Principal Sequence of Partitions (PSP)

Before presenting the MDA algorithm, it worth describing the concept of principal sequence of partitions (PSP). Firstly, the PSP contains all the parameters that are sufficient to characterize the variation of the value of the Dilworth truncation ˆf#

α(V) in α, which also gives a graphical interpretation of Theorem 5.3.3. Secondly, the

MDA in Section 5.5.2 is proposed based on the existing methods for determining the PSP in the literature due to the fact that the minimum sum-rate RACO(V) and

the fundamental partitionP∗ _{correspond to one of the critical/turning points in the} PSP. Thus, the optimality of the MDA algorithm is also proved based on the existing results on PSP.

We define the pairwise relationship between two partitions inΠ(V)as follows.

Definition 5.5.1(order). ForP,P0 ∈Π(V), we denoteP P0 if∃X0 ∈ P0 such that X ⊆X0for all X ∈ P;P = P0 _{ifP P}0 _{andP P}0_{;P ≺ P}0 _{ifP P}0 _{andP 6=P}0_.

In other words, P P0 if P is finer than P0 and P ≺ P0 if P is strictly finer thanP0_{. For example, for P = {{1, 2},{3},{4}} andP}0 _{= {{1, 2, 3},{4}}, we have}

P P0. In fact,P ≺ P0.

Theorem 5.5.2(PSP [167, 168]). fˆ#

α(V) = minP ∈Π(V) fα#[P]is a piecewise linear nonde- creasing curve inαwith p≤ |V| −1critical/turning points

H(V) =α0> α1 >α2> . . .>αp ≥0

that have the following properties.

(a) DenoteP_j the finest/minimal minimizers ofminP ∈Π(V) fα#j[P]. AllPjs form a partition

chain/sequence C_P:

{V}=P₀ P₁ . . . P_p ={{i}: i∈V}.

If α_j > α > α_j+1 for some j ∈ {0, . . . ,p−1}, the minimizer of minP ∈Π(V)fα#[P] is uniquelyP_j; Ifα < αp, the minimizer is uniquelyPp = {{i}: i ∈ V}; Ifα> α1, the

minimizer is uniquelyP₀={V};

(b) The gradient of fˆ_α#(V)is decreasing inα: The gradient is|Pp|= |V|initially; It changes

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 −8 −6 −4 −2 0 2 4 6 8 10 P1={{1},{2},{3}}=P∗ P0={{1, 2, 3}} α ˆf # (α V ) = min P ∈ Π ( V ) f # [α P ] ˆ f_α#(V) f_α#(V) =α

Figure 5.8: The value of ˆf_α#(V) as a function of α for the system in Example 5.2.1. α1 =RACO(V) = 7₂ andP1=P∗ ={{1},{2},{3}}according to Corollary 5.5.3.

Corollary 5.5.3. α1= RACO(V)andP1= P∗, i.e., the parameters in PSP that correspond

to the first critical pointα1constitute the solutions to the asymptotic CO problem.

Proof. According to Theorem 5.3.3, the base polyhedron B(f#

α,≤)is nonempty if and only ifα= fˆ_α#(V). In other words, B(f_α#,≤)6= ∅if and only if the value of αfalls in

the segment of the piecewise linear ˆf#

α(V)vs. αcurve where the gradient is 1, which, based on property (b) in Theorem 5.5.2, is when α ≥ α1. Then, the minimum sum-

rateRACO(V)is the smallest value ofαsuch thatB(f_α#,≤)6=∅, which isα1. The set of

all maximizers of (5.7a) coincides with the set of all minimizers of minP ∈Π(V) fα#1[P].

So, the minimal/finest minimizer P₁ is the minimal/finest maximizer of (5.7a), i.e.,

P1 equals to the fundamental partitionP∗.

Based on Corollary 5.5.3, solving the asymptotic minimum sum-rate problem reduces to determining the value ofα1andP1in the PSP. The proof of Corollary 5.5.3

is exemplified below.

Example 5.5.4. We show the plot fˆ#

α(V)inαfor the systems in Examples 5.2.1 and 5.3.11 in Figs. 5.8 and 5.9, respectively. It can be seen from both figures that fˆ_α#(V) is an in- creasing piecewise linear function in α. We discuss Fig. 5.3.11 based on Theorem 5.5.2 and

0 1 2 3 4 5 6 7 8 9 10 −25 −20 −15 −10 −5 0 5 10 15 P3={{1},{2},{3},{4},{5}} P2={{4, 5},{1},{2},{3}} P1={{1, 4, 5},{2},{3}}=P∗ P0={{1, 2, 3, 4, 5}} α ˆf # (α V ) = min P ∈ Π ( V ) f # [α P ] ˆ f_α#(V) f# α(V) =α

Figure 5.9: The value of ˆf_α#(V) as a function ofα for the system in Example 5.3.11. α1 =RACO(V) = 13₂ andP1 =P∗ = {{1, 4, 5},{2},{3}} according to Corollary 5.5.3.

In addition toα0andP0, there are three critical pointsα_jwithP_j, the minimal minimizers

ofminP ∈Π(V)fα#j[P], being

α0=10, P0 ={{1, 2, 3, 4, 5}}; α₁=6.5, P₁ ={{1, 4, 5},{2},{3}}; α2=6, P2 ={{4, 5},{1},{2},{3}}; α3=4, P3 ={{1},{2},{3},{4},{5}}.

We have the partition sequence CP: P0 P1 P2 P3. The gradient is: 5 when α ∈ [0,α3]; 4 when α ∈ [α3,α2]; 3 when α ∈ [α2,α₁]; 1 when α ∈ [α₁,α0]. In addition,

one can show that the minimizer of minP ∈Π(V) fα#[P] is uniquely P3 when α ∈ [0,α3),

P₂ when α ∈ (α3,α2), P1 when α ∈ (α2,α1) and P0 when α ∈ (α1,α0]. Here, α1 and

P₁ coincide with the minimum sum-rate RACO(V) = 13₂ and the fundamental partition

P∗ _{={{4, 5},{1},{2},{3}}in Example 5.3.11, respectively.}

In Fig. 5.9, we also plot the line f_α#(V) = α. It can be seen that f_α#(V)overlaps with

ˆ

f_α#(V), i.e., α = fˆ_α#(V), when α ∈ [α1,α0]. So, B(fα#,≤) 6= ∅when α ≥ α1 according to Theorem 5.3.3. α1=6.5, the minimal value ofαin the region[α1,α0], is the minimum sum-

rate RACO(V)and the corresponding partitionP1 ={{1, 4, 5},{2},{3}}is the fundamental

partitionP∗.

Consider the region when α< α1 = 6.5in Fig. 5.9. We have fα#(V) =α> fˆ

#

Algorithm 5.1:Modified Decomposition Algorithm (MDA)

input :the ground setV, an oracle that returns the value ofH(X)for a givenX⊆V output:αthat equals toRACO(V),P∗which is the fundamental partition and a rate

vectorrV in the optimal rate setRACO∗ (V) =B(fˆR#ACO(V),≤)

1 Initializeαaccording to Theorem 5.4.1:

α←maxi∈Vϕ({{i},V\ {i}}),ϕ({{i}:i∈V}) ;

2 find the minimal/finest minimizerP∗of min_{P ∈Π(V)}f_α#[P]and a rate vector

rV ∈B(fˆα#,≤);

3 αˆ ←ϕ(P∗); 4 whileαˆ 6=αdo 5 α←αˆ;

6 find the minimal/finest minimizerP∗of min_{P ∈Π(V)}f_α#[P]and a rate vector

rV ∈B(fˆα#,≤);

7 αˆ ←ϕ(P∗); 8 end

9 returnα,P∗andrV;

discussed in Section 5.3.3, the polyhedron P(f#

α,≤) does not intersect with the hyperplane

{rV ∈R|V|:r(V) =α}, i.e., B(f_α#,≤) =∅, in this region. It means that the minimum sum-

rate αis too low for attaining the omniscience in V. On the contrary, whenα ≥ α1 = 6.5,

P(f_α#,≤)intersects with the hyperplane {rV ∈ R|V|: r(V) = α}, i.e., B(f_α#,≤) = ∅. It

can be seen that the comparison between fˆ_α#(V)and f_α#(V)as functions ofαprovides another

interpretation of Theorem 5.3.3.