1 In the algorithm presented in the previous section,
2 the encoding power consumption is iteratively updated with
8 v_h^{(k)}.\dfrac{\ln(\sigma^2/D_h)}{\gamma p ^{2/3}} + \lambda_h^{(k)}p
11 The function inside the $\arg \min$ is stricly convex if and only if
12 $\lamda_h$ is not null. This asymptotic configuration may arrise due to
13 the definition of $\lambda_i$. Worth, in this case, the function is
14 stricly decreasing and the minimal value is obtained when $p$ is the infinity.
16 To prevent this configuration, we replace the objective function given
17 in equation~(\ref{eq:obj2}) by
19 \sum_{i \in N }q_i^2 +
20 \delta_x \sum_{h \in V, l \in L } .x_{hl}^2
21 + \delta_r\sum_{h \in V }\delta.R_{h}^2
22 + \delta_p\sum_{h \in V }\delta.P_{sh}^{\frac{8}{3}}.
25 In this equation we have first introduced new regularisation factors
26 (namely $\delta_x$, $\delta_r$, and $\delta_p$)
27 instead of the sole $\delta$.
28 This allows to further study the influence of each modification separately.
29 Next, the introduction of the rationnal exponent is motivated by the goal of
30 providing a stricly convex function.