Appendix
1.1 Proof of Theorem 2.1
Proof of Theorem 2.1
The three pair-wise constraints can be rewritten as compact form: \(Ax+By+Cz-b=0\), where \( A=\left( \begin{array}{c} A_1\\ A_2 \\ 0 \end{array}\right) \), \( B=\left( \begin{array}{c} B_1\\ 0 \\ B_3 \end{array}\right) \), and \( C=\left( \begin{array}{c} 0 \\ C_2\\ C_3 \end{array}\right) \). If we let \( \lambda =\left( \begin{array}{c} \lambda _1 \\ \lambda _2\\ \lambda _3 \end{array}\right) \), then the augmented Lagrangian function of (2.3) is:
$$\begin{aligned} L(x,y,z,\lambda ,\rho )= & {} F(x)+G(y)+H(z)+\lambda ^T(Ax+By+Cz-b)\nonumber \\&+\,\frac{\rho }{2} \Vert Ax+By+Cz-b\Vert _2^2. \end{aligned}$$
(4.1)
Since the augmented Lagrangian function always has a saddle point in the finite dimension case, there exists \((\hat{x},\hat{y},\hat{z},\hat{\lambda })\) such that: \(L(\hat{x},\hat{y},\hat{z},\lambda ,\rho ) \le L(\hat{x},\hat{y},\hat{z},\hat{\lambda },\rho ) \le L(x,y,z,\hat{\lambda },\rho )\) for any \((x,y,z,\lambda )\).
The left inequality shows that: \(A\hat{x}+B\hat{y}+C\hat{z}-b=0\), which combines with the right inequality shows that:
$$\begin{aligned} F(\hat{x})+G(\hat{y})+H(\hat{z})\le & {} F(x)+G(y)+H(z)+\langle \hat{\lambda },Ax+By+Cz-b\rangle \nonumber \\&+\,\frac{\rho }{2} \Vert Ax+By+Cz-b\Vert _2^2. \end{aligned}$$
(4.2)
Let \(x=\hat{x}+t(w-\hat{x}), 0<t<1\), we have \(F(x)-F(\hat{x}) \ge t[F(w)-F(\hat{x})] \) because of the convexity of F. Since the arbitrariness of (x, y, z), let \((x,y,z)=(x,\hat{y},\hat{z})\) then we get:
$$\begin{aligned} t[F(w)-F(\hat{x})]+t\langle \hat{\lambda },A(w-\hat{x})\rangle +\frac{t^2\rho }{2}\Vert A(w-\hat{x})\Vert _2^2 \ge 0. \end{aligned}$$
(4.3)
Divide by t on both side of the inequality and let \(t \rightarrow 0\) we will get:
$$\begin{aligned} F(w)-F(\hat{x})+\langle \hat{\lambda },A(w-\hat{x})\rangle \ge 0. \end{aligned}$$
(4.4)
Let \(w=x^{(n)}\),
$$\begin{aligned} F(x^{(n)})-F(\hat{x})+\langle \hat{\lambda },A(x^{(n)}-\hat{x})\rangle , \ge 0, \end{aligned}$$
(4.5)
Similarly, we can get:
$$\begin{aligned} G(y^{(n)})-G(\hat{y})+\langle \hat{\lambda },B(y^{(n)}-\hat{y})\rangle\ge & {} 0. \end{aligned}$$
(4.6)
$$\begin{aligned} H(z^{(n)})-H(\hat{z})+\langle \hat{\lambda },C(z^{(n)}-\hat{z})\rangle\ge & {} 0, \end{aligned}$$
(4.7)
Add the three inequalities above and we will get:
$$\begin{aligned}&F(x^{(n)})+G(y^{(n)})+H(z^{(n)})-F(\hat{x})-G(\hat{y})-H(\hat{z})\nonumber \\&\quad +\,\langle \hat{\lambda },A(x^{(n)}-\hat{x})+B(y^{(n)}-\hat{y})+C(z^{(n)}-\hat{z})\rangle \ge 0. \end{aligned}$$
(4.8)
Noted the update of x in Algorithm 1, we can get:
$$\begin{aligned}&F(x)+\frac{\rho }{2}\Vert Ax+By^{(n-1)}+Cz^{(n-1)-b}\Vert _2^2\nonumber \\&\quad +\,\langle \lambda ^{(n)},Ax+By^{(n-1)}+Cz^{(n-1)}-b\rangle \nonumber \\&\ge F(x^{(n)})+\frac{\rho }{2}\Vert Ax^{(n)}+By^{(n-1)}+Cz^{(n-1)-b}\Vert _2^2\nonumber \\&\quad +\,\langle \lambda ^{(n)},Ax^{(n)}+By^{(n-1)}+Cz^{(n-1)}-b\rangle \end{aligned}$$
(4.9)
for any x. Simple derivations of the inequality show that:
$$\begin{aligned}&F(x)-F(x^{(n)})+\langle A(x-x^{(n)}),\lambda ^{(n)}+\rho (Ax^{(n)}+By^{(n-1)}+Cz^{(n-1)}-b)\rangle \nonumber \\&\quad +\,\frac{\rho }{2}\Vert A(x-x^{(n)})\Vert _2^2 \ge 0. \end{aligned}$$
(4.10)
Here we use the similar skill, i.e., let \(x=x^{(n)}+t(w-x^{(n)}), 0<t<1\) and let \( t \rightarrow 0\):
$$\begin{aligned}&F(w)-F(x^{(n)})+\langle A(w-x^{(n)}),\lambda ^{(n)}\nonumber \\&\quad +\,\rho (Ax^{(n)}+By^{(n-1)}+Cz^{(n-1)}-b)\rangle \ge 0. \end{aligned}$$
(4.11)
Letting \(w=\hat{x}\), then we can obtain that:
$$\begin{aligned} F(\hat{x})\,{-}\,F(x^{(n)})\,{+}\,\langle A(\hat{x}-x^{(n)}),\lambda ^{(n)}\,{+}\,\rho (Ax^{(n)}\,{+}\,By^{(n-1)}+Cz^{(n-1)}-b)\rangle \ge 0. \end{aligned}$$
(4.12)
Similarly, following the order of subproblems in Algorithm 1 we can get:
$$\begin{aligned}&G(\hat{y})-G(y^{(n)})+\langle B(\hat{y}-y^{(n)}),\lambda ^{(n)}+\rho (Ax^{(n)}+By^{(n)}+Cz^{(n-1)}-b)\rangle \ge 0, \quad \nonumber \\ \end{aligned}$$
(4.13)
$$\begin{aligned}&H(\hat{z})-H(z^{(n)})+\langle C(\hat{z}-z^{(n)}),\lambda ^{(n)}+\rho (Ax^{(n)}+By^{(n)}+Cz^{(n)}-b)\rangle \ge 0.\quad \nonumber \\ \end{aligned}$$
(4.14)
Add the three inequalities and use the notations \(\bar{x}^{(n)}=x^{(n)}-\hat{x}, \bar{y}^{(n)}=y^{(n)}-\hat{y}, \bar{z}^{(n)}=z^{(n)}-\hat{z}\) we can get:
$$\begin{aligned}&F(\hat{x})+G(\hat{y})+H(\hat{z})-F(x^{(n)})-G(y^{(n)})-H(z^{(n)})\nonumber \\&\quad -\,\langle \lambda ^{(n)},A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\rangle -\rho \Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2\nonumber \\&\quad +\,\rho \langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)}+C\bar{z}^{(n)}-C\bar{z}^{(n-1)},A\bar{x}^{(n)} \rangle \nonumber \\&\quad +\,\rho \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},B\bar{y}^{(n)} \rangle \ge 0. \end{aligned}$$
(4.15)
Add (4.15) to (4.8) and denote \(\bar{\lambda }^{(n)}=\lambda ^{(n)}-\lambda \):
$$\begin{aligned}&-\langle \bar{\lambda }^{(n)},A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)} \rangle -\rho \Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2\nonumber \\&+\,\rho \langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)}+C\bar{z}^{(n)}-C\bar{z}^{(n-1)},A\bar{x}^{(n)} \rangle +\rho \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},B\bar{y}^{(n)} \rangle \ge 0.\nonumber \\ \end{aligned}$$
(4.16)
According to the update of \(\lambda \) in Algorithm 1, we know that:
$$\begin{aligned} \bar{\lambda }^{(n+1)}-\bar{\lambda }^{(n)}= & {} \lambda ^{(n+1)}-\lambda ^{(n)} \nonumber \\= & {} \rho (Ax^{(n)}+By^{(n)}+C^{(n)}-b)\nonumber \\= & {} \rho (A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}). \end{aligned}$$
(4.17)
Therefore,
$$\begin{aligned} |\bar{\lambda }^{(n)}|^2-|\bar{\lambda }^{(n+1)}|^2= & {} \langle \bar{\lambda }^{(n)}-\bar{\lambda }^{(n+1)},\bar{\lambda }^{(n)}+\bar{\lambda }^{(n+1)} \rangle \nonumber \\= & {} -\rho \langle A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}, 2\bar{\lambda }^{(n)}+\rho (A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}) \rangle \nonumber \\= & {} -\rho ^2\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2\nonumber \\&-\,2\rho \langle \bar{\lambda }^{(n)},A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)} \rangle \nonumber \\\ge & {} \rho ^2\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2 \nonumber \\&-\,2\rho ^2 \langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)}+C\bar{z}^{(n)}-C\bar{z}^{(n-1)},A\bar{x}^{(n)} \rangle \nonumber \\&-\,2\rho ^2 \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},B\bar{y}^{(n)} \rangle , \end{aligned}$$
(4.18)
where the last inequality follows from (4.16).
From the definition of (4.14) we can replace \(\hat{z}\) with \(z^{(n-1)}\):
$$\begin{aligned} H(z^{(n-1)})-H(z^{(n)})+\langle C(z^{(n-1)}-z^{(n)}),\lambda ^{(n)}+\rho (Ax^{(n)}+By^{(n)}+Cz^{(n)}-b) \rangle \ge 0. \end{aligned}$$
(4.19)
On the other hand, take the \((n-1)\)th iteration of the inequality (4.14) and let \(\hat{z}=z^{(n)}\):
$$\begin{aligned}&H(z^{(n)})-H(z^{(n-1)})+\langle C(z^{(n)}-z^{(n-1)}),\lambda ^{(n-1)}+\rho (Ax^{(n-1)}\nonumber \\&\quad +By^{(n-1)}+Cz^{(n-1)}-b) \rangle \ge 0. \end{aligned}$$
(4.20)
Add the above two inequalities and make some derivations of the inequality then we can see:
$$\begin{aligned}&-\rho \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},A\bar{x}^{(n)}+B\bar{y}^{(n)} \rangle \nonumber \\&\ge \rho \Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2+\rho \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},C\bar{z}^{(n-1)} \rangle . \end{aligned}$$
(4.21)
Take (4.21) into (4.18):
$$\begin{aligned} |\bar{\lambda }^{(n)}|^2-|\bar{\lambda }^{(n+1)}|^2\ge & {} \rho ^2\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2 -2\rho ^2 \langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)},A\bar{x}^{(n)} \rangle \nonumber \\&+\,2\rho ^2\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2+2\rho ^2 \langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},C\bar{z}^{(n-1)} \rangle .\nonumber \\ \end{aligned}$$
(4.22)
Since
$$\begin{aligned}&2\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2+2\langle C\bar{z}^{(n)}-C\bar{z}^{(n-1)},C\bar{z}^{(n-1)}\rangle =\nonumber \\&\Vert C\bar{z}^{(n)}\Vert _2^2-\Vert C\bar{z}^{(n-1)}\Vert _2^2+\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2 \end{aligned}$$
(4.23)
yields:
$$\begin{aligned} |\bar{\lambda }^{(n)}|^2-|\bar{\lambda }^{(n+1)}|^2\ge & {} \rho ^2\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2\nonumber \\&+\,\rho ^2(\Vert C\bar{z}^{(n)}\Vert _2^2-\Vert C\bar{z}^{(n-1)}\Vert _2^2)\nonumber \\&+\,\rho ^2\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2\nonumber \\&-\,2\rho ^2\langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)},A\bar{x}^{(n)} \rangle . \end{aligned}$$
(4.24)
Then
$$\begin{aligned}&(|\bar{\lambda }^{(n)}|^2+\rho ^2\Vert C\bar{z}^{(n-1)}\Vert _2^2)-(|\bar{\lambda }^{(n+1)}|^2+\rho ^2\Vert C\bar{z}^{(n)}\Vert _2^2) \nonumber \\&\ge \rho ^2\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2 +\rho ^2\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2 \nonumber \\&\quad -\,2\rho ^2\langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)},A\bar{x}^{(n)}\rangle . \end{aligned}$$
(4.25)
It is easy to see that:
$$\begin{aligned}&\Vert A\bar{x}^{(n)}+B\bar{y}^{(n)}+C\bar{z}^{(n)}\Vert _2^2-2(B\bar{y}^{(n)}-B\bar{y}^{(n-1)},A\bar{x}^{(n)}) \nonumber \\ \ge&\Vert A_1\bar{x}^{(n)}+B_1\bar{y}^{(n)}\Vert _2^2-2(B_1\bar{y}^{(n)}-B_1\bar{y}^{(n-1)},A_1\bar{x}^{(n)}) \nonumber \\ \ge&\Vert A_1\bar{x}^{(n)}+B_1\bar{y}^{(n-1)}\Vert _2^2+\Vert B_1\bar{y}^{(n)}\Vert _2^2-\Vert B_1\bar{y}^{(n-1)}\Vert _2^2. \end{aligned}$$
(4.26)
Combine (4.26) with (4.25) we can get:
$$\begin{aligned}&\left( |\bar{\lambda }^{(n)}|^2+\rho ^2\Vert C\bar{z}^{(n-1)}\Vert _2^2+\rho ^2\Vert B_1\bar{y}^{(n-1)}\Vert _2^2\right) \nonumber \\&\quad -\,\left( |\bar{\lambda }^{(n+1)}|^2+\rho ^2\Vert C\bar{z}^{(n)}\Vert _2^2+\rho ^2\Vert B_1\bar{y}^{(n)}\Vert _2^2\right) \nonumber \\ \ge&\rho ^2\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2+\rho ^2\Vert A_1\bar{x}^{(n)}+B_1\bar{y}^{(n-1)}\Vert _2^2. \end{aligned}$$
(4.27)
Then the right-hand side is nonnegative, that is to say, the nonnegative sequences \( \left\{ (|\bar{\lambda }^{(n)}|^2+\rho ^2\Vert C\bar{z}^{(n-1)}\Vert _2^2+\rho ^2\Vert B_1\bar{y}^{(n-1)}\Vert _2^2)\right\} \) are decreasing. Therefore, it converges while the nonnegative sequences \(\{\Vert C\bar{z}^{(n)}-C\bar{z}^{(n-1)}\Vert _2^2\} \) and \(\{\Vert A_1\bar{x}^{(n)}+B_1\bar{y}^{(n-1)}\Vert _2^2\}\) converge to zero when n goes to infinity. Therefore, we have:
$$\begin{aligned} \lim _{n\rightarrow \infty } C\bar{z}^{(n)}-C\bar{z}^{(n-1)}&=\lim _{n\rightarrow \infty } A_1\bar{x}^{(n)}+B_1\bar{y}^{(n-1)}. \end{aligned}$$
(4.28)
The last two limits yield that:
$$\begin{aligned} \lim _{n\rightarrow \infty } B_1\bar{y}^{(n)}-B_1\bar{y}^{(n-1)}=\lim _{n\rightarrow \infty } A_1\bar{x}^{(n)}+B_1\bar{y}^{(n-1)}=0 \end{aligned}$$
(4.29)
which, combines with the pair-wise structure of the constraints, shows that:
$$\begin{aligned} \lim _{n\rightarrow \infty } \langle B\bar{y}^{(n)}-B\bar{y}^{(n-1)},A\bar{x}^{(n)}\rangle = \lim _{n\rightarrow \infty } \langle B_1\bar{y}^{(n)}-B_1\bar{y}^{(n-1)},A_1\bar{x}^{(n)}\rangle = 0. \end{aligned}$$
(4.30)
Bring (4.28) and (4.30) into (4.8) and (4.15) and take the limit we can get:
$$\begin{aligned}&\limsup _{n\rightarrow \infty }F(x^{(n)})+G(y^{(n)})+H(z^{(n)}) \le F(\hat{x})+G(\hat{y})+H(\hat{z})\nonumber \\&\le \liminf _{n\rightarrow \infty }F(x^{(n)})+G(y^{(n)})+H(z^{(n)}) \end{aligned}$$
(4.31)
that is,
$$\begin{aligned} \lim _{n\rightarrow \infty }F(x^{(n)})+G(y^{(n)})+H(z^{(n)}) = F(\hat{x})+G(\hat{y})+H(\hat{z}). \end{aligned}$$
(4.32)
Then \(\{(x^{(n)},y^{(n)},z^{(n)})\}\) is a minimizing sequence of objective function, and it is convergent. Next, we need to prove the sequence \(\{(x^{(n)},y^{(n)},z^{(n)},\lambda ^{(n)})\}\) converge to the KKT point.
From the above equation, the sequence \(\{(x^{(n)},y^{(n)},z^{(n)},\lambda ^{(n)})\}\) has a convergent subsequence \(\{(x^{(n_k)},y^{(n_k)},z^{(n_k)},\lambda ^{(n_k)})\}\). Letting
$$\begin{aligned} \lim _{k\rightarrow \infty }\{(x^{(n_k)},y^{(n_k)},z^{(n_k)},\lambda ^{(n_k)})\} = (\tilde{x},\tilde{y},\tilde{z},\tilde{\lambda }). \end{aligned}$$
(4.33)
Then \((\tilde{x},\tilde{y},\tilde{z},\tilde{\lambda })\) is an optimal solution of the objective function and \(\tilde{x},\tilde{y}\) and \(\tilde{z}\) satisfy the constraint of problem (2.3):
$$\begin{aligned} A\tilde{x}+B\tilde{y}+C\tilde{z}=b. \end{aligned}$$
(4.34)
Considering the following x-subproblem:
$$\begin{aligned} x^{(n_k+1)}:=\arg \min \left\{ F(x)+\frac{\rho }{2}\Vert Ax+By^{(n_k)}+Cz^{(n_k)}-b-\frac{\lambda ^{(n_k)}}{\rho }\Vert _2^2\right\} .\qquad \end{aligned}$$
(4.35)
Its optimality condition is given by
$$\begin{aligned}&A^T(\lambda ^{(n_k)}-\rho (Ax^{(n_k+1)}+By^{(n_k+1)}+Cz^{(n_k+1)}-b)) \nonumber \\&\quad +\,\rho A^T(B(y^{(n_k+1)}-y^{(n_k)})+C(z^{(n_k+1)}-z^{(n_k)}))\in \partial F(x^{(n_k+1)}). \end{aligned}$$
(4.36)
Because of \(\lambda ^{(n_k+1)}=\lambda ^{(n_k)}-\rho (Ax^{(n_k+1)}+By^{(n_k+1)}+Cz^{(n_k+1)}-b)\). Then (4.36) can be rewritten as
$$\begin{aligned} A^T\lambda ^{(n_k+1)}+\rho (A_1^TB_1(y^{(n_k+1)}-y^{(n_k)})+A_2^TC_2(z^{(n_k+1)}-z^{(n_k)}))\in \partial F(x^{(n_k+1)}). \end{aligned}$$
(4.37)
By(4.28) and (4.29), and taking the limit for both sides that means \(k\rightarrow \infty \), then we can obtain
$$\begin{aligned} A^T\tilde{\lambda }\in \partial F(\tilde{x}). \end{aligned}$$
(4.38)
The same as the y-subproblem and z-subproblem and we can get:
$$\begin{aligned}&\displaystyle B^T\lambda ^{(n_k+1)}+\rho B_3^TC_3(z^{(n_k+1)}-z^{(n_k)})\in \partial G(y^{(n_k+1)}), \end{aligned}$$
(4.39)
$$\begin{aligned}&\displaystyle C^T\lambda ^{(n_k+1)}\in \partial H(z^{(n_k+1)}). \end{aligned}$$
(4.40)
Using the property of limit, we can know that \(\tilde{x},\tilde{y},\tilde{z}\) and \(\tilde{\lambda }\) satisfy the KKT conditions of original problem (2.3):
$$\begin{aligned} \left\{ \begin{array}{l} A^T\tilde{\lambda }\in \partial G(\tilde{x}),\\ B^T\tilde{\lambda }\in \partial G(\tilde{y}),\\ C^T\tilde{\lambda }\in \partial G(\tilde{z}),\\ A\tilde{x}+B\tilde{y}+C\tilde{z}=b. \end{array}\right. \end{aligned}$$
(4.41)
Therefore, the sequence generated by Algorithm 1 converges to the KKT point \((\tilde{x},\tilde{y},\tilde{z},\tilde{\lambda })\). The proof is done. \(\square \)
1.2 Lemma 1
Lemma 1
The pair-wise linear constraints in (2.1) or (2.2) have the following property:
$$\begin{aligned}&\left\| \sum _{i=1}^mA_i\bar{x}_i^{(n)}\right\| _2^2-2\sum _{i=1}^{m-2} \langle A_ix_i^{(n)},\sum _{j=i+1}^{m-1} (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)}) \rangle \nonumber \\&\ge \sum _{i=1}^{m-2}\sum _{j=i+1}^{m-1} \Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2+\Vert A_j^{(i,j)}\bar{x}_j^{(n)}\Vert _2^2-\Vert A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2.\nonumber \\ \end{aligned}$$
(4.42)
Proof
According to the pair-wise structure of the constraints, the L2-norm can be separated to match the inner product terms:
$$\begin{aligned}&\Vert \sum _{i=1}^mA_i\bar{x}_i^{(n)}\Vert _2^2 = \sum _{i=1}^{m-1}\sum _{j=i+1}^{m} \Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n)}\Vert _2^2 \nonumber \\&\ge \sum _{i=1}^{m-2}\sum _{j=i+1}^{m-1} \Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n)}\Vert _2^2 \nonumber \\&\langle A_ix_i^{(n)},(A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)}) \rangle =\langle A_i^{(i,j)}x_i^{(n)},(A_j\bar{x}_j^{(n)}-A_j^{(i,j)}\bar{x}_j^{(n-1)}) \rangle . \end{aligned}$$
(4.43)
For each (i, j), we have:
$$\begin{aligned}&\Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n)}\Vert _2^2 -2\langle A_i^{(i,j)}x_i^{(n)},(A_j\bar{x}_j^{(n)}-A_j^{(i,j)}\bar{x}_j^{(n-1)}) \rangle \nonumber \\&=\Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2+\Vert A_i^{(i,j)}x_i^{(n)}\Vert _2^2-\Vert A_i^{(i,j)}x_i^{(n-1)}\Vert _2^2. \end{aligned}$$
(4.44)
Combining the last two equations will lead to the inequality we want. \(\square \)
1.3 Proof of Theorem 2.2
Proof of Theorem 2.2
Similar to three-variable case, the augmented Lagrangian function of problem (2.9) always has a saddle point \((\bar{x}_1,\bar{x}_2,\ldots ,\bar{x}_m,\bar{\lambda })\) such that for any \((x_1,x_2,\ldots ,x_m,\lambda )\)
$$\begin{aligned} L(\bar{x}_1,\bar{x}_2,\ldots ,\bar{x}_m,\lambda ) \le L(\bar{x}_1,\bar{x}_2,\ldots ,\bar{x}_m,\bar{\lambda }) \le L(x_1,x_2,\ldots ,x_m,\bar{\lambda }). \end{aligned}$$
(4.45)
The same derivation in the proof of Theorem 2.1 shows that:
$$\begin{aligned}&\sum _{i=1}^m A_i\bar{x}_i-b=0, \end{aligned}$$
(4.46)
$$\begin{aligned}&\sum _{i=1}^m F_i(x_i^{(n)})- \sum _{i=1}^m F_i(\bar{x}_i)+\langle \bar{\lambda },\sum _{i=1}^m A_i\bar{x}_i^{(n)}\rangle \ge 0, \end{aligned}$$
(4.47)
where \(\bar{x}_i^{(n)}=x_i^{(n)}-\bar{x}_i, i=1,\ldots ,m\).
On the other hand, according to the order of subproblems, we have: (for any \(x_i\))
$$\begin{aligned}&F_i(x_i)-F_i(x_i^{(n)}) - \langle A_i\bar{x}_i^{(n)}, \lambda ^{(n)}+\rho (A_1x_1^{(n)}+ \nonumber \\&\cdots +\, A_ix_i^{(n)}+A_{i+1}x_{i+1}^{(n-1)}+\cdots +A_mx_m^{(n-1)}-b)\rangle \ge 0 \ \ i=1,\ldots ,m.\qquad \qquad \end{aligned}$$
(4.48)
Add all m inequalities, let \(x_i=\bar{x}_i\) and make some simplifications we will have:
$$\begin{aligned}&\sum _{i=1}^m F_i(\bar{x}_i)- \sum _{i=1}^m F_i(x_i^{(n)})-\langle \lambda ^{(n)},\sum _{i=1}^m A_i\bar{x}_i^{(n)}\rangle -\rho \left\| \sum _{i=1}^m A_i\bar{x}_i^{(n)}\right\| _2^2 \nonumber \\&\quad +\,\rho \sum _{i=1}^{m-1} \langle A_ix_i^{(n)},\sum _{j=i}^m (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)}) \rangle \ge 0. \end{aligned}$$
(4.49)
Add (4.49) to (4.47) and take the notation \(\bar{\lambda }^{(n)}=\lambda ^{(n)}-\bar{\lambda }\), we can see:
$$\begin{aligned}&-\langle \bar{\lambda }^{(n)},\sum _{i=1}^m A_i\bar{x}_i^{(n)}\rangle -\rho \left\| \sum _{i=1}^m A_i\bar{x}_i^{(n)}\right\| _2^2\nonumber \\&+\,\rho \sum _{i=1}^{m-1} \langle A_ix_i^{(n)},\sum _{j=i}^m (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)}) \rangle \ge 0. \end{aligned}$$
(4.50)
According to the update of Lagrangian multiplier, we have:
$$\begin{aligned} \bar{\lambda }^{(n+1)}-\bar{\lambda }^{(n)}=\lambda ^{(n+1)}-\lambda ^{(n)} =\rho \left( \sum _{i=1}^m A_ix_i^{(n)}-b\right) =\rho \sum _{i=1}^m A_i\bar{x}_i^{(n)}. \end{aligned}$$
(4.51)
Therefore,
$$\begin{aligned} |\bar{\lambda }^{(n)}|^2-|\bar{\lambda }^{(n+1)}|^2= & {} \langle \bar{\lambda }^{(n)}-\bar{\lambda }^{(n+1)},\bar{\lambda }^{(n)}+\bar{\lambda }^{(n+1)}\rangle \nonumber \\= & {} -\rho \langle \sum _{i=1}^m A_i\bar{x}_i^{(n)},2\bar{\lambda }^{(n)}+\rho \sum _{i=1}^m A_i\bar{x}_i^{(n)}\rangle \nonumber \\= & {} -\rho ^2\left\| \sum _{i=1}^m A_i\bar{x}_i^{(n)})\right\| _2^2-2\rho \langle \bar{\lambda }^{(n)},\sum _{i=1}^m A_i\bar{x}_i^{(n)}\rangle \nonumber \\\ge & {} \rho ^2\left\| \sum _{i=1}^m A_i\bar{x}_i^{(n)})\right\| _2^2\nonumber \\&-2\rho ^2\sum _{i=1}^{m-1} \langle A_ix_i^{(n)},\sum _{j=i+1}^m (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)})\rangle . \end{aligned}$$
(4.52)
Take \(x_i=\bar{x}_m\) in (4.48) we have:
$$\begin{aligned} F_m(\bar{x}_m)-F_m(x_m^{(n)})-\langle A_m\bar{x}_m^{(n)},\lambda ^{(n)}+\rho \sum _{i=1}^mA_i\bar{x}_i^{(n)}\rangle \ge 0. \end{aligned}$$
(4.53)
Let \(\bar{x}_m=x_m^{(n-1)}\) in (4.53) and take the \((n-1)\)th iteration. We will have:
$$\begin{aligned}&F_m(x_m^{(n-1)})-F_m(x_m^{(n)}) -\langle A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)},\lambda ^{(n)}+\rho \sum _{i=1}^mA_i\bar{x}_i^{(n)}\rangle \ge 0, \end{aligned}$$
(4.54)
$$\begin{aligned}&F_m(x_m^{(n)})-F_m(x_m^{(n-1)}) -\langle A_m\bar{x}_m^{(n-1)}-A_m\bar{x}_m^{(n)},\lambda ^{(n-1)}+\rho \sum _{i=1}^mA_i\bar{x}_i^{(n-1)}\rangle \ge 0.\nonumber \\ \end{aligned}$$
(4.55)
Add the two equations together and we will see:
$$\begin{aligned}&-\langle \lambda ^{(n)}-\lambda ^{(n-1)},A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)}\rangle \nonumber \\&-\,\rho \langle A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)},\sum _{i=1}^mA_i\bar{x}_i^{(n)}-\sum _{i=1}^mA_i\bar{x}_i^{(n-1)}\rangle \ge 0. \end{aligned}$$
(4.56)
Combining with the update of \(\lambda \) yields that:
$$\begin{aligned}&-\,\rho \langle \sum _{i=1}^m A_i\bar{x}_i^{(n-1)},A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)}\rangle \nonumber \\&-\,\rho \langle A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)},\sum _{i=1}^mA_i\bar{x}_i^{(n)}-\sum _{i=1}^mA_i\bar{x}_i^{(n-1)}\rangle \ge 0. \end{aligned}$$
(4.57)
Then we can see:
$$\begin{aligned}&|\bar{\lambda }^{(n)}|^2-|\bar{\lambda }^{(n+1)}|^2 \ge \rho ^2\Vert \sum _{i=1}^m A_i\bar{x}_i^{(n)})\Vert _2^2 +2\rho ^2\langle A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)},A_m\bar{x}_m^{(n)}\rangle \nonumber \\&\quad -\,2\rho ^2\sum _{i=1}^{m-2} \langle A_ix_i^{(n)},\sum _{j=i+1}^{m-1} (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)})\rangle . \end{aligned}$$
(4.58)
Take the identity
$$\begin{aligned}&2\langle A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)},A_m\bar{x}_m^{(n)}\rangle \nonumber \\&=\Vert A_m\bar{x}_m^{(n)}\Vert _2^2-\Vert A_m\bar{x}_m^{(n-1)}\Vert _2^2+\Vert A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)}\Vert _2^2 \end{aligned}$$
(4.59)
into (4.58):
$$\begin{aligned}&(|\bar{\lambda }^{(n)}|^2 +\rho ^2\Vert A_m\bar{x}_m^{(n-1)}\Vert _2^2) -(|\bar{\lambda }^{(n+1)}|^2 +\rho ^2\Vert A_m\bar{x}_m^{(n)}\Vert _2^2) \nonumber \\&\ge \rho ^2\Vert A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)}\Vert _2^2+\rho ^2\Vert \sum _{i=1}^mA_i\bar{x}_i^{(n)}\Vert _2^2 \nonumber \\&\quad -\,2\rho ^2\sum _{i=1}^{m-2} \langle A_ix_i^{(n)},\sum _{j=i+1}^{m-1} (A_j\bar{x}_j^{(n)}-A_j\bar{x}_j^{(n-1)}) \rangle . \end{aligned}$$
(4.60)
According to Lemma 1, we can get:
$$\begin{aligned}&(|\bar{\lambda }^{(n)}|^2+\rho ^2\Vert A_m\bar{x}_m^{(n-1)}\Vert _2^2+\rho ^2\sum _{i=1}^{m-2}\sum _{j=i}^{m-1} \Vert A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2) \nonumber \\&\quad -\,(|\bar{\lambda }^{(n+1)}|^2+\rho ^2\Vert A_m\bar{x}_m^{(n)}\Vert _2^2+\rho ^2\sum _{i=1}^{m-2}\sum _{j=i}^{m-1} \Vert A_j^{(i,j)}\bar{x}_j^{(n)}\Vert _2^2) \nonumber \\&\ge \rho ^2\Vert A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)}\Vert _2^2 +\rho ^2\sum _{i=1}^{m-2}\sum _{j=i+1}^{m-1} \Vert A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2.\nonumber \\ \end{aligned}$$
(4.61)
Then the nonnegative sequence
$$\begin{aligned} \left( |\bar{\lambda }^{(n)}|^2 +\rho ^2\Vert A_m\bar{x}_m^{(n-1)}\Vert _2^2+\rho ^2\sum _{i=1}^{m-2}\sum _{j=i}^{m-1} \Vert A_j^{(i,j)}\bar{x}_j^{(n-1)}\Vert _2^2\right) \end{aligned}$$
(4.62)
is decreasing and has a lower bound, which means it converges. Therefore, the nonnegative sequences of right-hand side in (4.61) have the limits 0, which yields:
$$\begin{aligned}&\lim _{n\rightarrow \infty } A_m\bar{x}_m^{(n)}-A_m\bar{x}_m^{(n-1)} =\lim _{n\rightarrow \infty } \sum _{i=1}^mA_i\bar{x}_i^{(n)} =0, \nonumber \\&\lim _{n\rightarrow \infty } A_i^{(i,j)}\bar{x}_i^{(n)}+A_j^{(i,j)}\bar{x}_j^{(n-1)}=0 \ (1 \le i<j \le m-1). \end{aligned}$$
(4.63)
Based on the structure of pair-wise constraints and the last two limits, we can get:
$$\begin{aligned} \lim _{n\rightarrow \infty } A_j^{(i,j)}\bar{x}_j^{(n-1)}-A_j^{(i,j)}\bar{x}_j^{(n-1)}=0, \ 1 \le i<j \le m-1 \end{aligned}$$
(4.64)
which means:
$$\begin{aligned} \lim _{n\rightarrow \infty } \sum _{i=1}^{m-2} \langle A_i^{(i,j)}x_i^{(n)},\sum _{j=i}^{m-1} (A_j^{(i,j)}\bar{x}_j^{(n)}-A_j^{(i,j)}\bar{x}_j^{(n-1)}) \rangle =0. \end{aligned}$$
(4.65)
Take (4.65) and the first two limits in (4.63) to (4.49) and take the superior limit:
$$\begin{aligned} \sum _{i=1}^m F_i(\bar{x}_i) \ge \limsup _{n\rightarrow \infty }\sum _{i=1}^m F_i(x_i^{(n)}). \end{aligned}$$
(4.66)
On the other hand, take the second limit in (4.63) to (4.47) and take the inferior limit:
$$\begin{aligned} \sum _{i=1}^m F_i(\bar{x}_i) \le \liminf _{n\rightarrow \infty }\sum _{i=1}^m F_i(x_i^{(n)}). \end{aligned}$$
(4.67)
Now we can say that Algorithm 2 converges to the minimum point:
$$\begin{aligned} \lim _{n\rightarrow \infty }\sum _{i=1}^m F_i(x_i^{(n)}) = \sum _{i=1}^m F_i(\bar{x}_i). \end{aligned}$$
(4.68)
Then \(\{(x_1^{(n)},x_2^{(n)},\ldots ,x_m^{(n)})\}\) is a minimizing sequence of objective function, and it is convergent. Next, we need to prove the sequence \(\{(x_1^{(n)},x_2^{(n)},\ldots ,x_m^{(n)},\lambda ^{(n)})\}\) converge to the KKT point.
From the above equation, the sequence \(\{(x_1^{(n)},x_2^{(n)},\ldots ,x_m^{(n)},\lambda ^{(n)})\}\) has a convergent subsequence \(\{(x_1^{(n_k)},x_2^{(n_k)},\ldots ,x_m^{(n_k)},\lambda ^{(n_k)})\}\). Letting
$$\begin{aligned} \lim _{k\rightarrow \infty }\{(x_1^{(n_k)},x_2^{(n_k)},\ldots ,x_m^{(n_k)},\lambda ^{(n_k)})\} = (\tilde{x}_1,\tilde{x}_2,\ldots ,\tilde{x}_m,\tilde{\lambda }). \end{aligned}$$
(4.69)
Then \((\tilde{x}_1,\tilde{x}_2,\ldots ,\tilde{x}_m,\tilde{\lambda })\) is an optimal solution of the objective function and \(x_i (i=1,2,\ldots ,m)\) satisfy the constraint of problem (2.3):
$$\begin{aligned} \sum _{i=1}^mA_i\tilde{x}_i=b. \end{aligned}$$
(4.70)
Considering the following \(x_i\)-subproblem:
$$\begin{aligned} x_i^{(n_k+1)}:=\arg \min _{x_i} \{F_i(x_i)+\frac{\rho }{2}\Vert A_ix_i+\sum _{j=1,j\ne i}^m A_jx_j^{(n_k)}-b-\frac{\lambda ^{(n_k)}}{\rho }\Vert _2^2\}.\qquad \end{aligned}$$
(4.71)
Its optimality condition is given by
$$\begin{aligned}&A_i^T\left( \lambda ^{(n_k)}-\rho (\sum _{i=1}^mA_ix_i^{(n_k+1)}-b)\right) \nonumber \\&\quad +\,\rho A_i^T\left( \sum _{j=1,j\ne i}^mA_j(x_j^{(n_k+1)}-x_j^{(n_k)})\right) \in \partial F(x^{(n_k+1)}).\qquad \end{aligned}$$
(4.72)
Because of \(\lambda ^{(n_k+1)}=\lambda ^{(n_k)}-\rho (\sum _{i=1}^mA_ix_i^{(n_k+1)}-b)\). Then (4.72) can be rewritten as
$$\begin{aligned} A_i^T\lambda ^{(n_k+1)}+\rho A_i^T\left( \sum _{j=1,j\ne i}^mA_j\left( x_j^{(n_k+1)}-x_j^{(n_k)}\right) \right) \in \partial F_i(x_i^{(n_k+1)}). \end{aligned}$$
(4.73)
By(4.63) and (4.64), and taking the limit for both sides that means \(k\rightarrow \infty \), then we can obtain
$$\begin{aligned} A_i^T\tilde{\lambda }\in \partial F_i(\tilde{x}_i). \end{aligned}$$
(4.74)
Using the same method, we can get
$$\begin{aligned} A_j^T\tilde{\lambda }\in \partial F_j(\tilde{x}_j),(j=1,2,\ldots ,m). \end{aligned}$$
(4.75)
Then we can know that \(\tilde{x}_i(i=1,\ldots ,m)\) and \(\tilde{\lambda }\) satisfy the KKT conditions of original problem (2.9):
$$\begin{aligned} \left\{ \begin{array}{l} A_i^T\tilde{\lambda }\in \partial F_j(\tilde{x}_i),(i=1,2,\ldots ,m),\\ \sum _{i=1}^mA_i\tilde{x}_i=b. \end{array}\right. \end{aligned}$$
(4.76)
Then \((\tilde{x}_1,\tilde{x}_2,\ldots ,\tilde{x}_m,\tilde{\lambda })\) is the KKT point, which completes the proof. \(\square \)