Denote x_{\xi}=\sum_{i=1}^n\xi_ix_i(0), x(t)=(x_1(t),\cdots,x_n(t))^T, and H(x(t))=(h(x_1(t)),
\cdots,h(x_n(t)))^T. Then equations (?) can be rewritten in the compact form as
\begin{aligned}
\dot{x}(t)=-LH(x(t))
\end{aligned}
Define a function as:
\begin{aligned}
V(x(t))=\sum\limits_{i=1}^n\xi_i\int_{0}^{x_i(t)}h(s)ds
\end{aligned}
Obviously, V(x(t))\geq 0 is radially unbounded, and V(x(t))=0 if and only if x(t)=0.
Denote B=(b_{ij})=(\Xi L+L^T\Xi)/2, where \Xi=\mathrm{diag}(\xi). It is easy to check that B is a symmetric matrix with zero row-sum, i.e., B can be regarded as a graph Laplacian of a bi-graph. Differentiating V(x(t)) and using Lemma 1, we have
\begin{aligned}
\dot{V}(x(t))&=-\sum\limits_{i=1}^n\xi_ih(x_i(t))\sum\limits_{j\in
\mathcal{N}_i}{l}_{ij}h(x_j(t))\\
&=-H(x(t))^T{\Xi L}H(x(t))=-H(x(t))^TBH(x(t))\\
&=\sum\limits_{i>j}b_{ij}(h(x_i(t))-h(x_j(t)))^2\leq 0,~~~~
(\mathrm{since}~~b_{ij}\le 0)
\end{aligned}
Thus, 0\le V(x(t))\leq V(x(0)), which implies x(t) is bounded for any t\geq 0. And the largest invariant subset (\Omega-limit set) for the equations (?) is
\begin{aligned}
\Omega=\{x:x_i(t)=x_j(t); i,j=1,\cdots,n\}
\end{aligned}
Now, we claim that for all i=1,\cdots,n, \lim\limits_{t\rightarrow\infty}x_{i}(t)=x_{\xi}.
In fact, if t_{m}\rightarrow\infty and for all i,j=1,\cdots,n, x_{i}(t_{m})\rightarrow\beta, then,
\begin{aligned}
x_{\xi}=
\lim\limits_{t\rightarrow\infty}\sum\limits_{i=1}^{n}\xi_ix_i(t_{m})
=\sum\limits_{i=1}^{n}\xi_i\beta=\beta
\end{aligned}
which means that x_{\xi} is the group decision of the consensus problem. Theorem 1 is proved completely.
Myrem. Let h(x_i(t))=\alpha x_i(t) with \alpha>0, The nonlinear function h(\cdot) becomes a linear function. therefore, Theorem 1 can be regarded as a generalization of the consensus problem under linear protocols. It also give a simple proof for the consensus problem under linear protocols, too.
Myrem. Assume (h(w_1)-h(w_2))/(w_1-w_2)\geq \alpha holds for \alpha>0 and any w_1\not=w_2\in R. In this case, we have
\begin{aligned}
\dot{V}(x(t))=\sum\limits_{i>j}b_{ij}(h(x_i(t))-h(x_j(t)))^2
\leq\alpha^2\sum\limits_{i>j}b_{ij}(x_i(t)-x_j(t))^2
\end{aligned}
and the consensus problem will be realized exponentially. Moreover, it seems that the nonlinear protocol can be realized faster than that under the linear protocol h(w)=\alpha w. Therefore, nonlinear protocols can be applied to calculate the average value of large-scale networks more effectively.