Can somebody provide an explanation of how network capacity region is calculated in this example?

64 Views Asked by At

In chapter 3 of book Stochastic Network Optimization with Application to Communication and Queueing Systems, a 2-queue wireless downlink is considered. I put the description in this figure.

I understand how to retrieve the upper bound of $\lambda_1$ and $\lambda_2$ but not sure for that of $\lambda_1 + \lambda_2$ and really don't know about $\lambda_1 + \lambda_2/2$. I think as following:

$\lambda_1 + \lambda_2 \leq 1*(P[S_1=1,S_2=0] + P[S_1=1,S_2=1]) + (P[S_1=0,S_2=1] + P[S_1=0,S_2=2]) + 1.5(P[S_1=1,S_2=2])$

However, I am still not sure if this is correct. So my question is how to calculate the network capacity region as shown in the Fig.3b of the attached image?

1

There are 1 best solutions below

5
On

The problem has two channels that vary independently over time with channel states $\{S_1(t)\}_{t=0}^{\infty}$ i.i.d. and $\{S_2(t)\}_{t=0}^{\infty}$ i.i.d. with \begin{align} &P[S_1(t)=0]=0.3, \quad P[S_1(t)=1]=0.7\\ &P[S_2(t)=0]=0.2, \quad P[S_2(t)=1]=0.5, \quad P[S_2(t)=2]=0.3 \end{align} Every step $t \in \{0, 1, 2, ...\}$ you observe $(S_1(t), S_2(t))$, which tells you the current state of both channels, and then choose to serve either channel 1 or channel 2. So the transmission variables on slot $t$ are $$ (b_1(t),b_2(t)) = \left\{ \begin{array}{ll} (S_1(t), 0) &\mbox{ if we serve channel 1 on slot $t$} \\ (0,S_2(t)) & \mbox{ if we serve channel 2 on slot $t$} \end{array} \right. \quad \quad (Eq. 1)$$

Let $\Lambda$ be the shaded region of Fig. 3.1. Let's show that the set of all expectations $E[(b_1(t),b_2(t))]$ that can be achieved on a particular slot $t$, considering all possible decision rules, does not depend on $t$, and is equal to the set $\Lambda$. The proof has an achievability part and a converse part.

Achievability:

Design stationary and randomized rules (so that under each rule, $\{(b_1(t), b_2(t))\}_{t=0}^{\infty}$ are just i.i.d. vectors) to achieve the corner points of $\Lambda$.

  • Rule 1 should achieve: $E[(b_1(t), b_2(t))] = (0.14, 1.10)$ for all $t$.

  • Rule 2 should achieve: $E[(b_1(t),b_2(t))] = (0.49, 0.75)$ for all $t$.

  • Rule 3 should achieve: $E[(b_1(t),b_2(t))] = (0.70, 0.33)$ for all $t$.

By taking probabilistic mixtures of these rules, we can design a rule to ensure $E[(b_1(t), b_2(t))]$ is any desired point on the Pareto boundary of $\Lambda$.

It is easy to design Rule 1: "Transmit on channel 2 whenever $S_2(t) \neq 0$, and otherwise transmit on channel 1." Rule 2 takes a bit more thought.

Converse:

Fix $t \in \{0, 1, 2, ...\}$ and let $(b_1(t), b_2(t))$ be decisions of any strategy that satisfies (Eq. 1). We want to show $E[(b_1(t),b_2(t))]\in \Lambda$. Recall that the expectation of a random vector is taken componentwise: $$ E[(b_1(t),b_2(t))] = (E[b_1(t)], E[b_2(t)]) $$ So we just show that $(E[b_1(t)], E[b_2(t)])$ satisfies the 6 linear inequality constraints that define the set $\Lambda$ in the figure. To do that, we show there are constants $c_1, c_2$ such that

1) $E[b_1(t)]\geq 0$.

2) $E[b_2(t)]\geq 0$.

3) $E[b_1(t)] \leq 0.7$.

4) $E[b_2(t)]\leq 1.10$.

5) $E[b_1(t)]+E[b_2(t)] \leq c_1$.

6) $2E[b_1(t)] + E[b_2(t)]\leq c_2$.

Inequalities 1-3 hold because, regardless of the decision rule, we have from (Eq. 1) that \begin{align} b_1(t) &\geq 0\\ b_2(t) &\geq 0\\ b_1(t) &\leq S_1(t) \end{align} So we take expectations of both sides to obtain Inequalities 1-3: \begin{align} E[b_1(t)]&\geq 0\\ E[b_2(t)]&\geq 0\\ E[b_1(t)]&\leq E[S_1(t)]=0.7 \end{align} Inequality 4 holds similarly (can you reprodce it?) Inequality 5 holds because from (Eq. 1) $$b_1(t)+b_2(t) \leq \max[S_1(t),S_2(t)]$$ and again we take expectations of both sides. Inequality 6 holds by a similar argument (can you reproduce it?)


If you want to directly connect this with the queueing you could, for $i \in \{1,2\}$: $$ Q_i(t+1) = \max[Q_i(t)-b_i(t),0] + a_i(t) \geq Q_i(t)-b_i(t)+a_i(t) \quad \forall t \in \{0, 1 ,2, ...\}$$ Thus $$Q_i(T)-Q_i(0) \geq \sum_{t=0}^{T-1}[a_i(t)-b_i(t)] \quad \forall T \in\{1, 2, 3, ...\}$$ Now take any service decision strategy. Assume $Q_1(0)=Q_2(0)=0$ for simplicity. We get for all positive integers $T$: \begin{align} 2Q_1(T)+Q_2(T) &\geq \sum_{t=0}^{T-1}\left[[2a_1(t)-2b_1(t)] + [a_2(t)-b_2(t)]\right]\\ &= \sum_{t=0}^{T-1}[(2a_1(t)+a_2(t)) - (2b_1(t)+b_2(t))] \end{align} Taking expectations and using $E[a_i(t)]=\lambda_i$ and $E[2b_1(t)+b_2(t)]\leq c_2$ gives $$ E[2Q_1(T)+Q_2(T)] \geq \sum_{t=0}^{T-1}[(2\lambda_1 + \lambda_2)-c_2] $$ Then $$ \frac{E[2Q_1(T)+Q_2(T)]}{T}\geq (2\lambda_1+\lambda_2) - c_2$$ If both queues are mean rate stable then the left-hand-side $\rightarrow 0$ as $T\rightarrow \infty$, which implies $2\lambda_1+\lambda_2 \leq c_2$. Therefore the inequality $2\lambda_1+\lambda_2\leq c_2$ is a necessary condition for both queues to be mean rate stable. If this condition is violated, it is impossible for any service strategy to stabilize both queues.