canonical. Although it can be done with a time complexity polynomial with $n$, we still think it is a little complicated. In our implementation, we will not obey this routine, and we will suppose $\stab(r_{\F_0}) = \stab(r_{\F_1}) = \{I\}$, then (iii) in Lemma~\ref{lem:norm} will be changed to \[\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) = \big((X\otimes I) \cdot w, node\big),\] if $w_0^\dag w_1 \neq w_1^\dag w_0$ or $v_0 \neq v_1$; and (iv) needs $w_0\cdot g_0$ and $w_1\cdot g_1$ to remain in the same order as $w_0$ and $w_1$. This damages the property of canonical to some extent, but it is not serious according to our experimental data. In fact, the complete canonicity is not a very necessary matter for simulation or for verification of quantum circuits since two quantum circuits are equivalent if and only if $tr(UV^\dag)=2^n$ and equivalent up to global phase if and only if $|tr(UV^\dag)|=2^n$. Here, $U$ and $V$ represent the matrix representation of the two circuits, $n$ is the number of qubits, $tr(UV^\dag)$ is a scalar and can be calculated by reversing one of the circuits and connecting the input and output of the two circuits.
% \input{Appendix}
\bibliographystyle{IEEEtran}
\bibliography{ref}
\section*{Appendix}\label{sec:appendix}
\input{Appendix}
\end{document}
% \documentclass{standalone}
% \usepackage{tikz}
\usetikzlibrary{backgrounds,fit,calc}
% \usepackage{xcolor}
\definecolor{tensorBlue}{HTML}{224f70}
% \definecolor{tensorBlueLight}{HTML}{4793ca}
% \definecolor{tensorBlue}{HTML}{6F8DA2}
\definecolor{tensorBlueLight}{HTML}{87B9DD}
\definecolor{nodeGray}{HTML}{888888}
\definecolor{nodeGrayLight}{HTML}{ACACAC}
\definecolor{edgeOrange}{HTML}{f18810}
\definecolor{edgeMagenta}{HTML}{f0116d}
\definecolor{edgeGreen}{HTML}{50c1b8}
% \begin{document}
\begin{tikzpicture}[
% Main node styles
tensor ring/.style args={#1/#2}{
circle, fill=#1, minimum size=4.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=#2, minimum size=4ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
small gray ring/.style={
circle, fill=nodeGray, minimum size=3.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=nodeGrayLight, minimum size=3.2ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
% Text styles
tensor text/.style={inner sep=0pt, font=\large},
small text/.style={inner sep=0pt, font=\small}
]
\begin{scope}
\node at (0,0) {$\lambda$};
\end{scope}
\begin{scope}[shift={(0.8,0)}]
\node at (0,0) {$\times$};
\end{scope}
\begin{scope}[shift={(1.8,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[small gray ring] (O1) at (-0.4,0.9) {};
\node[] (O1End) at (-0.8,1.6) {};
\node[small gray ring] (O2) at (0.4,0.9) {};
\node[] (O2End) at (0.8,1.6) {};
\node[small gray ring] (O3) at (0,-0.9) {};
\node[] (O3End) at (0,-1.6) {};
% Add text labels with appropriate styles
\node[tensor text] at (psi) {$\psi$};
\node[small text] at (O1) {$O_1$};
\node[small text] at (O2) {$O_2$};
\node[small text] at (O3) {$O_3$};
\draw[very thick,edgeGreen] (psi) -- (O1);
\draw[very thick,edgeMagenta] (psi) -- (O2);
\draw[very thick,edgeOrange] (psi) -- (O3);
\draw[very thick,edgeGreen] (O1End) -- (O1);
\draw[very thick,edgeMagenta] (O2End) -- (O2);
\draw[very thick,edgeOrange] (O3End) -- (O3);
\end{scope}
\begin{scope}[shift={(3,0)}]
\node at (0,0) {$\cong $};
\end{scope}
\begin{scope}[shift={(4.2,0)}]
% Create ring nodes with empty outer circles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[] (O1End) at (-0.8,1.6) {};
\node[] (O2End) at (0.8,1.6) {};
\node[] (O3End) at (0,-1.6) {};
% Add text labels in the inner circles
\node[tensor text] at (psi) {$\psi$};
\draw[very thick,edgeGreen] (psi) -- (O1End);
\draw[very thick,edgeMagenta] (psi) -- (O2End);
\draw[very thick,edgeOrange] (psi) -- (O3End);
\end{scope}
\end{tikzpicture}
% \end{document}
\section{Background}\label{sec:background}
\subsection{Quantum Computing}
Quantum computing leverages principles of quantum mechanics to perform computations that classical computers cannot efficiently handle. The fundamental unit of quantum information is the \textit{qubit}, which differs from a classical bit by existing in a superposition of states. A single-qubit state is generally represented as:
\begin{equation}
\label{eq:qubit}
\ket{ \varphi} := \alpha_0 \ket{0} + \alpha_1 \ket{1},
\end{equation}
where $\alpha_0$, $\alpha_1$ are complex amplitudes satisfying ${\left| \alpha_0 \right|^2} + {\left| \alpha_1 \right|^2} = 1$. Alternatively, a single-qubit state can be represented as a column vector $[\alpha_0, \alpha_1]^\intercal$ (where $^\intercal$ denotes the transpose). In general, an $n$-qubit quantum state can be expressed as a $2^n$-dimensional vector $[\alpha_0, \alpha_1, \dots, \alpha_{2^n-1}]^\intercal$ in Hilbert space.
% \begin{figure}
% \centerline{$\mathit{X\ gate}: \ \
% \begin{tikzcd}[column sep=0.4cm,row sep=0.4cm]
% &\qw & \gate{X} &\qw &\qw \\
% \end{tikzcd}\ \ \ \ \ \ \ \ \ \ \ \ \ \
% \left[\begin{array}{cccc}
% 0 & 1\\
% 1 & 0\\
% \end{array}\right]
% $}
% \centerline{
% % $\mathit{P(\theta)\ gate}:
% % \begin{tikzcd}[column sep=0.4cm,row sep=0.4cm]
% % &\qw & \gate{P(\theta)} &\qw &\qw \\
% % \end{tikzcd}\ \ \ \ \ \
% % \left[\begin{array}{cccc}
% % 1 & 0\\
% % 0 & e^{2\pi i\theta}\\
% % \end{array}\right]
% % $
% $\mathit{P\ gate}: \ \
% \begin{tikzcd}[column sep=0.4cm,row sep=0.4cm]
% &\qw & \gate{P} &\qw &\qw \\
% \end{tikzcd}\ \ \ \ \ \ \ \ \ \ \ \ \ \
% \left[\begin{array}{cccc}
% 1 & 0\\
% 0 & \omega^2\\
% \end{array}\right]
% $
% }
% \centerline{
% $\mathit{CZ\ gate}:
% \begin{tikzcd}[column sep=0.4cm,row sep=0.4cm]
% &\qw & \ctrl{1} &\qw &\qw \\
% &\qw & \gate{Z} &\qw &\qw \\
% \end{tikzcd}\ \ \ \ \
% \left[\begin{array}{cccc}
% 1 & 0 &0 &0\\
% 0 & 1 &0 &0\\
% 0 & 0 &1 &0\\
% 0 & 0 &0 &-1\\
% \end{array}\right]
% $}
% \caption{Matrix representations of $X$, $P$, and $CZ$ gates. Here, $\omega$ is an $2N$-th root of the unity $\omega = e^{\pi i/N}$. To emphasize the degree of the root of the unity, we also denote it as $P_N$. If not specified, $\omega = e^{\pi i/8}$. }
% \label{fig:gates}
% \end{figure}
Quantum computation is performed using quantum gates, which are represented by unitary matrices acting on qubits. Common quantum gates include the Pauli gates ($X$, $Y$, $Z$), the Hadamard gate ($H$), and controlled operations such as the Controlled $Z$ ($CZ$) gate and the Controlled Not (CNOT) gate (cf.~\cite{nielsen2010quantum}). Each quantum gate has a unique unitary matrix representation in a predefined orthonormal basis.
%\delete{, see Fig.~\ref{fig:gates} for some illustrations}.
Of particular importance to this paper is the Phase gate
$$P = \left[\begin{array}{cccc}
1 & 0\\
0 & \omega^2\\
\end{array}\right],
$$
which is defined with a parameter $\omega \equiv e^{\pi i/N}$, the $2N$-th root of the unity. To emphasize, we also denote $P$ as $P_N$. In this paper, if not specified, $N=8$, $\omega = e^{\pi i/8}$, and $P, P^2, P^4$ correspond to gates $T, S, Z$, respectively.
The outcome of applying a quantum gate to an input state is determined by multiplying the corresponding unitary matrix with the vector representing the input quantum state. For instance, the result of applying an $X$ gate to the input state $\left[{\alpha _0}, {\alpha _1}\right]^\intercal$ is given by $\left[{\alpha _1}, {\alpha _0}\right]^\intercal$. In a broader context, an $n$-qubit quantum gate is represented by a $2^n\times 2^n$-dimensional unitary transformation matrix.
A quantum circuit is composed of qubits and a sequence of quantum gates. When a quantum circuit operates on an input state, the quantum gates are applied sequentially. The functionality of an $n$-qubit quantum circuit is described by a $2^n\times 2^n$-dimensional unitary transformation matrix.
\begin{figure}[h]
\centerline{
\begin{tikzcd}[column sep=0.35cm,row sep=0.35cm]
\lstick{$q_3\ $}&\qw &\gate{H} &\qw &\qw &\qw &\ctrl{1} &\qw &\ctrl{2} &\qw&\qw\\
\lstick{$q_2\ $}&\qw &\gate{H} &\qw &\ctrl{1} &\qw &\gate{S} &\qw &\qw &\qw&\qw\\
\lstick{$q_1\ $}&\qw &\gate{H} &\qw &\gate{Z} &\qw &\qw &\qw &\gate{Y} &\qw&\qw\\
\end{tikzcd}
}
\caption{A quantum circuit with 3 qubits and 6 gates.}
\label{fig:qc}
\end{figure}
\begin{exam}\label{exp:qc}
Consider the circuit shown in Fig.~\ref{fig:qc}. This circuit consists of 3 Hadamard gates, a controlled-$Z$ gate, a controlled-$S$ gate, and a controlled-$Y$ gate. Starting from the input state $\ket{000}$, the output state of this circuit is $\frac{1}{2\sqrt{2}}(\ket{000}+ \ket{001}+\ket{010}-\ket{011}-i\ket{100}+i\ket{101}-\ket{110}-\ket{111})$.
\end{exam}
\subsection{Tensors and Tensor Networks}
Tensor networks are a powerful framework for representing and manipulating high-dimensional data. They have become indispensable in areas such as quantum circuit simulation, quantum many-body physics, and quantum chemistry, where they enable efficient data compression and scalable computation.
Beyond physics, tensor-network techniques have also been applied to %quantum-inspired
machine learning and data modeling, offering both computational efficiency and physical interpretability.
%Tensor network is a useful tool for many tasks, such as quantum circuit simulation and quantum many-body problems. It helps cut the cost of data processing and can be used to discover new phases in quantum matter, predict chemical reactions, and train quantum-inspired machine-learning models.
%A \emph{tensor} is a multidimensional array associated with a set of indices. In quantum computing, we assume each index takes a value in $\{0,1\}$. A tensor with an index set $S=\{x_n, \ldots, x_1\}$ can be viewed as a mapping $\phi: \{0,1\}^S \to \mathbb{C}$, where $\mathbb{C}$ represents the field of complex numbers. For clarity, we denote such a tensor as $\phi_{x_n\ldots x_1}$ or $\phi_{\vec{x}}$, and its value on the evaluation $\{x_i \mapsto a_i, 1\leq i\leq n\}$ as $\phi_{x_n\ldots x_1}(a_n, \ldots, a_1)$, or simply $\phi_{\vec{x}}(\vec{a})$ or even $\phi(\vec{a})$ when unambiguous. The \emph{rank} of a tensor refers to the number $n$ of its indices. Scalars, 2-dimensional vectors, and $2\times 2$ matrices are considered as rank 0, rank 1, and rank 2 tensors, respectively. An $n$-qubit quantum state can be regarded as a tensor with $n$ indices.
A \emph{tensor} can be formally defined as a mapping from a set of indices to the complex numbers. In quantum computing, each index typically takes values in $\{0,1\}$. Given an index set $S = \{x_n, \ldots, x_1\}$, a tensor is a function $\phi : \{0,1\}^S \rightarrow \mathbb{C}$, where $\mathbb{C}$ denotes the field of complex numbers. For clarity, we denote such a tensor as $\phi_{x_n \ldots x_1}$ or $\phi_{\vec{x}}$,
and its value under the assignment $\{x_i \mapsto a_i \mid 1 \le i \le n\}$ as
$\phi_{x_n \ldots x_1}(a_n, \ldots, a_1)$, or simply $\phi_{\vec{x}}(\vec{a})$ or $\phi(\vec{a})$ when unambiguous.
%Equivalently, a tensor can be represented as a multidimensional array whose entries correspond to the complex values assigned to each binary configuration of its indices.
The \emph{rank} of a tensor is the number $n$ of its indices.
Scalars, two-dimensional vectors, and $2\times 2$ matrices correspond to rank-0, rank-1, and rank-2 tensors, respectively.
An $n$-qubit quantum state can be regarded as a rank-$n$ tensor.
The primary operation between tensors is \emph{contraction}. Given tensors $\gamma_{\vec{x}, \vec{z}}$ and $\xi_{\vec{y}, \vec{z}}$ sharing a common index set $\vec{z}$, their contraction results in a new tensor $\phi_{\vec{x}, \vec{y}}$ defined as:
% \znote{For tensor contraction, it could happen that we have a single tensor and want to contract two of its indices. So the above definition which introduces two tensors may not be appropriate.}
\begin{equation}\label{eq:contdef}
\phi_{\vec{x}, \vec{y}}(\vec{a}, \vec{b})=\sum_{\vec{c}\in \{0,1\}^{\vec{z}}}{\gamma_{\vec{x}, \vec{z}}(\vec{a}, \vec{c})\cdot \xi_{\vec{y}, \vec{z}}}(\vec{b}, \vec{c}).
\end{equation}
Another useful operation is \emph{slicing}, analogous to the cofactor operation for Boolean functions. For a tensor $\phi$ with index set $S=\{x, x_n,\ldots,x_1\}$, the slicing with respect to $x = c$ (where $c\in \{0,1\}$) is a tensor $\phi_{x=c}$ over $S'=\{x_n, \ldots, x_1\}$ given by:
\begin{align}
\phi_{x=c}(\vec{a}):= \phi(c, \vec{a})
\end{align}
for any $\vec{a} \in \{0,1\}^n$. The \textit{negative} and \textit{positive} slicings of $\phi$ with respect to $x$ are denoted as $\phi_{x=0}$ and $\phi_{x=1}$, respectively. Suppose $\phi$ does not rely on $x$; that is $x$ is not in the index set of $\phi$. We denote $\phi_{x=0}=\phi_{x=1}=\phi$.
A \emph{tensor network} is represented by an undirected graph $G=(V, E)$ with zero or multiple open edges, where each vertex $v$ in $V$ represents a tensor, and each edge corresponds to a shared index between two adjacent tensors. By contracting connected tensors (i.e., vertices in $V$) in an arbitrary order, we obtain a rank $m$ tensor, where $m$ is the number of open edges of $G$. This tensor, which is independent of the contraction order, is called the tensor representation of the tensor network. For further details, see~\cite{markov2008simulating} and~\cite{biamonte2019lectures}.
%Quantum circuits are natural examples of tensor networks. In this paper, we use applications on quantum circuits as our main examples.
Quantum circuits provide natural examples of tensor networks. In this paper, we use quantum-circuit applications as the primary examples to illustrate our framework.
\begin{exam}\label{exp:tensor}
The circuit shown in Fig.~\ref{fig:qc} can be represented as a tensor network consisting of three rank-2 tensors and three rank-4 tensors. Contracting these tensors with a rank-3 tensor representing the input state $\ket{000}$ yields a rank-3 tensor $\phi_{x_3x_2x_1}$ as follows, which represents the output state as shown in Example~\ref{exp:qc}. Here, we use $x_3,\ x_2,\ x_1$ to represent the indices corresponding to the output of qubits $q_3,\ q_2,\ q_1$.
% \begin{table}[H]
% \centering
% \begin{tabular}{|c|c|c|c|c|c|c|c|c|} \hline
% $x_3x_2x_1$& 000& 001& 010& 011& 100& 101& 110& 111\\ \hline
% $\phi$& $\frac{1}{2\sqrt{2}}$& $\frac{1}{2\sqrt{2}}$& $\frac{1}{2\sqrt{2}}$& $\frac{-1}{2\sqrt{2}}$& $\frac{-i}{2\sqrt{2}}$& $\frac{i}{2\sqrt{2}}$& $\frac{-1}{2\sqrt{2}}$& $\frac{-1}{2\sqrt{2}}$\\ \hline
% \end{tabular}
% % \caption{}
% \label{tab:tensor}
% \end{table}
\begin{table}[H]
\centering
\setlength{\tabcolsep}{4pt} % Adjust column spacing for better fit
\renewcommand\arraystretch{1.6} % Adjust row height for better readability
\begin{tabular}{c|cccccccc}
\( x_3x_2x_1 \) & 000 & 001 & 010 & 011 & 100 & 101 & 110 & 111 \\ \hline
\( \phi \) & \(\frac{1}{2\sqrt{2}}\) & \(\frac{1}{2\sqrt{2}}\) & \(\frac{1}{2\sqrt{2}}\) & \(\frac{-1}{2\sqrt{2}}\) & \(\frac{-i}{2\sqrt{2}}\) & \(\frac{i}{2\sqrt{2}}\) & \(\frac{-1}{2\sqrt{2}}\) & \(\frac{-1}{2\sqrt{2}}\) \\
\end{tabular}
\label{tab:tensor}
\end{table}
% \begin{table}[H]
% \centering
% \begin{tabular}{c|c} \hline
% $x_3x_2x_1$&$\phi$\\ \hline
% 000&$\frac{1}{2\sqrt{2}}$ \\
% 001&$\frac{1}{2\sqrt{2}}$\\
% 010&$\frac{1}{2\sqrt{2}}$\\
% 011&$\frac{-1}{2\sqrt{2}}$\\
% 100&$\frac{-i}{2\sqrt{2}}$ \\
% 101&$\frac{i}{2\sqrt{2}}$ \\
% 110&$\frac{-1}{2\sqrt{2}}$ \\
% 111&$\frac{-1}{2\sqrt{2}}$ \\ \hline
% \end{tabular}
% % \caption{}
% \label{tab:tensor}
% \end{table}
%\begin{table}[H]
% \centering
% \renewcommand\arraystretch{1.8}
% \begin{tabular}{c|c|c|c|c } \hline
% $x_3x_2x_1$&$\phi$ & & $x_3x_2x_1$&$\phi$\\ \hline
% 000&$\frac{1}{2\sqrt{2}}$ & & 100 &$\frac{-i}{2\sqrt{2}}$\\
% &&&&\\
% 001&$\frac{1}{2\sqrt{2}}$ & &101&$\frac{i}{2\sqrt{2}}$\\
% &&&&\\
% 010&$\frac{1}{2\sqrt{2}}$ & &110&$\frac{-1}{2\sqrt{2}}$\\
% &&&&\\
% 011&$\frac{-1}{2\sqrt{2}}$ & &111&$\frac{-1}{2\sqrt{2}}$ \\ \hline
% 100&$\frac{-i}{2\sqrt{2}}$ & & & \\
% 101&$\frac{i}{2\sqrt{2}}$ & & & \\
% 110&$\frac{-1}{2\sqrt{2}}$ & & & \\
% 111&$\frac{-1}{2\sqrt{2}}$ & & & \\ \hline
% \end{tabular}
% \caption{}
% \label{tab:tensor1}
%\end{table}
\end{exam}
%\subsection{TDD and LIMDD}
\subsection{Local Invertible Maps and $XP$-Operators}\label{sec:xplim}
\begin{defn}[Local Invertible Maps \cite{vinkhuijzen2023limdd}]\label{def:lim}
% Let $\mathcal{O}_n = \{ \lambda \cdot O_n \otimes \cdots \otimes O_1 \ |\ \lambda \in \mathbb{C}, O_i\ be\ invertible\ 2\times 2\ matrices\}$.
An $n$-qubit Local Invertible Map (LIM) is an operator $O$ of the form
\begin{equation}\label{eq:limO}
O = \lambda O_n \otimes \cdots \otimes O_1,
\end{equation}
where each $O_i$ is an invertible $2\times 2$ matrix and $\lambda \in \mathbb{C}$. The set of all such maps is denoted as $\mathcal{O}_n$, and the set of all LIMs is defined as
\begin{equation}\label{eq:limCalO}
\mathcal{O} = \bigcup_{n\in \mathbb{N}}{\mathcal{O}_n}.
\end{equation}
\end{defn}
% \magenta{Check if these notations are used in the paper.}
Let $O=\lambda \cdot O_n\otimes \cdots \otimes O_1$ be an LIM. We write $c(O)=\lambda$.
% \delete{We introduce the following notations:}
% \begin{itemize}
% \item \delete{$c(O)=\lambda$,}
% \item \delete{$O[i]=O_i$ for $1\leq i\leq n$,}
% \item \delete{$O[i)= O_i\otimes \cdots \otimes O_1$ for $1\leq i\leq n$.}
% \end{itemize}
Suppose $\lambda = r\cdot e^{2\pi i \theta}$ for some real number $r \in \mathbb{R}$. We call $\theta$ the angle of $\lambda$, denoted as $\texttt{ang}(\lambda)=\theta$.
In practice, the operators are restricted to a smaller group for computational efficiency. Let $G = \{g_1,\cdots, g_m\}$ be a set of LIMs. We denote the group generated by $G$ as $\langle G\rangle$.
Let $\mathcal{G}$ be a subgroup of $\mathcal{O}_n$. For an $n$-qubit quantum state $\ket{\psi}$, an operator $g\in \mathcal{G}$ is a $\mathcal{G}$-stabilizer of $\ket{\psi}$ if $g\ket{\psi}=\ket{\psi}$. The $\mathcal{G}$-stabilizer group of $\ket{\psi}$ is defined as
\begin{equation}\label{eq:stab_state}
\stab(\ket{\psi}) = \{g\in \mathcal{G}\ |\ g\ket{\psi} =\ket{\psi}\}.
\end{equation}
%\addt{Shortly, in Sec.~\ref{sec:tensoriso}, we will see how this concept and notation can be extended to general tensors.}
%\red{This concept and notation extend to general tensors. We defer the discussion to Sec.~\ref{sec:tensoriso}.}
%\magenta{Explain how $O$ is applied on $\ket{\psi}$?}
%where $g$ is a $\mathcal{G}$-stabilizer of a rank $n$ tensor $\varphi$ if $g\cdot\varphi=\varphi$ \addt{[This may need further explanation]}.
An $n$-qubit LIM $O=\lambda O_n \otimes \cdots \otimes O_1$ is a Pauli operator if $\lambda\in \set{1,-1,i,-i}$ and each $O_i$ is $I$ or a Pauli gate $X,Y,Z$.
$XP$-operators are more general than Pauli operators. An $n$-qubit $XP$-operator of precision $N$ (where $\omega=e^{\pi i/N}$) is an operator of the form:
\begin{equation}
XP_N(p|\bold{x}|\bold{z}) := \omega^p \bigotimes_{0\leq i<n}X^{\bold{x}[i]}P^{\bold{z}[i]},
\end{equation}
% \znote{Define $\omega$ and $P$?}
where $p$, $\bold{x}[i]$, and $\bold{z}[i]$ are integers satisfying:
\begin{itemize}
\item $0\leq p<2N$,
\item $0\leq \bold{x}[i]<2$,
\item $0\leq \bold{z}[i]<N$.
\end{itemize}
We call $p$, $\bold{x}$, and $\bold{z}$ the phase component, $X$-component, and $Z$-component of the $XP$-operator, respectively. It is easy to see that two $XP$-operators $XP_N(p|\bold{x}|\bold{z})$ and $XP_N(p'|\bold{x'}|\bold{z'})$ are identical if and only if $p=p'$, $\bold{x}=\bold{x'}$, and $\bold{z}=\bold{z'}$.
As established in \cite[Proposition 4.1]{webster_xp_2022}, $XP$-operators possess useful group properties. Examples include elegant characterizations of the unit operator, multiplication, and inverses of $XP$-operators. Of particular importance is the following result: Let $G = \{g_1,\cdots, g_m\}$ be a set of $XP$-operators. There exists a unique set of diagonal operators $S_Z:=\{ B_j:0 \leq j <s\}$ and non-diagonal operators $ S_X:=\{ A_i:0\leq i<r\}$, called the \emph{canonical generators} of $\langle G\rangle$, such that all group elements $g \in \langle G\rangle$ can be expressed as $g=S_X^{\bold{a}} S_Z^{\bold{b}}$, where $\bold{a} \in \mathbb{Z}_2^{|S_X|}$ and $\bold{b} \in \mathbb{Z}_2^{|S_Z|}$. Furthermore, two sets of $XP$-operators of precision $N$ generate the same group if and only if they have the same canonical generators.
We denote by \(\mathcal{XP}_N\) the group generated by \(XP\)-operators with precision \( N \). When the precision \( N \) is not specified, we use the shorthand \(\mathcal{XP}\).
%\magenta{Introduce the notion of stabilizer and stabilizer group here?}
\subsection{LIMDD for Quantum State Representation}
%Although the circuit model provides an intuitive representation of quantum computation, its
Quantum circuit simulation is typically computationally intensive. LIMDD \cite{vinkhuijzen2023limdd} offers a compact representation of quantum states, enabling efficient simulation. Within this framework, quantum states that differ only by a local invertible transformation on each qubit can share the same memory space.
\begin{defn}[Quantum State Isomorphism \cite{vinkhuijzen2023limdd}]
Two $n$-qubit quantum states $\ket{\Phi}$ and $\ket{\Psi}$ are said to be \emph{isomorphic} if $\ket{\Phi} = O \ket{\Psi}$ for some $O \in \mathcal{O}_n$.
\end{defn}
With the notion of quantum state isomorphism in place, we now introduce Local Invertible Map Decision Diagrams.
\begin{defn}[LIMDD \cite{vinkhuijzen2023limdd}]\label{def:limdd}
Let $\mathcal{G}$ be a subgroup of $\mathcal{O}$. An $n$-qubit $\mathcal{G}$-LIMDD is a rooted, directed acyclic graph (DAG) that represents an $n$-qubit quantum state. Formally, a LIMDD is a 6-tuple $(V, E, \idx, \low, \high, \w)$ where:
\begin{itemize}
\item $V$ is a finite set of nodes, consisting of non-terminal nodes $V_{NT}$ and a terminal node $v_T$ labeled with integer 1;
\item $\idx: V_{NT} \rightarrow [n]$ assigns each node a qubit index in $[n]$;
\item $\low$ and $\high$ are mappings in $V_{NT} \rightarrow V$ that assign each non-terminal node its 0-successor and 1-successor, respectively;
\item $E = \{(v, \low(v)), (v, \high(v)) : v\in V_{NT}\}$ is the set of edges, where $(v, \low(v))$ and $(v, \high(v)) $ are called the low-edge and high-edge of $v$, respectively. For simplicity, we assume the root node $r_\F$ has a unique incoming edge, denoted $e_r$, which has no source node;
\item $\w: E \rightarrow \mathcal{G}$ is a function that labels edges with LIMs.
\end{itemize}
By overloading the Dirac notation, the semantics of the terminal node is defined to be $\ket{v_T} = 1$, the semantics of an edge $e$, directing to a node $v$, is defined to be
$$
\ket{e} = w(e) \cdot \ket{v},
$$
and the semantics of a non-terminal node $v$ is defined to be
$$
\ket{v} = \ket{0}\otimes \ket{(v,\low(v))} + \ket{1}\otimes \ket{(v,\high(v))}.
$$
\end{defn}
In the LIMDD framework \cite{vinkhuijzen2023limdd} and implementation \cite{vinkhuijzen2023efficient}, the group $\mathcal{G}$ is chosen as the Pauli group, i.e., the group of Pauli operators. By leveraging isomorphisms among quantum states, LIMDDs achieve greater compactness than QMDDs and TDDs in representing quantum states.
\begin{exam}\label{exp:LIMDD}
\addt{
Fig.~\ref{fig:run_exp}\,(a) illustrates the LIMDD representation of the 3-qubit quantum state introduced in Example~\ref{exp:qc}.
In the diagram, dotted red lines denote low edges, solid blue lines denote high edges, and the purple edge indicates the incoming edge.
The terminal node represents the constant value~1. For clarity, edge weights equal to $1$ or $I$ are omitted there.
The node labeled $x_1$ corresponds to the state
$\ket{+} = \ket{0} + \ket{1}$.
The left node labeled $x_2$ represents the 2-qubit state
$\ket{0}\ket{+} + \ket{1}(Z\ket{+})
= \ket{0}\ket{+} + \ket{1}\ket{-}$; and the right node labeled $x_2$ represents
$\ket{0}\ket{+} + \ket{1}(iZ\ket{+})
= \ket{0}\ket{+} + i\ket{1}\ket{-}$.
The node labeled $x_3$ represents the 3-qubit state
\begin{align*}
\ket{\Phi}=&\ket{0}(\ket{0}\ket{+}+\ket{1}\ket{-})
+ \ket{1}(iZ\!\otimes\! Z)(\ket{0}\ket{+}+i\ket{1}\ket{-}) \\
= & \ket{0}\ket{0}\ket{+}+\ket{0}\ket{1}\ket{-}
+ i\ket{1}\ket{0}\ket{-}+\ket{1}\ket{1}\ket{+}.
\end{align*}
Finally, the entire LIMDD represents the normalized state
\begin{align*}
& \frac{1}{2\sqrt{2}}\,
(Z\!\otimes\! I\!\otimes\! I) \ket{\Phi} \\
%(\ket{0}\ket{0}\ket{+}
%+\ket{0}\ket{1}\ket{-}
%+i\ket{1}\ket{0}\ket{-}
%+\ket{1}\ket{1}\ket{+}) \\
= & \frac{1}{2\sqrt{2}}
(
\ket{0}\ket{0}\ket{+}
+\ket{0}\ket{1}\ket{-}
-i\ket{1}\ket{0}\ket{-}
-\ket{1}\ket{1}\ket{+}
).
\end{align*}
}
% \addt{Fig.~\ref{fig:run_exp} \addt{(a)} gives the LIMDD representation of the 3-qubit quantum state shown in Example~\ref{exp:qc} \addt{, where dotted red lines represent the low edges, solid blue lines represent the high edges and the incoming edge is represented as purple}. The terminal node represents a constant 1. The node labeled with $x_1$ represents a single qubit quantum state $\ket{+}=\ket{0}+\ket{1}$. The left node labeled with index $x_2$ represents a two-qubit quantum state $\ket{0}\ket{+}+\ket{1}(Z\ket{+})=\ket{0}\ket{+}+\ket{1}\ket{-}$. Similarly, the right node labeled with index $x_2$ represents a two-qubit quantum state $\ket{0}\ket{+}+\ket{1}(iZ\ket{+})=\ket{0}\ket{+}+i\ket{1}\ket{-}$. The node labeled with index $x_3$ represent a 3-qubit state $\ket{0}(\ket{0}\ket{+}+\ket{1}\ket{-})+\ket{1}(iZ\otimes Z(\ket{0}\ket{+}+i\ket{1}\ket{-}))=\ket{0}\ket{0}\ket{+}+\ket{0}\ket{1}\ket{-}+i\ket{1}\ket{0}\ket{-}+\ket{1}\ket{1}\ket{+}))$. Finally, the quantum state represented by the whole LIMDD is $\frac{1}{2\sqrt{2}}Z\otimes I \otimes I (\ket{0}\ket{0}\ket{+}+\ket{0}\ket{1}\ket{-}+i\ket{1}\ket{0}\ket{-}+\ket{1}\ket{1}\ket{+})))=\frac{1}{2\sqrt{2}}(\ket{0}\ket{0}\ket{+}+\ket{0}\ket{1}\ket{-}-i\ket{1}\ket{0}\ket{-}-\ket{1}\ket{1}\ket{+})))$. Note that weights 1 or $I$ are omitted in the figure.}
\end{exam}
\subsection{TDD for Tensor Network Representation}
%介绍TDD
%Although tensor networks offer flexibility to handle various tasks and reduce computational complexity, high-rank tensors still require substantial memory.
%Tensor Decision Diagrams (TDDs)
TDDs provide a data structure to represent and manipulate tensors compactly and efficiently. Their core concept employs a binary branching structure to map data elements to distinct paths, enabling compression by sharing common branches for repeated data. The formal definition is as follows:
\begin{defn}[TDD \cite{hong2020tensor}]
A \emph{Tensor Decision Diagram} (TDD) $\F$ over a set of indices $S$ is a rooted and weighted DAG
$\F = (V, E, \idx, \low, \high, \w)$ such that:
\begin{itemize}
\item $V$, $V_{NT}$, $v_T$, $\low$, $\high$, and $E$ are interpreted analogously to their roles in LIMDDs, representing the nodes, non-terminal nodes, the terminal node, 0- and 1-successor mappings, and edges. In particular, we write $r_\F$ and $e_r$ for the root node and the incoming edge of $\F$.
% \item $V$ is a finite set of nodes which is partitioned into non-terminal nodes $V_{NT}$ and a terminal node $v_T$ labeled with integer 1. Denote by $r_\F$ the unique root node of $\F$;
\item $\idx: V_{NT} \rightarrow S$ assigns each non-terminal node an index in $S$;
% \item $\val: \set{v_T} \rightarrow \mathbb{C}$ assigns the terminal node a complex value;
% \item both $low$ and $high$ are mappings in $V_{NT} \rightarrow V$, which assign each non-terminal node with its 0- and 1-successors, respectively;
% \item $E = \{(v, \low(v)), (v, \high(v)) : v\in V_{NT}\}$ is the set of edges, where $(v, \low(v))$ and $(v, \high(v))$ are called the low- and high-edges of $v$, respectively. For simplicity, we also assume the root node $r_\F$ has a unique incoming edge, denoted $e_r$, which has no source node;
\item $\w: E\rightarrow \mathbb{C}$ assigns each edge a complex weight. In particular, $\w(e_r)$ is called the weight of $\F$, denoted $w_\F$.
\end{itemize}
\end{defn}
Each node $v$ of a TDD $\F$ corresponds to a tensor $\bphi(v)$.
If $v=v_T$ is the terminal node, then $\bphi(v)=1$; %\delete{$\bphi(v):=\val(v)$ is a rank 0 tensor, i.e., a constant};
otherwise,
\begin{equation}\label{eq:tdd-intp}
\bphi(v) := w_0\cdot \overline{x}_v \cdot \bphi(\low(v))
+w_1 \cdot x_v \cdot \bphi(\high(v)),
\end{equation}
where $x_v=\idx(v)$, and $w_0$ and $w_1$ are the weights on the low- and high-edges of $v$, respectively. The tensor represented by $\F$ itself is defined to be
\begin{align}\label{eq:Phi_of_TDD}
\bphi(\F) := w_\F \cdot \bphi(r_\F).
\end{align}
%Recall here that $r_\F$ and $w_\F $ are the root node and the weight of $\F$, respectively.
\begin{figure}
\centering
\subfigure[]{
\resizebox{0.155\textwidth}{!}{
\input{figures/run_exp_limdd1.tex} }
}
\subfigure[]{%
\resizebox{0.15\textwidth}{!}{\input{figures/run_exp_tdd1.tex}}
}
\subfigure[]{
\hspace{2em}
\resizebox{0.091\textwidth}{!}{
\input{figures/run_exp_limtdd1.tex} }
}
\caption{ An intuitive comparison among (a) LIMDD, (b) TDD, and (c) LimTDD. All decision diagrams represent the output state resulted from simulating the circuit shown in Fig.~\ref{fig:qc} with input state $\ket{000}$. The dotted red lines correspond to low edges and the solid blue lines correspond to high edges. The amplitude of the state can be obtained by multiplying the weights along a path. When $Z$ or $P$ is encountered, a phase $-1$ or $e^{2\pi i/8}$ should be added when going along the high edges. Here, $P = P_8$.
% \magenta{Explain the lines, how to get the values, and $P$. Take a specific path, e.g., (blue, blue, red) as an example.}
%A more specific explanation can be seen in the corresponding part in the main text.
}
% The TDD (a), LIMDD (b), and LimTDD (c) representations of the output state \blue{resulted} from simulating the circuit shown in Fig.~\ref{fig:qc} with input state $\ket{000}$.
\label{fig:run_exp}
\end{figure}
\begin{exam}
Fig.~\ref{fig:run_exp}\,(b) shows the TDD of the tensor introduced in Example~\ref{exp:tensor}.
The tensors represented by the nodes can be constructed from bottom to top, following the same procedure as in Example~\ref{exp:LIMDD}.
The value of each tensor element is obtained by multiplying the weights along the corresponding path from the root to the terminal node.
For instance, the leftmost path, denoted by the binary string $000$, yields
\[
\phi_{x_3x_2x_1}(000)
= \tfrac{1}{2\sqrt{2}} \times 1 \times 1 \times 1
= \tfrac{1}{2\sqrt{2}}.
\]
% Fig.~\ref{fig:run_exp} \addt{(b)} gives the TDD representation of the tensor shown in Example~\ref{exp:tensor}. \addt{The tensors represented by the nodes can be constructed from bottom to top, similar to the procedure displayed in Example \ref{exp:LIMDD}.} The values of the tensor can be obtained by multiplying the weights along the paths. For example, the leftmost path gives the value $\phi_{x_3x_2x_1}(000)=\frac{1}{2\sqrt{2}}\times 1 \times 1 \times 1=\frac{1}{2\sqrt{2}}$.
\end{exam}
% \documentclass{article}
% \usepackage[x11names, svgnames, rgb]{xcolor} % Color support
% \usepackage[utf8]{inputenc} % Input encoding
% \usepackage{tikz} % For creating graphics
% \usetikzlibrary{arrows.meta, shapes, positioning} % TikZ libraries for arrows and shapes
% \usepackage{amsmath} % For mathematical formulas
% \begin{document}
% \pagestyle{empty}
\begin{tikzpicture}[
>=Latex, % Arrow tip type
line join=bevel,
node distance=3cm, % Distance between nodes
every node/.style={minimum size=2cm, very thick, font=\Huge}, % Node style
every path/.style={line width=3pt} % Path style
]
% Nodes
\node (x3) [circle, draw, red, text=black] {$x_3$};
\node (x20) [circle, draw, red, below left=of x3, text=black] {$x_2$};
\node (x21) [circle, draw, red, below right=of x3, text=black] {$x_2$};
\node (x1) [circle, draw, red, below right=of x20, text=black] {$x_1$};
\node (t) [circle, draw, red, below=of x1, text=black] {1};
\node (none) [above=of x3] {};
% Edges
\draw[->, red,dotted] (x3) -- (x20) node[midway, left] {};
\draw[->, blue] (x3) -- (x21) node[midway, right,text=black] {$iZ\otimes Z$};
\draw[->, red , dotted] (x20) to[bend right] node[midway, left] {} (x1);
\draw[->, blue] (x20) to[bend left] node[near start, right,text=black] {$Z$} (x1);
\draw[->, red , dotted] (x1) to[bend right] node[midway, left] {} (t);
\draw[->, blue] (x1) to[bend left] node[midway, right] {} (t);
\draw[->, red , dotted] (x21) to[bend left] node[midway, right] {} (x1);
\draw[->, blue] (x21) to[bend right] node[midway, right,text=black] {$iZ$} (x1);
\draw[->, custompurple] (none) -- (x3) node[midway, right,text=black] {$\frac{1}{2\sqrt{2}} Z\otimes I \otimes I$};
\end{tikzpicture}
% \end{document}
% \documentclass{article}
% \usepackage{xcolor}
% \usepackage[utf8]{inputenc}
% \usepackage{tikz}
% \usetikzlibrary{arrows.meta, positioning, shapes}
% \begin{document}
% \pagestyle{empty}
\begin{tikzpicture}[
>=Latex, % Arrow tip type
line join=bevel,
node distance=3cm, % Distance between nodes
every node/.style={minimum size=2cm, very thick, font=\Huge}, % Node style
every path/.style={line width=3pt} % Path style
]
% Nodes
\node (x3) [circle, draw, red, text=black] {$x_3$};
\node (x20) [below left=of x3, circle, draw, red, text=black] {$x_2$};
\node (x21) [below right=of x3, circle, draw, red, text=black] {$x_2$};
\node (x10) [below=of x20, circle, draw, red, text=black] {$x_1$};
\node (x11) [below=of x21, circle, draw, red, text=black] {$x_1$};
\node (t) [below right=of x10,circle, draw, red, text=black] {1};
\node (none) [above=of x3] {};
% Edges
\draw [->,red, dotted] (x3) -- (x20);
\draw [->,blue] (x3) -- (x21) node[midway,right,text=black,text=black] {$-i$};
\draw [->,red, dotted] (x20) to[bend right] (x10);
\draw [->,blue] (x20) -- (x11);
\draw [->,red, dotted] (x10) to[bend right] (t);
\draw [->,blue] (x10) -- (t);
\draw [->,red, dotted] (x11) to[bend left] (t);
\draw [->,blue] (x21) -- (x10) node[pos = 0.9, right, text=black] {$-i$};
\draw [->,red, dotted] (x21) to[bend left] (x11);
\draw[->, custompurple] (none) -- (x3) node[midway,text=black, right,text=black] {$\frac{1}{2\sqrt{2}}$};
\draw [->, blue] (x11) -- (t) node[midway, above, text=black] {$-1$};
\end{tikzpicture}
% \end{document}
% \documentclass{article}
% \usepackage{xcolor}
% \usepackage[utf8]{inputenc}
% \usepackage{tikz}
% \usetikzlibrary{arrows.meta, positioning, shapes.geometric}
% \begin{document}
% \pagestyle{empty}
\begin{tikzpicture}[
>=Latex, % Arrow tip type
line join=bevel,
node distance=3cm, % Distance between nodes. Note the added comma here!
every node/.style={minimum size=2cm, very thick, font=\Huge}, % Node style
every path/.style={line width=3pt} % Path style
]
% Nodes
\node (x3) [circle, draw, red, text=black] {$x_3$};
\node (x2) [below=of x3, circle, draw, red, text=black] {$x_2$};
\node (x1) [below=of x2, circle, draw, red, text=black] {$x_1$};
\node (t) [below=of x1, circle, draw, red, text=black] {1};
\node (none) [above=of x3] {}; % Not sure what this node is for, seems not used in any edges?
% Edges
\draw [->, red, dotted] (x3) to[bend right] (x2);
\draw [->, blue] (x3) to[bend left] node[midway, right, text=black] {$P^{6} \otimes P^{4}$} (x2);
\draw [->, red, dotted] (x2) to[bend right] (x1);
\draw [->, blue] (x2) to[bend left] node[midway, right, text=black] {$P^{4}$} (x1);
\draw [->, red, dotted] (x1) to[bend right] (t);
\draw [->, blue] (x1) to[bend left] node[midway, right] {} (t);
\draw [->, blue, custompurple] (none) -- (x3) node[midway, right, text=black] {$\frac{1}{2\sqrt{2}} P^{6} \otimes I \otimes I$};
\end{tikzpicture}
% \end{document}
% \documentclass{standalone}
% \usepackage{tikz}
\usetikzlibrary{backgrounds,fit,calc}
% \usepackage{xcolor}
\definecolor{tensorBlue}{HTML}{224f70}
% \definecolor{tensorBlueLight}{HTML}{4793ca}
% \definecolor{tensorBlue}{HTML}{6F8DA2}
\definecolor{tensorBlueLight}{HTML}{87B9DD}
\definecolor{nodeGray}{HTML}{888888}
\definecolor{nodeGrayLight}{HTML}{ACACAC}
\definecolor{edgeOrange}{HTML}{f18810}
\definecolor{edgeMagenta}{HTML}{f0116d}
\definecolor{edgeGreen}{HTML}{50c1b8}
% \begin{document}
\begin{tikzpicture}[
% Main node styles
tensor ring/.style args={#1/#2}{
circle, fill=#1, minimum size=4.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=#2, minimum size=4ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
small gray ring/.style={
circle, fill=nodeGray, minimum size=3.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=nodeGrayLight, minimum size=3.2ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
tiny gray ring/.style={
circle, fill=nodeGray, minimum size=2.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=nodeGrayLight, minimum size=2.4ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
% Text styles
tensor text/.style={inner sep=0pt, font=\large},
small text/.style={inner sep=0pt, font=\small},
tiny text/.style={inner sep=0pt, font=\tiny},
]
\begin{scope}
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (phi) at (0,0) {};
\node[small gray ring] (O1) at (-0.4,0.9) {};
\node[] (O1End) at (-0.8,1.6) {};
\node[small gray ring] (O2) at (0.4,0.9) {};
\node[] (O2End) at (0.8,1.6) {};
\node[small gray ring] (O3) at (0,-0.9) {};
\node[] (O3End) at (0,-1.6) {};
% Add text labels with appropriate styles
\node[tensor text] at (phi) {$\phi$};
\node[small text] at (O1) {$O_1$};
\node[small text] at (O2) {$O_2$};
\node[small text] at (O3) {$O_3$};
\draw[very thick,edgeGreen] (phi) -- (O1);
\draw[very thick,edgeMagenta] (phi) -- (O2);
\draw[very thick,edgeOrange] (phi) -- (O3);
\draw[very thick,edgeGreen] (O1End) -- (O1);
\draw[very thick,edgeMagenta] (O2End) -- (O2);
\draw[very thick,edgeOrange] (O3End) -- (O3);
\end{scope}
\begin{scope}[shift={(1.5,0)}]
\node at (0,0) {$+$};
\end{scope}
\begin{scope}[shift={(3,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[small gray ring] (O4) at (-0.4,0.9) {};
\node[] (O4End) at (-0.8,1.6) {};
\node[small gray ring] (O5) at (0.4,0.9) {};
\node[] (O5End) at (0.8,1.6) {};
\node[small gray ring] (O6) at (0,-0.9) {};
\node[] (O6End) at (0,-1.6) {};
% Add text labels with appropriate styles
\node[tensor text] at (psi) {$\psi$};
\node[small text] at (O4) {$O_4$};
\node[small text] at (O5) {$O_5$};
\node[small text] at (O6) {$O_6$};
\draw[very thick,edgeGreen] (psi) -- (O4);
\draw[very thick,edgeMagenta] (psi) -- (O5);
\draw[very thick,edgeOrange] (psi) -- (O6);
\draw[very thick,edgeGreen] (O4End) -- (O4);
\draw[very thick,edgeMagenta] (O5End) -- (O5);
\draw[very thick,edgeOrange] (O6End) -- (O6);
\end{scope}
\begin{scope}[shift={(4.5,0)}]
\node at (0,0) {$\cong $};
\end{scope}
\begin{scope}[shift={(6,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (phi) at (0,0) {};
\node[tiny gray ring] (O1) at (-0.3,0.6) {};
\node[tiny gray ring] (O4dagger) at (-0.6,1) {};
\node[] (O4daggerEnd) at (-0.8,1.6) {};
\node[tiny gray ring] (O2) at (0.3,0.6) {};
\node[tiny gray ring] (O5dagger) at (0.6,1) {};
\node[] (O5daggerEnd) at (0.8,1.6) {};
\node[tiny gray ring] (O3) at (0,-0.65) {};
\node[tiny gray ring] (O6dagger) at (0,-1.2) {};
\node[] (O6daggerEnd) at (0,-1.75) {};
% Add text labels with appropriate styles
\node[tensor text] at (phi) {$\phi$};
\node[tiny text] at (O1) {$O_1$};
\node[tiny text] at (O4dagger) {$O_4^\dagger$};
\node[tiny text] at (O2) {$O_2$};
\node[tiny text] at (O5dagger) {$O_5^\dagger$};
\node[tiny text] at (O3) {$O_3$};
\node[tiny text] at (O6dagger) {$O_6^\dagger$};
\draw[very thick,edgeGreen] (phi) -- (O1);
\draw[very thick,edgeMagenta] (phi) -- (O2);
\draw[very thick,edgeOrange] (phi) -- (O3);
\draw[very thick,edgeGreen] (O1) -- (O4dagger);
\draw[very thick,edgeGreen] (O4daggerEnd) -- (O4dagger);
\draw[very thick,edgeMagenta] (O2) -- (O5dagger);
\draw[very thick,edgeMagenta] (O5daggerEnd) -- (O5dagger);
\draw[very thick,edgeOrange] (O3) -- (O6dagger);
\draw[very thick,edgeOrange] (O6daggerEnd) -- (O6dagger);
\end{scope}
\begin{scope}[shift={(7.5,0)}]
\node at (0,0) {$+ $};
\end{scope}
\begin{scope}[shift={(9,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[] (O4End) at (-0.8,1.6) {};
\node[] (O5End) at (0.8,1.6) {};
\node[] (O6End) at (0,-1.6) {};
% Add text labels with appropriate styles
\node[tensor text] at (psi) {$\psi$};
\draw[very thick,edgeGreen] (psi) -- (O4End);
\draw[very thick,edgeMagenta] (psi) -- (O5End);
\draw[very thick,edgeOrange] (psi) -- (O6End);
\end{scope}
\end{tikzpicture}
% \end{document}
% \documentclass{standalone}
% \usepackage{tikz}
\usetikzlibrary{backgrounds,fit,calc}
% \usepackage{xcolor}
\definecolor{tensorBlue}{HTML}{224f70}
% \definecolor{tensorBlueLight}{HTML}{4793ca}
% \definecolor{tensorBlue}{HTML}{6F8DA2}
\definecolor{tensorBlueLight}{HTML}{87B9DD}
\definecolor{nodeGray}{HTML}{888888}
\definecolor{nodeGrayLight}{HTML}{ACACAC}
\definecolor{edgeOrange}{HTML}{f18810}
\definecolor{edgeMagenta}{HTML}{f0116d}
\definecolor{edgeGreen}{HTML}{50c1b8}
\definecolor{edgeDarkGreen}{HTML}{2c847d}
\definecolor{edgeDarkMagenta}{HTML}{8d315f}
% \begin{document}
\begin{tikzpicture}[
% Main node styles
tensor ring/.style args={#1/#2}{
circle, fill=#1, minimum size=4.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=#2, minimum size=4ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
small gray ring/.style={
circle, fill=nodeGray, minimum size=3.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=nodeGrayLight, minimum size=3.2ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
tiny gray ring/.style={
circle, fill=nodeGray, minimum size=2.8ex, inner sep=0pt,
append after command={
\pgfextra{\let\mytikzlastnode\tikzlastnode}
node[circle, fill=nodeGrayLight, minimum size=2.4ex, inner sep=0pt] at (\mytikzlastnode) {}
}
},
% Text styles
tensor text/.style={inner sep=0pt, font=\large},
small text/.style={inner sep=0pt, font=\small},
tiny text/.style={inner sep=0pt, font=\tiny},
]
\begin{scope}
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (phi) at (0,0) {};
\node[small gray ring] (O1) at (-0.9,0.4) {};
\node[] (O1End) at (-1.6,0.8) {};
\node[small gray ring] (O2) at (-0.9,-0.4) {};
\node[] (O2End) at (-1.6,-0.8) {};
\node[small gray ring] (O3) at (0.9,0) {};
\node[] (O3End) at (1.6,0) {};
% Add text labels with appropriate styles
\node[tensor text] at (phi) {$\phi$};
\node[small text] at (O1) {$O_1$};
\node[small text] at (O2) {$O_2$};
\node[small text] at (O3) {$O_3$};
\draw[very thick,edgeGreen] (phi) -- (O1);
\draw[very thick,edgeMagenta] (phi) -- (O2);
\draw[very thick,edgeOrange] (phi) -- (O3);
\draw[very thick,edgeGreen] (O1End) -- (O1);
\draw[very thick,edgeMagenta] (O2End) -- (O2);
\draw[very thick,edgeOrange] (O3End) -- (O3);
\end{scope}
\begin{scope}[shift={(2,0)}]
\node at (0,0) {$\times$};
\end{scope}
\begin{scope}[shift={(4,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[small gray ring] (O4) at (-0.9,0) {};
\node[] (O4End) at (-1.6,0) {};
\node[small gray ring] (O5) at (0.9,0.4) {};
\node[] (O5End) at (1.6,0.8) {};
\node[small gray ring] (O6) at (0.9,-0.4) {};
\node[] (O6End) at (1.6,-0.8) {};
% Add text labels with appropriate styles
\node[tensor text] at (psi) {$\psi$};
\node[small text] at (O4) {$O_4$};
\node[small text] at (O5) {$O_5$};
\node[small text] at (O6) {$O_6$};
\draw[very thick,edgeOrange] (psi) -- (O4);
\draw[very thick,edgeDarkGreen] (psi) -- (O5);
\draw[very thick,edgeDarkMagenta] (psi) -- (O6);
\draw[very thick,edgeOrange] (O4End) -- (O4);
\draw[very thick,edgeDarkGreen] (O5End) -- (O5);
\draw[very thick,edgeDarkMagenta] (O6End) -- (O6);
\end{scope}
\begin{scope}[shift={(6,0)}]
\node at (0,0) {$\cong $};
\end{scope}
\begin{scope}[shift={(8,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (phi) at (0,0) {};
\node[] (O1End) at (-1.6,0.8) {};
\node[] (O2End) at (-1.6,-0.8) {};
\node[tiny gray ring] (O3) at (0.65,0) {};
\node[tiny gray ring] (O4) at (1.2,0) {};
\node[] (O4End) at (1.75,0) {};
% Add text labels with appropriate styles
\node[tensor text] at (phi) {$\phi$};
\node[tiny text] at (O3) {$O_3$};
\node[tiny text] at (O4) {$O_4$};
\draw[very thick,edgeOrange] (phi) -- (O3);
\draw[very thick,edgeGreen] (O1End) -- (phi);
\draw[very thick,edgeMagenta] (O2End) -- (phi);
\draw[very thick,edgeOrange] (O3) -- (O4);
\draw[very thick,edgeOrange] (O4End) -- (O4);
\end{scope}
\begin{scope}[shift={(10,0)}]
\node at (0,0) {$\times$};
\end{scope}
\begin{scope}[shift={(12,0)}]
% Create ring nodes with different styles
\node[tensor ring=tensorBlue/tensorBlueLight] (psi) at (0,0) {};
\node[] (O4End) at (-1.6,0) {};
\node[] (O5End) at (1.6,0.8) {};
\node[] (O6End) at (1.6,-0.8) {};
% Add text labels with appropriate styles
\node[tensor text] at (psi) {$\psi$};
\draw[very thick,edgeOrange] (psi) -- (O4End);
\draw[very thick,edgeDarkGreen] (psi) -- (O5End);
\draw[very thick,edgeDarkMagenta] (psi) -- (O6End);
\end{scope}
\end{tikzpicture}
% \end{document}
%\section*{Appendix}\label{sec:appendix}
\subsection{Operations of XP-Operators}\label{sec:appxp}
Understanding the group generated by XP-operators requires characterizing the unit operator, multiplication, and inverses of XP-operators. We will need the following notation: for any integer vector $\bold{z} \in \mathbb{Z}^n$, we write $D_N(\bold{z}) := XP_N(\sum_i{\bold{z}[i]}|\bold{0}|{-}\bold{z})$, denoting the antisymmetric operator for $\bold{z}$.
First, we note that
$$I = XP_N(0|\bold{0}|\bold{0}),\quad -I=XP_N(N|\bold{0}|\bold{0}).$$
Second, suppose $\bold{u}_1 = (p_1|\bold{x_1}|\bold{z_1})$ and $\bold{u_2} = (p_2|\bold{x_2}|\bold{z_2})$. The multiplication of two XP-operators $XP_N(\bold{u_1})$ and $XP_N(\bold{u_2})$ can then be computed as:
\begin{equation}\label{eq:XPD1}
XP_N(\bold{u_1})XP_N(\bold{u_2}) = XP_N(\bold{u}_1+\bold{u}_2)D_N(2\bold{x_2}\bold{z_1}),
\end{equation}
where addition and multiplication of vectors are performed component-wise and with proper modulos. That is:
\begin{eqnarray*}
(\bold{a}+\bold{b})[j] &=& \bold{a}[j]+\bold{b}[j] \mod{\bold{m}[j]},\\
(\bold{a}\bold{b})[j] &=& \bold{a}[j]\bold{b}[j] \mod{\bold{m}[j]},
\end{eqnarray*}
where $\bold{m}$ is a $(2n+1)$-dimensional vector such that
\[
\bold{m}[j]=\begin{cases}
2N, & \text{if $j=0$}\\
2, & \text{if $1\leq j\leq n$}\\
N, & \text{if $n+1\leq j\leq 2n$}\\
\end{cases}
\]
Lastly, the inverse of the operator $XP_N(p|\bold{x}|\bold{z})$ is given by:
\begin{equation}
\label{eq:XPD2}
XP_N(-p|\bold{x}|-\bold{z})D_N(-2\bold{x}\bold{z}).
\end{equation}
Furthermore, since the $X$-components of $D_N(2\bold{x_2}\bold{z_1})$ and $D_N(-2\bold{x}\bold{z})$ are $\bold{0}$, right multiplication of these operators in \eqref{eq:XPD1} and \eqref{eq:XPD2} only modifies the phase and $Z$-components of the original XP-operator.
% \section{Calculate $\stab(v)$}\label{sec:stab}
% In this subsection, we show how to calculate a generating set for $\stab(v)$ for an internal node $v$ and calculate $\min_{g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})} {g_0^{\dag} \cdot w \cdot g_1}$.
% Suppose $v$ is a node with two successors $v_0$, $v_1$, and the label on the high edge is $h$; then there are two cases. First, if $v_0 \neq v_1$, then only operators of the form $P^z \otimes g$ can be in $\stab(v)$, and $g$ must stabilize $v_0$ and $\omega^{2z} h^{\dagger} g h$ must stabilize $v_1$ for some $z$. Otherwise, operators of the form $P^z \otimes g$ and $XP^z \otimes g$ all can be in $\stab(v)$. More specifically:
% (1) If $v_0 \neq v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\}$;
% (2) If $v_0 = v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\} \cup \{ XP^z \otimes g | hg\in \stab(v_0)\ and\ \omega^{2z} g h \in \stab(v_1) \}$.
% For the rank 1 cases, the tensor represented by a normalized node $v$ can only be $[1,r]^\intercal$, where $r \in \mathbb{R}$, and $0 \leq r \leq 1$. Then, there are three cases:
% (1) If $r=0$, then $\stab(v) = \langle P \rangle$;
% (2) If $r=1$, then $\stab(v) = \langle X \rangle$;
% (3) $\stab(v) = \emptyset$, otherwise.
% \blue{Suppose} $G_0$ and $G_1$ are the generating sets of $\stab(v_0)$ and $\stab(v_1)$ ignoring their phase components. To calculate a generating set for $\stab(v)$, we only need to calculate $\langle G_0 \rangle \cap \langle h G_1 h^\dag \rangle$ and then add a suitable phase for each item. Note that the intersection of two groups of XP Stabilizers can be calculated by solving a system of linear equations $S_{X0}^{\vec{a_0}} S_{Z0}^{\vec{b_0}} = S_{X1}^{\vec{a_1}} S_{Z1}^{\vec{b_1}}$ in the canonical generating set form, and the time complexity is polynomial with \addt{the length of the operators}. The minimum value of $ {g_0^{\dag} \cdot w \cdot g_1}$ for $g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})$ can also be calculated using their canonical generating set form.
% Note that in lines 9 and 10 of Algorithm~\ref{alg:LimTDD_norm}, we are supposed to find the minimal element of the set \[\{g_0^{\dag} \cdot w_{\F_0}^\dag \cdot w_{\F_1} \cdot g_1 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\}\] and \[\{g_1^{\dag} \cdot w_{\F_1}^\dag \cdot w_{\F_0} \cdot g_0 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\},\] which is useful for making the representation canonical. Although it can be done with a time complexity polynomial with $n$, we still think it is a little complicated. In our implementation, we will not obey this routine, and we will suppose $\stab(r_{\F_0}) = \stab(r_{\F_1}) = \{I\}$, then (iii) in Lemma~\ref{lem:norm} will be changed to \[\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) = \big((X\otimes I) \cdot w, node\big),\] if $w_0^\dag w_1 \neq w_1^\dag w_0$ or $v_0 \neq v_1$; and (iv) needs $w_0\cdot g_0$ and $w_1\cdot g_1$ to remain in the same order as $w_0$ and $w_1$. This damages the property of canonical to some extent, but it is not serious according to our experimental data. In fact, the complete canonicity is not a very necessary matter for simulation or for verification of quantum circuits since two quantum circuits are equivalent if and only if $tr(UV^\dag)=2^n$ and equivalent up to global phase if and only if $|tr(UV^\dag)|=2^n$. Here, $U$ and $V$ represent the matrix representation of the two circuits, $n$ is the number of qubits, $tr(UV^\dag)$ is a scalar and can be calculated by reversing one of the circuits and connecting the input and output of the two circuits.
\subsection{Calculate $\stab(v)$}\label{sec:stab}
We show how to calculate a generating set for $\stab(v)$ for an internal node $v$, and how to calculate $\min_{g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})} {g_0^{\dag} \cdot w \cdot g_1}$.
Suppose $v$ is a node with two successors $v_0$, $v_1$, and the label on the high edge is $h$; then there are two cases. First, if $v_0 \neq v_1$, then only operators of the form $P^z \otimes g$ can be in $\stab(v)$, and $g$ must stabilize $v_0$ and $\omega^{2z} h^{\dagger} g h$ must stabilize $v_1$ for some $z$. Otherwise, operators of the form $P^z \otimes g$ and $XP^z \otimes g$ all can be in $\stab(v)$. More specifically:
(1) If $v_0 \neq v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\}$;
(2) If $v_0 = v_1$, $\stab(v) = \{ P^z \otimes g | g\in \stab(v_0)\ and\ \omega^{2z} h^{\dagger} g h \in \stab(v_1)\} \cup \{ XP^z \otimes g | hg\in \stab(v_0)\ and\ \omega^{2z} g h \in \stab(v_1) \}$.
For the rank 1 cases, the tensor represented by a normalized node $v$ can only be $[1,r]^\intercal$, where $r \in \mathbb{R}$, and $0 \leq r \leq 1$. Then, there are three cases:
(1) If $r=0$, then $\stab(v) = \langle P \rangle$;
(2) If $r=1$, then $\stab(v) = \langle X \rangle$;
(3) $\stab(v) = \emptyset$, otherwise.
Suppose $G_0$ and $G_1$ are the generating sets of $\stab(v_0)$ and $\stab(v_1)$ ignoring their phase components. To calculate a generating set for $\stab(v)$, we only need to calculate $\langle G_0 \rangle \cap \langle h G_1 h^\dag \rangle$ and then add a suitable phase for each item. Note that the intersection of two groups of XP Stabilizers can be calculated by solving a system of linear equations $S_{X0}^{\vec{a_0}} S_{Z0}^{\vec{b_0}} = S_{X1}^{\vec{a_1}} S_{Z1}^{\vec{b_1}}$ in the canonical generating set form, and the time complexity is polynomial with the length of the operators. The minimum value of $ {g_0^{\dag} \cdot w \cdot g_1}$ for $g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})$ can also be calculated using their canonical generating set form.
Note that in lines 9 and 10 of Algorithm~\ref{alg:LimTDD_norm}, we are supposed to find the minimal element in the following sets
\begin{align*}
\{g_0^{\dag} \cdot w_{\F_0}^\dag \cdot w_{\F_1} \cdot g_1 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\} \\
\{g_1^{\dag} \cdot w_{\F_1}^\dag \cdot w_{\F_0} \cdot g_0 \ |\ g_0 \in \stab(r_{\F_0}),g_1 \in \stab(r_{\F_1})\},
\end{align*}
which is useful for making the representation canonical. Although it can be done with a time complexity polynomial with $n$, we still think it is a little complicated. In our implementation, we will not obey this routine, and we will suppose $\stab(r_{\F_0}) = \stab(r_{\F_1}) = \{I\}$, then (iii) in Lemma~\ref{lem:norm} will be changed to \[\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) = \big((X\otimes I) \cdot w, node\big),\] if $w_0^\dag w_1 \neq w_1^\dag w_0$ or $v_0 \neq v_1$; and (iv) needs $w_0\cdot g_0$ and $w_1\cdot g_1$ to remain in the same order as $w_0$ and $w_1$. This damages the property of canonical to some extent, but it is not serious according to our experimental data. In fact, the complete canonicity is not a very necessary matter for simulation or for verification of quantum circuits since two quantum circuits are equivalent if and only if $tr(UV^\dag)=2^n$ and equivalent up to global phase if and only if $|tr(UV^\dag)|=2^n$. Here, $U$ and $V$ represent the matrix representation of the two circuits, $n$ is the number of qubits, $tr(UV^\dag)$ is a scalar and can be calculated by reversing one of the circuits and connecting the input and output of the two circuits.
\subsection{Proof of Some Theorems and Results}\label{sec:proof}
% \begin{lem}
\textbf{Lemma 1.}\textit{
Suppose $v$ is a node in a LimTDD with form
\[\Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}.\]
Let $g_i$ be a $XP$-stabilizer of $\vecc{v_i}$ $(i=0,1)$, and $0\leq k <N$.
Then the following LimTDDs represent the same tensor $\Phi(v)$:
\[\Big( 1 ,\ \Circled{v_{0}}\overset{w_0\cdot g_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1\cdot g_1}\Circled{v_{1}}\Big),\]
\[\Big(P^ k \otimes (w_0\cdot g_0),\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{ \omega^{-2k}\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1}\Circled{v_{1}}\Big),\]
\[\Big( X \otimes (w_1\cdot g_1),\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{g_1^\dagger\cdot w_1^\dagger\cdot w_0\cdot g_0}\Circled{v_{0}}\Big).\]
% \end{lem}
}
\begin{proof}
According to the semantics of LimTDD given in Definition \ref{def:limtdd}, $\Big( 1 ,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$ represents a tensor whose vectorization is $\vecc{v} = [(w_0\vecc{v_0})^\intercal,(w_1\vecc{v_1})^\intercal]^\intercal$.
Similarly, the vectorization of the tensor represented by $\Big( 1 ,\ \Circled{v_{0}}\overset{w_0\cdot g_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1\cdot g_1}\Circled{v_{1}}\Big)$ is $\vecc{v'} = [(w_0g_0\vecc{v_0})^\intercal,(w_1g_1\vecc{v_1})^\intercal]^\intercal$. Since $g_i$ is a stabilizer of $\vecc{v_i}$, that is $g_i\vecc{v_i} = \vecc{v_i}$, $\vecc{v'}=\vecc{v}$.
The vectorization of the tensor represented by $\Big(P^ k \otimes (w_0\cdot g_0),\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{ \omega^{-2k}\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1}\Circled{v_{1}}\Big)$ is $\vecc{v''} = P^ k \otimes (w_0\cdot g_0)\cdot [(\vecc{v_0})^\intercal,(\omega^{-2k}\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1\vecc{v_1})^\intercal]^\intercal = P^ k \otimes I\cdot [(w_0\cdot g_0\vecc{v_0})^\intercal,(\omega^{-2k}\cdot w_1\cdot g_1\vecc{v_1})^\intercal]^\intercal$. Since the function of the $P^k$ gate is to add a phase $\omega^{2k}$ to the second part of the vector, $\vecc{v''}=\vecc{v'}$.
The vectorization of the tensor represented by $\Big( X \otimes (w_1\cdot g_1),\ \Circled{v_{1}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{g_1^\dagger\cdot w_1^\dagger\cdot w_0\cdot g_0}\Circled{v_{0}}\Big)$ is $\vecc{v'''} = X \otimes (w_1\cdot g_1)\cdot [(\vecc{v_1})^\intercal,(g_1^\dagger\cdot w_1^\dagger\cdot w_0\cdot g_0\vecc{v_0})^\intercal]^\intercal = X \otimes I\cdot [(w_1\cdot g_1\vecc{v_1})^\intercal,(w_0\cdot g_0\vecc{v_0})^\intercal]^\intercal$. Since the function of the $X$ gate is to exchange the first and second parts of the vector, $\vecc{v'''}=\vecc{v'}$.
\end{proof}
\textbf{Lemma 2.}\textit{
Suppose $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = \xrightarrow[]{w}\Circled{v}$. Then we have the following new local normalizations:
\begin{itemize}
\item[(i)] $\locnorm\Big(x,\xrightarrow[]{O\cdot w_0}\Circled{v_0},\xrightarrow[]{O\cdot w_1}\Circled{v_1}\Big) = \xrightarrow[]{I\otimes O\cdot w}\Circled{v}$, for any $O \in \mathcal{XP}$;
\item[(ii)] $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{\omega^{2k}w_1}\Circled{v_1}\Big) = \xrightarrow[]{(P^k\otimes I)\cdot w}\Circled{v}$;
\item[(iii)] $\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big) =\xrightarrow[]{(X\otimes I) \cdot w}\Circled{v}$, if $v_0\neq v_1$ or $w_0 \neq w_1$;
\item[(iv)] $\locnorm\Big(x,\xrightarrow[]{w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{w_1 \cdot g_1}\Circled{v_1}\Big) = \xrightarrow[]{w}\Circled{v}$, for any $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$.
\end{itemize}
}
\begin{proof}
For convenience, we only prove the case when $v_0< v_1$ and $w_0\neq w_1$. Other cases can be easily extended.
Let $g_0^\star$ and $g_1^\star$, be two operators in $\stab(v_0)$ and $\stab(v_1)$ that minimize $\{g_0'^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1'\}$ for $g_0'\in \stab(v_0), g_1'\in \stab(v_1)$.
Suppose $w_{temp} =g_0^{\star\dagger}\cdot w_0^\dagger\cdot w_1\cdot g_1^\star =c\cdot e^{i\theta}\omega^{2a} O'$, where $c \in \mathbb{R}$, $0\leq \theta <2\cdot \texttt{ang}(\omega)$, $a\in \mathbb{N}$, and $c(O')=1$. Note that $\texttt{ang}(\omega)$ is the angle of $\omega$, and $c(O')$ is the coefficient of $O'$. Suppose $\omega = e^{2\pi i/N}$, $\texttt{ang}(\omega) = 2\pi/N$. Then $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = \Big( w ,\ \Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{c\cdot e^{i\theta} O'}\Circled{v_{1}}\Big)$, where $\idx(v)=x$, and $w = P^a\otimes (w_0\cdot g_0^\star)$.
Then, for $\locnorm\Big(x,\xrightarrow[]{O\cdot w_0}\Circled{v_0},\xrightarrow[]{O\cdot w_1}\Circled{v_1}\Big)$, according to the procedure of Algorithm \ref{alg:LimTDD_norm}, we have $w_{temp}' = g_0^{\star\dagger}\cdot w_0^\dagger\cdot O^\dagger \cdot O \cdot w_1\cdot g_1 = w_{temp}$. Thus, the high-edge weight will still be $c\cdot e^{i\theta} O'$, but the weight of the incoming edge will becomes $P^k\otimes (O\cdot w_0\cdot g_0^\star) = (I\otimes O)\cdot w$.
For $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{\omega^{2k}w_1}\Circled{v_1}\Big)$, $w_{temp}'' = \omega^{2k}\cdot g_0^{\star\dagger}\cdot w_0^\dagger\cdot O^\dagger \cdot O \cdot w_1\cdot g_1 = \omega^{2k} w_{temp}$. Then $\omega^{2k}$ will become a $P^k\otimes I$ operator applied on the incoming edge. Thus, the result will become $\xrightarrow[]{(P^k\otimes I)\cdot w}\Circled{v}$.
For $\locnorm\Big(x,\xrightarrow[]{w_1}\Circled{v_1},\xrightarrow[]{w_0}\Circled{v_0}\Big)$, since $v_0<v_1$, we will exactly exchange the two sub-LimTDDs and enter Algorithm \ref{alg:LimTDD_norm} with $\locnorm\Big(x,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big)$ and add a $X\otimes I$ operator on the weight of the incoming edge to indicate that the two successors have been changed. Thus, the result will be $\xrightarrow[]{(X\otimes I) \cdot w}\Circled{v}$.
For the last case, $\locnorm\Big(x,\xrightarrow[]{w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{w_1 \cdot g_1}\Circled{v_1}\Big)$, since $g_0 \in \stab(v_0)$ and $g_1\in \stab(v_1)$, then $\min_{g_0'\in \stab(v_0), g_1'\in \stab(v_1)} {g_0'}^\dagger\cdot g_0^\dagger\cdot w_0^\dagger\cdot w_1\cdot g_1\cdot g_1' = w_{temp}$. Thus, the result will still be $\xrightarrow[]{w}\Circled{v}$.
\end{proof}
\textbf{Lemma 3.}\textit{
Let $\F$ be the LimTDD of tensor $\phi_{x_n\cdots x_1}$. For $c\in \{0,1\}$, define $\F_{x_i=c} = \texttt{Slicing}(\F,x_i,c)$. Then, $\F_{x_i=c}$ is the LimTDD representation of the tensor $\phi_{x_i=c}$.
}
\begin{proof}
We only prove the case when $x_i = x_n$, i.e., the top index of the LimTDD, and $c=0$. Suppose $\F$ is in the form $\Big( \lambda\cdot X^{b_n}\cdot P^{z_n} \otimes w,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$, representing the tensor $\phi_{x_n\cdots x_1}$. Here, $b_n \in {0,1}$, $z_n \in \{0,\cdots N-1\}$, where $N$ is the precision of the XP-operators, and $w$ is an operator with $c(w)=1$. Then the vectorization of the tensor $\phi$ is $\vecc{\phi} = \lambda\cdot X^{b_n}\cdot P^{z_n} \otimes w \cdot [(w_0\vecc{v_0})^\intercal, (w_1\vecc{v_1})^\intercal]^\intercal = X^{b_n} \cdot [(\lambda\cdot w\cdot w_0\vecc{v_0})^\intercal, (\lambda \omega^{2z_n}\cdot w\cdot w_1\vecc{v_1})^\intercal]^\intercal$. Thus, if $b_n = 0$, $\vecc{\phi_{x_i=0}} = (\lambda\cdot w\cdot w_0\vecc{v_0})^\intercal$; else $b_n =1$ and $\vecc{\phi_{x_i=0}} = (\lambda\cdot \omega^{2z_n}\cdot w\cdot w_1\vecc{v_1})^\intercal$.
Suppose $b_n = 0$, then the LimTDD obtained through the Slicing procedure given in Algorithm \ref{alg:LimTDD_Slicing} is $\xrightarrow[]{\lambda\cdot w\cdot w_0}\Circled{v_0}$. If $b_n = 1$, then the LimTDD obtained is $\xrightarrow[]{\lambda \cdot \omega^{2z_n}\cdot w\cdot w_1}\Circled{v_1}$. Both cases represent the tensor $\phi_{x_i=0}$.
\end{proof}
\textbf{Lemma 4.}\textit{
Let $\F$ be the LimTDD representation of tensor $\phi_{x_n\cdots x_1}$. Then
$\locnorm(x_n,\F_{x_n=0},\F_{x_n=1}) = \F$.
}
\begin{proof}
Suppose $\F$ is in the form $\Big( \lambda\cdot X^{b_n}\cdot P^{z_n} \otimes w,\ \Circled{v_{0}}\overset{w_0}{\dashleftarrow}\Circled{v}\xrightarrow[]{w_1}\Circled{v_{1}}\Big)$, representing the tensor $\phi_{x_n\cdots x_1}$. Here, we only consider the case that $b_n=0$, $v_0 < v_1$. Then, according to the Slicing procedure given in Algorithm \ref{alg:LimTDD_Slicing}, $\F_{x_n=0} = \xrightarrow[]{\lambda\cdot w\cdot w_0}\Circled{v_0}$ and $\F_{x_n=1} = \xrightarrow[]{\lambda\cdot \omega^{2z_n}\cdot w\cdot w_1}\Circled{v_1}$. Then, according to Lemma \ref{lem:norm} (i) and (ii), $\locnorm(x_n,\F_{x_n=0},\F_{x_n=1}) = \xrightarrow[]{\lambda \cdot P^{z_n}\otimes w}\Circled{v} = \F$.
\end{proof}
\textbf{Theorem 1.}\textit{
Suppose $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ are two tensors with the same index set. Let $\F$ and $\G$ be the $N$-LimTDD representations of $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ with the same index order. The following two conditions are equivalent:
\begin{enumerate}
\item $\phi_{x_n\cdots x_1}$ and $\gamma_{x_n\cdots x_1}$ are $\mathcal{XP_N}$-isomorphic.
\item $r_\F =r_\G$ and $w_\F = O \cdot w_\G \cdot g$ for some $O\in \mathcal{XP_N}$ and some $g \in \stab(r_\G)$.
\end{enumerate}
}
\begin{proof}
$2\Rightarrow 1$ follows directly from the definition. We prove $1\Rightarrow 2$ by induction on the tensor rank.
For rank 0 tensors, the claim holds trivially. Now, assume it holds for rank $n-1$, and consider the rank-$n$ case. Given that $\vecc{\phi_{x_n\cdots x_1}} = O \vecc{\gamma_{x_n\cdots x_1}}$ with $O= X^{b_n}P^{z_n}\otimes w$, we show $r_\F=r_\G$, and $w_\F = O \cdot w_\G \cdot g$ for some $g \in \stab(\Phi(r_\G))$.
Suppose $\G_{x_n=0} = \xrightarrow[]{w_0}\Circled{v_0}$ and $\G_{x_n=1} = \xrightarrow[]{w_1}\Circled{v_1}$. Then $\G = \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = (w_\G, r_\G)$, because $\G$ is normalized.
According to the value of $b_n$, there are two cases to examine:
\textbf{Case 1}: If $b_n = 0$, then the components satisfy:
$$\Phi(\F_{x_n=0}) = \phi_{x_n=0} = w \gamma_{x_n=0} = w \Phi(\G_{x_n=0}),$$
$$\Phi(\F_{x_n=1}) = \phi_{x_n=1} = \omega^{2z_n}\cdot w \gamma_{x_n=1} = \omega^{2z_n}\cdot w \Phi(\G_{x_n=1}).$$
By the induction hypothesis, there exist $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$ such that:
\[\F_{x_n=0} = \xrightarrow[]{w\cdot w_0 \cdot g_0}\Circled{v_0}\ \text{and}\ \F_{x_n=1} = \xrightarrow[]{\omega^{2z_n} w\cdot w_1\cdot g_1}\Circled{v_1}.\]
Because
\[\F = \locnorm\Big(x_n,\xrightarrow[]{w\cdot w_0 \cdot g_0}\Circled{v_0},\xrightarrow[]{ \omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1}\Big),\]
applying Lemma~\ref{lem:norm} in the order (i), (ii), and (iv), we conclude $\F \cong \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big) = \G$.
\textbf{Case 2}: If $b_n=1$, then the components satisfy:
$$\Phi(\F_{x_n=0}) = \phi_{x_n=0} = \omega^{2z_n}\cdot w \gamma_{x_n=1} = \omega^{2z_n}\cdot w \Phi(\G_{x_n=1}),$$
$$\Phi(\F_{x_n=1}) = \phi_{x_n=1} = w \gamma_{x_n=0} = w \Phi(\G_{x_n=0}).$$
There exist $g_0 \in \stab(v_0)$ and $g_1 \in \stab(v_1)$ such that:
\[\F_{x_n=0} = \xrightarrow[]{\omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1}\ \text{and}\ \F_{x_n=1} = \xrightarrow[]{w \cdot w_0\cdot g_0}\Circled{v_0}.\]
Because
\[\F = \locnorm\Big(x_n,\xrightarrow[]{\omega^{2z_n} w\cdot w_1 \cdot g_1}\Circled{v_1},\xrightarrow[]{w \cdot w_0\cdot g_0}\Circled{v_0}\Big),\]
applying Lemma~\ref{lem:norm} in the order (i), (ii), (iii), and (iv), we conclude
\[\F \cong \locnorm\Big(x_n,\xrightarrow[]{w_0}\Circled{v_0},\xrightarrow[]{w_1}\Circled{v_1}\Big)=G.\]
This means that, for both cases, $r_\F =r_\G$. Since $w_\F \cdot \Phi(r_\F) =\vecc{\phi_{x_n\cdots x_1}} = O\cdot \vecc{\gamma_{x_n\cdots x_1}} =O\cdot w_\G \cdot \Phi(r_\G)$, we have $w_\G^\dag \cdot O^\dag \cdot w_\F \cdot \Phi(r_\G) = \Phi(r_\G)$. Let $g=w_\G^\dag \cdot O^\dag \cdot w_\F$. Then $g \in \stab(\Phi(r_\G))$ and $w_\F = O \cdot w_\G \cdot g$.
\end{proof}
\textbf{Corollary 1.}\textit{
For any tensor $\phi$ and any $N\geq 0$, let $\F$ be its $(N+1)$-LimTDD representation and $\G$ its $N$-LimTDD representation. Then, $size(\F)\leq size(\G)$, where $size(\F)$ and $size(\G)$ denote the number of nodes in $\F$ and $\G$, respectively.
}
\begin{proof}
This is a direct corollary of Theorem \ref{thm:iso_nodes}
and the fact that $\mathcal{XP_N} \subset \mathcal{XP}_{\mathcal{N}+1}$.
\end{proof}
\textbf{Theorem 2.}(LimTDD vs. TDD \& LIMDD)\textit{
\begin{enumerate}
\item Let $\F$ be the LimTDD representation and $\G$ the TDD representation of the same tensor $\phi_{x_n\cdots x_1}$ with the same index order. Then, $size(\F)\leq size(\G)$.
\item Let $\F$ be the N-LimTDD ($N\geq 2$) representation and $\G$ the LIMDD representation of the same quantum state $\ket{\phi_{x_n\cdots x_1}}$ with the same index (qubit) order. Then, $size(\F)\leq size(\G)$.
\end{enumerate}
}
\begin{proof}
Denote $\mathcal{XP}_0 = \{c\cdot I, c\in \mathbb{C}\}$. Since TDD only allows two tensors that differ only by a complex number to share the same node. Thus, TDD can be seen as 0-LimTDD, and when representing quantum states LIMDD can be seen as 2-LimTDD. Thus, this is a direct corollary of Corollary \ref{cor:size}.
\end{proof}
% \subsection{Pseudocode of some algorithms}\label{sec:code}
% \begin{algorithm}
% \caption{$\locnorm(x,\F_0,\F_1)$}
% \begin{algorithmic}[1]
% \Require{Two normalized LimTDDs $\xrightarrow[]{w_i}\Circled{v_i}$ ($i=0,1)$; and an index $x$. We require $v_0\leq v_1$}.
% %\Require{Two normalized LimTDDs $\F_i$ with root $v_i$ and weight $w_i$ ($i=0,1)$; and an index $x$. We require $v_0\leq v_1$. }
% \Ensure{A normalized LimTDD $\Big(\tilde{w},\Circled{v_{0}}\overset{I}{\dashleftarrow}\Circled{v}\xrightarrow[]{\hat{w}}\Circled{v_{1}}\Big)$}.
% \State $b_n = 0$\hfill \# \texttt{Are $v_0,v_1$ exchanged?}
% %\If{$v_1 < v_0$ }
% %\State $\F_0, \F_1 = \F_1, \F_0$
% %\State $v_0, v_1 = v_1, v_0$
% % \znote{what is $r_{\F_1}$?}
% %\State $b_n \leftarrow 1$
% %\EndIf
% \If{$w_0=w_1=0$ }
% \State \Return $\texttt{make\_dd}(0,x,0,v_0,0,v_1)$
% %\hfill \# \texttt{a DD representing the 0 tensor}
% \EndIf
% \State $w_\text{temp} \leftarrow \min_{g_0 \in \stab(v_0),g_1 \in \stab(v_1)} {g_0^{\dag} \cdot w_0^\dag \cdot w_1 \cdot g_1}$
% \State $w_\text{temp}' \leftarrow \min_{g_0 \in \stab(v_0),g_1 \in \stab(v_1)} {g_1^{\dag} \cdot w_1^\dag \cdot w_0 \cdot g_0}$
% \If{$v_1 = v_0$ and $w_\text{temp}' < w_\text{temp}$}
% \State $b_n \leftarrow 1$
% \State $g_1^\star \leftarrow \text{the stabilizer of } v_1 \text{ yielding } w_\text{temp}'$
% \State $w \leftarrow w_1\cdot g_1^\star$ \hfill %\# \texttt{$g_1^\star$ is the stabilizer of $v_1$ which yields $w_\text{temp}'$}
% \State $w'_1 \leftarrow w_\text{temp}'$
% \Else
% \State $g_0^\star \leftarrow \text{the stabilizer of } v_0 \text{ yielding } w_\text{temp}$
% \State $w \leftarrow w_0 \cdot g_0^\star$ \hfill
% %\# \texttt{$g_0^\star$ is the stabilizer of $v_0$ which yields $w_\text{temp}$}
% %\State $w'_0 \leftarrow I$
% \State $w'_1 \leftarrow w_\text{temp}$
% % \znote{so $w_0$ is set to $I$ all the time?}
% \EndIf
% %\State $w'_0 \leftarrow I$
% % \State $\theta \leftarrow \texttt{ang}(w_1)$
% \State $k \leftarrow \lfloor \texttt{ang}(w_1\addt{'})/2\omega \rfloor$
% %\State $w'_1 \leftarrow w'_1/\omega^{2k}$
% %\State $w \leftarrow X^{b_n}P^k \otimes w$
% %\State $dd \leftarrow \texttt{make\_dd}(w,x,w'_0,v_0,w'_1,v_1)$
% \State $dd \leftarrow \texttt{make\_dd}(X^{b_n}P^k \otimes w,x,I,v_0,w'_1/\omega^{2k},v_1)$
% \State \Return $\texttt{check\_unique\_table}(dd)$
% \end{algorithmic}
% \label{alg:LimTDD_norm}
% \end{algorithm}
% \begin{algorithm}
% \caption{$\LimTDD\_generate(\phi)$}
% \begin{algorithmic}[1]
% \Require{A tensor $\phi$ over a linearly ordered index set $S$.}
% \Ensure{The normalized LimTDD of $\phi$.}
% %\If{the rank of $\phi$ is 0}
% \If{$\phi \equiv c$ is a constant}
% % \State \Return the trivial LimTDD with weight $c$
% \State \Return $\xrightarrow[]{\ c\ }\Circled{v_T}$
% \EndIf
% \State $x\leftarrow$ the smallest index of $\phi$
% % \State $dd\leftarrow$ an empty LimTDD
% % \State $dd.root \leftarrow$ a new node $v$ with index $x$
% \State $r_0 \leftarrow \LimTDD\_generate(\phi_{x=0})$
% \State $r_1 \leftarrow \LimTDD\_generate(\phi_{x=1})$
% % \If{$l>h$}
% % \State $l, h = h,l$
% % \State $b_n = 1$
% % \EndIf
% % \State $dd = normlise(l,h)$
% % \State $\w(dd) = X^{b_n}\w(dd)$
% % \State \Return $\texttt{check\_unique\_table}(dd)$
% \State \Return $\locnorm(x,r_0,r_1)$
% \end{algorithmic}
% \label{alg:LimTDD_generate}
% \end{algorithm}
% \begin{algorithm}
% \caption{$\texttt{Slicing}(\F,x,c)$}
% \begin{algorithmic}[1]
% \Require{A LimTDD $\F$ with weight $w_\F = \lambda X^{b_n}P^{z_n}\otimes w$ representing the tensor $\phi$; an index $x$ and a value $c\in\set{0,1}$.}
% \Ensure{The sub-LimTDD of $\F$ that represents $\phi_{x=c}$.}
% \If{$\F$ is a trivial LimTDD}
% \State \Return $\F$
% \EndIf
% \State $x^{\prime}\leftarrow \idx(r_{\F})$
% \If{$x < x^{\prime}$}
% \State \Return $\F$
% \EndIf
% \If{$x = x^{\prime}$}
% \State $dd\leftarrow$ an empty LimTDD
% \State $v_0 \leftarrow \low(r_\F)$
% \State $v_1 \leftarrow \high(r_\F)$
% \State $w_0 \leftarrow \w(r_\F,\low(r_\F))$
% \State $w_1 \leftarrow \w(r_\F,\high(r_\F))$
% \If{$c=0$}
% \State $dd.node = v_{b_n}$
% \State $dd.weight = \lambda \cdot \omega^{2z b_n} \cdot w \cdot w_{b_n}$
% \Else
% \State $dd.node = v_{1-b_n}$
% \State $dd.weight = \lambda \cdot \omega^{2z (1-b_n)} \cdot w \cdot w_{1-b_n}$
% \EndIf
% \State \Return $dd$
% \EndIf
% \If{$x > x^{\prime}$}
% \State $r_0 \leftarrow \texttt{Slicing}(\texttt{Slicing}(\F,x^{\prime},0),x,c)$
% \State $r_1 \leftarrow \texttt{Slicing}(\texttt{Slicing}(\F,x^{\prime},1),x,c)$
% \State \Return $\locnorm(x^{\prime},r_0,r_1)$
% \EndIf
% \end{algorithmic}
% \label{alg:LimTDD_Slicing}
% \end{algorithm}
% \begin{algorithm}
% \caption{$\add(\F, \G)$}
% \begin{algorithmic}[1]
% \Require{
% Two LimTDDs $\F$ and $\G$}.
% \Ensure {The LimTDD representation of $\bphi(\F)+\bphi(\G)$}.
% \If{$r_\F= r_\G$ and $\w(\F) = c \cdot \w(\G)$}
% \State $dd\leftarrow \F$
% \State $dd.weight\leftarrow w_\F\ + \ w_\G$
% \State \Return $dd$
% \EndIf
% \State $l \leftarrow$ the smaller one of $\F$ and $\G$
% \State $h \leftarrow$ the bigger one of $\F$ and $\G$
% \State $temp\_w \leftarrow \w(h)$
% \State $\w(l) \leftarrow \w(h)^\dag\cdot \w(l)$
% \State $\w(h) \leftarrow I$
% \State $dd\leftarrow \texttt{find\_computed\_table}(l,h,+)$
% \If {$dd$ is empty}
% \State $x\leftarrow$ the smaller index of $r_\F$ and $r_\G$
% \State $r_0 \leftarrow \add(l_{x=0},h_{x=0})$
% \State $r_1 \leftarrow \add(l_{x=1},h_{x=1})$
% \State $dd\leftarrow \locnorm(x,r_0,r_1)$
% \EndIf
% \State $dd.weight\leftarrow temp\_w \cdot dd.weight$
% \State \Return $dd$
% \end{algorithmic}
% \label{alg:LimTDD_Add}
% \end{algorithm}
% \begin{algorithm}
% \caption{$\cont(\F, \G, var)$}
% \begin{algorithmic}[1]
% \Require {Two LimTDDs $\F$ and $\G$; the set $var$ of indices to be contracted.}
% \Ensure {The contracted LimTDD over $var$.}
% \If{both $\F$ and $\G$ are trivial}
% \State $dd\leftarrow \F$
% \State $dd.weight \leftarrow w_\F \cdot w_\G \cdot 2^{len(var)}$
% \State \Return $dd$
% \EndIf
% \State $temp\_w_0 =$ operators in $\w(\F)$ applied to open indices
% \State $temp\_w_1 =$ operators in $\w(\G)$ applied to open indices
% \State $\w(\G) = w(\G)/temp\_w_1$
% \State $\w(\F) = w(\G)^\dag\cdot \w(\F)/temp\_w_0$
% \State $\w(\G) = I$
% \State $dd\leftarrow \texttt{find\_computed\_table}(\F,\G,\times )$
% \If {$dd$ is empty}
% \If{$\G$ is identity}
% \State $dd\leftarrow \F$
% \Else
% \State $x\leftarrow$ the smaller index of $r_\F$ and $r_\G$
% \If{$x\in var$}
% \State $r_0 \leftarrow \cont(\F_{x=0},\G_{x=0},var\backslash\{x\})$
% \State $r_1 \leftarrow \cont(\F_{x=1},\G_{x=1},var\backslash\{x\})$
% \State $dd\leftarrow \add(r_0,r_1)$
% \Else
% \State $r_0 \leftarrow \cont(\F_{x=0},\G_{x=0},var)$
% \State $r_1 \leftarrow \cont(\F_{x=1},\G_{x=1},var)$
% \State $dd \leftarrow \locnorm(x,r_0,r_1)$
% \EndIf
% \EndIf
% \EndIf
% \State $dd.weight = temp\_w_0\cdot temp\_w_1\cdot dd.weight$
% \State \Return $dd$
% \end{algorithmic}
% \label{alg:LimTDD_Contraction}
% \end{algorithm}
% Generated by IEEEtran.bst, version: 1.14 (2015/08/26)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus \fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl
\bibitem{ran2020quantum}
S.-J. Ran, E.~Tirrito, C.~Peng, X.~Chen, L.~Tagliacozzo, G.~Su, and M.~Lewenstein, ``Quantum entanglement simulation inspired by tensor network,'' in \emph{Tensor Network Contractions: Methods and Applications to Quantum Many-Body Systems}, 2020, pp. 131--146.
\bibitem{haller2023quantum}
L.~Haller, W.-T. Xu, Y.-J. Liu, and F.~Pollmann, ``Quantum phase transition between symmetry enriched topological phases in tensor-network states,'' \emph{Physical Review Research}, vol.~5, no.~4, p. 043078, 2023.
\bibitem{ma2019tensorized}
X.~Ma, P.~Zhang, S.~Zhang, N.~Duan, Y.~Hou, M.~Zhou, and D.~Song, ``A tensorized transformer for language modeling,'' \emph{Advances in Neural Information Processing Systems}, vol.~32, 2019.
\bibitem{fishman2022itensor}
M.~Fishman, S.~White, and E.~M. Stoudenmire, ``The itensor software library for tensor network calculations,'' \emph{SciPost Physics Codebases}, p. 004, 2022.
\bibitem{ganahl2019tensornetwork}
M.~Ganahl, A.~Milsted, S.~Leichenauer, J.~Hidary, and G.~Vidal, ``Tensornetwork on tensorflow: Entanglement renormalization for quantum critical lattice models,'' \emph{arXiv preprint arXiv:1906.12030}, 2019.
\bibitem{hong2021approximate}
X.~Hong, M.~Ying, Y.~Feng, X.~Zhou, and S.~Li, ``Approximate equivalence checking of noisy quantum circuits,'' in \emph{Proceedings of the 58th ACM/IEEE Design Automation Conference}, 2021, pp. 637--642.
\bibitem{10069104}
X.~Hong, Y.~Feng, S.~Li, and M.~Ying, ``Equivalence checking of dynamic quantum circuits,'' in \emph{Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design}, 2022, pp. 1--8.
\bibitem{ftdd10528748}
Q.~Zhang, M.~Saligane, H.-S. Kim, D.~Blaauw, G.~Tzimpragos, and D.~Sylvester, ``Quantum circuit simulation with fast tensor decision diagram,'' in \emph{2024 25th International Symposium on Quality Electronic Design (ISQED)}, 2024, pp. 1--8.
\bibitem{vinkhuijzen2023limdd}
L.~Vinkhuijzen, T.~Coopmans, D.~Elkouss, V.~Dunjko, and A.~Laarman, ``{LIMDD}: A decision diagram for simulation of quantum computing including stabilizer states,'' \emph{Quantum}, vol.~7, p. 1108, 2023.
\bibitem{vinkhuijzen2023efficient}
L.~Vinkhuijzen, T.~Grurl, S.~Hillmich, S.~Brand, R.~Wille, and A.~Laarman, ``Efficient implementation of {LIMDDs} for quantum circuit simulation,'' in \emph{International Symposium on Model Checking Software}, 2023, pp. 3--21.
\bibitem{webster_xp_2022}
M.~A. Webster, B.~J. Brown, and S.~D. Bartlett, ``\BIBforeignlanguage{en}{The {XP} {Stabiliser} {Formalism}: a {Generalisation} of the {Pauli} {Stabiliser} {Formalism} with {Arbitrary} {Phases}},'' \emph{\BIBforeignlanguage{en}{Quantum}}, vol.~6, p. 815, Sep. 2022.
\bibitem{fujita1997multi}
M.~Fujita, P.~C. McGeer, and J.-Y. Yang, ``Multi-terminal binary decision diagrams: An efficient data structure for matrix representation,'' \emph{Formal Methods in System Design}, vol.~10, pp. 149--169, 1997.
\bibitem{vrudhula1996edge}
S.~B. Vrudhula, M.~Pedram, and Y.-T. Lai, ``Edge valued binary decision diagrams,'' \emph{Representations of Discrete Functions}, pp. 109--132, 1996.
\bibitem{tafertshofer1997factored}
P.~Tafertshofer and M.~Pedram, ``Factored edge-valued binary decision diagrams,'' \emph{Formal Methods in System Design}, vol.~10, pp. 243--270, 1997.
\bibitem{kisa2014probabilistic}
D.~Kisa, G.~Van~den Broeck, A.~Choi, and A.~Darwiche, ``Probabilistic sentential decision diagrams.'' in \emph{KR}, 2014.
\bibitem{bahar1997algebric}
R.~I. Bahar, E.~A. Frohm, C.~M. Gaona, G.~D. Hachtel, E.~Macii, A.~Pardo, and F.~Somenzi, ``Algebric decision diagrams and their applications,'' \emph{Formal Methods in System Design}, vol.~10, pp. 171--206, 1997.
\bibitem{sanner2005affine}
S.~Sanner and D.~McAllester, ``Affine algebraic decision diagrams ({AADD}s) and their application to structured probabilistic inference,'' in \emph{Proceedings of the 19th International Joint Conference on Artificial Intelligence}, 2005, p. 1384–1390.
\bibitem{wilson2005decision}
N.~Wilson, ``Decision diagrams for the computation of semiring valuations,'' in \emph{Proceedings of the 19th International Joint Conference on Artificial Intelligence}, 2005, p. 331–336.
\bibitem{viamontes2003improving}
G.~F. Viamontes, I.~L. Markov, and J.~P. Hayes, ``Improving gate-level simulation of quantum circuits,'' \emph{Quantum Information Processing}, vol.~2, no.~5, pp. 347--380, 2003.
\bibitem{niemann2015qmdds}
P.~Niemann, R.~Wille, D.~M. Miller, M.~A. Thornton, and R.~Drechsler, ``{QMDD}s: Efficient quantum function representation and manipulation,'' \emph{IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems}, vol.~35, no.~1, pp. 86--99, 2015.
\bibitem{wang2008xqdd}
S.-A. Wang, C.-Y. Lu, I.-M. Tsai, and S.-Y. Kuo, ``An {XQDD-based} verification method for quantum circuits,'' \emph{IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences}, vol.~91, no.~2, pp. 584--594, 2008.
\bibitem{sistla2024cflobdds}
M.~A. Sistla, S.~Chaudhuri, and T.~Reps, ``{CFLOBDD}s: Context-free-language ordered binary decision diagrams,'' \emph{ACM Transactions on Programming Languages and Systems}, vol.~46, no.~2, pp. 1--82, 2024.
\bibitem{viamontes2004high}
G.~F. Viamontes, I.~L. Markov, and J.~P. Hayes, ``High-performance {QuIDD-based} simulation of quantum circuits,'' in \emph{Proceedings Design, Automation and Test in Europe Conference and Exhibition}, vol.~2.\hskip 1em plus 0.5em minus 0.4em\relax IEEE, 2004, pp. 1354--1355.
\bibitem{burgholzer2020advanced}
L.~Burgholzer and R.~Wille, ``Advanced equivalence checking for quantum circuits,'' \emph{IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems}, vol.~40, no.~9, pp. 1810--1824, 2020.
\bibitem{burgholzer2021qcec}
------, ``{QCEC}: A {JKQ} tool for quantum circuit equivalence checking,'' \emph{Software Impacts}, vol.~7, p. 100051, 2021.
\bibitem{DBLP:conf/dac/MatoHW24}
K.~Mato, S.~Hillmich, and R.~Wille, ``Mixed-dimensional qudit state preparation using edge-weighted decision diagrams,'' in \emph{Proceedings of the 61st {ACM/IEEE} Design Automation Conference}, 2024, pp. 50:1--50:6.
\bibitem{DBLP:conf/qsw/WilleBFKMPQRSSSSB24}
R.~Wille, L.~Berent, T.~Forster, J.~Kunasaikaran, K.~Mato, T.~Peham, N.~Quetschlich, D.~Rovara, A.~Sander, L.~Schmid, D.~Sch{\"{o}}nberger, Y.~Stade, and L.~Burgholzer, ``The {MQT} handbook : {A} summary of design automation tools and software for quantum computing,'' in \emph{{IEEE} International Conference on Quantum Software}, 2024, pp. 1--8.
\bibitem{fargier2014knowledge}
H.~Fargier, P.~Marquis, A.~Niveau, and N.~Schmidt, ``A knowledge compilation map for ordered real-valued decision diagrams,'' in \emph{Proceedings of the AAAI Conference on Artificial Intelligence}, vol.~28, no.~1, 2014.
\bibitem{vinkhuijzen2024knowledge}
L.~Vinkhuijzen, T.~Coopmans, and A.~Laarman, ``A knowledge compilation map for quantum information,'' \emph{arXiv preprint arXiv:2401.01322}, 2024.
\bibitem{nielsen2010quantum}
M.~A. Nielsen and I.~L. Chuang, \emph{Quantum computation and quantum information}.\hskip 1em plus 0.5em minus 0.4em\relax Cambridge University Press, 2010.
\bibitem{markov2008simulating}
I.~L. Markov and Y.~Shi, ``Simulating quantum computation by contracting tensor networks,'' \emph{SIAM Journal on Computing}, vol.~38, no.~3, pp. 963--981, 2008.
\bibitem{biamonte2019lectures}
J.~Biamonte, ``Lectures on quantum tensor networks,'' \emph{arXiv preprint arXiv:1912.10049}, 2019.
\bibitem{hong2020tensor}
X.~Hong, X.~Zhou, S.~Li, Y.~Feng, and M.~Ying, ``A tensor network based decision diagram for representation of quantum circuits,'' \emph{ACM Transactions on Design Automation of Electronic Systems (TODAES)}, vol.~27, no.~6, pp. 1--30, 2022.
\bibitem{quetschlich2023mqtbench}
N.~Quetschlich, L.~Burgholzer, and R.~Wille, ``{{MQT Bench}}: Benchmarking software and design automation tools for quantum computing,'' \emph{{Quantum}}, 2023.
\bibitem{xin2025limtddqsp}
X.~Hong, C.~Li, A.~Dai, S.~Li, S.~Ying, and M.~Ying, ``Quantum state preparation based on {L}im{TDD},'' in \emph{2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD)}, 2025, pp. 1--8.
\end{thebibliography}