Appendix
1.1 Proof of Lemma 1
Proof
We only prove that the set Λ
q
p
is non-empty since the corresponding result for the set Λ
*q
p
can be shown similarly.
Let us define Λ
q+
p
as the subset of Λ
q
p
with additional constraint λ
j
i
≥ 0:
$$ \Uplambda_p^{q+}=\left\{\sum_{i\in M_p^q(\alpha,q)}\lambda_i^q\hbox{x}_i^q+\sum_{i\in M_p^q(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}\in \Uplambda_p^q, \lambda_i^j\geq 0,j=q,q+1\right\}. $$
(42)
Next we shall show that the set Λ
q
p
is non-empty by proving the subset Λ
q+
p
≠ ∅. In order to prove that \(\Uplambda_p^{q+}\neq \varnothing\) , it is sufficient to prove that there exists a vector λ such that
$$ \lambda_i^q=\mu{\frac{C-\alpha_i^q}{\alpha_p^q}},\quad i\in M_p^q(\alpha,q), $$
(43)
$$ \lambda_i^{q+1}=\mu{\frac{\alpha_i^{q+1}}{\alpha_p^q}},\quad i\in M_p^q(\alpha^*,q+1), $$
(44)
$$ 0\leq \mu\leq 1, $$
(45)
$$ \sum_{i\in M_p^q (\alpha,q)}\lambda_i^q+\sum_{i\in M_p^q(\alpha^*,q+1)}\lambda_i^{q+1}=1, $$
(46)
because it is straightforward to show that when a vector λ satisfies (43)–(46), we have
$$ \sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^q\hbox{x}_i^q+\sum\limits_{i\in M_p^q(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}\in \Uplambda_p^{q+}, $$
(47)
and therefore Λ
q+
p
≠ ∅. Now we prove that the vector λ satisfying (43)–(46) exists. Taking into account equations (43) and (44), we rewrite constraint (46) as
$$ 1={\frac{\mu}{\alpha_p^q}}\left[\sum_{i\in M_p^q(\alpha,q)}(C-\alpha_i^q)+\sum_{i\in M_p^q(\alpha^*,q+1)}\alpha_i^{q+1}\right]. $$
(48)
Thus, it is sufficient to show that the value of μ given by equation (48) satisfies constraint (45). For this purpose, define
$$ \Updelta=\sum_{i\in M(\alpha,q)}(C-\alpha_i^q)+\sum_{i\in M(\alpha^*,q+1)}\alpha_i^{q+1}, $$
(49)
$$ =\sum_{i\in M(\alpha,q)}C-\sum_{i\in M(\alpha,q)}\alpha_i^q+\sum_{i\in M(\alpha^*,q+1)}\alpha_i^{q+1}. $$
(50)
Noting that
$$ \begin{aligned} 0&=\sum_{i\in M(\alpha^*,q+1)}\alpha_i^{q+1}+\sum_{i\in N(\alpha^*,q+1)}\alpha_i^{q+1}-\sum_{i\in M(\alpha,q)}\alpha_i^{q}-\sum_{i\in N(\alpha,q)}\alpha_i^{q}\\ &=\sum_{i\in M(\alpha^*,q+1)}\alpha_i^{q+1}-\sum_{i\in M(\alpha,q)}\alpha_i^{q}+\sum_{i\in N(\alpha^*,q+1)}C-\sum_{i\in N(\alpha,q)}C, \end{aligned} $$
(51)
and combining equation (50) and (51), we get
$$ \Updelta=\sum_{i\in M(\alpha,q)}C-\sum_{i\in N(\alpha^*,q+1)}C+\sum_{i\in N(\alpha,q)}C=C\tau, $$
(52)
where τ is a integer.
Since equation (49) gives Δ > 0, we have
$$ \Updelta\geq C. $$
(53)
Rewrite equation (48) as
$$ 1={\frac{\mu}{\alpha_p^q}}(\Updelta-(C-\alpha_p^q)), $$
or
$$ \mu={\frac{\alpha_p^q}{\Updelta-(C-\alpha_p^q)}}. $$
Taking into account inequality (53), we finally get 0 ≤ μ ≤ 1. Thus the set Λ
q+
p
is not empty. Namely, the set Λ
q
p
is not empty. \(\square\)
1.2 Proofs of Lemma 2 and Lemma 3
We only give the proof of Lemma 2 since the proof of Lemma 3 is similar.
Proof
Recalling (10), the vector w can be expressed as
$$ \hbox{w}=\sum_{j\neq q}\sum_{i=1}^{l^j}(\alpha_i^{*j}-\alpha_i^j)\hbox{x}_i^j+\sum_{i\neq q}(\alpha_i^{*q}-\alpha_i^q)+(\alpha^{*q}_p-\alpha_p^q)\hbox{x}_p^q. $$
(54)
Now we endeavor to replace support vector x
q
p
with a linear combination of the remaining margin support vectors about α in the qth class points and margin support vectors about α* in the q + 1th class points; this gives:
$$ \hbox{x}_p^q\approx \sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^q\hbox{x}_i^q+\sum\limits_{i\in M_p^{q}(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}=\tilde{\hbox{x}}_p^q. $$
Taking this replacement yields an approximate expression for w
$$ \begin{aligned} \tilde{\hbox{w}}&=\sum_{j\neq q}\sum_{i=1}^{l^j}(\alpha_i^{*j}-\alpha_i^j)\hbox{x}_i^j+\sum_{i\neq q}(\alpha_i^{*q}-\alpha_i^q)+(\alpha^{*q}_p-\alpha_p^q)\left[\sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^q\hbox{x}_i^q+\sum\limits_{i\in M_p^{q}(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}\right], \\ &=\sum_{j\neq q,q+1}\sum_{i=1}^{l^j}(\alpha_i^{*j}-\alpha_i^j)\hbox{x}_i^j+\sum_{i\notin M_p^q(\alpha,q),i\neq p}(\alpha_i^{*q}-\alpha_i^{q})\hbox{x}_i^{q}+\sum_{i\notin M_p^q(\alpha^*,q+1)}(\alpha_i^{*q+1}-\alpha_i^{q+1})\hbox{x}_i^{q+1}\\ &\quad+\sum_{i\in M _p^q(\alpha,q)}\left[\underbrace{(\alpha_i^{*q}-\alpha_i^q)+(\alpha_p^{*q}-\alpha_p^q)\lambda_i^q}_{\tilde{\alpha}_i^{*q}-\tilde{\alpha}_i^q}\right]\hbox{x}_i^q +\sum_{i\in M _p^q(\alpha^*,q+1)}\left[\underbrace{(\alpha_i^{*q+1}-\alpha_i^{q+1})+(\alpha_p^{*q}-\alpha_p^{q})\lambda_i^{q+1}}_{\tilde{\alpha}_i^{*q+1}-\tilde{\alpha}_i^{q+1}}\right]\hbox{x}_i^{q+1}, \end{aligned} $$
Setting
$$ \begin{aligned} \,&\tilde{\alpha}_i^{q}=\alpha_i^{q}+\lambda_i^{q}\alpha_p^{q},\ \tilde{\alpha}_i^{*q}=\alpha_i^{*q}+\lambda_i^{q}\alpha_p^{*q},\quad i\in M_p^q(\alpha,q), \\ \,&\tilde{\alpha}_i^{q+1}=\alpha_i^{q+1}-\lambda_i^{q+1}\alpha_p^{*q},\quad \tilde{\alpha}_i^{*q+1}=\alpha_i^{*q+1}-\lambda_i^{q+1}\alpha_p^{q},\quad i\in M_p^q(\alpha^*,q+1), \\ \,&\tilde{\alpha}_i^q=\alpha_i^q,\tilde{\alpha}_i^{*q}=\alpha_i^{*q},\quad i\notin M_p^q(\alpha,q),\\ \,&\tilde{\alpha}_i^{q+1}=\alpha_i^{q+1},\tilde{\alpha}_i^{*q+1}=\alpha_i^{*q+1},\quad i\notin M_p^q(\alpha^*,q+1), \\ \,&\tilde{\alpha}_i^j=\alpha_i^j,\tilde{\alpha}_i^{*j}=\alpha_i^{*j},\quad j=1,\ldots,q-1,q+2,\ldots,k,\quad i=1,\ldots,l^j, \end{aligned} $$
we have
$$ \begin{aligned} &0\leq \alpha_i^{q}+\lambda_i^q\alpha_p^{q}\leq C,\quad 0\leq \alpha_i^{*q}+\lambda_i^{q}\alpha_p^{*q}\leq C, \quad i\in M _p^q(\alpha,q), \\ &0\leq \alpha_i^{q+1}-\lambda_i^{q+1}\alpha_p^{*q}\leq C,\quad 0\leq\alpha_i^{*q+1}-\lambda_i^{q+1}\alpha_p^{q}\leq C ,\quad i\in M _p^q(\alpha^*,q+1), \\ &\sum_{i\in M _p^q(\alpha,q)}\tilde{\alpha}_i^q+\sum_{i\in N _p^q(\alpha,q)}\alpha_i^q=\sum_{i\in M _p^q(\alpha^*,q+1)}\tilde{\alpha}_i^{*q+1}+\sum_{i\in N _p^q(\alpha^*,q+1)}\alpha_i^{*q+1}. \end{aligned} $$
The above equalities imply that
$$ \sum_{i\in V_p^q(\alpha,q)}\alpha_i^{q}+\alpha_p^q\sum_{i\in M_p^q(\alpha,q)}\lambda_i^{q}=\sum_{i\in V _p^q(\alpha^*,q+1)}\alpha_i^{*q+1}-\alpha_p^{q}\sum_{i\in M _p^q(\alpha^*,q+1)}\lambda_i^{q+1}. $$
(55)
According to the constraint (8) of the dual problem (7)–(9) again, we get
$$ \sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^{q}+\sum\limits_{i\in M_p^q(\alpha^*,q+1)}\lambda_i^{q+1}=1, \lambda_p^q=-1. $$
(56)
So, by (55) and (56), \(\tilde{\alpha}^{(*)}\) is a feasible solution of dual problem (7)–(9) for the training set \({T_p^q=T\setminus \{(\hbox{x}_p^q,y_p^q)\}.}\)
\(\square\)
1.3 Proof of Lemma 4
Proof
1. For being left out the point x
q
p
, we consider its margin support vector about α. Consider the following optimization problem
$$ W(\alpha^{(*)}):=\max_{\alpha^{(*)}}\quad\sum_{j,i}(\alpha_i^j+\alpha_i^{*j})-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\alpha_i^{*j}-\alpha_i^j)(\alpha_{i'}^{*j'}-\alpha_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}), $$
(57)
$$ \hbox{s.t.}\quad\sum_{i=1}^{l^j}\alpha_i^j=\sum_{i=1}^{l^{j+1}}\alpha_i^{*j+1},\quad j=1,2,\ldots,k-1, $$
(58)
$$0\leq \alpha_i^j,\alpha_i^{*j}\leq C,\quad j=1,\ldots,k,\quad i=1,\ldots,l^j, $$
(59)
$$\alpha_p^q=\alpha_p^{*q}=0. $$
(60)
Assuming that α
(*)q
p
is the optimal solution of the optimization problem (57)–(60) and α(*) is the optimal solution of the dual problem (7)–(9), then the following inequalities hold:
$$ W(\alpha_p^{(*)q})\geq W(\alpha^{(*)}-\delta^{(*)}), $$
(61)
$$ W(\alpha^{(*)})\geq W(\alpha_p^{(*)q}+\gamma^{(*)}), $$
(62)
where δ(*) satisfies the following conditions:
$$ \begin{aligned} \,&0\leq \alpha_i^j-\delta_i^j\leq C, \quad 0\leq \alpha_i^{*j}-\delta_i^{*j}\leq C, \quad j=1,\ldots,k,\quad i=1,\ldots,l^j, \\ \,&\sum_{i=1}^{l^j}\delta_i^j=\sum_{i=1}^{l^{j+1}}\delta_i^{*j+1},\quad j=1,2,\ldots,k-1, \\ &\delta_i^{*1}=0,\quad i=1,2,\ldots,l^1,\delta_i^k=0,\quad i=1,2,\ldots,l^k, \\ \,&\delta_p^q=\alpha_p^q,\quad \delta_p^{*q}=\alpha_p^{*q}, \end{aligned} $$
and γ(*) satisfies the following conditions:
$$ 0\leq \alpha_{pi}^{qj}+\gamma_i^j\leq C,\quad 0\leq \alpha_{pi}^{*qj}+\gamma_i^{*j}\leq C,\quad j=1,\ldots,k,\quad i=1,\ldots,l^j, $$
(63)
$$ \sum_{i=1}^{l^q}\gamma_i^q=\sum_{i=1}^{l^{q+1}}\gamma_i^{*q+1}, $$
(64)
$$ \alpha_i^q=0\Rightarrow \gamma_i^q=0,\quad i\neq p,\quad \alpha_i^{*q+1}=0\Rightarrow \gamma_i^{*q+1}=0, $$
(65)
$$ \gamma_i^j=0,j\neq q,i=1,\ldots,l^j, \quad \gamma_i^{*j}=0,\quad j\neq q+1,\quad i=1,\ldots,l^j, $$
(66)
$$ \gamma_i^{*1}=0,\quad i=1,2,\ldots,l^1,\quad \gamma_i^k=0,\quad i=1,2,\ldots,l^k. $$
(67)
From inequality (61) we obtain
$$ W(\alpha^{(*)})-W(\alpha_p^{(*)q})\leq W(\alpha^{(*)})-W(\alpha^{(*)}-\delta^{(*)}). $$
(68)
Combining inequality (68) with (62), we get
$$ I_1=W(\alpha_p^{(*)q}+\gamma^{(*)})-W(\alpha_p^{(*)q})\leq W(\alpha^{(*)})-W(\alpha^{(*)}-\delta^{(*)})=I_2. $$
(69)
Next we calculate both the left hand side I
1 and the right hand side I
2 of inequality (69). First, for I
1, we have
$$ \begin{aligned} I_1&=W(\alpha_p^{(*)q}+\gamma^{(*)})-W(\alpha_p^{(*)q})\\ &=\sum_{j,i}(\alpha_{pi}^{qj}+\gamma_i^j+\alpha_{pi}^{*qj}+\gamma_i^{*j})-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\alpha_{pi}^{*qj}+\gamma_i^{*j}-\alpha_{pi}^{qj}-\gamma_i^j)(\alpha_{pi'}^{*qj'}+\gamma_{i'}^{*j'}\\ &-\alpha_{pi'}^{qj'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'})-\sum_{j,i}(\alpha_{pi}^{qj}+\alpha_{pi}^{*qj})+\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})(\alpha_{pi'}^{*qj'}-\alpha_{pi'}^{qj'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}),\\ &=\sum_{j,i}(\gamma_i^j+\gamma_i^{*j})-\sum_{j,i}\sum_{j',i'}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'})-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\gamma_i^{*j}-\gamma_i^j)(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}),\\ &=\sum_{j,i}(\gamma_i^j+\gamma_i^{*j})-\sum_{j,i}(\gamma_{i}^{*j}-\gamma_{i}^{j})(w_p^q\cdot \hbox{x}_i^j)-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\gamma_i^{*j}-\gamma_i^j)(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}). \end{aligned} $$
According to equalities (63)–(67), we rewrite the above expression as
$$ \begin{aligned} I_1&=\sum_{i=1}^{l^{q+1}}\gamma_i^{*q+1}[1-(w_p^q\cdot \hbox{x}_i^{q+1})]+\sum_{i=1}^{l^{q}}\gamma_i^{q}[1+(w_p^q\cdot \hbox{x}_i^q)] -\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\gamma_i^{*j}-\gamma_i^j)(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}),\\ &=\sum_{i=1}^{l^{q+1}}\gamma_i^{*q+1}[1-(w_p^q\cdot \hbox{x}_i^{q+1})+b'_q]+\sum_{i=1}^{l^{q}}\gamma_i^{q}[1+(w_p^q\cdot \hbox{x}_i^q)-b'_q] \\ &\quad -\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\gamma_i^{*j}-\gamma_i^j)(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}). \end{aligned} $$
Taking into account that γ(*) satisfies the conditions (63)–(67), the following two equalities hold:
$$ \begin{aligned} \,&\gamma_i^{*q+1}[1-(w_p^q\cdot \hbox{x}_i^{q+1})+b'_q]=0, \\ \,&\gamma_i^{q}[1+(w_p^q\cdot \hbox{x}_i^q)-b'_q]=0,\quad i\neq p. \end{aligned} $$
So we obtain
$$ I_1=\gamma_p^{q}[1+(\hbox{x}_p^q\cdot \hbox{x}_i^q)-b'_q]-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\gamma_i^{*j}-\gamma_i^j)(\gamma_{i'}^{*j'}-\gamma_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}). $$
(70)
Now let us define vector γ(*) as follows:
$$ \gamma_p^q=\gamma_s^{*q+1}=a, $$
(71)
$$ \gamma_i^j=0,\quad j= q,\quad i\neq p, \gamma_i^{*j}=0,\quad j= q+1,\quad i\neq s, $$
(72)
$$ \gamma_i^j=0,\quad j\neq q,\quad i=1,\ldots,l^j,\quad \gamma_i^{*j}=0,\quad j\neq q+1,\quad i=1,\ldots,l^j, $$
(73)
where a is some constant and α
*q+1
s
∈ (0, C). Substituting (71)–(73) into (70), we get
$$ \begin{aligned} I_1&=a\left[(\hbox{x}_p^q\cdot \hbox{x}_p^q)-b'_q+1\right]-{\frac{a^2} 2}\|\hbox{x}_p^{q}-\hbox{x}_s^{q+1}\|^2\\ &\geq a\left[(\hbox{x}_p^q\cdot \hbox{x}_p^q)-b'_q+1\right]-{\frac{a^2} 2}D_{q,q+1}^2. \end{aligned} $$
(74)
where D
q,q+1 is the diameter of the minimum sphere containing the qth class points and the q + 1th class points in the training set T. Now choose the value a* by maximizing the expression (74)
$$ a^*=\frac{\left[(\hbox{x}_p^q\cdot \hbox{x}_p^q)-b'_q+1\right]}{D_{q,q+1}^2}. $$
Putting this expression back into (74), we get
$$ I_1\geq \frac{{\left[(w_p^q\cdot \hbox{x}_p^q)-b'_q+1\right]^2}}{2D_{q,q+1}^2}. $$
Since, according to our assumption, when the LOO procedure commits an error at the point x
q
p
, the following inequality holds
$$ (\hbox{x}_p^q\cdot \hbox{x}_p^q)-b'_q > 0, $$
we obtain
$$ I_1\geq \frac{1}{2D_{q,q+1}^2}. $$
But we need to fulfill the condition a ≤ C. Thus, if a* > C, we replace a by C in equality (74) and we get
$$ \begin{aligned} I_1&\geq C\left[(w_p^q\cdot \hbox{x}_p^q)-b'_q+1\right]-\frac{{C^2}}{2}D_{q,q+1}^2\\ &=CD_{q,q+1}^2\left(a^*-\frac{C}{2}\right)\\ &\geq CD_{q,q+1}^2\frac{{a^*}}{2}\\ &=C\left[(w_p^q\cdot \hbox{x}_p^q)-b'_q+1\right],\\ &\geq \frac{{C}}{2}. \end{aligned} $$
Finally, we have
$$ I_1\geq \frac{1}{ 2}\min\left(C,\frac{1}{D_{q,q+1}^2}\right). $$
(75)
Now we estimate the right hand side I
2 of the inequality (69). According to Lemma 2, we choose
$$ \begin{aligned} \,&\delta_i^{q}=-\lambda_i^{q}\alpha_p^{q},\quad \delta_i^{*q}=-\lambda_i^{q}\alpha_p^{*q},\quad i\in M_p^q(\alpha,q), \\ \,&\delta_i^{q+1}=\lambda_i^{q+1}\alpha_p^{*q},\quad \delta_i^{*q+1}=\lambda_i^{q+1}\alpha_p^{q},\quad i\in M_p^q(\alpha^*,q+1), \\ \,&\delta_i^{q}=\delta_i^{*q}=0,\quad i\notin M_p^q(\alpha,q), \\ \,&\delta_i^{q+1}=\delta_i^{*q+1}=0,\quad i\notin M_p^q(\alpha^*,q+1), \\ \,&\delta_i^{j}=\delta_i^{*j}=0, \quad j\neq q,q+1,\quad i=1,\ldots,l^j, \\ \,& \delta_p^q=\alpha_p^q,\quad \delta_p^{*q}=\alpha_p^{*q}, \end{aligned} $$
where
$$ \sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^q\hbox{x}_i^q+\sum\limits_{i\in M_p^{q}(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}\in \Uplambda_p^q. $$
Then the right hand side I
2 of the inequality (69) is expressed as
$$ \begin{aligned} I_2&=W(\alpha^{(*)})-W(\alpha^{(*)}-\delta^{(*)}) \\ &=\sum_{j,i}(\alpha_i^j+\alpha_i^{*j})-\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\alpha_i^{*j}-\alpha_i^j)(\alpha_{i'}^{*j'}-\alpha_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'})-\sum_{j,i}(\alpha_i^j-\delta_i^j+\alpha_i^{*j}-\delta_i^{*j}) \\ &\quad+\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\alpha_i^{*j}-\delta_i^{*j}-\alpha_i^j+\delta_i^j)(\alpha_{i'}^{*j'}-\delta_{i'}^{*j'}-\alpha_{i'}^{j'}+\delta_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}) \\ &=\sum_{j,i}(\delta_i^j+\delta_i^{*j})-\sum_{j,i}\sum_{j',i'}(\alpha_i^{*j}-\alpha_i^j)(\delta_{i'}^{*j'}-\delta_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'})+\frac{1}{2}\sum_{j,i}\sum_{j',i'}(\delta_i^{*j}-\delta_i^j)(\delta_{i'}^{*j'}-\delta_{i'}^{j'})(\hbox{x}_i^j\cdot \hbox{x}_{i'}^{j'}) \\ &=-(\alpha_p^q+\alpha_p^{*q})\left[\sum_{i\in M(\alpha,q)\cup \{p\}}\lambda_i^{q}-\sum_{i\in M(\alpha^*,q+1)}\lambda_i^{q+1}\right]\\ &\quad+(\alpha_p^q+\alpha_p^{*q})\left[\sum_{i\in M(\alpha,q)\cup \{p\}}\lambda_i^{q}(w\cdot \hbox{x}_i^{q})+\sum_{i\in M(\alpha^*,q+1)}\lambda_i^{q+1}(w\cdot \hbox{x}_i^{q+1})\right]\\ &\quad+\frac{{(\alpha_p^{*q}-\alpha_p^q)^2}}{2}\left\|\hbox{x}_p^q-\sum\limits_{i\in M_p^q(\alpha,q)}\lambda_i^{q}\hbox{x}_i^{q-1}+\sum\limits_{i\in M_p^{q}(\alpha^*,q+1)}\lambda_i^{q+1}\hbox{x}_i^{q+1}\right\|^2.\\ \end{aligned} $$
Since the definition of Λ
q
p
implies that
$$ \sum_{i\in M_p^q(\alpha,q)}\lambda_i^{q}+ \sum_{i\in M_p^q(\alpha^*,q+1)}\lambda_i^{q+1}=1, \lambda_p^q=-1, $$
(76)
We have
$$ \begin{aligned} I_2&=-(\alpha_p^q+\alpha_p^{*q})\sum_{i\in M_p^q(\alpha,q)\cup \{p\}}\lambda_i^{q}[1-(w\cdot \hbox{x}_i^{q})]\\ &\quad+(\alpha_p^q+\alpha_p^{*q})\sum_{i\in M(\alpha^*,q+1)}\lambda_i^{q+1}[1+(w\cdot \hbox{x}_i^{q+1})]+\frac{{(\alpha_p^{*q}-\alpha_p^q)^2}}{2}S^2(p,q),\\ &=-(\alpha_p^q+\alpha_p^{*q})\sum_{i\in M_p^q(\alpha,q)\cup \{p\}}\lambda_i^{q}[1-(w\cdot \hbox{x}_i^{q})+b_q]\\ &\quad+(\alpha_p^q+\alpha_p^{*q})\sum_{i\in M(\alpha^*,q+1)}\lambda_i^{q+1}[1+(w\cdot \hbox{x}_i^{q+1})-b_q]+\frac{{(\alpha_p^{*q}-\alpha_p^q)^2}}{2}S^2(p,q),\\ &=\frac{{(\alpha_p^{*q}-\alpha_p^q)^2}}{2}S^2(p,q). \end{aligned} $$
(77)
Combining the equalities (69), (75) and (77), we obtain
$$ (\alpha_p^{*q}-\alpha_p^q)^2S^2(p,q)\geq\min \left(C,\frac{1}{D_{q,q+1}^2}\right), $$
(78)
where D
q,q+1 is the diameter of the minimum sphere containing both the qth class points and the q + 1th class points in the training set T.
2. Similarly to the process of deriving (78), for being left out the point x
q
p
, we can get the following inequality by considering its margin support vector about α*:
$$ (\alpha_p^{*q}-\alpha_p^q)^2S^{*2}(p,q)\geq\min \left(C,\frac{1}{D_{q-1,q}^2}\right), $$
(79)
where D
q−1,q
is the diameter of minimum sphere containing the q−1th class points and the qth class points in the training set T. \(\square\)
1.4 Proof of Lemma 5
Proof
Suppose that α(*) is the optimal solution of the problem (7)–(9). It is sufficient to study the following three cases, respectively:
1. The case α
q
p
= α
*q
p
= 0: Being left out the points (x
q
p
, y
q
p
) is not a support vector. Then the object function value of the problem (7)–(9) is equal to that of the problem (34)–(36), namely, W(α(*)) = W
q
p
(α
(*)q
p
). So the decision function does not change after left out the point (x
q
p
, y
q
p
). So the point (x
q
p
, y
q
p
) is not counted as a leave one out error.
2. The case α
q
p
> 0: Being left out the point (x
q
p
, y
q
p
) is a support vector. Starting from the solution α
(*)q
p
of the problem (34)–(36), a feasible points β(*) of the problem (7)–(9) can be constructed by
$$ \begin{aligned} \beta_i^j&=\left\{\begin{array}{ll} \alpha_{pi}^{qj},& j=1,\ldots,q-1,q+1,\ldots,k,\quad i=1,\ldots,l^j,\\ \alpha_{pi}^{qq},& \alpha_{pi}^{qq}=0,\quad \alpha_{pi}^{qq}=C,\\ \alpha_{pi}^{qq}-\nu_i^q,& i\in M(\alpha_p^{(*)q},q),\\ \alpha_p^q, & j=q,\quad i=p,\\ \end{array}\right.\\ \beta_i^{*j}&=\left\{\begin{array}{ll} \alpha_{pi}^{*qj},& j=1,\ldots,q-1,q+1,\ldots,k,\quad i=1,\ldots,l^j,\\ \alpha_{pi}^{*qq},& \alpha_{pi}^{*qq}=0,\quad \alpha_{pi}^{*qq}=C,\\ \alpha_{pi}^{*qq}-\nu_i^{*q},& i\in M(\alpha_p^{(*)q},q),\\ \alpha_p^{*q}, & j=q,\quad i=p,\\ \end{array}\right. \end{aligned} $$
where M(α
(*)q
p
, q) is the set of margin support vector about α(*) for the problem (34)–(36), ν
q
i
and ν
*q
i
, respectively, satisfy conditions
$$ \sum_{i\in M(\alpha_p^{(*)q},q)}\nu_i^q=\alpha_p^q, $$
(80)
and
$$ \sum_{i\in M(\alpha_p^{(*)q},q)}\nu_i^{*q}=\alpha_p^{*q}. $$
(81)
It is easy to see that
$$ \begin{aligned} \,&0\leq \beta_i^j,\quad \beta_i^{*j}\leq C,\quad j=1,\ldots,k,\quad i=1,\ldots,l^j, \\ \,&\sum_{i=1}^{l^j}\beta_i^j=\sum_{i=1}^{l^{j+1}}\beta_i^{*j+1},\quad j=1,\ldots,k-1. \end{aligned} $$
Thus β(*) is a feasible solution of the problem (7)–(9). After a series of transformations, W(β(*)) can be written as
$$ \begin{aligned} W(\beta^{(*)})=&W_p^q(\alpha_p^{(*)q})+(\alpha_p^{*q}+\alpha_p^q)-\frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2K(x_p^q,x_p^q)\\ &-(\alpha_p^{*q}-\alpha_p^q)\sum_{(j,i)\in I\setminus\{(q,p)\}}(\alpha_{pi}^{*qq}-\alpha_{pi}^{qq})K(x_p^q,x_i^j)\\ &+\sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)\left[1+\sum_{(j',i')\in I\setminus\{(q,p)\}}(\alpha_{pi'}^{*qj'}-\alpha_{pi'}^{qj'})K(x_i^q,x_{i'}^{j'})\right]\\ &+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)K(x_p^q,x_i^q)\\ &-\frac{1}{ 2}\sum_{i\in M(\alpha_p^{(*)q},q)}\sum_{i'\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)(\nu_{i'}^{*q}-\nu_{i'}^q)K(x_i^q,x_{i'}^q). \end{aligned} $$
(82)
According to the assumption in the lemma, there exists at least a α
j
i
∈ (0, C), j = 1,…, k in each class points. Therefore,
$$ 1+\sum_{(j',i')\in I\setminus\{(q,p)\}}(\alpha_{pi'}^{*qj'}-\alpha_{pi'}^{qj'})K(x_i^q,x_{i'}^{j'})=b'_q. $$
(83)
From the equalities (80) and (81), we get
$$ \sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)=(\alpha_p^{*q}-\alpha_p^q). $$
(84)
So we rewrite the equality (82) as
$$ \begin{aligned} \,&(\alpha_p^{*q}-\alpha_p^q) {\left [\sum_{(j,i)\in I\setminus\{(q,p)\}}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})K(x_p^q,x_i^j)-b'_q\right]} \\ \,&\quad=-W(\beta^{(*)})+W_p^q(\alpha_p^{(*)q})+(\alpha_p^{*q}+\alpha_p^q)-\frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2K(x_p^q,x_p^q)\\ \,&\quad\quad-\frac{1}{ 2}\sum_{i\in M(\alpha_p^{(*)q},q)}\sum_{i'\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)(\nu_{i'}^{*q}-\nu_{i'}^q)K(x_i^q,x_{i'}^q),\\ \,&\quad\quad+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)K(x_p^q,x_i^q). \end{aligned} $$
(85)
Similarly, let us construct a feasible solution γ(*) of the problem (34)–(36) based on the solution α(*) of the problem (7)–(9) by
$$ \gamma_i^j= \left\{\begin{array}{ll} \alpha_i^j,& j=1,\ldots,q-1,q+1,\ldots,k, i=1,\ldots,l^j,\\ \alpha_i^q,& \alpha_i^q=0,\quad \alpha_i^q=C,\\ \alpha_i^q+\mu_i^q,& i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}, \end{array}\right. $$
and
$$ \gamma_i^{*j}= \left\{\begin{array}{ll} \alpha_i^{*j},& j=1,\ldots,q-1,q+1,\ldots,k,\quad i=1,\ldots,l^j,\\ \alpha_i^{*q},& \alpha_i^q=0 \hbox{ or } \alpha_i^q=C,\\ \alpha_i^{*q}+\mu_i^{*q},& i\in M(\alpha^{(*)},q)\setminus \{(q,p)\},\\ \end{array}\right. $$
where μ
q
i
≥ 0 and μ
*q
i
≥ 0, respectively, satisfy conditions
$$ \sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\mu_i^q=\alpha_p^q; $$
and
$$ \sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\mu_i^{*q}=\alpha_p^{*q}. $$
It is easy to see that
$$ \begin{aligned} &0\leq \gamma_i^j,\quad \gamma_i^{*j}\leq C, \quad (j,i)\in I\setminus \{(q,p)\},\\ &\sum_{i=1}^{l^j}\gamma_i^j=\sum_{i=1}^{l^{j+1}}\gamma_i^{*j+1},\quad (j,i)\in I\setminus \{(q,p)\}. \end{aligned} $$
So γ(*) is a feasible solution to the problem (34)–(36). After a series of transformations, W
q
p
(γ(*)) can be written as
$$ \begin{aligned} W_p^q(\gamma^{(*)})=&W(\alpha^{(*)})+(\alpha_p^{*q}-\alpha_p^q)+\frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2K(x_p^q,x_p^q)\\ &+(\alpha_p^{*q}-\alpha_p^q)\sum_{(j,i)\in I\setminus \{(q,p)\}}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)\\ &-\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)\left[1+\sum_{j',i'}(\alpha_{i'}^{*j'}-\alpha_{i'}^{j'})K(x_i^q,x_{i'}^{j'})\right]\\ &-\frac{1}{ 2}\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\sum_{i'\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)(\mu_{i'}^{*q}-\mu_{i'}^q)K(x_i^q,x_i^q)\\ &+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)K(x_p^q,x_i^q). \end{aligned} $$
(86)
According to the assumption that there exists a α
j
i
∈ (0, C), j = 1,…, k in each class points at least, then we have
$$ 1+\sum_{j',i'}(\alpha_{i'}^{*j'}-\alpha_{i'}^{j'})K(x_i^q,x_{i'}^{j'})=b_q. $$
So we rewrite the equality (86) as
$$ \begin{aligned} -W(\alpha^{(*)})=&-W_p^q(\gamma^{(*)})+(\alpha_p^{*q}-\alpha_p^q)+\frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2K(x_p^q,x_p^q)\\ &+(\alpha_p^{*q}-\alpha_p^q)\left[\sum_{(j,i)\in I\setminus \{(q,p)\}}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)-b_q\right]\\ &-\frac{1}{ 2}\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\sum_{i'\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)(\mu_{i'}^{*q}-\mu_{i'}^q)K(x_i^q,x_i^q)\\ &+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)K(x_p^q,x_i^q). \end{aligned} $$
(87)
Due to W(α(*)) ≥ W(β(*)), W
q
p
(α
(*)q
p
) ≥ W
q
p
(γ(*)) and the equalities (85) and (87), we get
$$ \begin{aligned} \,&(\alpha_p^{*q}-\alpha_p^q) {\left [\sum_{(j,i)\in I\setminus\{(q,p)\}}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})K(x_p^q,x_i^j)-b'_q \right]} \geq (\alpha_p^{*q}-\alpha_p^q) {\left [\sum_{(j,i)\in I\setminus \{(q,p)\}}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)-b_q \right]} \\ \,&\quad-\frac{1}{ 2}\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\sum_{i'\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)(\mu_{i'}^{*q}-\mu_{i'}^q)K(x_i^q,x_i^q)\\ \,&\quad+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)K(x_p^q,x_i^q) \\ \,&\quad-\frac{1}{ 2}\sum_{i\in M(\alpha_p^{(*)q},q)}\sum_{i'\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)(\nu_{i'}^{*q}-\nu_{i'}^q)K(x_i^q,x_{i'}^q)+(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)K(x_p^q,x_i^q). \end{aligned} $$
According to the definitions of ν
q
i
, ν
*q
i
, i ∈ M(α
(*)q
p
, q) and \({\mu_i^q,\mu_i^{*q},i\in M(\alpha^{(*)},q)\setminus \{(q,p)\},}\) we know
$$ \begin{aligned} &(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)K(x_p^q,x_i^q)\geq 0,\\ &(\alpha_p^{*q}-\alpha_p^q)\sum_{i\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)K(x_p^q,x_i^q)\geq 0. \end{aligned} $$
Furthermore since
$$ \begin{aligned} &\frac{1}{ 2}\sum_{i\in M(\alpha_p^{(*)q},q)}\sum_{i'\in M(\alpha_p^{(*)q},q)}(\nu_i^{*q}-\nu_i^q)(\nu_{i'}^{*q}-\nu_{i'}^q)K(x_i^q,x_{i'}^q)\leq \frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2R^2,\\ &\frac{1}{ 2}\sum_{i\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}\sum_{i'\in M(\alpha^{(*)},q)\setminus \{(q,p)\}}(\mu_i^{*q}-\mu_i^q)(\mu_{i'}^{*q}-\mu_{i'}^q)K(x_i^q,x_i^q)\leq \frac{1}{ 2}(\alpha_p^{*q}-\alpha_p^q)^2R^2, \end{aligned} $$
where R
2 = max {K(x
j
i
, x
j
i
)|j = 1,…, k, i = 1,…, l
j}, we arrive that
$$ \begin{aligned} \,&(\alpha_p^{*q}-\alpha_p^q)\left[\sum_{(j,i)\in I\setminus\{(q,p)\}}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})K(x_p^q,x_i^j)-b'_q\right] \\ \,&\quad\geq (\alpha_p^{*q}-\alpha_p^q)\left[\sum_{(j,i)\in I\setminus \{(q,p)\}}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)-b_q\right]-(\alpha_p^{*q}-\alpha_p^q)^2R^2, \end{aligned} $$
namely,
$$ \begin{aligned} \,&\left[\sum_{(j,i)\in I\setminus\{(q,p)\}}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})K(x_p^q,x_i^j)-b'_q\right] \\ \,&\quad\leq \left[\sum_{j,i}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)-b_q\right]-(\alpha_p^{*q}-\alpha_p^q)(K(x_p^q,x_p^q)+R^2). \end{aligned} $$
3. The case α
*q
p
> 0: Following the argumentation in the case (2), an LOO error can occur only if
$$ \begin{aligned} \,&-\left[\sum_{(j,i)\in I\setminus \{(q,p)\}}(\alpha_{pi}^{*qj}-\alpha_{pi}^{qj})K(x_p^q,x_i^j)-b'_{q-1}\right]\\ \,&\quad\leq-\left[\sum_{j,i}(\alpha_i^{*j}-\alpha_i^j)K(x_p^q,x_i^j)- b_{q-1}\right]+(\alpha_p^{*q}-\alpha_p^q)(K(x_p^q,x_p^q)+R^2). \end{aligned} $$
\(\square\)
1.5 Computation of the S-span
In this subsection, we address the problem of computing S
2(q, p) and S
*2(q, p) appeared in Definitions 3 and 4. They can be obtained by solving a quadratic programming, respectively. We only discuss the computation of S
2(q, p) since the computation of S
*2(q, p) can be shown similarly.
The following lemma gives an equivalent description to Definition 3.
Lemma 6
Introducing the following notation:
$$ \{\lambda_i^q, i\in M_p^q(\alpha,q)\}=\{\lambda_1,\ldots,\lambda_r\}, $$
(88)
$$ \{\lambda_i^{q+1}, i\in M_p^q(\alpha^*,q+1)\}=\{\lambda_{r+1},\ldots,\lambda_s\}, $$
(89)
$$ \{x_i^q,i\in M_p^q(\alpha,q)\}=\{x_1,\ldots,x_r\}, $$
(90)
$$ \{x_i^{q+1},i\in M_p^q(\alpha^*,q+1)\}=\{x_{r+1},\ldots,x_s\}, $$
(91)
$$ \{\alpha_i^q,\alpha_i^{*q},i\in M_p^q(\alpha,q)\}=\{\alpha_1,\ldots,\alpha_r,\alpha^*_1,\ldots,\alpha^*_r\}, $$
(92)
$$ \{\alpha_i^{q+1},\alpha_i^{*q+1},i\in M_p^q(\alpha^*,q+1)\}=\{\alpha_{r+1},\ldots,\alpha_s,\alpha^*_{r+1},\ldots,\alpha^*_s\}, $$
(93)
Definition 3 is equivalent to the following description: For any margin support vector
x
q
p
about α, its S-span is
$$ S^2(q,p):=\min\{\|x_p^q-\tilde{x}_p^q\|^2|\tilde{x}_p^q\in\Uplambda_p^q\}, $$
(94)
where
\( \Uplambda_p^q:= \left\{\sum\limits_{i=1}^s\lambda_ix_i\right\},\)
subject to constraints
$$ \begin{aligned} \,&0\leq \alpha_i+\lambda_i\alpha_p^{q}\leq C,\quad 0\leq \alpha_i^*+\lambda_i\alpha_p^{*q}\leq C, \quad i=1,\ldots,r,\\ \,&0\leq \alpha_i-\lambda_i\alpha_p^{*q}\leq C,\quad 0\leq\alpha_i^*-\lambda_i\alpha_p^{q}\leq C ,\quad i=r+1,\ldots,s, \\ \,& \sum_{i=1}^s\lambda_i=1,\quad\lambda_p^q=-1. \end{aligned} $$
Proof
It is sufficient to substitute (88)–(93) into Definition 3. \(\square\)
The following theorem shows that, in Definition 3, S
2(q, p) can be obtained by solving a quadratic programming.
Theorem 3
S
2(q, p) defined in Definition 3 can be obtained by solving the following quadratic programming:
$$ \min_\lambda \quad \sum_{i=1}^s\sum_{j=1}^s\lambda_i\lambda_j(x_i\cdot x_j)-2\sum_{i=1}^s\lambda_i(x_p^q\cdot x_i), $$
(95)
$$ \hbox{s.t.}\quad -{\frac{\alpha_i}{\alpha_p^q}}\leq \lambda_i\leq {\frac{C-\alpha^*_i}{\alpha_p^{*q}}},\quad -{\frac{\alpha^*_i} {\alpha_p^{*q}}}\leq \lambda_i\leq {\frac{C-\alpha^*_i} {\alpha_p^{*q}}},\quad i=1,\ldots,r, $$
(96)
$$ {\frac{\alpha_i-C}{\alpha_p^{*q}}}\leq \lambda_i\leq {\frac{\alpha_i}{\alpha_p^{*q}}},{\frac{\alpha^*_i-C} {\alpha_p^{q}}}\leq \lambda_i\leq {\frac{\alpha^*_i} {\alpha_p^{q}}},\quad i=r+1,\ldots,s, $$
(97)
$$ \sum_{i=1}^s\lambda_i=1. $$
(98)
Proof
We want to minimize the quantity in (94) with respect to {λ
i
}
$$ \begin{aligned} \|x_p^q-\tilde{x}_p^q\|^2&=((x_p^q-\tilde{x}_p^q)\cdot (x_p^q-\tilde{x}_p^q)) \\ &=(x_p^q\cdot x_p^q)+(\tilde{x}_p^q\cdot \tilde{x}_p^q)-2(x_p^q\cdot\tilde{x}_p^q) \\ &=(x_p^q\cdot x_p^q)+\sum_{i=1}^s\sum_{j=1}^s\lambda_i\lambda_j(x_i\cdot x_j)-2\left(x_p^q\cdot \left(\sum_{i=1}^s\lambda_ix_i\right)\right), \end{aligned} $$
subject to constraints
$$ \begin{aligned} \,&-{\frac{\alpha_i}{\alpha_p^q}}\leq \lambda_i\leq {\frac{C-\alpha^*_i} {\alpha_p^{*q}}},\quad -{\frac{\alpha^*_i}{\alpha_p^{*q}}}\leq \lambda_i\leq {\frac{C-\alpha^*_i}{\alpha_p^{*q}}},\quad i=1,\ldots,r, \\ \,&{\frac{\alpha_i-C}{\alpha_p^{*q}}}\leq \lambda_i\leq {\frac{\alpha_i} {\alpha_p^{*q}}},\quad {\frac{\alpha^*_i-C}{\alpha_p^{q}}}\leq \lambda_i\leq {\frac{\alpha^*_i}{\alpha_p^{q}}},\quad i=r+1,\ldots,s, \\ \,&\sum_{i=1}^s\lambda_i=1, \end{aligned} $$
namely, solving the quadratic programming (95)–(98). \(\square\)
In order to compute S
2(p, q) and S
*2(p, q) in Theorem 1, we only need to replace the inner product (x · x′) in (95)–(98) by the kernel function K(x, x′).