Rate-Optimal Streaming Codes Under an Extended Delay Profile for Three-Node Relay Networks With Burst Erasures

Extracted source text · arXiv 2511.06882

\documentclass[journal]{IEEEtran}
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{mathtools}
\usepackage{bm}
\usepackage{adjustbox}
\usepackage[table]{xcolor}
\usepackage{makecell}
\usepackage{booktabs} 
\usepackage{hhline}
\usepackage{diagbox}       
\usepackage{multirow}
\usepackage{array}
\usepackage{float}
\usepackage{lscape}
\usepackage{graphicx}
\usepackage{nicematrix} 
\usepackage{tikz}
\usetikzlibrary{matrix,fit,calc,arrows.meta,bending,positioning,arrows}
\usepackage[ruled,vlined]{algorithm2e}
\usepackage{algpseudocode}
\usepackage[most]{tcolorbox}

\usepackage[caption=false,font=normalsize,labelfont=sf,textfont=sf]{subfig}
\usepackage{textcomp}
\usepackage{url}
\usepackage{verbatim}
\usepackage{cite}
\usepackage{mathrsfs}
\newcommand{\blue}[1]{{\color[rgb]{0.2,0,0.8} #1}}
\newcommand{\red}[1]{{\color[rgb]{0.75,0,0} #1}}
\newtheorem{theorem}{Theorem}         
\newtheorem{lemma}[theorem]{Lemma}             
\newtheorem{definition}{Definition}    
\newtheorem{example}{Example}          
\newtheorem{corollary}[theorem]{Corollary}      
\newtheorem{remark}{Remark}           
\newtheorem{construction}{Construction}
\newtheorem{proposition}[theorem]{Proposition}
\setlength{\parindent}{1em}
\setcounter{MaxMatrixCols}{20}


\hyphenation{op-tical net-works semi-conduc-tor IEEE-Xplore}
\begin{document}
	
	\title{Rate-Optimal Streaming Codes Under an Extended Delay Profile for   Three-Node Relay Networks With Burst Erasures}
	
	\author{Zhipeng Li, Wenjie Ma}
	
	\maketitle
	
	\begin{abstract}
		
		This paper investigates streaming codes for three-node relay networks under burst packet erasures with a delay constraint $T$. In any sliding window of $T+1$ consecutive packets, the source-to-relay and relay-to-destination channels may introduce burst erasures of lengths at most $b_1$ and $b_2$, respectively.     Let $u = \max\{b_1, b_2\}$ and $v = \min\{b_1, b_2\}$.  Singhvi et al. proposed a construction achieving the optimal  rate when $u\mid (T-u-v)$. In this paper, we present an extended delay profile method that attains the optimal rate under a relaxed constraint $\frac{T - u - v}{2u - v} \leq \left\lfloor \frac{T - u - v}{u} \right\rfloor$ and it strictly cover restriction $u\mid (T-u-v)$. %Furthermore,  we demonstrate that the optimal rate for streaming codes is not achievable when $0< T-u-v<v$ under the convolutional code framework.
		
	\end{abstract}
	
	\begin{IEEEkeywords}
		Streaming codes, burst erasures, relay networks, maximum achievable rate, low-latency communication. 
	\end{IEEEkeywords}
	
	\newcommand{\dbox}[1]{
		\begin{tikzpicture}[baseline=(X.base)]
			\node[
			draw=red,
			dashed,
			line width=0.4pt,
			inner sep=2pt,
			rounded corners=2pt,
			text width=0.95\linewidth,
			align=center,
			font=\scriptsize
			] (X) {#1};
		\end{tikzpicture}%
	}
	

	\newenvironment{proof}[1][Proof]{
		\par                            
		\normalfont
		\topsep6pt                       
		\trivlist                        
		\item[\hskip1em\hskip\labelsep\textit{#1:}] 
		\ignorespaces                        
	}{%
		\nobreak\hfill$\square$          
		\endtrivlist                    
	}
	
	
	\section{Introduction}
\IEEEPARstart{T}{he} growing demand for real-time applications, such as interactive multimedia, virtual reality, industrial automation, and autonomous driving, has imposed stringent requirements for low-latency and highly reliable data transmission \cite{7870757}. Among these, Ultra-Reliable Low-Latency Communication  has emerged as a critical enabler in 5G and beyond wireless networks, demanding both ultra-high reliability and minimal end-to-end latency \cite{ChinaUnicomURLLC}.

The concept of point-to-point streaming codes, characterized by parameters $(b, T)$, was first introduced by Martinian and Sundberg \cite{1337126}. They studied the problem of correcting a burst erasure of length at most $b$ under a strict decoding-delay constraint $T$. In their encoding framework, redundancy is strategically incorporated into trailing data to facilitate the recovery of lost packets. While \cite{1337126} primarily considered point-to-point burst erasure channels, subsequent work extended the analysis to channels that experience both burst and arbitrary erasures \cite{7593328}. In this model, every sliding window of length $w$ contains either a single burst erasure of at most $b$ symbols or up to $a$ arbitrary erasures. Under a delay constraint $T = w - 1$, such channels are termed $(a, b, T)$ delay-constrained sliding-window channels \cite{9047134}. This channel model has motivated extensive research, with a central goal being the construction of rate-optimal codes. In particular, \cite{8621051} established the capacity and presented optimal constructions for channels with burst and arbitrary erasures. Subsequent efforts have focused on implementations with practical advantages: \cite{9578987} achieved rate-optimality over a smaller field via an implicit construction, which was later made explicit in \cite{8917664}.

Further studies have explored various other aspects of streaming codes, including theoretical bounds and code constructions \cite{4557355,6567095,6620379,7839973,9611442}. These contributions include delay-optimal code designs \cite{6567095,7839973}, layered constructions for reduced delay \cite{4557355}, and simple yet effective coding schemes \cite{9611442}.

In many practical scenarios, such as multimedia streaming, satellite networks, and vehicular communications, data transmission often involves an intermediate relay node to enhance reliability and reduce latency \cite{6129370}. This motivates the study of streaming codes in three-node relay networks, where a source communicates with a destination via a relay. In such networks, burst erasures may occur independently on both the \emph{source-to-relay} (SR) and \emph{relay-to-destination} (RD) links, significantly complicating the code design and decoding process.

Three-node relay networks over erasure channels were initially studied in \cite{8835153}, which analyzed constructions where the SR and RD links experience $N_1$ and $N_2$ arbitrary erasures, respectively. Subsequently, \cite{10064107} proposed a construction that achieves a higher rate compared to \cite{8835153} through an adaptive relaying strategy. Later, work by Singhvi et al. \cite{9834645} established an optimal non-adaptive rate for networks subject to both burst or arbitrary erasures, in which the SR and RD links experience $b_1$ or $a_1$, and $b_2$ or $a_2$ erasures, respectively, under a delay constraint $T$. This optimal rate is given by
\begin{equation}\label{ob}
	\min \left\{ \dfrac{T - b_1 - a_2 + 1}{T - b_1 + b_2 - a_2 + 1}, \dfrac{T - b_2 - a_1 + 1}{T - b_2 + b_1 - a_1 + 1} \right\}.
\end{equation}
The authors also present a code that achieves the optimal rate under the condition $\max\{b_1, b_2\} \mid (T - b_1 - b_2 - \max\{a_1, a_2\} + 1)$. This bound corresponds to a channel model where links experience only burst erasures, which is captured by setting $a_1 = a_2 = 1$, Under this  model, the general bound in \eqref{ob} simplifies to
\[
\min \left\{ \dfrac{T - b_1}{T - b_1 + b_2}, \dfrac{T - b_2}{T - b_2 + b_1} \right\}.
\]
The construction in this paper is attainable when $\max\{b_1, b_2\} \mid (T - b_1 - b_2)$. For the symmetric case ($b_1 = b_2 = b$), prior work \cite{10764771} demonstrated coding schemes that achieve rates approaching $\frac{T - b}{T}$ asymptotically as the message size increases. In contrast, for the asymmetric case ($b_1 \neq b_2$), we proposed a construction under the constraint $T \geq b_1 + b_2 + \frac{b_1 b_2}{|b_1 - b_2|}$ in \cite{li2025rateoptimalstreamingcodesthreenode}. In these works, the SR link can be viewed as a point-to-point streaming code with parameters $(b_1, T - b_2)$, and the RD link as a point-to-point streaming code with parameters $(b_2, T - b_1)$, however, this interpretation is not entirely accurate (see Lemma~\ref{lemma:relay-decoding}). Streaming codes over multi-hop relay networks were studied in \cite{9814088}, and multi-access relay networks were investigated in \cite{9916294}.
	
		\begin{table*}[!t]
		\centering
		\small
		\caption{Comparison of Burst Erasure Streaming  Code Constructions over $\mathbb{F}_2$}
		\label{tab:comparison}
		\begin{tabular}{|l|c|c|c|c|}
			\hline
			\textbf{Construction} & \textbf{Parameter Constraints} & \textbf{Rate-Optimal?} & \textbf{Method} &\textbf{Delay}\\
			\hline
			Singhvi et al.~\cite{9834645} & $u\mid T-u-v$ & Yes & DE & Delay profile\\
			\hline
			Ramkumar et al.~\cite{10764771} & $b_1 = b_2$ & Asymptotically & DE &Delay profile \\
			\hline
			Zhipeng Li et al. \cite{li2025rateoptimalstreamingcodesthreenode} & $T\geq b_1+b_2+\frac{b_1b_2}{|b_1-b_2|}$  & Yes & Non-DE &Delay profile\\
			\hline
			This paper &$\frac{T-u-v}{2u-v}\leq \lfloor \frac{T-u-v}{u}\rfloor$  & Yes & Non-DE &Extended delay profile(Definition \ref{d2})\\
			\hline
		\end{tabular}
		
		\vspace{0.2cm}
		\begin{minipage}{\textwidth}
			\begin{remark}
				Our Construction in \cite{li2025rateoptimalstreamingcodesthreenode} is originally defined under a broader constraint $\frac{T-u}{v}\geq \lfloor \frac{T-v}{u}\rfloor+1$.
			\end{remark}
		\end{minipage}
	\end{table*}
	\begin{figure*}[htbp]
		\centering
		\includegraphics[width=0.8\textwidth]{parameter_pairs_comparison_T80.pdf}
		\caption{Comparison of valid parameter pairs for different constructions. Here, $b_1, b_2 \in [1, 5]$ and $b_1 + b_2 \leq T$. The total number of pairs for each $T$ is shown as a dashed black line. It can be observed that construction in this paper achieves full coverage in some cases, such as at $T=60$.}
		\label{fig:parameter_comparison}
	\end{figure*}
	
	\subsection{Our Contribution}
 We propose an extended delay profile (Definition \ref{d2}) and derive rate-optimal streaming codes for three-node relay networks with burst erasures under the constraint $\frac{T-u-v}{2u-v} \leq \left\lfloor \frac{T-u-v}{u} \right\rfloor$. This constraint completely covers the restriction $u\mid T-u-v$ in \cite{9834645}, thereby generalizing the construction in \cite{li2025rateoptimalstreamingcodesthreenode}. Moreover, several conclusions from \cite{li2025rateoptimalstreamingcodesthreenode} can be derived as special cases of our framework.

Furthermore, in the non-adaptive case, the optimal rate is given by $\min \{ \frac{T-b_1}{T-b_1+b_2}, \frac{T-b_2}{T-b_2+b_1} \}$. We prove, however, that this optimal rate for streaming codes is unattainable within the convolutional code framework when $0 < T - u - v < v$. This result characterizes the fundamental limitations of code constructions under such delay constraints.
 
 The comprehensive comparison of code parameters is presented in Table~\ref{tab:comparison}. To further illustrate the coverage advantage of our construction, Figure~\ref{fig:parameter_comparison} provides a quantitative comparison of valid parameter pairs. Specifically, the $x$-axis represents the parameter $T$ and the $y$-axis represents the number of valid $(b_1, b_2)$ pairs, where $1\leq b_1,b_2\leq 5$.
 

\subsection{Organization of this paper}
This paper is organized as follows. Section II presents the encoding framework of streaming codes and the optimal rate for three-node burst erasure-correcting streaming codes with parameters $(b_1, b_2, T)$.

Section III introduces our main construction under the constraint
\[\frac{T-u-v}{2u-v} \leq \left\lfloor \frac{T-u-v}{u} \right\rfloor\].

Section IV presents a simplified version of the construction from \cite{li2025rateoptimalstreamingcodesthreenode} under the constraint $T \geq b_1 + b_2 + \frac{b_1 b_2}{|b_1 - b_2|}$. To facilitate a direct comparison with our new proposed construction, we have reformulated the presentation of the key arguments.

Section V presents two illustrative examples for the cases $b_1 < b_2$ and $b_1 > b_2$. Under our proposed constraint, these examples demonstrate specific parameter  where our construction achieves the optimal rate, whereas the existing constructions in \cite{li2025rateoptimalstreamingcodesthreenode} and \cite{9834645} cannot.

Section VI establishes a fundamental limitation: we prove that the optimal rate is unattainable when $0 < T - u - v < v$, characterizing the boundaries of achievable performance under such delay constraints.
	
	\section{Problem Setup}
	\subsection{Notation}
	In this paper, $\mathbb{N}$ denotes the set of non-negative integers, and $\mathbb{N}_+$ denotes the set of positive integers. For $i, j \in \mathbb{N}$, we denote $[i:j] = \{ k \in \mathbb{N} : i \leq k \leq j\}$ and $[i] = \{ k \in \mathbb{N} : 1 \leq k \leq i\}$. 
	$I_n$ denotes the $n \times n$ identity matrix. 
	The $k$-dimensional row vector space over $\mathbb{F}$ is denoted by $\mathbb{F}^k$, and the $k \times n$ matrix space over $\mathbb{F}$ is denoted by $\mathbb{F}^{k \times n}$. For $P \in \mathbb{F}^{k \times n}$, $P(i,j)$ denotes the element at position $(i,j)$ of matrix $P$. For $s, t, a, b \in \mathbb{N}_+$ and $S \subseteq \mathbb{Z}$, let $\{B_i : i \in S\}$ be a set of $s \times t$ matrices and $h$ be a map from $[a] \times [b]$ to $S$. Then the notation $B = (B_{h(i,j)})_{i \in [a],\, j \in [b]}$ denotes that $B$ is an $sa \times tb$ matrix which block at position $(i,j)$ is $B_{h(i,j)}$. For integers $m\in \mathbb{N}$ and $n\in\mathbb{N}_+$, $m\mod n$ denotes the remainder of $m$ modulo $n$, which ranges in $[n]$.
	\subsection{Point-to-Point Streaming Codes Model }
Let $b,T \in \mathbb{N}_+$. A point-to-point \emph{streaming code} (SC) with parameters $(b,T)$ facilitates the transmission of source messages $\{S[t]\}_{t=0}^{\infty}$ over a channel susceptible to burst erasures of maximum length $b$. At each time instant $t$, the encoder generates a transmitted packet $X[t]$ that typically contains the source message $S[t]$ along with parity information derived from previous source messages. The channel introduces erasure bursts such that in any sliding window of $T+1$ consecutive packets, there occurs a burst erasure of length at most $b$. The code ensures that the destination recovers each source message $S[t]$ by its deadline $t+T$.

Within the convolutional coding framework \cite{MaISIT25}, the $(b,T)$ streaming code operates in the following manner. At each time instant $t\in\mathbb{N}$, the encoded packet $X[t]=(S[t],P[t])\in\mathbb{F}^n$ comprises a message packets $S[t] \in \mathbb{F}^k$ and a parity packets $P[t]\in\mathbb{F}^{n-k}$. The encoding process employs matrices $G_i\in\mathbb{F}^{k\times n}$, $i\in \mathbb{Z}$, satisfying
\begin{align}
	&(X[0], X[1], \dots, X[t])\notag\\
	&=(S[0], S[1], \dots, S[t])\begin{pNiceMatrix}
		G_0 & G_1 & \cdots & G_t\\
		& G_0 & \cdots & G_{t-1}\\
		\Block{2-2}{\mathbf{0}}  &  & \ddots & \vdots\\
		& & & G_0
	\end{pNiceMatrix}\label{eq-encoding}\\
	&=(S[0], S[1], \dots, S[t])( G_{j-i} )_{i,j\in [t+1]} \;.\notag
\end{align}

Given the systematic nature of the code, the encoding matrices take the form
\begin{equation*}
	G_i=\begin{cases}
		(I_k\mid P_0), & i=0\\
		(\mathbf{0}_{k\times k}\mid P_i), & \text{otherwise}
	\end{cases}
\end{equation*}
where $I_k$ is the $k\times k$ identity matrix, $P_i\in \mathbb F^{k\times (n-k)}$ for $i\in \mathbb Z$, and $P_i=\mathbf{0}$ when $i\notin [0,T]$. Consequently, the parity components are generated as
\begin{equation}\label{eq-parity}
	P[t]=\sum_{i=0}^T S[t-i]P_i\;.
\end{equation}
The code rate is defined as $\frac{k}{n}$.

During transmission, encoded packets may experience erasures. Even when $X[t]$ is erased, the messages contained in $S[t]$ remains recoverable from subsequently received packets via the relationship in (\ref{eq-parity}). This recovery, however, introduces a decoding delay. To characterize the worst-case delay for $S[t]$, consider the scenario where packets $X[t],\dots,X[t+b-1]$ are erased. The next potential erasure occurs no earlier than time $t+b+T$, ensuring successful reception of packets $X[t+b],\dots,X[t+b+T-1]$. This yields the linear system for recovering the erased messages
\begin{equation}\label{eq-recovery}
	\begin{split}
		(\bar{P}[t+b], \bar{P}[t+b+1], \dots, \bar{P}[t+b+T-1]) \\
		= \left( S[t], S[t+1], \dots, S[t+b-1] \right)\mathbf{P},
	\end{split}
\end{equation}
where $\bar{P}[i]$ represents the modified parity packet after excluding contributions from successfully received messages, and
\begin{equation}\label{eq-matrixP}
	\resizebox{0.9\linewidth}{!}{$
		\mathbf{P} = 
		\begin{pNiceMatrix}
			P_{b} & P_{b+1} & \ldots & P_{T} & 0& \cdots & 0 \\
			P_{b-1} & P_{b} & \cdots & P_{T-1} & P_{T} & \cdots &0    \\
			\vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots  \\
			P_1 & P_2 & \cdots & P_{T-b+1} & P_{T-b+2} & \cdots & P_T
		\end{pNiceMatrix}
		$}
\end{equation}

Notably, different coordinates of the message packets $S[t]=(S_1[t],\dots,S_k[t])\in\mathbb{F}^k$ may experience distinct decoding delays. This motivates the following definition.

\begin{definition}\label{def-delay-profile}
	The {\it delay profile} of a $(b,T)$ streaming code is a sequence $(d_1, \ldots, d_k) \in \mathbb{N}^k$ where for every $t\in\mathbb{N}$ and $i \in [k]$, the symbol $S_i[t]$ is guaranteed to be recovered by time $t + d_i$. It follows that $d_i \leq T$ for all $i\in [k]$.
\end{definition}
	
	
	\subsection{Three-Node Burst Erasures-Correcting Streaming Codes Model}\label{key}
A \emph{three-node burst erasure-correcting streaming code} (TBSC) is characterized by parameters $(b_1,b_2,T)$, where the communication from the source to the destination involves two channels: the source-to-relay (SR) channel and the relay-to-destination (RD) channel. The SR channel is subject to burst erasures of up to $b_1$ consecutive packets within any sliding window of $T+1$ packets, while the RD channel allows burst erasures of at most $b_2$ consecutive packets in any sliding window of $T+1$ packets. The overall transmission from the source to the destination must adhere to a maximum delay of $T$ time slots. It is inherently required that $T \geq b_1 + b_2$.

	Assume that the source generates a message packets $\{S[t]\}_{t=0}^{\infty}$ at time $t$,  where $S[t] \in \mathbb{F}^k$ to the destination.
In papers \cite{10764771}, \cite{li2025rateoptimalstreamingcodesthreenode} and \cite{9814088}, 
it has been suggested that streaming codes over the SR link can be modeled as a point-to-point
SC with parameters $(b_1, T - b_2)$. This interpretation implies that the relay node $r$
must recover the source message $S[t]$ by time $t + T - b_2$. However, this view is not entirely accurate. We therefore introduce the following lemma
\begin{lemma}	\label{lemma:relay-decoding}
In a three-node relay network with parameters $(b_1, b_2, T)$, any streaming code with delay $T$ must satisfy the following necessary condition for each source message $S[t]$: the relay node $r$ must have recovered, by time $t + T - b_2$, either $S[t]$ itself or a linear combination of $S[t]$ and prior messages, i.e., $S[t] + \sum\limits_{i < t}  S[i]$.

In the latter case, the dependency on previous messages implies that $S[t]$ cannot be recovered independently at the relay. Nonetheless, the code must still guarantee that $S[t]$ is ultimately decodable at the destination by its deadline $t+T$.
\end{lemma}
	
	\begin{proof}
		Suppose relay $r$ fails to recover $S[t]$ or such a linear combination by time $t + T - b_2$. Consider an erasure pattern where the RD link experiences a burst erasure of length $b_2$ from time $t + T - b_2 + 1$ to $t + T$. During this period, destination $d$ receives no packets from $r$.
		
		Since $d$ must decode $S[t]$ by time $t + T$, it must rely solely on packets received from $r$ by time $t + T - b_2$. For $d$ to decode $S[t]$ using these packets and its knowledge of prior messages $S[0], \ldots, S[t-1]$, the packets transmitted by $r$ by time $t + T - b_2$ must contain sufficient information about $S[t]$.
		
		This information must be computable by $r$ at time $t + T - b_2$. Since node $r$ could transmit received messages $S[t]$ or $S[t] + \sum\limits_{i < t}S[i]$. 
		
		Therefore, to enable $d$ to decode $S[t]$ with delay $T$, $r$ must have recovered either $S[t]$ or a linear combination $S[t] + \sum\limits_{i < t}S[i]$ by time $t + T - b_2$.
	\end{proof}
	
Hence, the decoding constraint at the relay implies that the SR link can be modeled as a \emph{special point-to-point channel model} with parameters $(b_1, T - b_2)$, where $S[t]$ (or a linear combination containing it) must be recovered by time $t + T - b_2$.  Nevertheless, the point-to-point sc model with parameters $(b_1, T - b_2)$ from the previous section remains applicable. The key refinement lies in the decoding capability of the matrix $P$ in \eqref{eq-matrixP}: it must now be able to decode not necessarily the message $S[t]$ itself, but a linear combination that includes $S[t]$, as stipulated by Lemma \ref{lemma:relay-decoding}. A similar reasoning applies to the RD link  can be modeled as a special point-to-point SC with parameters $(b_2, T-b_1)$.
 All references to point-to-point SCs in this context imply this specialized model that accommodates linear combinations of messages.
 This special channel model will be used to characterize the three key components of the TBSCs

\begin{itemize}
	\item {\it The SR transmission}, realized by a $(b_1,T-b_2)$ point-to-point SC encoding from $\{S[t]\}_{t=0}^{\infty}$ to $\{X[t]\}_{t=0}^{\infty}$. At time $t$, node $S$ applies an encoding function $E^{(s)}$ to generate a coded packet
	\begin{align*}
		X[t] = (S[t], P[t]) = E^{(s)}(S[0],\ldots,S[t]) \in \mathbb{F}^n,
	\end{align*}
	where $P[t] \in \mathbb{F}^{n-k}$ is the parity packet. The SR link is subject to burst erasures of length at most $b_1$.
	
	\item {\it The decoding and relay map}: $D^{(r)}$, which decoding the received packets $Y^{(s)}[0], Y^{(s)}[1], \ldots, Y^{(s)}[t]$ and mapping to a packet $R[t] \in \mathbb{F}^n$, where
	\begin{equation*}
		Y^{(s)}[t] = \begin{cases}
			\perp, & \text{if } X[t] \text{ is erased}; \\
			X[t], & \text{otherwise}.
		\end{cases}
	\end{equation*}
	The relay applies function $D^{(r)}$ to generate $R[t]$ based on the received packets up to time $t$.
	
	\item {\it The RD transmission}, realized by a $(b_2,T-b_1)$ point-to-point SC encoding from $\{R[t]\}_{t=0}^{\infty}$ to $\{Z[t]\}_{t=0}^{\infty}$ using encoding function $E^{(r)}$. Node $r$ transmits $Z[t]$ to node $d$, where the RD link is subject to burst erasures of length at most $b_2$. The destination receives packets
	\begin{equation*}
		Y^{(r)}[t] = \begin{cases}
			\perp, & \text{if } Z[t] \text{ is erased}; \\
			Z[t], & \text{otherwise}.
		\end{cases}
	\end{equation*}
\end{itemize}

The complete TBSC is characterized by the encoding and decoding functions
\begin{align*}
	X[t] &= E^{(s)}(S[0],\ldots,S[t]); \\
	R[t] &= D^{(r)}(Y^{(s)}[0],\ldots,Y^{(s)}[t]); \\
	Z[t] &= E^{(r)}(R[0],\ldots,R[t]); \\
	\tilde{S}[t-T] &= D^{(d)}(Y^{(r)}[0],\ldots,Y^{(r)}[t]).
\end{align*}

A TBSC $\mathcal{C}$ constructed for parameters $(b_1,b_2,T)$ must tolerate burst erasures of length at most $b_1$ in the SR link and $b_2$ in the RD link. We say that a  $(b_1,b_2,T)$ TBSC is a valid TBSC if $\tilde{S}[t] = S[t]$ for all $t \in \mathbb{N}$.

The overall transmission framework is illustrated in Fig. \ref{fig:relay}.
\tikzset{
	node style/.style = {
		draw, circle, minimum size=1.5cm, thick,
		font=\large\bfseries,
	},
	label node/.style = {
		font=\normalsize,
		inner sep=2pt
	}
}
\begin{figure}[htbp]
	\centering
	\resizebox{1\columnwidth}{!}{  
		\begin{tikzpicture}[>=stealth, node distance=2cm] 
			\node[draw, circle, minimum size=1.2cm] (s) {$\mathsf{s}$};
			\node[draw, circle, minimum size=1.2cm, right=of s] (r) {$\mathsf{r}$};
			\node[draw, circle, minimum size=1.2cm, right=of r] (d) {$\mathsf{d}$};
			
			\node[left=0.1cm of s.west, anchor=east, yshift=0.2cm] 
			{$ S[t] \in \mathbb{F}^k$};  
			
			\draw[->] (s) -- 
			node[above, pos=0.3] {\small $ X[t] \in \mathbb{F}^{n}$}  
			node[below, pos=0.5, align=center, text width=2.5cm, yshift=-0.2cm] 
			{\small Burst erasure of length at most $b_1$} (r);
			
			\draw[->] (r) -- 
			node[above, pos=0.3] {\small ${Z[t]} \in \mathbb{F}^{n}$}  
			node[below, pos=0.5, align=center, text width=2.5cm, yshift=-0.2cm] 
			{\small Burst erasure of length at most $b_2$} (d);
			
			\node[right=0.2cm of d.east, anchor=west, yshift=0.2cm] 
			{\small ${ \tilde{S}[t-T] }\in \mathbb{F}^k$};
		\end{tikzpicture}
	}
	\caption{Three-node relay networks with burst erasure $(b_1,b_2,T)$} 
	\label{fig:relay}
\end{figure}

In conventional point-to-point SCs, each message symbol $S[t]$ is recovered individually with a fixed delay. However, in the three-node relay setting, the decoding constraints at the relay node may necessitate a more flexible recovery pattern. Specifically, the relay may recover linear combinations of current and prior messages rather than individual symbols. This observation motivates the following generalization of delay profiles
	\begin{definition}\label{d2}
		Consider the  message packets $(s_1[t], s_2[t], \ldots, s_k[t])$ at time $t$. For each component $s_i[t]$ ($1 \leq i \leq k$), suppose the receiver can only recover a linear combination of $s_i[t]$ with prior message symbols at time $t + t_i$, specifically
		\(
		s_i[t] + m_i(t),
		\)
		where $m_i(t)$ denotes a linear combination of source symbols $\{ s_j[t_j] \mid t_j < t, j \in [k] \}$ (dependent only on information available strictly before time $t$). 
		
		We define the sequence $(t_1, \dots, t_k)$ as an \emph{extended delay profile} with symbols
		\[
		\Big( s_1[t] + m_1[t],\ s_2[t] + m_2[t],\ \dots,\ s_k[t] + m_k[t] \Big).
		\]
	\end{definition}
	

According to the types of delay profiles employed on the SR and RD links, we can categorize TBSCs into three cases. 
The first case uses  delay profiles on both links, the second case uses an extended delay profile on the SR link and a  delay profile on the RD link, and the third case uses a  delay profile on the SR link and an extended delay profile on the RD link. 
We now present these three cases in order.

\noindent \textit{Case 1:}  Delay Profiles on Both Links

A TBSC is \emph{delay-profile-based} when delay profiles are utilized on both the SR and RD links. 
The SR link has delay profile $(t_1, t_2, \ldots, t_k)$, while the RD link has delay profile $(t'_1, t'_2, \ldots, t'_k)$, 
with the delay constraint $T$ satisfying $t_i + t'_i \leq T$ for all $i \in [k]$. We have the following lemma for this case
\begin{lemma} \label{case1}
	Consider a code designed such that the SR link employs a delay profile $(t_1, t_2, \ldots, t_k)$ and the RD link employs a delay profile $(t'_1, t'_2, \ldots, t'_k)$. If $t_i + t'_i \leq T$ holds for all $i \in [k]$, then the destination can recover each source symbol within delay $T$.
\end{lemma}

\begin{proof}
	At source node $S[t] = (s_1[t], s_2[t], \ldots, s_k[t])$ by time $t$. Each symbol $s_i[t]$ is received at the relay by time $t + t_i$. 
	Thus by time $t$, the relay possesses $s_i[t - t_i]$ for all $i \in [k]$. 
	The relay constructs $R[t] = (r_1[t], r_2[t], \ldots, r_k[t])$ with $r_i[t] = s_i[t - t_i]$. 
	Each $r_i[t]$ arrives at the destination by time $t + t'_i$. 
	Therefore, the source symbol $s_i[\tau]$ generated at time $\tau$ is recovered by $\tau + t_i + t'_i \leq \tau + T$.
\end{proof}

\noindent \textit{Case 2:} Extended Delay Profile on SR Link and  Delay Profile on RD Link

A TBSC is \emph{extended-delay-profile-based} when extended delay profile are utilized on  the SR or RD links. 
We now consider a more complex scenario where the SR link employs an extended delay profile while the RD link maintains a  delay profile.

\begin{lemma}\label{case2}
Suppose a code is implemented where the SR link uses an extended delay profile $(t_1,t_2,\ldots,t_k)$ for symbols $\big( s_1[t] + m_1(t), \dots, s_k[t] + m_k(t) \big)$, with each $m_i(t)$ being a linear combination of previously available source symbols. Further, suppose the RD link uses a delay profile $(t'_1,t'_2,\ldots,t'_k)$ for symbols $(r_1[t], \ldots, r_k[t])$. Then, provided that $t_i + t'_i \leq T$ holds for all $i \in [k]$, it follows that all source symbols can be recovered at the destination within the maximum delay $T$.
\end{lemma}

\begin{proof}
The source node encodes message $S[t] = (s_1[t], \ldots, s_k[t])$, available at time $t$, into a packet $X[t] = (S[t], P[t]) \in \mathbb{F}^n$. Here, $P[t]$ is produced by \eqref{eq-recovery} to provide the relay with the extended symbols $s_i[t] + m_i(t)$ (where $m_i(t)$ is a linear combination of earlier symbols) by time $t + t_i$ for each $i$. Upon recovery of $s_i[t - t_i] + m_i(t - t_i)$ by time $t$, the relay constructs $r_i[t] = s_i[t - t_i] + m_i(t - t_i)$ and assembles the transmission packet $R[t] = (r_1[t], \ldots, r_k[t])$.
	
  Relay encoder $r_i[t]$ and transmit over the RD link and arrives at the destination by time $t + t'_i$, delivering
	\[
	s_i[t - t_i] + m_i(t - t_i).
	\]
	Given $t_i + t'_i \leq T$, this arrival occurs by $(t - t_i) + (t_i + t'_i) \leq (t - t_i) + T$.
	
	Now, consider the recovery of $s_i[t - t_i]$ at the destination
	\begin{itemize}
		\item When $t - t_i = 0$, we have $m_i(0) = 0$ (no prior symbols exist), so $s_i[0]$ is directly obtained.
		\item When $t - t_i > 0$, $m_i(t - t_i)$ is a linear combination of source symbols from times before $t - t_i$. 
		Due to the delay constraint $T$, all these symbols have already been recovered at the destination 
		through prior transmissions. Therefore, $s_i[t - t_i]$ can be recovered by subtracting 
		the known linear combination $m_i(t - t_i)$.
	\end{itemize}
	
	In this cases, $s_i[t - t_i]$ is recovered by time $(t - t_i) + T$, satisfying the delay constraint.
\end{proof}

\noindent \textit{Case 3:}  Delay Profile on SR Link and Extended Delay Profile on RD Link

We now consider the third configuration where the SR link employs a  delay profile while the RD link utilizes an extended delay profile. To properly define this case, we first introduce the following concept
\begin{definition}\label{def:eds}
	An extended delay profile with symbols at the relay node satisfies the \emph{extended delay separable} property if the symbols can be partitioned into two distinct types:
	

	
	Type 1 (Independent): Symbols of the form $r_j[t]$ for $j \in A_I$, which depend solely on the current symbol without any linear combination with past symbols.
	
		Type 2 (Dependent): Symbols of the form $r_i[t] + m_i(t)$ for $i \in A_D$, where $m_i(t)$ is a linear combination of symbols ${r_j[t-d] \mid d > 0, j \in A_I}$ from strictly previous time instants.
		
	Here, $A_D$ and $A_I$ form a partition of the index set $[k]$, i.e., $A_D \cup A_I = [k]$ and $A_D \cap A_I = \emptyset$.
	
	For example, the extended delay profile with symbols $(r_1[t]+r_3[t-1]+r_4[t-1], r_2[t]+r_3[t-2], r_3[t], r_4[t])$ satisfies the extended delay separable property with $A_D = \{1,2\}$ and $A_I = \{3,4\}$.
\end{definition}
\begin{lemma} \label{case3}
Given a code where the SR link employs a delay profile $(t_1, \ldots, t_k)$ for the symbols $(s_1[t], \ldots, s_k[t])$, and the RD link employs an extended delay profile for symbols $\big( r_1[t] + m_1(t), \dots, r_{k_1}[t] + m_{k_1}(t), r_{k_1+1}[t], \ldots, r_k[t] \big)$ satisfying the extended delay separable property with $A_D = {1, \ldots, k_1}$ and $A_I = {k_1+1, \ldots, k}$, then, under the condition $t_i + t'_i \leq T$ for all $i \in [k]$, all source symbols can be recovered at the destination within the maximum delay $T$.
\end{lemma}
\begin{proof}
	Without loss of generality, assume \( k_1 = 1 \) and the RD extended delay profile has symbols 
	$(r_1[t] + m_1(t), r_2[t], \ldots, r_k[t])$, 
	where $m_1(t)$ is a linear combination of $\{r_j[t-d] \mid d > 0, j \in \{2,\ldots,k\}\}$.
	
At the source node, $S[t] = (s_1[t], s_2[t], \ldots, s_k[t])$ is available by time $t$. 
Each symbol $s_i[t]$ arrives at the relay by time $t + t_i$. 
By time $t$, the relay has received $s_i[t - t_i]$ for all $i \in [k]$.

The relay constructs the symbols for the RD link in a specific order that relies on the extended delay separable property. First, for $i = 2, \ldots, k$, the relay sets $r_i[t] = s_i[t - t_i]$. These symbols are constructed directly from the received source symbols and belong to set $A_I$, meaning they are pure symbols without any linear combinations. Then, the relay constructs $m_1(t)$ as a linear combination of the available symbols $\{r_j[t-d] \mid d > 0, j \in \{2,\ldots,k\}\}$. Since these $r_j[t-d]$ were received at earlier time instants, they are already available at time $t-d$. Finally, the relay constructs $r_1[t] = s_1[t - t_1] + m_1(t)$, which belongs to set $A_D$.


At the destination, by time $t + t'_1$, the symbol $r_1[t] + m_1(t)$ is received. Substituting the expression for $r_1[t]$, we have
\[
r_1[t] + m_1(t) = (s_1[t - t_1] + m_1(t)) + m_1(t) = s_1[t - t_1]
\]
Thus, $s_1[t - t_1]$ is recovered at time $t + t'_1$. Since $t_1 + t'_1 \leq T$, this recovery occurs by time $(t - t_1) + T$.

For $i = 2, \ldots, k$, the symbols $s_i[t - t_i]$ are directly recovered from $r_i[t]$ at time $t + t'_i \leq (t - t_i) + T$.

Therefore, all source symbols are recovered within the maximum delay $T$.
\end{proof}
	
	\subsection{Optimal Rate }
The following theorem pertains to the non-adaptive case and presents a partial result of Theorem 1 in \cite{9834645}.
	\begin{theorem}
	The rate $R$ of any $(b_1,b_2,T)$ TBSC is bounded by
		\begin{align}  
			R &\leq  R(b_1,b_2,T) \notag \\ 
			&= \begin{cases}
				\min \biggl\{ \dfrac{T-b_1}{T-b_1+b_2}, \dfrac{T-b_2}{T-b_2+b_1} \biggr\} 
				& \text{if } T \geq b_1+b_2, \\
				0 & \text{otherwise}.
			\end{cases}
		\end{align}
	\end{theorem}
	
	A $(b_1,b_2,T)$ TBSC  is rate-optimal, if it is of rate $R(b_1,b_2,T)$. 
	
	
	\section{ Constructions of Binary Rate-Optimal Streaming Codes with constraint $\frac{T-u-v}{2u-v}\leq \lfloor \frac{T-u-v}{u}\rfloor$} \label{KEY CONSTRUCTION}


In this section, we construct binary rate-optimal $(b_1,b_2,T)$ TBSCs for a wide parameter region. 
Firstly, we assume that $b_1 < b_2$ and let $T - b_1 = pb_2 + q$ with $0 \leq q < b_2$. The optimal rate is given by
\[
\frac{T - b_1}{T - b_1 + b_2} = \frac{p b_2 + q}{(p + 1) b_2 + q}.
\]

We construct the SR code and RD code separately under the constraint $\frac{T - b_2 - b_1}{2b_2 - b_1} \leq \lfloor \frac{T - b_2 - b_1}{b_1} \rfloor$, using an extended-delay-profile-based construction.

\subsection{Construction for SR Code}

According to \ref{key}, an SR link is modeled as a $(b_1, T - b_2)$ point-to-point SC that recovers $S[t]$ or $S[t] + \sum\limits_{i < t} S[i]$ by time $t + T - b_2$.

The encoding matrix $\mathbf{P}_{\mbox{\tiny SR}}$ is obtained by substituting parameters $b_1$ and $T - b_2$ into equation (\ref{eq-matrixP}), with its detailed structure visualized in Fig.~\ref{figure PSR}. We now define the constituent $(T - b_1) \times b_2$ submatrices $P_i$ for $i \in [0, T - b_1]$

\begin{itemize}
	\item[(i)] For $i = jb_1$ where $j \in [p - 1]$
	\[
	P_{jb_1} = 
	\begin{pmatrix}
		\mathbf{0}_{(p - j)b_2 + q} \\
		I_{b_2} \\                     
		\mathbf{0}_{(j - 1)b_2}
	\end{pmatrix}\footnote{Hereafter, the subscript of the zero matrix $\mathbf{0}$ denotes its row dimension, while the column dimension is omitted as it is clear from context. In some cases, the subscript is omitted entirely when the matrix dimensions are unambiguous.}
	\]
	
	\item[(ii)] For $i = (p - 1)b_1 + j$ where $j \in [b_1 - 1]$
	
	If $q \geq b_1$, then $P_i = 0$;
	
	If $q < b_1$, define $d_i = b_2 + (j \bmod q)$ and $e_i = q + (j \bmod (b_1 - q))$, then set $P_i(d_i, e_i) = 1$ with all other entries zero.
	
	\item[(iii)] For $i = pb_1$
	\[
	P_{pb_1} = 
	\begin{pmatrix}
		\mathbf{0}_{b_2} \\
		I_q & \mathbf{0}_q \\
		\mathbf{0}_{(p - 1)b_2}
	\end{pmatrix}
	\]
	
	\item[(iv)] For $i = pb_1 + q \leq T - b_2=(p-1)b_2+b_1+q$
	\[
	P_{pb_1 + q} = 
	\begin{pmatrix}
		I_{b_2} \\
		\mathbf{0}_{pb_2 + q}
	\end{pmatrix}
	\]
\end{itemize}

To clarify the global organization, we partition $\mathbf{P}_{\mbox{\tiny SR}}$ into $p + 1$ block matrices $\mathbf{P}_0, \mathbf{P}_1, \ldots, \mathbf{P}_{p - 1}, \mathbf{P}_p$. For $0 \leq i \leq p - 2$, the blocks are

\begin{equation}
	\mathbf{P}_i = 
	\scalebox{0.85}{$\displaystyle
		\begin{pNiceMatrix}
			P_{(i + 1)b_1} & P_{(i + 1)b_1 + 1} & \ldots & P_{(i + 2)b_1 - 1} \\
			P_{(i + 1)b_1 - 1} & P_{(i + 1)b_1} & \ldots & P_{(i + 2)b_1 - 2} \\
			\vdots & \vdots & \ddots & \vdots \\
			P_{ib_1 + 1} & \ldots & \ldots & P_{(i + 1)b_1} 
		\end{pNiceMatrix}
		$}.
\end{equation}

Each $\mathbf{P}_i$ forms a $b_1 \times b_1$ block matrix.

The next block is defined as
\begin{equation}\label{eq9}
	\mathbf{P}_{p - 1} = 
	\begin{pmatrix}
		P_{pb_1} & \cdots & P_{pb_1 + q - 1} \\
		\vdots & \vdots & \vdots \\
		P_{(p - 1)b_1 + 1} & \cdots & P_{(p - 1)b_1 + q}
	\end{pmatrix},
\end{equation}
forming a $b_1 \times q$ block matrix.

The final block captures the remaining components
\begin{equation}
	\mathbf{P}_p = 
	\begin{pmatrix}
		P_{pb_1 + q} & 0 & \cdots & 0 \\
		P_{pb_1 + q - 1} & P_{pb_1 + q} & \cdots & 0 \\
		\vdots & \vdots & \ddots & \vdots \\
		P_{(p - 1)b_1 + q + 1} & \cdots & \cdots & P_{pb_1 + q}
	\end{pmatrix},
\end{equation}.

Having established the complete structure of the encoding matrix $\mathbf{P}{\mbox{\tiny SR}}$, we now proceed to analyze its key algebraic properties. The invertibility of certain submatrices plays a crucial role in ensuring the decoding capability of SR codes. In particular, we focus on the block $\mathbf{P}_{p-1}$ and prove the following fundamental result.

	\begin{lemma}\label{lemma:P_invertible}
		After deleting the first $b_2$ rows, the last $(p-1)b_2$ rows, 
		and the last $b_2 - b_1$ columns from $\mathbf{P}_{p-1}$, 
		the resulting $b_1 q \times b_1 q$ binary matrix is invertible.
	\end{lemma}
	\begin{proof}
		For simplicity, we still use the notation \( P_i \) to denote the reduced matrix.  
		From (ii) and (iii),  
		\[
		P_{pb_1 + q} = \begin{pmatrix} I_q & \mathbf{0} \end{pmatrix}, \quad 
		P_i = \begin{pmatrix} \mathbf{0} & \tilde{\mathbf{P}}_i \end{pmatrix}, (p-1)b_1 + 1 \leq i < pb_1.
		\]  
		, where $\tilde{P}_i$ denotes the the right $b_1-q$ columns. Put the first $q$ columns of all blocks together and perform invertible column transformations. We obtain the \( b_1q \times b_1q \) matrix  
		\[
		\begin{pmatrix}
			I_{q^2} & \mathbf{0} \\
			\mathbf{0} & \tilde{\mathbf{P}}
		\end{pmatrix},
		\quad \text{where} \quad
		\tilde{\mathbf{P}} = \begin{pmatrix}
			\tilde{P}_{pb_1 - q} & \cdots & \tilde{P}_{pb_1 - 1} \\
			\vdots & \ddots & \vdots \\
			\tilde{P}_{(p-1)b_1 + 1} & \cdots & \tilde{P}_{(p-1)b_1 + q}
		\end{pmatrix}.
		\]  
		Note that $\tilde{P}_i$, $1\leq i<b_2$, has only one entry being $1$, and zeros elsewhere. Moreover, from (ii) one can see that $\tilde{\mathbf{P}}$ is actually a permutation matrix (i.e., each row and each column has exactly one $1$) and therefore is invertible. The lemma follows immediately.
	\end{proof}
	\begin{figure*}
		$ \mathbf{P}_{\mbox{\tiny SR}}$ = 
		\scalebox{0.85}{$\displaystyle 
			\begin{pNiceArray}{c|cc|c|c|c}
				\mathbf{P}_0 & \ldots && \mathbf{P}_{p-2} & \mathbf{P}_{p-1} & \mathbf{P}_p
			\end{pNiceArray}
			$}
		\vspace{2ex}
		
		\scalebox{0.85}{$\displaystyle 
		\begin{pNiceArray}{c|cc|c}
			\mathbf{P}_{0} &\ldots && \mathbf{P_{p-3}}
		\end{pNiceArray}
		$} =\\
		
	\scalebox{0.85}{$\displaystyle 
		\begin{pNiceMatrix}
			\Block[fill=red!15,rounded-corners]{}{P_{b_1}} & P_{b_1+1} & \ldots & P_{2b_1-1} & \Block[fill=red!15,rounded-corners]{}{P_{2b_1}} & P_{2b_1+1} & \ldots & P_{3b_1-1} & ~\cdots~ & \Block[fill=red!15,rounded-corners]{}{P_{(p-2) b_1}} & P_{(p-2) b_1+1} & \cdots & P_{(p-1) b_1-1}\\
			P_{b_1-1} & \Block[fill=red!15,rounded-corners]{}{P_{b_1}} & \ldots & P_{2b_1-2} & P_{2b_1-1} & \Block[fill=red!15,rounded-corners]{}{P_{2b_1}} & \ldots & P_{3b_1-2} & \cdots & P_{(p-2) b_1-1} & \Block[fill=red!15,rounded-corners]{}{P_{(p-2) b_1}} & \cdots & P_{(p-1) b_1-2} \\
			\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
			P_1 & \ldots & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{b_1}} & P_{b_1+1} & \ldots & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{2b_1}} & \cdots & P_{(p-3) b_1+1} & P_{(p-3) b_1+2} & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{(p-2) b_1}}
		\end{pNiceMatrix}
		$}
		\vspace{2ex}
		
		\text{when $q<b_1$,} 
		\scalebox{0.85}{$\displaystyle 
			\begin{pNiceArray}{c|c|c}
				\mathbf{P}_{p-2} & \mathbf{P}_{p-1} & \mathbf{P_{p}}
			\end{pNiceArray}
			$} 
		= \\
		
		
		\scalebox{0.85}{$\displaystyle 
			\begin{pNiceArray}{ccccc|cccc|ccccc}
				\Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1}} & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1+1}} & \ldots & \ldots& \Block[fill=red!15,rounded-corners]{}{P_{pb_1-1}} 
				& \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} & P_{pb_1+1} & \ldots & P_{pb_1+q-1} 
				& \Block[fill=red!15,rounded-corners]{}{P_{pb_1+q}} & \ldots & P_{(p+1)b_1-1} & \ldots & 0 \\
				
				& \Block{2-1}{\ddots} 
				&\Block{2-1}{\ddots} & &\Block{2-1}{\vdots} & \Block[fill=red!15,rounded-corners]{}{P_{pb_1-1}} & \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} 
				& & \Block{2-1}{\vdots} & \Block{2-1}{\vdots} & \Block{2-1}{\ddots} & & & \\
				
				& & & & & \Block{2-1}{\vdots} & \Block{2-1}{\vdots} & \ddots 
				& & & & & & \\
				
				& & & & & & & & \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} & P_{pb_1+1} & & & & \\
				
				& & & \Block{2-1}{\ddots}&\Block{2-1}{\ddots} & \Block{2-1}{\vdots} & \Block{2-1}{\vdots} 
				& & \Block{2-1}{\vdots} & \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} 
				& & & \Block{2-1}{\ddots} & \\
				
				& & & & & & & & & & \ddots & & & \\
				
				& & & & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1}} & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1+1}} & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1+2}} 
				& \ldots & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1+q+1}} & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{pb_1+q}}
			\end{pNiceArray}
			$}
		
		\vspace{2ex}
		
		\text{when $b_1\leq q<b_2$,} 
		\scalebox{0.85}{$\displaystyle 
			\begin{pNiceArray}{c|c|c}
				\mathbf{P}_{p-2} & \mathbf{P}_{p-1} & \mathbf{P_{p}}
			\end{pNiceArray}
			$} 
		= \\
		
		\scalebox{0.85}{$\displaystyle 
			\begin{pNiceArray}{ccccc|ccccccc|ccccc}
				\Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1}} & P_{(p-1)b_1+1} & \ldots & \ldots& P_{pb_1-1} 
				& \Block[fill=red!15,rounded-corners]{}{P_{pb_1}} & \ldots & \ldots& \ldots&P_{(p+1)b_1-1} & \ldots & P_{pb_1+q-1} 
				& \Block[fill=red!15,rounded-corners]{}{P_{pb_1+q}} &  &  &  & \\
				
				& \Block{2-1}{\ddots} 
				&\Block{2-1}{\ddots} & &\Block{2-1}{\vdots} &  & \Block{2-1}{\ddots}
				&&&& & \Block{2-1}{\vdots} & \Block{2-1}{\vdots} & \Block{2-1}{\ddots} & & & \\
				
				&&& & & & & &  &  &
				& & & & & & \\
				
				&&&& & & & & & & &  &  & & & & \\
				
				& & & \Block{2-1}{\ddots}&\Block{2-1}{\ddots} &  & 
				& &\Block{2-1}{\ddots}&\Block{2-1}{\vdots}& & \Block{2-1}{\vdots} & \Block{2-1}{\vdots}
				&\Block{2-1}{\ddots} & & \Block{2-1}{\ddots} & \\
				
				&&&& & & & & & & & & &  & & & \\
				
				& & & & \Block[fill=red!15,rounded-corners]{}{P_{(p-1)b_1}} & P_{(p-1)b_1+1} & \ldots
				&\ldots&& \Block[fill=red!15,rounded-corners]{}{P_{pb_1}}& \ldots & P_{(p-1)b_1+q} & P_{(p-1)b_1+q+1} & \ldots & \ldots & \ldots & \Block[fill=red!15,rounded-corners]{}{P_{pb_1+q}}
			\end{pNiceArray}
			$}
		\vspace{2ex}
		  \caption{Structure of the encoding matrix $\mathbf{P}_{\mbox{\tiny SR}}$ for the SR code. The red blocks indicate assigned (non-zero) submatrices, while all other entries are zero. The matrix is clearly partitioned into $p+1$ block matrices $\mathbf{P}_0, \mathbf{P}_1, \ldots, \mathbf{P}_p$, which directly corresponds to the $p+1$ segments in the delay profile of the SR code. }\label{figure PSR}
	\end{figure*}
	\newcommand{\tikzmark}[1]{\tikz[remember picture] \node[inner sep=0pt, outer sep=0pt] (#1) {\ensuremath{#1}};}
	
	Based on the algebraic structure in Lemma \ref{lemma:P_invertible}, we now characterize the delay properties of the SR code. 
	\begin{lemma}\label{delay SR}
		The SR code defined in (i)-(iv) satisfies the following extended delay profile
		\begin{IEEEeqnarray*}{rCl}
			\IEEEeqnarraymulticol{3}{l}{
				\begin{array}{c}
					\Bigl( 
					\underbrace{pb_1+q,\ldots,pb_1+q}_{b_2},\,
					\underbrace{pb_1,\ldots,pb_1}_{q},\\  
					\qquad \underbrace{(p-1)b_1,\ldots,(p-1)b_1}_{b_2},\,
					\ldots,\,\\
					\qquad \underbrace{2b_1,\ldots,2b_1}_{b_2}
					\underbrace{b_1,\ldots,b_1}_{b_2}
					\Bigr)
				\end{array}
			}
		\end{IEEEeqnarray*}, and it has extended delay separable property.
	\end{lemma}
	\begin{proof}
		We partition the decoding matrix $\mathbf{P}_{\mbox{\tiny{SR}}}$  into $p+1$ blocks, denoted as $\mathbf{P}_0, \mathbf{P}_1, \ldots, \mathbf{P}_p$. We analyze the decoding process from matrix $\mathbf{P_0}$ to matrix $\mathbf{P_p}$.
		
		It suffices to analyze the delay of $s_j[i]$ for $i\in [t,t+b_1-1]$ and $j\in [T-b_1]$.  The matrix $P_{\mbox{\tiny{SR}}}$ is depicted in Fig.~\ref{figure PSR}.
		\begin{enumerate}
			\item $2b_2+q< j\leq pb_2+q=T-b_1$ 
			
			The sub-matrices $\mathbf{P}_0,\ldots,\mathbf{P}_{p-3}$ are all diagonal matrices in the block sense. According to (i), the matrix $P_{b_1}$ contains $I_{b_2}$ in its bottom block. Then the last $b_2$ symbols, where $(p-1)b_2+q<j\leq pb_2+q$, will be recovered at time $b_1$. For $1\leq l\leq p-2$, $(p-l)b_2+q< j\leq (p-l+1)b_2+q$ it is clear that $s_j[i]$ will be recovered at time $lb_1+i $.
			\item $b_2+q<j\leq 2b_2+q$ 
			
			According to Fig.~\ref{figure PSR}, when $b_1\leq q <b_2$, $\mathbf{P}_{p-2}$ is also the diagonal matrix in the block sense. Then for $b_2+q<j\leq 2b_2+q $, $s_j[i]$ will be recovered at time $(p-1)b_1+i$. 
			
			When $q<b_1$, we employ the extended delay profile. 
			Essentially, it receives a column sub-matrices of the matrix $\hat{\mathbf{P}}_{p-2}$ as composite information, where 
			\[
			\hat{\mathbf{P}}_{p-2}=\begin{pmatrix}
				P_{pb_1-1} &\\
				\vdots &P_{pb_1-1}\\
				\vdots &\vdots &\ddots\\
				P_{(p-1)b_1}&\vdots&\vdots &P_{pb_1-1}\\
				&\ddots &\vdots &\vdots \\
				&& \ddots &\vdots\\
				&&&P_{(p-1)b_1}
			\end{pmatrix}
			\]
			Since every $T+1$ consecutive packets contain at most $b_1$ burst erasures, we have recovered the messages before time $0$. We can then populate the decoding matrix $\mathbf{P}_{p-2}$ into $\hat{\mathbf{P}}_{p-2}$ using these previously recovered messages.
			
			The first block-column of matrix $\hat{\mathbf{P}}_{p-2}$ is received at time $(p-1)b_1$, with subsequent block-columns representing the messages combined at successive time slots. Then $s_j[i] + m_j[i]$ will be received at time $(p-1)b_1 + i$, where $m_j[i]$ represents the messages combined before time $i$. 
			
			As defined in (ii) for the SR code, we know that each matrix $P_{l}$ has only a single $1$ at position $(e_l, d_l)$, where $(p-1)b_1 + 1 \leq l \leq pb_1 - 1$. The expression $m_j[i]$ can be written as:
			\[
			m_j[i] = \sum\limits_{l:e_l = j - b_2 - q} s_{d_l}[i - (l - (p-1)b_1)].
			\]
		Moreover, the combined symbol $s_j[i] + m_j[i]$ satisfies the extended delay separable property (Definition~\ref{def:eds}). The term $m_j[i]$ is generated by the matrix sequence from $P_{pb_1-1}$ to $P_{(p-1)b_1+1}$, which, according to construction (ii), combines source symbols $s_{d_l}[i]$ with indices $b_2 < d_l \leq b_2 + q$. At 4), we will demonstrate that such $s_{d_l}[i]$ will be recovered separately.
		 Then we get the dependent set is $A_D = [b_2 + q + 1, 2b_2 + q]$, comprising symbols recovered as linear combinations, while the independent set is $A_I = [1, T - b_1] \setminus A_D$, comprising symbols recovered individually. Since $d_l \in A_I$, the linear combination $m_j[i]$ depends only on independently recovered symbols from strictly prior time instants, thus fulfilling the extended delay separable property.
			\item $b_2<j\leq b_2+q$ 
			
			We consider the decoding matrix $\mathbf{P}_{p-1}$. According to the structure of $\mathbf{P}_{p-1}$ in Fig.~\ref{SR-2}, when $b_1 \leq q < b_2$, $s_j[i]$ are recovered at time $pb_1 + i$. When $q < b_1$, the recovery of $s_j[i]$ depends on the invertible of the reduced matrix $\mathbf{P}_{p-1}$ in Lemma~\ref{lemma:P_invertible}. Since this matrix is invertible, all $s_j[i]$ are recovered upon receiving packets $P[t+pb_1]$ through $P[t+pb_1+q-1]$, corresponding to full reception of $\mathbf{P}_{p-1}$. For $i \in [t+q, t+b_2-1]$, the recovery delay is within $pb_1$. For $i \in [t, t+q-1]$, the construction (ii),(iii) imply only $P_{pb_1}$ contains $I_q$ in its first $q$ columns (the first $q$ columns of $P_{(p-1)b_1+1}$ to $P_{pb_1-1}$ are zeros). Thus $s_j[i]$ is recovered at time $pb_1 + i$ with delay $pb_1$.
			
			\item $1\leq j\leq b_2$
			
			We consider the decoding matrix $\mathbf{P}_{p}$. When $b_1 \leq q < b_2$, the matrix $\mathbf{P}_p$ is block-diagonal, enabling the recovery of $s_j[i]$ at time $pb_1 + q + i$. When $q < b_1$, the matrix $\mathbf{P}_p$ contains non-zero sub-blocks in its lower triangular part. These correspond to symbols $s_j[i]$ where $b_2 < j \leq b_2 + q$, which have already been recovered in case 3). Consequently, $s_j[i]$ will be recovered at time $pb_1 + q + i$ with a delay of $pb_1 + q$.
		\end{enumerate}
	\end{proof}
	\subsection{Construction for RD Code}
	\begin{figure*}[ht]
		\begin{equation}\label{eqRDP}
			\mathbf{P}_{\mbox{\tiny RD}} = 
			\scalebox{0.85}{$\displaystyle 
				\begin{pNiceArray}{cccc|ccccc|c|cc|c}
					\Block[fill=red!15,rounded-corners]{}{P'_{b_2}} & P'_{b_2+1}&\ldots&P'_{b_2+q-1}&\Block[fill=red!15,rounded-corners]{}{P'_{b_2+q}}&\ldots&P'_{2b_2-1}&\ldots
					&P'_{2b_2+q-1}&\Block{7-1}{\mathbf{P}'_2}&\Block{7-2}{\ldots}&&\Block{7-1}{\mathbf{P}'_p}\\
					\Block[fill=red!15,rounded-corners]{}{P'_{b_2-1}}&\Block[fill=red!15,rounded-corners]{}{P'_{b_2}}&&
					\Block{2-1}{\vdots}&\Block{2-1}{\vdots}&\Block{2-1}{\ddots} &&&\Block{2-1}{\vdots}&&&&
					\\\Block{2-1}{\vdots}&\Block{2-1}{\vdots}&\ddots&&&&&&&&&&\\
					&& &\Block[fill=red!15,rounded-corners]{}{P'_{b_2}}&P'_{b_2+1}&&\Block{2-2}{\ddots}&&\Block{2-1}{\vdots}&&&&\\
					\Block{2-1}{\vdots}&\Block{2-1}{\vdots}&&\Block{2-1}{\vdots}&\Block[fill=red!15,rounded-corners]{}{P'_{b_2}}&&&
					&&&&\\
					&&&& &\ddots&&&&&&&\\
					\Block[fill=red!15,rounded-corners]{}{P'_1}& \Block[fill=red!15,rounded-corners]{}{P'_2}&\ldots&\ldots&
					\Block[fill=red!15,rounded-corners]{}{P'_{q+1}}&\ldots&\Block[fill=red!15,rounded-corners]{}{P'_{b_2}}&\ldots
					&\Block[fill=red!15,rounded-corners]{}{P'_{b_2+q}}&&&
				\end{pNiceArray}
				$}
		\end{equation}
	\end{figure*}
	
According to \ref{key},  RD link is modeled as a $(b_2, T - b_1)$ point-to-point SC that recovers $S[t]$ or $S[t] + \sum\limits_{i < t} S[i]$ by time $t + T - b_1$.

The construction proceeds by defining the constituent $(T-b_1)\times b_2$ binary matrices $P'_i$ for $i\in[0,T-b_1]$, which collectively form the encoding matrix $\mathbf{P}_{\mbox{\tiny RD}}$. These submatrices are constructed according to the following

\begin{itemize}
	\item[(I)]For $i\in[b_2-1]$, let $d_i=i\mod q$ and $e_i=q+(i\mod (b_2-q))$. Then we set $P'_i(d_i,e_i)=1$, with all remaining entries being zero. 
	
	\item[(II)] For $i=b_2$
	\[
	P'_{b_2} =
	\begin{pmatrix}
		I_{q} & \mathbf{0} \\
		\multicolumn{2}{@{}c@{\hspace{2\arraycolsep}}}{\mathbf{~~~~0}_{p b_2 }}
	\end{pmatrix}.
	\]
	
	\item[(III)]  For $i=jb_2+q$ where $j\in[p]$ 
	\[
	P'_{j  b_2 + q} = 
	\begin{pmatrix}
		\mathbf{0}_{( j - 1) b_2 + q }  \\
		I_{b_2} \\                
		\mathbf{0}_{(p-j) b_2 }
	\end{pmatrix}.
	\]
	
	\item[(IV)] For all other indices not covered by the previous cases, we set $P'_i=\mathbf{0}$.
\end{itemize}

Following the organizational framework established for the SR code, we partition the complete encoding matrix $\mathbf{P}_{\mbox{\tiny RD}}$ into $p+1$ coherent block matrices $\mathbf{P}'_0, \mathbf{P}'_1, \ldots, \mathbf{P}'_{p}$. This partitioning reveals the underlying recursive structure and facilitates the subsequent delay analysis.

For $1\leq i\leq p$
\begin{equation}
	\mathbf{P}'_i= \scalebox{0.85}{$\displaystyle\begin{pNiceMatrix}
			P'_{ib_2+q} & P'_{ib_2+q+1} & \ldots & P'_{(i+1)b_2+q-1} \\
			P'_{ib_2+q-1} & P'_{ib_2+q} & \ldots & P'_{(i+1)b_2+q-2} \\
			\vdots & \vdots & \ddots & \vdots \\
			P'_{(i-1)b_2+q+1} & \ldots & \ldots & P'_{ib_2+q}
		\end{pNiceMatrix}$}\;.
\end{equation}
Each $\mathbf{P}'_i$ forms a $b_2 \times b_2$ block matrix.

Of particular importance is submatrix $\mathbf{P}'_{0}$, which captures the initial segment of the encoding structure
\begin{equation}\label{eq9}
	\mathbf{P}'_{0} = \begin{pmatrix}
		P'_{b_2} & \cdots & P'_{b_2+q-1} \\
		\vdots & \vdots & \vdots \\
		P'_{1} & \cdots & P'_{q}
	\end{pmatrix}.
\end{equation}
This $b_2 \times q$ block matrix.

We now analyze the structural properties of the RD encoding matrix $\mathbf{P}_{\mbox{\tiny RD}}$. The following lemma establishes the invertibility of the  submatrix $\mathbf{P}'_{0}$, which plays a crucial role in the initial decoding phases.

	\begin{lemma}  \label{invertible}
		By deleting the bottom $T-b_1-q$ rows from each block $P'_i$ in $\mathbf{P}'_{0}$, the resulting matrix becomes an invertible $b_2q\times b_2q$  binary matrix.
	\end{lemma}  
	\begin{proof}  
		We only need to prove the resulting matrix is invertible. For simplicity, we still use the notation $P'_i$ to denote the reduced block after removing the bottom $T-b_1-q$ rows from $P'_i$. First, according to (II) and (III), one can see that
		$$P'_{b_2}=(I_q~\mathbf{0}),\;\;\;\;P'_i=(\mathbf{0}~\tilde{P}_i), \;1\leq i<b_2,$$
		where $\tilde{P}_i$ denotes the right $b_2-q$ columns. Put the first $q$ columns of all blocks together and perform invertible column transformations, then the resulting $b_2q\times b_2q$ matrix becomes
		$$\begin{pmatrix}
			I_{q^2}&\mathbf{0}\\\mathbf{0}&\tilde{\mathbf{P}}
		\end{pmatrix}, \mbox{~~~where~}
		\tilde{\mathbf{P}}=\begin{pmatrix}
			\tilde{P}_{b_2-q} & \cdots & \tilde{P}_{b_2-1} \\
			\vdots & \vdots & \vdots \\
			\tilde{P}_{1} & \cdots & \tilde{P}_{q}
		\end{pmatrix}.$$
		Note that $\tilde{P}_i$, $1\leq i<b_2$, has only one entry being $1$, and zeros elsewhere. Moreover, from (III) one can see that $\tilde{\mathbf{P}}$ is actually a permutation matrix (i.e., each row and each column has exactly one $1$) and therefore is invertible. The lemma follows immediately.
	\end{proof}
	
Based on the invertibility established in Lemma \ref{invertible}, we now characterize the delay profile of the RD code.
	
	\begin{lemma}\label{RD}
		The RD code defined in (I)-(IV) satisfies the following  delay profile
		\begin{IEEEeqnarray*}{rCl}
			\IEEEeqnarraymulticol{3}{l}{
				\begin{array}{c}
					\Bigl(
					\underbrace{b_2,\ldots,b_2}_{q},\,
					\underbrace{b_2\!+\!q,\ldots,b_2\!+\!q}_{b_2},\, \underbrace{2b_2\!+\!q,\ldots,2b_2\!+\!q}_{b_2},\\  
					\ldots,\,
					\underbrace{p b_2+q,\ldots,p b_2+q}_{b_2} 
					\Bigr)
				\end{array}
			}\IEEEeqnarraynumspace 
		\end{IEEEeqnarray*}
	\end{lemma}
	\begin{proof}
		We partition the decoding matrix $\mathbf{P}_{\mbox{\tiny{RD}}}$  into $p+1$ blocks, denoted as $\mathbf{P'}_0, \mathbf{P'}_1, \ldots, \mathbf{P'}_p$. We analyze the decoding process from matrix $\mathbf{P'}_0$ to matrix $\mathbf{P'}_p$.
		Similar to Lemma \ref{delay SR}, it suffices to analyze the delay of $s_j[i]$ for $i\in[t, t+b_2-1]$ and $j\in[T-b_1]$ from the matrix $\mathbf{P}_{\mbox{\tiny RD}}$ displayed in (\ref{eqRDP}). It is accomplished in three cases below according to the value of $j$.
		\begin{enumerate}
			\item $1\leq j \leq q$. 
			
			The proof is similar as Lemma~\ref{delay SR} case 3).
			The recovery of the first $q$ symbols for all $S[i]$, $i\in[t, t+b_2-1]$ , depends on the matrix $\mathbf{P}'_0$ in (\ref{eq9}). Note the nonzero rows of $\mathbf{P}'_0$ correspond exactly to these symbols. Due to Lemma \ref{invertible}, these symbols are uniquely determined at receiving the packets $P[t+b_2], P[t+b_2+1],...,P[t+b_2+q-1]$. That is, all these symbols can be recovered by time $t+b_2+q-1$. Particularly, for $i\in[t+q,t+b_2-1]$, the recovery delay of $s_j[i]$ is thus within $b_2$. For $i\in[t, t+q-1]$, the column block of $\mathbf{P}'_0$ that contains $P'_{b_2}$ in the $(i-t+1)$-th row block contains the identity $I_q$ which corresponds exactly to the symbols $s_j[i]$, $1\leq j\leq q$, and zeros elsewhere. This implies a direct recovery of $s_j[i]$ at time $t+b_2-1+(i-t+1)=i+b_2$, so the delay is $b_2$. Therefore, for all $i\in[t, t+b_2-1]$, the first $q$ symbols of $S[i]$ can be recovered with delay $b_2$.
			
			\item $q<j\leq b_2+q$.
			
			As in Lemma~\ref{delay SR}, case 4), the proof is analogous.
			Consider the matrix $\mathbf{P}'_1$. It can be seen that $\mathbf{P}'_{1}$ is a $b_2\times b_2$ lower triangular matrix in the block sense with $P'_{b_2+q}$'s in the diagonal line. Moreover, these $P'_{b_2+q}$'s contain the identity matrices $I_q$ which correspond exactly to the symbols $s_j[i]$ for $q<j\leq b_2+q$ and all $i\in[t, t+b_2-1]$. Although there are nonzero entries below the $P'_{b_2+q}$'s, they all correspond to the first $q$ symbols of $S[i]$ which have been recovered before time $t+b_2+q$ in case 1). Therefore, for $q<j\leq b_2+q$, $s_j[i]$ can be recovered at time $i+b_2+q$.
			
			\item $b_2+q< j\leq T-b_1=pb_2+q$.
			
			Partition the interval $[b_2+q, pb_2+q]$ into $\bigcup_{l=2}^p[(l\!-\!1)b_2\!+\!q,lb_2\!+\!q\!-\!1]$.
			Note the sub-matrices $\mathbf{P}'_2,...,\mathbf{P}'_p$ are all diagonal matrices in the block sense. 
			It is easy to see that for $2\leq l\leq p$, $\mathbf{P}'_l$ indicates the symbols $s_j[i]$ for $j\in[(l\!-\!1)b_2\!+\!q,lb_2\!+\!q\!-\!1]$ can be recovered at time $i+lb_2+q$.
		\end{enumerate}
	\end{proof}
	

	
\subsection{Construction of Rate-Optimal Codes}

Having established the individual constructions for SR and RD codes, we now demonstrate how they can be integrated to form a complete $(b_1,b_2,T)$ TBSC that achieves the optimal rate. This construction corresponds to Case 2 in section \ref{key}, where the SR link employs an extended delay profile while the RD link utilizes  delay profile.

The key to achieving optimal rate lies in the careful coordination between the SR and RD delay profiles. As established in Lemma \ref{case2}, when the sum of corresponding delays in the SR and RD profiles does not exceed $T$ for all positions, the destination can recover all source symbols within the maximum delay constraint. The following lemma verifies that our specific construction satisfies this critical condition while achieving the optimal rate of $\frac{T-b_1}{T-b_1+b_2}$.

\begin{lemma}\label{Construction-2}
	The SR and RD codes construct a $(b_1,b_2,T)$ TBSC, where $b_1 < b_2$ and the constraint 
	\[
	\frac{T-b_1-b_2}{2b_2-b_1} \leq \left\lfloor \frac{T-b_1-b_2}{b_2} \right\rfloor
	\]
	must hold. This code achieves the optimal rate $R(b_1,b_2,T)$.
\end{lemma}

\begin{proof}
	
		The SR code satisfies the extended delay profile given by Lemma~\ref{delay SR}
	\begin{IEEEeqnarray*}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl( 
				\underbrace{pb_1+q,\ldots,pb_1+q}_{b_2},\,
				\underbrace{pb_1,\ldots,pb_1}_{q},\\  
				\qquad \underbrace{(p-1)b_1,\ldots,(p-1)b_1}_{b_2},\,
				\ldots,\,\\
				\qquad \underbrace{2b_1,\ldots,2b_1}_{b_2},\,
				\underbrace{b_1,\ldots,b_1}_{b_2}
				\Bigr)
			\end{array}
		}
	\end{IEEEeqnarray*}
	
	The delay profile of the RD code is given by Lemma~\ref{RD}
	\begin{IEEEeqnarray*}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl(
				\underbrace{b_2,\ldots,b_2}_{q},\,
				\underbrace{b_2\!+\!q,\ldots,b_2\!+\!q}_{b_2},\, \underbrace{2b_2\!+\!q,\ldots,2b_2\!+\!q}_{b_2},\\  
				\ldots,\,
				\underbrace{p b_2+q,\ldots,p b_2+q}_{b_2} 
				\Bigr)
			\end{array}
		}\IEEEeqnarraynumspace 
	\end{IEEEeqnarray*}
	
	To apply Lemma \ref{case2}, we need to verify that the sum of corresponding delays in the SR and RD profiles does not exceed $T$. We begin by examining the first positions of both delay profiles
	\begin{equation}
		b_2 + (pb_1 + q) \leq b_1 + pb_2 + q = T
	\end{equation}
	
	The constraint emerges from ensuring that all such delay sums satisfy the maximum delay constraint. Starting from $b_2 + q + p b_1 + q \leq T = b_1 + p b_2 + q$. Substituting $q = T - b_1 - p b_2$ into the inequality $b_2 + q + p b_1 + q \leq b_1 + p b_2 + q$ yields $T \leq (2p - 1) b_2 - (p - 2) b_1$. Then, $T - b_1 - b_2 \leq (2p - 2) b_2 - (p - 1) b_1 = (p - 1)(2b_2 - b_1)$, leading to the constraint
	\[
	\frac{T - b_1 - b_2}{2b_2 - b_1} \leq \left\lfloor \frac{T - b_1 - b_2}{b_2} \right\rfloor
	\]
	
	For the general case where $0 \leq l \leq p-1$, we have
	\[
	(p-l)b_1 + (l+1)b_2 + q \leq b_1 + pb_2 + q = T
	\]
	
	This confirms that for all positions in the delay profiles, the sum of SR and RD delays is bounded by $T$. According to Lemma~\ref{case3}, every source message will be recovered within delay $T$, which means this is a rate optimal $(b_1,b_2,T)$ TBSC. 
\end{proof}

Based on the SR and RD code constructions presented previously, we now integrate them to form a complete TBSC. The following theorem shows that under appropriate parameter constraints, this construction achieves optimal rate while satisfying the delay requirements.

	\begin{theorem}
		The SR and RD codes construct a $(b_1,b_2,T)$ TBSC. Under the  constraint 
		\[ 
		\frac{T-u-v}{2u-v}\leq \lfloor \frac{T-u-v}{u}\rfloor
		\]
		this code achieves the optimal rate $R(b_1,b_2,T)$.
	\end{theorem}
	\begin{proof}
		We have constructed rate-optimal codes under the constraint $b_1 < b_2$ with $\dfrac{T - b_1 - b_2}{2b_2 - b_1} \leq \left\lfloor \dfrac{T - b_1 - b_2}{b_2} \right\rfloor$, 
		and similarly for $b_1 > b_2$, denote $T-b_2=p'b_1+q'$ where $0< q'\leq b_1$. We have constraint $\dfrac{T - b_1 - b_2}{2b_1 - b_2} \leq \left\lfloor \dfrac{T - b_1 - b_2}{b_1} \right\rfloor$.
		
		The constructions proceed as follows
		
		\textit{SR Link Construction:} 
	This situation achieves the same optimal rate $R = \frac{T - b_2}{T - b_2 + b_1}$ as a $(b_2, b_1, T)$ TBSC. Consequently, we employ the RD code from the $(b_2, b_1, T)$ TBSC construction, whose delay profile is given by...
		\begin{IEEEeqnarray*}{rCl}
			\IEEEeqnarraymulticol{3}{l}{
				\Biggl( 
				\underbrace{b_1, \ldots, b_1}_{q'},\ 
				\underbrace{b_1 + q', \ldots, b_1 + q'}_{b_1},\ 
				\underbrace{2b_1 + q', \ldots, 2b_1 + q'}_{b_1}, 
			}
			\\
			\IEEEeqnarraymulticol{3}{l}{\hspace{2em} 
				\ldots,\ 
				\underbrace{p' b_1 + q', \ldots, p' b_1 + q'}_{b_1}
				\Biggr)
			}
		\end{IEEEeqnarray*}
		\textit{RD Link Construction:} 
		Implement the SR code for $(b_2, b_1, T)$ TBSC, which delay profile is
		\begin{IEEEeqnarray*}{rCl}
			\IEEEeqnarraymulticol{3}{l}{
				\begin{array}{c}
					\Bigl( 
					\underbrace{p'b_2+q_1,\ldots,p'b_2+q_1}_{b_1},\,
					\underbrace{p'b_2,\ldots,p'b_2}_{q'}, \\ 
					\qquad, \underbrace{(p'-1)b_2,\ldots,(p'-1)b_2}_{b_1},\,
					\ldots,\,\\
					\qquad \underbrace{2b_2,\ldots,2b_2}_{b_1},
					\underbrace{b_2,\ldots,b_2}_{b_1}
					\Bigr)
				\end{array}
			}
		\end{IEEEeqnarray*}
		
This situation for $b_1 > b_2$  corresponds to section \ref{key} Case 3, where the SR link employs delay profile while the RD link utilizes  extended delay profile. By Lemma \ref{delay SR}, the RD code construction satisfies the extended delay separable property required in Lemma \ref{case3}. Moreover, Lemma \ref{Construction-2} ensures that the sum of corresponding delays between the SR and RD profiles is bounded by $T$. Thus, all conditions of Lemma \ref{case3} are met, guaranteeing recovery of all source messages within delay $T$. This completes the proof for $b_1 > b_2$, and together with the $b_1 < b_2$ case, establishes the theorem.

	\end{proof}


\section{Constructions of Binary Rate-Optimal Streaming Codes with constraint \(T \geq b_1 + b_2 + \frac{b_1 b_2}{|b_1 - b_2|}\) }
The construction discussed here was initially proposed in our previous work \cite{li2025rateoptimalstreamingcodesthreenode}. 
In this paper, we introduce a slightly modified and simplified version to facilitate a unified comparison. 

%Specifically, in \cite{li2025rateoptimalstreamingcodesthreenode}, we expressed the  constraint as $T - b_1 = p b_2 + q$ with $0 < q \leq b_2$, and derived the condition $\frac{T - u}{v} \geq \lceil \frac{T - v}{u} \rceil$. We now adopt $T - b_1 = p b_2 + q$ with $0 \leq q < b_2$, but with constraint $\frac{T - u}{v} \geq \lfloor \frac{T - v}{u} \rfloor + 1$. 

Additionally, the RD link reflects the RD codes presented earlier in section III. We only construct SR-2 code at SR link. Firstly, we simplify this restriction.

\begin{lemma}
	Let \( b_1, b_2, T \in \mathbb{N}_+ \).  If 
	\[
	T \geq b_1 + b_2 + \frac{b_1 b_2}{|b_1 - b_2|},
	\]
	then 
	\[
	\frac{T - u}{v} \geq \left\lfloor \frac{T - v}{u} \right\rfloor + 1.
	\]
\end{lemma}

\begin{proof}
	Starting from the assumption:
	\[
	T \geq b_1 + b_2 + \frac{b_1 b_2}{|b_1 - b_2|}.
	\]
	Note that \( |b_1 - b_2| = u - v \), \( b_1 + b_2 = u + v \), and \( b_1 b_2 = u v \). Substituting these, we obtain:
	\[
	T \geq u + v + \frac{u v}{u - v}.
	\]
	Multiply both sides by \( u - v \) :
	\[
	(u - v)T \geq (u + v)(u - v) + u v = u^2 - v^2 + u v.
	\]
	Rearranging terms:
	\[
	u T - v T \geq u^2 - v^2 + u v.
	\]
	Divide both sides by \( u v \) :
	\[
	\frac{T}{v} - \frac{T}{u} \geq \frac{u}{v} - \frac{v}{u} + 1.
	\]
	Rearranging again:
	\[
	\frac{T - u}{v} \geq \frac{T - v}{u} + 1.
	\]
	Since \( \left\lfloor \frac{T - v}{u} \right\rfloor \leq \frac{T - v}{u} \), we have:
	\[
	\frac{T - u}{v} \geq \frac{T - v}{u} + 1 \geq \left\lfloor \frac{T - v}{u} \right\rfloor + 1,
	\]
	which completes the proof.
\end{proof}

We can therefore conclude that this condition is satisfied for arbitrarily large values of T.
We assume that \( b_1 < b_2 \). The optimal rate is given by
$
\frac{T - b_1}{T - b_1 + b_2} = \frac{p b_2 + q}{(p + 1) b_2 + q}.
$

We will construct SR-2 code  under the constraint \(\frac{T -  b_2}{b_1} \geq \left\lfloor \frac{T - b_1}{ b_2} \right\rfloor + 1\) at SR link. This construction is a delay-profile-based construction.

\subsubsection{Construction for SR-2 Code}

\begin{figure*}
	\centering
	
	$	\mathbf{P}_{\mbox{\tiny SR-2}}$ = 
	\scalebox{0.85}{$\displaystyle 
		\begin{pNiceMatrix}
			\Block[fill=red!15]{}{P_{b_1}} & P_{b_1+1} & \ldots & P_{2b_1-1} &\Block[fill=red!15]{}{P_{2b_1}} & P_{2b_1+1} & \ldots & P_{3b_1-1} & \cdots & \Block[fill=red!15]{}{P_{(p+1)b_1}} & 0 &\cdots & 0 \\            
			P_{b_1-1} & \Block[fill=red!15]{}{P_{b_1}} & \ldots & P_{2b_1-2} &P_{2b_1-1} & \Block[fill=red!15]{}{P_{2b_1}} & \ldots & P_{3b_1-2} & \cdots & P_{(p+1)b_1-1} & \Block[fill=red!15]{}{P_{(p+1)b_1}} & \cdots & 0 \\
			\vdots & \vdots & \ddots & \vdots &\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
			P_1 & \ldots & \ldots & \Block[fill=red!15]{}{P_{b_1}} &P_{b_1+1} & \ldots & \ldots & \Block[fill=red!15]{}{P_{2b_1}} & \cdots & P_{pb_1+1} & P_{(p+1)b_1+2} & \ldots & \Block[fill=red!15]{}{P_{(p+1)b_1}}
		\end{pNiceMatrix}
		$}
\caption{Structure of the encoding matrix $\mathbf{P}_{\mbox{\tiny SR-2}}$ for the SR-2 code. The red blocks indicate the positions containing assigned non-zero submatrices $P_i$, while all other entries are zero matrices. } \label{eq-SR2}
\end{figure*}

According to~(\ref{eq-matrixP}), we construct matrix $ \mathbf{P}_{\mbox{\tiny SR-2}} $ where each submatrix $ P_i \in \mathbb{F}_2^{(p b_2 + q) \times b_2} $ for $ 0 \leq i \leq (p-1) b_2 + b_1 + q $. Fig.~\ref{eq-SR2} provides a detailed explanation of the assignment of matrix  $ \mathbf{P}_{\mbox{\tiny SR-2}} $.

\begin{itemize}
	\item[(a)] For $i=jb_1$, $j\in[p]$, set \[
	P_{j  b_1} = 
	\begin{pmatrix}
		\mathbf{0}_{(p-j)b_2 + q)} \\
		I_{b_2} \\                 
		\mathbf{0}_{(j-1) b_2}  
	\end{pmatrix}.
	\]
	
	\item[(b)] For $i=(p+1)b_1$, set
	\[
	P_{(p+1) b_1} =
	\begin{pmatrix}
		I_{q} & \mathbf{0}_{q } \\
		\multicolumn{2}{c}{\mathbf{0}_{p b_2}}
	\end{pmatrix}.
	\]
	Note due to the constraint $\frac{T-b_2}{b_1}\geq p+1$, it follows that $T-b_2\geq (p+1)b_1$.
	\item[(c)] For other cases, set $P_i=\mathbf{0}$.
\end{itemize}



\begin{lemma}\label{SR-2}
	The SR-2 code satisfies the following delay profile
	\begin{IEEEeqnarray}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl( 
				\underbrace{(p + 1) b_1,\ldots,(p + 1) b_1}_{q},\,\underbrace{pb_1,\ldots,pb_1}_{b_2}, \\ 
				\quad  \ldots, \underbrace{2 b_1,\ldots,2 b_1}_{b_2},
				\underbrace{ b_1,\ldots, b_1}_{b_2}
				\Bigr)
			\end{array}
		}
		\IEEEeqnarraynumspace
	\end{IEEEeqnarray}
\end{lemma}

\begin{proof}
	
	It suffices to analyze the delay of $s_j[i]$ for $i\in[t, t+b_1-1]$ and $j\in[T-b_1]$ from the linear system (\ref{eq-matrixP}). According to (i)-(iii), the matrix $\mathbf{P}$, here denoted as $\mathbf{P}_{\mbox{\tiny SR}}$, is displayed in (\ref{eq-SR2}). Note only the blocks highlighted in red contain nonzero entries, while the remaining blocks are all zero blocks. Then, due to the structure of $P_{b_1}$ defined in (i), one can see the last $b_2$ symbols of $S[i]$, $i\in[t, t+b_1-1]$, can be recovered with delay $b_1$. Similarly, the structure of $P_{2b_1}$ demonstrates that the preceding $b_2$ symbols (i.e., the block immediately before the last recovered symbols) can be recovered with delay $2b_1$, and so on. The lemma is proved.
	
\end{proof}


\subsubsection{Construction of Rate-Optimal Codes}

Having presented the constructions for SR-2 and RD codes, we now establish their combined performance in the following lemma.

\begin{lemma}\label{b1b2}
	The SR-2 and RD codes construct a TBSC with parameters $(b_1,b_2,T)$. Under the constraint $\frac{T-b_2}{b_1} \geq \lfloor \frac{T - b_1}{b_2} \rfloor + 1=p + 1$ and $b_1<b_2$, this code  achieves the optimal rate $R(b_1,b_2,T)$.
\end{lemma}
\begin{proof}
	The delay profile of the SR-2 code is given by Lemma~\ref{SR-2}
	\begin{IEEEeqnarray}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl( 
				\underbrace{(p + 1) b_1,\ldots,(p + 1) b_1}_{q},\,\underbrace{pb_1,\ldots,pb_1}_{b_2}, \\ 
				\quad  \ldots, \underbrace{2 b_1,\ldots,2 b_1}_{b_2},
				\underbrace{ b_1,\ldots, b_1}_{b_2}
				\Bigr)
			\end{array}
		}
		\IEEEeqnarraynumspace
	\end{IEEEeqnarray}
	
	The delay profile of the RD code is given by Lemma~\ref{RD}.
	\begin{IEEEeqnarray*}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl(
				\underbrace{b_2,\ldots,b_2}_{q},\,
				\underbrace{b_2\!+\!q,\ldots,b_2\!+\!q}_{b_2},\, \underbrace{2b_2\!+\!q,\ldots,2b_2\!+\!q}_{b_2},\\  
				\ldots,\,
				\underbrace{p b_2+q,\ldots,p b_2+q}_{b_2} 
				\Bigr)
			\end{array}
		}\IEEEeqnarraynumspace 
	\end{IEEEeqnarray*}
	
	Adding the corresponding positions of these two delay profiles together gives the combined delay for each position. According to  $T = (p + 1) b_1 + b_2$  for all $0 \leq i \leq p-1$, the combined delay at positions corresponding to the $i[b_2]$-th SR block and the $i[b_2]$-th RD block is
	\[
	(p-i)b_2 + (i+1)b_1 + q \leq p b_2 + b_1 + q_1 = T.
	\]
	Additionally, the condition $(p+1)b_1 + b_2 \leq T$ is satisfied. 
	This is a delay-profile-based construction and according to lemma~\ref{case1} the resulting a rate-optimal $(b_1,b_2,T)$ TBSCs.
\end{proof}
\begin{theorem}
	The SR-2 and RD codes construct a $(b_1,b_2,T)$ TBSC. Under the more general constraint
	\(
	\frac{T - u}{v} \geq \left\lfloor \frac{T - v}{u} \right\rfloor + 1,
	\)
	this code achieves the optimal rate $R(b_1,b_2,T)$.
\end{theorem}
\begin{proof}
	Assume that $b_1 > b_2$. Denote $T-b_2=p'b_1+q'$ where $0< q'\leq b_1$. Under the condition $\frac{T - b_1}{b_2} \geq p' + 1$, we construct the TBSC as follows
	
	\textit{SR Link Construction:} Implement the RD code with  $(b_2, b_1, T)$ TBSCs, which has the delay profile
	\begin{IEEEeqnarray*}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl(
				\underbrace{b_1,\ldots,b_1}_{q'},\,
				\underbrace{b_1\!+\!q',\ldots,b_1\!+\!q'}_{b_1},\, \underbrace{2b_1\!+\!q',\ldots,2b_1\!+\!q'}_{b_1},\\  
				\ldots,\,
				\underbrace{p' b_1+q',\ldots,p' b_1+q'}_{b_1} 
				\Bigr)
			\end{array}
		}\IEEEeqnarraynumspace 
	\end{IEEEeqnarray*}

	
	\textit{RD Link Construction:} Implement the SR code with  $(b_2, b_1, T)$ TBSCs, which has the delay profile
	
	\begin{IEEEeqnarray}{rCl}
		\IEEEeqnarraymulticol{3}{l}{
			\begin{array}{c}
				\Bigl( 
				\underbrace{(p' + 1) b_2,\ldots,(p' + 1) b_2}_{q_2},\,\underbrace{p'b_2,\ldots,p'b_2}_{b_1}, \\ 
				\quad  \ldots, \underbrace{2 b_2,\ldots,2 b_2}_{b_1},
				\underbrace{ b_2,\ldots, b_2}_{b_1}
				\Bigr)
			\end{array}
		}
		\IEEEeqnarraynumspace
	\end{IEEEeqnarray}
	
	For each symbol index $i$, the sum of the $i$-th delay value in the SR link profile and the $i$-th delay value in the RD link profile satisfies $ \leq T$.
	According to Lemma~\ref{case1}, the source messages can be recovered within maximum delay $T$. 
	
	Combining this result with the rate constraint established in Lemma~\ref{b1b2} completes the proof.
\end{proof}

\begin{figure*}[htpb]
	\centering
	\newcommand{\textsoftbf}[1]{\textcolor{black!80}{\textbf{#1}}}
	\newcommand{\bmsoft}[1]{\textcolor{black!80}{\bm{#1}}}
	\setlength{\tabcolsep}{1.5pt} 
	
	% ====== 第一个表格 ======
	\begingroup
	\renewcommand{\arraystretch}{1.0}
	\small
	\begin{tabular}{|*{9}{>{\centering\arraybackslash}p{13mm}|}}
		\hline
		\diagbox[innerwidth=13mm,height=3.5ex]{\scriptsize\textbf{Symbol}}{\scriptsize\textbf{Time}}&0& 1&2 &3 &4 &5 &6 &7 \\
		\hline
		\bmsoft{$s_1[t]$}& \textcolor{red}{$s_1[0]$} & $s_1[1]$ & $s_1[2]$ & $s_1[3]$ & $s_1[4]$ & $s_1[5]$ & $s_1[6]$ & $s_1[7]$ \\
		\hline
		\bmsoft{$s_2[t]$}& \textcolor{red}{$s_2[0]$} & $s_2[1]$ & $s_2[2]$ & $s_2[3]$ & $s_2[4]$ & $s_2[5]$ & $s_2[6]$ & $s_2[7]$ \\
		\hline
		\bmsoft{$s_3[t]$}& \textcolor{red}{$s_3[0]$} & $s_3[1]$ & $s_3[2]$ & $s_3[3]$ & $s_3[4]$ & $s_3[5]$ & $s_3[6]$ & $s_3[7]$ \\
		\hline
		\bmsoft{$s_4[t]$}& \textcolor{red}{$s_4[0]$} & $s_4[1]$ & $s_4[2]$ & $s_4[3]$ & $s_4[4]$ & $s_4[5]$ & $s_4[6]$ & $s_4[7]$ \\
		\hline
		\bmsoft{$s_5[t]$} & \textcolor{red}{$s_5[0]$} & $s_5[1]$ & $s_5[2]$ & $s_5[3]$ & $s_5[4]$ & $s_5[5]$ & $s_5[6]$ & $s_5[7]$ \\
		\hline
		\bmsoft{$s_6[t]$} & \textcolor{red}{$s_6[0]$} & $s_6[1]$ & $s_6[2]$ & $s_6[3]$ & $s_6[4]$ & $s_6[5]$ & $s_6[6]$ & $s_6[7]$ \\
		\hline
		\bmsoft{$s_7[t]$} & \textcolor{red}{$s_7[0]$} & $s_7[1]$ & $s_7[2]$ & $s_7[3]$ & $s_7[4]$ & $s_7[5]$ & $s_7[6]$ & $s_7[7]$ \\
		\hline
		\bmsoft{$s_8[t]$} & \textcolor{red}{$s_8[0]$} & $s_8[1]$ & $s_8[2]$ & $s_8[3]$ & $s_8[4]$ & $s_8[5]$ & $s_8[6]$ & $s_8[7]$ \\
		\hline
		\bmsoft{$s_9[t]$} & \textcolor{red}{$s_9[0]$} & $s_9[1]$ & $s_9[2]$ & $s_9[3]$ & $s_9[4]$ & $s_9[5]$ & $s_9[6]$ & $s_9[7]$ \\
		\hline 
		\rowcolor{gray!15} 
		\raisebox{-1.5ex}{\bmsoft{$p_1[t]$}} & 
		\scriptsize$s_6[-3]+s_5[-6]+s_1[-7]$ & 
		\scriptsize$s_6[-2]+s_5[-5]+s_1[-6]$ & 
		\scriptsize$s_6[-1]+s_5[-4]+s_1[-5]$ & 
		\scriptsize$\textcolor{red}{s_6[0]}+s_5[-3]+s_1[-4]$ & 
		\scriptsize$s_6[1]+s_5[-2]+s_1[-3]$ & 
		\scriptsize$s_6[2]+s_5[-1]+s_1[-2]$ & 
		\scriptsize$s_6[3]+\textcolor{red}{s_5[0]}+s_1[-1]$ & 
		\scriptsize$s_6[4]+s_5[1]+\textcolor{red}{s_1[0]}$ \\ 
		\hline
		\rowcolor{gray!15}
		\raisebox{-1.5ex}{\bmsoft{$p_2[t]$}} &
		\dbox{\scriptsize$s_7[-3]+s_5[-4]+$} &
		\dbox{\scriptsize$s_7[-2]+s_5[-3]+$} &
		\dbox{\scriptsize$s_7[-1]+s_5[-2]+$} &
		\dbox{\scriptsize$\textcolor{red}{s_7[0]}+s_5[-1]+$} & 
		\dbox{\scriptsize$s_7[1]+\textcolor{red}{s_5[0]}+$} &
		\dbox{\scriptsize$s_7[2]+s_5[1]+$} &
		\dbox{\scriptsize$s_7[3]+s_5[2]+$} &
		\dbox{\scriptsize$s_7[4]+s_5[3]+$} \\
		\rowcolor{gray!15}
		& \scriptsize$s_2[-7]$ & \scriptsize$s_2[-6]$ & \scriptsize$s_2[-5]$ &
		\scriptsize$s_2[-4]$ & \scriptsize$s_2[-3]$ & \scriptsize$s_2[-2]$ &
		\scriptsize$s_2[-1]$ & \scriptsize\textcolor{red}{$s_2[0]$} \\
		\hline
		\rowcolor{gray!15}
		\raisebox{-1.5ex}{\bmsoft{$p_3[t]$}} &
		\dbox{\scriptsize$s_8[-3]+s_5[-5]+$} &
		\dbox{\scriptsize$s_8[-2]+s_5[-4]+$} &
		\dbox{\scriptsize$s_8[-1]+s_5[-3]+$} &
		\dbox{\scriptsize$\textcolor{red}{s_8[0]}+s_5[-2]+$} & 
		\dbox{\scriptsize$s_8[1]+s_5[-1]+$} &
		\dbox{\scriptsize$s_8[2]+\textcolor{red}{s_5[0]}+$} &
		\dbox{\scriptsize$s_8[3]+s_5[1]+$} &
		\dbox{\scriptsize$s_8[4]+s_5[2]+$} \\
		\rowcolor{gray!15}
		& \scriptsize$s_3[-7]$ & \scriptsize$s_3[-6]$ & \scriptsize$s_3[-5]$ &
		\scriptsize$s_3[-4]$ & \scriptsize$s_3[-3]$ & \scriptsize$s_3[-2]$ &
		\scriptsize$s_3[-1]$ & \scriptsize\textcolor{red}{$s_3[0]$} \\
		\hline
		\rowcolor{gray!15} 
		\raisebox{-0.4ex}{\bmsoft{$p_4[t]$}} & 
		\scriptsize$s_9[-3]+s_4[-7]$ & 
		\scriptsize$s_9[-2]+s_4[-6]$ & 
		\scriptsize$s_9[-1]+s_4[-5]$ & 
		\scriptsize$\textcolor{red}{s_9[0]}+s_4[-4]$ & 
		\scriptsize$s_9[1]+s_4[-3]$ & 
		\scriptsize$s_9[2]+s_4[-2]$ & 
		\scriptsize$s_9[3]+s_4[-1]$ & 
		\scriptsize$s_9[4]+\textcolor{red}{s_4[0]}$ \\ 
		\hline
	\end{tabular}
	\endgroup
	
	\vspace{1em}
	\textbf{(a) Symbols transmitted by node s from time 0 to 7}
	
	\smallskip
	\caption{
		The messages inside the red dashed box are treated as linear combinations for transmission. Suppose a burst erasure of length 3 occurs, resulting in the loss of messages at times $0$, $1$, and $2$. Since messages before time $0$ and from time $3$ to $7$ are not lost, we can apply the decoding analysis from Example 1. Accordingly, the red marked symbols of $S[0]$ can be received  with the delay profile $(7,7,7,7,6,3,3,3,3)$. Following the decoding analysis in Example 1, the combined symbol 
		\(
		(s_1[0], s_2[0], s_3[0], s_4[0], s_5[0], s_6[0], s_7[0] + s_5[-1], s_8[0] + s_5[-2], s_9[0])
		\)
		can be recovered by the known symbols at relay node, as specified by the same delay profile.
	}
	\label{fig:table_s}
\end{figure*}

\section{Example}
	\subsection{Example 1} 
	Consider the example of $b_1=3$, $b_2=4$. Construction in \cite{li2025rateoptimalstreamingcodesthreenode} with constraint $\frac{T-u}{v}\geq \lfloor \frac{T-v}{u}\rfloor+1$ can construct rate-optimal TBSCs under $T \in \{10,13,14\}$ or $T \geq 16$. The construction of this paper has the constraint that $T \in \{7,11,12\}$ or $T \geq 15$, which fills the gap of construction in \cite{li2025rateoptimalstreamingcodesthreenode}. Thus, we can construct rate-optimal TBSCs for $T=7$ and $T \geq 10$. And in Section V, we prove that the optimal rate is not achievable for $T = 8$ and $T = 9$. The construction in \cite{9834645} can only construct rate-optimal TBSCs under the constraint that $b_2 \mid (T - b_1 - b_2)$, that is, $T \in \{7,11,15,19,\ldots\}$.
	
	For parameters $b_1=3$, $b_2=4$, and $T=12$, the construction presented here is the only known one capable of achieving the optimal rate. The detailed transmission processes of the SR and RD links are illustrated in Fig.~\ref{fig:table_s} and Fig.~\ref{fig:table_r}, respectively.
	Since $b_1<b_2$, we employ an extended delay profile at the SR link. The optimal rate is given by \[
	R=\frac{T-b_1}{T-b_1+b_2}=\frac{9}{13}
	\].
	

	\subsubsection{SR link construction}
	
Source messages are given by $\{S[t]\}_{t=0}^{\infty}$, where $S[t] \in \mathbb{F}^9$ at time $t$, and the transmitted symbols is $X[t] = (S[t], P[t]) \in \mathbb{F}^{13}$. The parity packets $P[t]$ are generated by a systematic convolutional code satisfying
\[
(P[0], P[1], \ldots, P[t]) = (S[0], S[1], \ldots, S[t])  (P_{j-i})_{i,j \in [t+1]},
\]
where the matrix $(P_{j-i})_{i,j \in [t+1]}$ is a block-lower triangular matrix of the form
\[
(P_{j-i})_{i,j \in [t+1]} = 
\begin{pNiceMatrix}
	P_0 & P_1 & \ldots & P_t \\
	& P_0 & \cdots & P_{t-1} \\
	\Block{2-2}{\mathbf{\Huge 0}}& & \ddots & \vdots \\
	& & & P_0
\end{pNiceMatrix},
\]
with $P_i \in \mathbb{F}_2^{9 \times 4}$.

For the SR link, the effective delay is $T - b_2 = 8$, which implies that each source message $S[t]$ (or a linear combination of $S[t]$ with prior messages $S[i]$ for $i < t$, i.e., $S[t] + \sum\limits_{i < t} S[i]$) must be recovered by the relay nodes at time $t+8$.

Suppose a burst erasure of length 3 occurs starting at time $0$, causing the loss of $S[0]$, $S[1]$, and $S[2]$. These packets (or equivalent linear combinations) must be recovered by times $8$, $9$, and $10$, respectively.

To enable recovery, we consider the parity packets received between times $3$ and $10$, and remove the known contributions from $S[3]$ to $S[10]$. This process results in an effective matrix for decoding
	\[
	P_{\text{\tiny{SR}}} = \begin{pmatrix}
		P_3 & P_4 & P_5 & P_6  & P_7 & P_8 & 0 & 0 \\
		P_2 & P_3 & P_4 & P_5 & P_6 & P_7 & P_8 & 0 \\
		P_1 & P_2 & P_3 & P_4 & P_5 & P_6 & P_7 & P_8
	\end{pmatrix}
	\]
	where $P_i \in \mathbb{F}_2^{9 \times 4}$.
	
 Set
	\[
	P_3 = \begin{pmatrix}
		\mathbf{0}_5 \\
		I_4
	\end{pmatrix}, 
	P_4 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		0 & 1 & 0 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix}, 
	P_5 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		0 & 0 & 1 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix},
	\]
	\[
	P_6 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		1 & 0 & 0 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix}, \quad
	P_7 = \begin{pmatrix}
		I_4 \\
		\mathbf{0}_5  
	\end{pmatrix}.
	\]
	
We partition the matrix $P_{\text{\tiny{SR}}}$ into blocks and obtain the following matrices for decoding
\[
\mathbf{P}_0 = \begin{pmatrix}
	P_3 & P_4 & P_5 \\
	P_2 & P_3 & P_4 \\
	P_1 & P_2 & P_3
\end{pmatrix}, \quad
\mathbf{P}_1 = \begin{pmatrix}
	P_6 \\
	P_5 \\
	P_4
\end{pmatrix}
\]
\\
\[
\mathbf{P}_3 = \begin{pmatrix}
	P_7 & P_8 & 0 &0 \\
	P_6 & P_7 & P_8&0 \\
	P_5 & P_6 & P_7& P_8
\end{pmatrix}.
\]

This structure implies that the decoding process consists of three steps, where the messages $S[i]$ for $i \in [3]$ are decoded separately using matrices $\mathbf{P}_0$, $\mathbf{P}_1$, and $\mathbf{P}_2$ at times $i+3$, $i+6$, and $i+7$, respectively.

We now analyze the decoding process for each matrix.

\paragraph{Decoding with $\mathbf{P}_0$}
Since $P_4 \neq \mathbf{0}$ and $P_5 \neq \mathbf{0}$, we consider an extended delay profile for $\mathbf{P}_0$. According to the channel model, the messages at times $-2$ and $-1$ are not lost. Thus, we reconstruct $\mathbf{P}_0$ to obtain $\mathbf{\hat{P}}_0$
\[
\mathbf{\hat{P}}_0 = \begin{pmatrix}
	P_5 & 0 & 0 \\
	P_4 & P_5 & 0 \\
	P_3 & P_4 & P_5 \\
	0 & P_3 & P_4 \\
	0 & 0 & P_3
\end{pmatrix}.
\]

At time $3$, we can recover the first column of $\mathbf{\hat{P}}_0$, which provides combined information for transmission. By examining the positions of $1$ in the first column of this block matrix, we obtain $s_7[0] + s_5[-1]$, $s_8[0] + s_5[-2]$, and $s_9[0]$ at time $3$. Similarly, for $i \in [3]$, we recover $s_7[i] + s_5[i-1]$, $s_8[i] + s_5[i-2]$, and $s_9[i]$ at time $i+3$.
\paragraph{Decoding with $\mathbf{P}_1$}

By removing the first four rows, the last four rows, and the last column from matrices $P_4$, $P_5$, and $P_6$, we obtain $P'_4$, $P'_5$, and $P'_6$, respectively. This yields
\[
\mathbf{\hat{P}}_1 = \begin{pmatrix}
	P'_6 \\
	P'_5 \\
	P'_4
\end{pmatrix} =
\begin{pmatrix}
	1 & 0 & 0 \\
	0 & 0 & 1 \\
	0 & 1 & 0
\end{pmatrix}.
\]

Clearly, $\mathbf{\hat{P}}_1$ is invertible. Therefore, we can recover $s_5[0]$, $s_5[1]$, and $s_5[2]$ at time $6$.

\paragraph{Decoding with $\mathbf{P}_2$}
Only matrices $P_5$, $P_6$, and $P_7$ contain nonzero entries. However, $P_5$ and $P_6$ only have a $1$ in the position corresponding to $s_5[i]$, which was already recovered in Step b). Thus, we can recover $s_1[i]$, $s_2[i]$, $s_3[i]$, and $s_4[i]$ at time $i+7$.

Consequently, SR codes achieves an extended delay profile of $(7,7,7,7,6,3,3,3,3)$ with symbols
\(
(s_1[t], s_2[t], s_3[t], s_4[t], s_5[t], s_6[t], s_7[t] + s_5[t-1], s_8[t] + s_5[t-2], s_9[t])\).
The detailed transmission process of the SR link from time $0$ to $7$ is illustrated in Fig.~\ref{fig:table_s}.

	\subsubsection{RD Link Construction}
	
	\begin{figure*}[htpb]
		\centering
		\newcommand{\textsoftbf}[1]{\textcolor{black!80}{\textbf{#1}}}
		\newcommand{\bmsoft}[1]{\textcolor{black!80}{\bm{#1}}}
		\setlength{\tabcolsep}{1.5pt} 
		
		\begingroup
		\small
		\renewcommand{\arraystretch}{1.1}
		\begin{tabular}{|*{11}{>{\centering\arraybackslash}p{13mm}|}}
			\hline
			\diagbox[innerwidth=13mm,height=3.5ex]{\scriptsize\textbf{Symbol}}{\scriptsize\textbf{Time}} & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 \\
			\hline
			\bmsoft{$r_1[t]$} &$s_1[-4]$ & $s_1[-3]$ & $s_1[-2]$ & $s_1[-1]$ & $\textcolor{red}{s_1[0]}$ & $s_1[1]$ & $s_1[2]$ & $s_1[3]$ & $s_1[4]$ & $s_1[5]$ \\
			\hline
			\bmsoft{$r_2[t]$} &$s_2[-4]$ & $s_2[-3]$ & $s_2[-2]$ & $s_2[-1]$ & $\textcolor{red}{s_2[0]}$ & $s_2[1]$ & $s_2[2]$ & $s_2[3]$ & $s_2[4]$ & $s_2[5]$ \\
			\hline
			\bmsoft{$r_3[t]$} &$s_3[-4]$ & $s_3[-3]$ & $s_3[-2]$ & $s_3[-1]$ & $\textcolor{red}{s_3[0]}$ & $s_3[1]$ & $s_3[2]$ & $s_3[3]$ & $s_3[4]$ & $s_3[5]$ \\
			\hline
			\bmsoft{$r_4[t]$} &$s_4[-4]$ & $s_4[-3]$ & $s_4[-2]$ & $s_4[-1]$ & $\textcolor{red}{s_4[0]}$ & $s_4[1]$ & $s_4[2]$ & $s_4[3]$ & $s_4[4]$ & $s_4[5]$ \\
			\hline
			\bmsoft{$r_5[t]$} & $s_5[-3]$ & $s_5[-2]$ & $s_5[-1]$ & $\textcolor{red}{s_5[0]}$ & $s_5[1]$ & $s_5[2]$ & $s_5[3]$ & $s_5[4]$ & $s_5[5]$ & $s_5[6]$ \\
			\hline
			\bmsoft{$r_6[t]$} & $\textcolor{red}{s_6[0]}$ & $s_6[1]$ & $s_6[2]$ & $s_6[3]$ & $s_6[4]$ & $s_6[5]$ & $s_6[6]$ & $s_6[7]$ & $s_6[8]$ & $s_6[9]$ \\
			\hline
			\bmsoft{$r_7[t]$} & $s_7[0]+s_5[-1]$ & $s_7[1]+\textcolor{red}{s_5[0]}$ & $s_7[2]+s_5[1]$ &  $s_7[3]+s_5[2]$ &  $s_7[4]+s_5[3]$ &  $s_7[5]+s_5[4]$ & $s_7[6]+s_5[5]$ & $s_7[7]+s_5[6]$ & $s_7[8]+s_5[7]$ &  $s_7[9]+s_5[8]$ \\
			\hline
			\bmsoft{$r_8[t]$} & $s_8[0]+s_5[-2]$ & $s_8[1]+s_5[-1]$ &  $s_8[2]+\textcolor{red}{s_5[0]}$ &  $s_8[3]+s_5[1]$ &  $s_8[4]+s_5[2]$ & $s_8[5]+s_5[3]$ &  $s_8[6]+s_5[4]$ &  $s_8[7]+s_5[5]$ &  $s_8[8]+s_5[6]$ &  $s_8[9]+s_5[7]$ \\
			\hline
			\bmsoft{$r_9[t]$} &$\textcolor{red}{s_9[0]}$ & $s_9[1]$ & $s_9[2]$ & $s_9[3]$ & $s_9[4]$ & $s_9[5]$ & $s_9[6]$ & $s_9[7]$ & $s_9[8]$ & $s_9[9]$ \\
			\hline
			\rowcolor{gray!15}
			\raisebox{-1.5ex}{\bmsoft{$p'_1[t]$}} &  \scriptsize$s_1[-8]+s_2[-9]+s_6[-9]$ & \scriptsize$s_1[-7]+s_2[-8]+s_6[-8]$ & \scriptsize$s_1[-6]+s_2[-7]+s_6[-7]$ & \scriptsize$s_1[-5]+s_2[-6]+s_6[-6]$ & \scriptsize$s_1[-4]+s_2[-5]+s_6[-5]$ & \scriptsize$s_1[-3]+s_2[-4]+s_6[-4]$ & \scriptsize$s_1[-2]+s_2[-3]+s_6[-3]$ & \scriptsize$s_1[-1]+s_2[-2]+s_6[-2]$ & \scriptsize$\textcolor{red}{s_1[0]}+s_2[-1]+s_6[-1]$ & \scriptsize$s_1[1]+\textcolor{red}{s_2[0]}+\textcolor{red}{s_6[0]}$ \\
			\hline
			\rowcolor{gray!15}
			\raisebox{-2.5ex}{\bmsoft{$p'_2[t]$}} & \scriptsize$s_1[-5]+s_3[-9]+s_7[-9]+s_5[-10]$ & \scriptsize$s_1[-4]+s_3[-8]+s_7[-8]+s_5[-9]$ & \scriptsize$s_1[-3]+s_3[-7]+s_7[-7]+s_5[-8]$ & \scriptsize$s_1[-2]+s_3[-6]+s_7[-6]+s_5[-7]$ & \scriptsize$s_1[-1]+s_3[-5]+s_7[-5]+s_5[-6]$ & \scriptsize$\textcolor{red}{s_1[0]}+s_3[-4]+s_7[-4]+s_5[-5]$ & \scriptsize$s_1[1]+s_3[-3]+s_7[-3]+s_5[-4]$ & \scriptsize$s_1[2]+s_3[-2]+s_7[-2]+s_5[-3]$ & \scriptsize$s_1[3]+s_3[-1]+s_7[-1]+s_5[-2]$ & \scriptsize$s_1[4]+\textcolor{red}{s_3[0]}+\textcolor{red}{s_7[0]}+s_5[-1]$ \\
			\hline
			\rowcolor{gray!15}
			\raisebox{-1.5ex}{\bmsoft{$p'_3[t]$}} &  \scriptsize$s_1[-6]+s_4[-9]+s_8[-9]+s_5[-11]$ & \scriptsize$s_1[-5]+s_4[-8]+s_8[-8]+s_5[-10]$ & \scriptsize$s_1[-4]+s_4[-7]+s_8[-7]+s_5[-9]$ & \scriptsize$s_1[-3]+s_4[-6]+s_8[-6]+s_5[-8]$ & \scriptsize$s_1[-2]+s_4[-5]+s_8[-5]+s_5[-7]$ & \scriptsize$s_1[-1]+s_4[-4]+s_8[-4]+s_5[-6]$ & \scriptsize$\textcolor{red}{s_1[0]}+s_4[-3]+s_8[-3]+s_5[-5]$ & \scriptsize$s_1[1]+s_4[-2]+s_8[-2]+s_5[-4]$ & \scriptsize$s_1[2]+s_4[-1]+s_8[-1]+s_5[-3]$ & \scriptsize$s_1[3]+\textcolor{red}{s_4[0]}+\textcolor{red}{s_8[0]}+s_5[-2]$ \\
			\hline
			\rowcolor{gray!15}
			\raisebox{-1.5ex}{\bmsoft{$p'_4[t]$}} & \scriptsize$s_1[-7]+s_5[-8]+s_9[-9]$ & \scriptsize$s_1[-6]+s_5[-7]+s_9[-8]$ & \scriptsize$s_1[-5]+s_5[-6]+s_9[-7]$ & \scriptsize$s_1[-4]+s_5[-5]+s_9[-6]$ & \scriptsize$s_1[-3]+s_5[-4]+s_9[-5]$ & \scriptsize$s_1[-2]+s_5[-3]+s_9[-4]$ & \scriptsize$s_1[-1]+s_5[-2]+s_9[-3]$ & \scriptsize$\textcolor{red}{s_1[0]}+s_5[-1]+s_9[-2]$ & \scriptsize$s_1[1]+\textcolor{red}{s_5[0]}+s_9[-1]$ & \scriptsize$s_1[2]+s_5[1]+\textcolor{red}{s_9[0]}$ \\
			\hline
		\end{tabular}
		\endgroup
		
		\vspace{1em}
		\textbf{(b) Symbols transmitted by node r from time 3 to 12}
		
		\vspace{0.5em}
	\caption{
		The relay node transmits $(R[t],P'[t])$ at time $t$, where 
\( R[t] = \big( s_1[t-7], s_2[t-7], s_3[t-7], s_4[t-7], s_5[t-6], s_6[t-3], \)
\( \smash{s_7[t-3] + s_5[t-4], s_8[t-3] + s_5[t-5], s_9[t-3]} \big) \).
	The red marked symbols of $S[0]$ can be verified that all its elements are received by time $12$. Our analysis in Example 1 demonstrates that $S[0]$ can be successfully recovered. Combined with Fig.~\ref{fig:table_s}, this constructs a rate-optional streaming code with parameters $b_1 = 3$, $b_2 = 4$, and $T = 12$. The SR code has an extended delay profile $(7,7,7,7,6,3,3,3,3)$ with symbols  $(s_1[t], s_2[t], s_3[t], s_4[t], s_5[t], s_6[t], s_7[t] + s_5[t-1], s_8[t] + s_5[t-2], s_9[t])$, while the RD code has delay profile $(4,5,5,5,5,9,9,9,9)$. Summing the two delay profiles yields a resultant delay under $12$.}\label{fig:table_r}
	\end{figure*}
Assume that the relay transmits $(R[t], P'[t])$ at time $t$, where $R[t] \in \mathbb{F}^9$ and $P'[t] \in \mathbb{F}^4$.
For the RD link, the effective delay is $T-b_1=9$.
The decoding process requires the following matrix
\[
P'_{\mbox{\tiny{RD}}} = 
\begin{pmatrix}
	P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & P'_9 & 0 & 0 & 0 \\
	P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & P'_9 & 0 & 0 \\
	P'_2 & P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & P'_9 & 0 \\
	P'_1 & P'_2 & P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & P'_9 
\end{pmatrix}.
\]

	
	Set
	\[
	P'_1 = \begin{pmatrix}
		0&1&0&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}, \quad
	P'_2 = \begin{pmatrix}
		0&0&1&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}
	\]
	\[
	P'_3 =   \begin{pmatrix}
		0&0&0&1\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix} ,\quad 
	P'_4 = \begin{pmatrix}
		1&0&0&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}
	\]
	\[
	P'_5 = \begin{pmatrix}
		\mathbf{0}_1 \\
		I_4 \\
		\mathbf{0}_4 
	\end{pmatrix}, \quad
	P'_9 = \begin{pmatrix}
		\mathbf{0}_5 \\
		I_4
	\end{pmatrix}.
	\]
	
We partition $P'_{\mbox{\tiny{RD}}}$ into three blocks
\[
\mathbf{P}'_0 = \begin{pmatrix}
	P'_4 \\ P'_3 \\ P'_2 \\ P'_1
\end{pmatrix}, \quad 
\mathbf{P}'_1 = \begin{pmatrix}
	P'_5 & P'_6 & P'_7 & P'_8 \\
	P'_4 & P'_5 & P'_6 & P'_7 \\
	P'_3 & P'_4 & P'_5 & P'_6 \\
	P'_2 & P'_3 & P'_4 & P'_5
\end{pmatrix}\]
\[ 
\mathbf{P}'_2 = \begin{pmatrix}
	P'_9 & 0 & 0 & 0 \\
	P'_8 & P'_9 & 0 & 0 \\
	P'_7 & P'_8 & P'_9 & 0 \\
	P'_6 & P'_7 & P'_8 & P'_9
\end{pmatrix}.
\]

Following an analysis similar to that of the SR link construction, this RD link achieves a delay profile of $(4,5,5,5,5,9,9,9,9)$.
The detailed transmission process of the RD link from time $3$ to $12$ is illustrated in Fig.~\ref{fig:table_r}.
	
	\subsubsection{The decoding and relay map}
	The source node transmits messages $S[t]$ at time $t$ and has an extended delay profile $(7,7,7,7,6,3,3,3,3)$ with symbols
	$(s_1[t], s_2[t], s_3[t], s_4[t], s_5[t], s_6[t], s_7[t] + s_5[t-1], s_8[t] + s_5[t-2], s_9[t])$. 
	Then, let \begin{align*}
		R[t] &= \Big(r_1[t],r_2[t],r_3[t],r_4[t],r_5[t],r_6[t],r_7[t],r_8[t],r_9[t]\Big)\\
		&=\Big(s_1[t-7], s_2[t-7], s_3[t-7], s_4[t-7],  \\
		& \qquad s_5[t-6],s_6[t-3], s_7[t-3] + s_5[t-4],\\
		&  \qquad s_8[t-3] + s_5[t-5],s_9[t-3]\Big)
	\end{align*}
	
	The relay symbol $R[t]$ is well-defined because at time $t$, the relay has access to all the required source symbols: 
	$s_1[t-7], s_2[t-7], s_3[t-7], s_4[t-7]$ , $s_5[t-6]$, $s_6[t-3]$, $s_7[t-3]+s_5[t-4]$, $s_8[t-3]+s_5[t-5]$, $s_9[t-3]$, which are available due to the respective delay constraints.
	
	The RD link has a delay profile $(4,5,5,5,5,9,9,9,9)$.
	
	{Delay Analysis}
	\begin{itemize}
		\item {Nodes 1-6 and 9:} The delay is clearly within $T=12$.
		\item {Node 7:} For $s_7[t-3] + s_5[t-4]$, it will be received at time $t+9$, and $s_5[t-4]$ has been received at time $t+7$ by node 5. 
		Thus, we can recover $s_7[t-3]$ with a delay of $12$. 
		\item {Node 8:} Similarly,  For $s_8[t-3] + s_5[t-5]$, it will be received at time $t+9$, and $s_5[t-5]$ has been received at time $t+6$ by node 5. 
	\end{itemize}
	
	
	
	\subsection{Example 2}
	Consider the system with parameters $b_1=4$, $b_2=3$, and $T=12$. Since $b_1 > b_2$, we employ an extended delay profile at the RD link. The rate upper bound is given by
	\[
	R = \frac{T - b_2}{T - b_2 + b_1} = \frac{9}{13}.
	\]
	
	\subsubsection{SR Link Construction}
	The decoding matrix at the SR link is
	\[
	P_{\text{SR}} = \begin{pmatrix}
		P_4 & P_5 & P_6 & P_7 & P_8 & P_9 & 0 & 0 & 0 \\
		P_3 & P_4 & P_5 & P_6 & P_7 & P_8 & P_9 & 0 & 0 \\
		P_2 & P_3 & P_4 & P_5 & P_6 & P_7 & P_8 & P_9 & 0 \\
		P_1 & P_2 & P_3 & P_4 & P_5 & P_6 & P_7 & P_8 & P_9 
	\end{pmatrix}.
	\]
	
	Similar as Example 2 for RD link construction, define the component matrices  as
	\[
	P_1 = \begin{pmatrix}
		0&1&0&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}, \quad
	P_2 = \begin{pmatrix}
		0&0&1&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}
	\]
	\[
	P_3 =   \begin{pmatrix}
		0&0&0&1\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix} ,\quad 
	P_4 = \begin{pmatrix}
		1&0&0&0\\
		\multicolumn{4}{c}{\mathbf{0}_{8}}
	\end{pmatrix}
	\]
	This construction yields an SR link delay profile of $(4,5,5,5,5,9,9,9,9)$.
	
	\subsubsection{RD Link Construction}
	The decoding matrix at the RD link is
	\[
	P'_{\text{RD}} = \begin{pmatrix}
		P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & 0 & 0 \\
		P'_2 & P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8 & 0 \\
		P'_1 & P'_2 & P'_3 & P'_4 & P'_5 & P'_6 & P'_7 & P'_8
	\end{pmatrix}.
	\]
	
	Similar as the example 2 for SR link, let
	\[
	P'_3 = \begin{pmatrix}
		\mathbf{0}_5 \\
		I_4
	\end{pmatrix}, \quad
	P'_4 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		0 & 1 & 0 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix}\]\\
	\[P'_5 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		0 & 0 & 1 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix},
	P'_6 = \begin{pmatrix}
		\multicolumn{4}{c}{\mathbf{0}_{4}} \\
		1 & 0 & 0 & 0 \\
		\multicolumn{4}{c}{\mathbf{0}_{4}} 
	\end{pmatrix},
	P'_7 = \begin{pmatrix}
		I_4 \\
		\mathbf{0}_5  
	\end{pmatrix}.
	\]
	This results in an extended delay profile $(7,7,7,7,6,3,3,3,3)$ with symbols
	$
	\Big( r_1[t],\, r_2[t],\, r_3[t],\, r_4[t],\, r_5[t],\, r_6[t],\, r_7[t] + r_5[t-1],\, r_8[t] + r_5[t-2],\, r_9[t] \Big).
	$
		\begin{figure*}[htpb]
		\centering
		\label{eqRDP}
		$\mathbf{P}_{\mbox{\tiny SR}} =$ 
		$\displaystyle 
		\begin{pNiceArray}{cccccccccc}
			P_{b_1} & P_{b_1+1} & \ldots & P_{b_1+q-1} & P_{b_1+q} && & \Block{2-2}{\mathbf{\Huge 0}} &  \\
			P_{b_1-1} & P_{b_1} & & \vdots & \vdots & \ddots & & &  \\
			\vdots & \vdots & \ddots & & & & & & \\
			& & & P_{b_1} & P_{b_1+1} & & \ddots & &  \\
			\vdots & \vdots & & \vdots & P_{b_1} & & & & \\
			& & & & & \ddots & & & \\
			P_1 & P_2 & \ldots & \ldots & P_{q+1} & \ldots & P_{b_1} & \ldots & P_{b_1+q}
		\end{pNiceArray}
		$
		\caption{Matrix $\mathbf{P}_{\mbox{\tiny{SR}}} $ when $\lfloor \frac{T - v}{u} \rfloor = 1$.}
		\label{p=1}
	\end{figure*}
	\subsubsection{The decoding and relay map}
	The source node transmits messages $S[t]$ at time $t$, which has a delay profile $(4,5,5,5,5,9,9,9,9)$. The relay input is
	\begin{align*}
		R[t] &=\Big(r_1[t],r_2[t],r_3[t],r_4[t],r_5[t],r_6[t],r_7[t],r_8[t],r_9[t]\Big)\\
		&=\Big(s_1[t-4],s_2[t-5],s_3[t-5],s_4[t-5],
		\\ & \qquad s_5[t-5],s_6[t-9],s_7[t-9]+s_5[t-6],
		\\ &\qquad r_8[t-9]+s_5[t-7],r_9[t-9]\Big),
	\end{align*}
	which has an extended delay profile $(7,7,7,7,6,3,3,3,3)$.
	
	The construction of $R[t]$ is well-defined since the relay can recover all required source symbols by time $t$
	$s_1[t-4]$, $s_2[t-5]$, $s_3[t-5]$, $s_4[t-5]$, $s_5[t-5]$, $s_6[t-9]$, $s_7[t-9]$, $s_8[t-9]$, and $s_9[t-9]$. 
	Additionally, the symbols $s_5[t-6]$ and $s_5[t-7]$ are available as they were received at earlier time instances $t-1$ and $t-2$, respectively. 
	Therefore, the relay node can successfully construct $R[t]$ at time $t$.


{Delay Analysis}
\begin{itemize}
	\item {Nodes 1-6 and 9:} For $i\in[6]$ and $i=9$, the symbols $r_i[t]$ are successfully recovered by the relay node at time $t$, as ensured by the delay profile of the SR link. The resulting  delay satisfies the constraint of $T=12$.
	
	\item {Node 7:} At the destination node, the received combination $r_7[t] + r_5[t-1]$ is processed as follows
	\[
	s_7[t-9] + s_5[t-6] + s_5[t-6] = s_7[t-9].
	\]
The symbol $s_7[t-9]$ is recovered at time $t+3$, satisfying the delay constraint of 12.
	
	\item {Node 8:} Similarly, for node 8, the operation $r_8[t] + r_5[t-2]$ yields
	\[
	s_8[t-9] + s_5[t-7] + s_5[t-7] = s_8[t-9].
	\]
	The symbol $s_8[t-9]$ is also recovered at time $t+3$, satisfying the delay constraint of 12.
\end{itemize}

\section{Analysis of Achievability }

In Example 1, a code has been constructed with parameters $b_1 = 3$, $b_2 = 4$ for both $T = 7$ and all $T \geq 10$. However, no code construction can achieve the optimal rate for $T = 8$ or $T = 9$. Similarly, in Example 3, a code with parameters $b_1 = 2$, $b_2 = 3$ has been constructed for both $T = 7$ and all $T \geq 10$, yet no such construction attains the optimal rate  when $T = 6$.

These examples suggest that, in certain scenarios, the optimal rate of   $(b_1, b_2, T)$ TBSCs is not achievable. Although a complete proof remains elusive, we have observed its validity through analysis.

When $b_1 < b_2$, we construct the encoding matrices $\mathbf{P}_{\mbox{\tiny SR}}$ and $\mathbf{P}_{\mbox{\tiny RD}}$ by substituting parameters $(b_1, T-b_2)$ and $(b_2, T-b_1)$ into equation (\ref{eq-matrixP}), respectively. This gives $\mathbf{P}_{\mbox{\tiny SR}} = (P_{b+j-i})_{i \in [1,b_1], j \in [1,T-b_2]}$, where each $P_l \in \mathbb{F}^{(T-b_1) \times b_2}$ and $P_l = 0$ for $l > T-b_2$. Similarly, we have $\mathbf{P}_{\mbox{\tiny RD}} = (P'_{b+j-i})_{i \in [1,b_2], j \in [1,T-b_1]}$, where each $P'_l \in \mathbb{F}^{(T-b_1) \times b_2}$ and $P'_l = 0$ for $l > T-b_1$.

Then we get the matrix $\mathbf{P}_{\mbox{\tiny{RD}}}$ is a square matrix of size $(T - b_1)b_2 \times (T - b_1)b_2$. To ensure recovery of all lost packets, we establish the following lemma

\begin{lemma}
	A SC with parameters $(b_1, b_2, T)$ achieving the optimal rate requires the matrix $P'_{T - b_1}$ to have full column rank.
\end{lemma}

\begin{proof}
	Viewing the RD link as a point-to-point SC with parameters $(b_2, T - b_1)$, the parity messages at the relay satisfy the relation from equation \eqref{eq-recovery}
	\begin{equation}
		\begin{split}
			(\bar{P}[t + b_2], \bar{P}[t + b_2 + 1], \ldots, \bar{P}[t + b_2 + T - b_1 - 1]) \\
			= (S[t], S[t+1], \ldots, S[t + b_2 - 1])  \mathbf{P}_{\mbox{\tiny{RD}}}.
		\end{split}
	\end{equation}
Since $\mathbf{P}{\mbox{\tiny{RD}}}$ is square, recovery of the erased messages $S[t], S[t+1], \ldots, S[t + b_2 - 1]$ is equivalent to the invertibility of $\mathbf{P}{\mbox{\tiny{RD}}}$. Considering the block structure of $\mathbf{P}{\mbox{\tiny{RD}}}$, the $(T-b_1+1)$-th column block has $P'_{T - b_1}$ only in the $b_2$-th row block, with all other row blocks being zero matrices. Therefore, for $\mathbf{P}{\mbox{\tiny{RD}}}$ to be invertible, $P'_{T - b_1}$ must possess full column rank.
\end{proof}

As $P'_{T - b_1}$ has full column rank, the RD link effectively carries $b_2$ message symbols under a delay constraint of $T - b_1$. Consequently, in the RD link, there must be at least $b_2$ message symbols that are constrained to a delay of exactly $b_1$. Suppose we define
\[
P_{b_1} = 
\begin{bmatrix}
	I_{b_2} \\
	\mathbf{0}_{(T - b_2 - b_1) }
\end{bmatrix}.
\]
Denote $T-b_1=pb_2+q$ where $0\leq q<b_2$. 
\begin{theorem}
	Consider a $(b_1, b_2, T)$ TBSC  with parameters satisfying $0 < T - b_1 - b_2 < b_1$ and $b_1 < b_2$. Then, no code construction can achieve the optimal rate under the convolutional code framework.
\end{theorem}

\begin{proof}
	
Under these conditions, let $T = b_1 + b_2 + q$ with $0 < q < b_1$. The construction presented in the previous section achieves this bound when $b_2 \leq q \leq b_1$. For the case where $0 < q < b_1$, we examine the configuration shown in Fig.~\ref{p=1}.
	
	If the matrices $P'_{b_2 + 1}, \ldots, P'_{b_2 + q_2}$ are zero, then the last $q$ symbols of the message $S[i]$ for $1 \leq i \leq b$ cannot be recovered within the required delay $T - b_2 = b_1 + q$. Consequently, the overall delay would exceed $T$.
	
	If the matrices $P'_{b_2 + 1}, \ldots, P'_{b_2 + q_2}$ are nonzero, each column of the matrix  $\mathbf{P}_{\mbox{\tiny{SR}}}$ is regarded as a combined information transmission, in a manner analogous to the extended delay profile. However,  the last $q$ symbols of $S[i]$ or $S[i] + \sum\limits_{j < i}S[i]$ remain unrecoverable within a delay of $b_1 + q$. Matrix $P_{b_1}$ is nonzero and lies below matrix $P_l$ for $b_1 + 1 \leq l \leq b_1 + q$. Since the messages recovered from time $b_1 + 1$ to $b_1 + q$ will include contributions of the form $\sum_{j > i} S[j]$, which contradicts Lemma~\ref{lemma:relay-decoding}. Thereby preventing the recovery of $S[i]$ at the destination within the delay constraint $T$.
	
\end{proof}

The case for $b_1 > b_2$ follows by symmetry from the analysis of $b_1 < b_2$. Consequently, the optimal rate is unattainable under the constraint $0 < T - u - v < v$ within the convolutional code framework. However, this limitation may be overcome with an adaptive strategy. 
\section{CONCLUSION}

This paper proposes a construction that achieves the optimal rate under the constraint \(
\frac{T-u-v}{2u-v} \leq \left\lfloor \frac{T-u-v}{u} \right\rfloor,
\) by introducing an extended delay profile. Additionally, we have only provided a proof sketch demonstrating that the  optimal rate is unachievable under the condition $0 < T - u - v < v$ in the convolutional code framework and non-adaptive case.



\bibliographystyle{IEEEtran}
\bibliography{references}
\end{document}
% 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{7870757}
A.~Badr, A.~Khisti, W.-T. Tan, and J.~Apostolopoulos, ``Perfecting protection
  for interactive multimedia: A survey of forward error correction for
  low-delay interactive applications,'' \emph{IEEE Signal Processing Magazine},
  vol.~34, no.~2, pp. 95--113, 2017.

\bibitem{ChinaUnicomURLLC}
\BIBentryALTinterwordspacing
{China Unicom}, ``{5G URLLC Technology White Paper},'' Online, 2022, version
  3.0. [Online]. Available: \url{https://www.digitalelite.cn/h-nd-4033.html}
\BIBentrySTDinterwordspacing

\bibitem{1337126}
E.~Martinian and C.-E. Sundberg, ``Burst erasure correction codes with low
  decoding delay,'' \emph{IEEE Transactions on Information Theory}, vol.~50,
  no.~10, pp. 2494--2502, 2004.

\bibitem{7593328}
A.~Badr, P.~Patil, A.~Khisti, W.-T. Tan, and J.~Apostolopoulos, ``Layered
  constructions for low-delay streaming codes,'' \emph{IEEE Transactions on
  Information Theory}, vol.~63, no.~1, pp. 111--141, 2017.

\bibitem{9047134}
M.~N. Krishnan, D.~Shukla, and P.~V. Kumar, ``Rate-optimal streaming codes for
  channels with burst and random erasures,'' \emph{IEEE Transactions on
  Information Theory}, vol.~66, no.~8, pp. 4869--4891, 2020.

\bibitem{8621051}
S.~L. Fong, A.~Khisti, B.~Li, W.-T. Tan, X.~Zhu, and J.~Apostolopoulos,
  ``Optimal streaming codes for channels with burst and arbitrary erasures,''
  \emph{IEEE Transactions on Information Theory}, vol.~65, no.~7, pp.
  4274--4292, 2019.

\bibitem{9578987}
E.~Domanovitz, S.~L. Fong, and A.~Khisti, ``An explicit rate-optimal streaming
  code for channels with burst and arbitrary erasures,'' \emph{IEEE
  Transactions on Information Theory}, vol.~68, no.~1, pp. 47--65, 2022.

\bibitem{8917664}
M.~Nikhil~Krishnan, V.~Ramkumar, M.~Vajha, and P.~Vijay~Kumar, ``Simple
  streaming codes for reliable, low-latency communication,'' \emph{IEEE
  Communications Letters}, vol.~24, no.~2, pp. 249--253, 2020.

\bibitem{4557355}
E.~Martinian and M.~Trott, ``Delay-optimal burst erasure code construction,''
  in \emph{2007 IEEE International Symposium on Information Theory}, 2007, pp.
  1006--1010.

\bibitem{6567095}
A.~Badr, A.~Khisti, W.-T. Tan, and J.~Apostolopoulos, ``Streaming codes for
  channels with burst and isolated erasures,'' in \emph{2013 Proceedings IEEE
  INFOCOM}, 2013, pp. 2850--2858.

\bibitem{6620379}
D.~Leong, A.~Qureshi, and T.~Ho, ``On coding for real-time streaming under
  packet erasures,'' in \emph{2013 IEEE International Symposium on Information
  Theory}, 2013, pp. 1012--1016.

\bibitem{7839973}
N.~Adler and Y.~Cassuto, ``Burst-erasure correcting codes with optimal average
  delay,'' \emph{IEEE Transactions on Information Theory}, vol.~63, no.~5, pp.
  2848--2865, 2017.

\bibitem{9611442}
S.~Bhatnagar, B.~Chakraborti, and P.~V. Kumar, ``Streaming codes for handling a
  combination of burst and random erasures,'' in \emph{2021 IEEE Information
  Theory Workshop (ITW)}, 2021, pp. 1--6.

\bibitem{6129370}
C.~Kachris and I.~Tomkos, ``A survey on optical interconnects for data
  centers,'' \emph{IEEE Communications Surveys \& Tutorials}, vol.~14, no.~4,
  pp. 1021--1036, 2012.

\bibitem{8835153}
S.~L. Fong, A.~Khisti, B.~Li, W.-T. Tan, X.~Zhu, and J.~Apostolopoulos,
  ``Optimal streaming erasure codes over the three-node relay network,''
  \emph{IEEE Transactions on Information Theory}, vol.~66, no.~5, pp.
  2696--2712, 2020.

\bibitem{10064107}
G.~K. Facenda, M.~N. Krishnan, E.~Domanovitz, S.~L. Fong, A.~Khisti, W.-T. Tan,
  and J.~Apostolopoulos, ``Adaptive relaying for streaming erasure codes in a
  three node relay network,'' \emph{IEEE Transactions on Information Theory},
  vol.~69, no.~7, pp. 4345--4360, 2023.

\bibitem{9834645}
S.~Singhvi, R.~Gayathri, and P.~V. Kumar, ``Rate-optimal streaming codes over
  the three-node decode-and-forward relay network,'' in \emph{2022 IEEE
  International Symposium on Information Theory (ISIT)}, 2022, pp. 1957--1962.

\bibitem{10764771}
V.~Ramkumar, M.~Vajha, and M.~Nikhil~Krishnan, ``Streaming codes for three-node
  relay networks with burst erasures,'' \emph{IEEE Transactions on Information
  Theory}, vol.~71, no.~1, pp. 348--359, 2025.

\bibitem{li2025rateoptimalstreamingcodesthreenode}
\BIBentryALTinterwordspacing
Z.~Li, W.~Ma, and Z.~Zhang, ``Rate-optimal streaming codes over three-node
  relay networks with burst erasures,'' 2025. [Online]. Available:
  \url{https://arxiv.org/abs/2509.06912}
\BIBentrySTDinterwordspacing

\bibitem{9814088}
E.~Domanovitz, A.~Khisti, W.-T. Tan, X.~Zhu, and J.~Apostolopoulos,
  ``State-dependent symbol-wise decode and forward codes over multihop relay
  networks,'' \emph{IEEE Transactions on Information Theory}, vol.~68, no.~11,
  pp. 7077--7095, 2022.

\bibitem{9916294}
G.~K. Facenda, E.~Domanovitz, A.~Khisti, W.-T. Tan, and J.~Apostolopoulos,
  ``Streaming erasure codes over multi-access relayed networks,'' \emph{IEEE
  Transactions on Information Theory}, vol.~69, no.~2, pp. 860--885, 2023.

\bibitem{MaISIT25}
W.~Ma and Z.~Zhang, ``Binary rate-optimal streaming codes with minimal code
  length,'' in \emph{2025 IEEE International Symposium on Information Theory},
  2025.

\end{thebibliography}