Supporting Appendix S7: mathematical demonstration Theorem 1: The problem 𝑛 2 𝑖𝑛𝑡 𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑓(𝑉) = ∑𝑖=1 (𝑣𝑖𝑛𝑡,𝑖 ) (3) Subject to the constraints: 𝑁𝑉 = 0 𝑙 𝑙 𝑉𝑖𝑛𝑡,𝑚𝑖𝑛 ≤ 𝑉𝑖𝑛𝑡 ≤ 𝑉𝑖𝑛𝑡,𝑚𝑎𝑥 𝑙 𝑙 𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 ≤ 𝑉𝑒𝑥𝑡 ≤ 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 and admits a unique solution. Proof: The stoichiometry matrix N can be reordered to be partitioned into 2×2 blocks 𝑁 𝑁 = ( 11 𝑁21 0 ) 𝑁22 where 𝑁11 is a 𝑚𝐼,𝑖𝑛𝑡 x𝑛𝑖𝑛𝑡 matrix, 𝑁21 is a 𝑚𝐸,𝑖𝑛𝑡 x𝑛𝑖𝑛𝑡 matrix and 𝑁22 is a 𝑚𝐸,𝑖𝑛𝑡 x𝑛𝑒𝑥𝑡 matrix so that the 𝑑𝑋𝐼,𝑖𝑛𝑡 𝑁 steady state constraints 𝑁𝑉 = 0 can be rewritten by (𝑑𝑋𝑑𝑡 ) = ( 11 𝐸,𝑖𝑛𝑡 𝑁21 𝑑𝑡 0 ) (𝑉𝑖𝑛𝑡 ) = 0 𝑁22 𝑉𝑒𝑥𝑡 𝑁11 𝑉𝑖𝑛𝑡 = 0 ⟺ { . 𝑁21 𝑉𝑖𝑛𝑡 = −𝑁22 𝑉𝑒𝑥𝑡 We denote by (𝑛𝑟𝑠 𝑖,𝑗 ) the matrix coefficients of 𝑁𝑟𝑠 . 𝑙 𝑙 From 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ≤ 𝑉𝑒𝑥𝑡 ≤ 𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 and 𝑁21 𝑉𝑖𝑛𝑡 = −𝑁22 𝑉𝑒𝑥𝑡 , we derive 𝑙 𝑙 𝑙 𝑙 𝑔𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ) ≤ 𝑘𝑗 (𝑉𝑖𝑛𝑡 ) ≤ ℎ𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 ) for 𝑗 = 1 … 𝑚𝐸,𝑖𝑛𝑡 where 𝑘𝑗 is the following linear function: 𝑛𝑖𝑛𝑡 𝑘𝑗 (𝑉𝑖𝑛𝑡 ) = ∑ 𝑛21 𝑗,𝑖 𝑣𝑖𝑛𝑡,𝑖 𝑖=1 and 𝑔𝑗 and ℎ𝑗 are defined by 𝑛𝑒𝑥𝑡 𝑙 𝑙 𝑔𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ) 𝑙 𝑙 = ∑ min (−𝑛22 𝑗,𝑖 𝑣𝑒𝑥𝑡,𝑚𝑖𝑛,𝑖 , −𝑛22 𝑗,𝑖 𝑣𝑒𝑥𝑡,𝑚𝑎𝑥,𝑖 ) 𝑖=1 𝑛𝑒𝑥𝑡 𝑙 𝑙 𝑙 𝑙 ℎ𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 , −𝑛22 𝑗,𝑖 𝑣𝑒𝑥𝑡,𝑚𝑎𝑥,𝑖 ) ) = ∑ max (−𝑛22 𝑗,𝑖 𝑣𝑒𝑥𝑡,𝑚𝑖𝑛,𝑖 𝑖=1 ∗ ∗ ∗ )𝑡 Then the problem (3) is equivalent to find 𝑉 ∗ = (𝑉𝑖𝑛𝑡 , 𝑉𝑒𝑥𝑡 such that 𝑉𝑖𝑛𝑡 minimizes 𝑛 2 𝑖𝑛𝑡 𝑓(𝑉𝑖𝑛𝑡 ) = ∑𝑖=1 (𝑣𝑖𝑛𝑡,𝑖 ) (4) subject to the constraints: 𝑁11 𝑉𝑖𝑛𝑡 = 0 𝑙 𝑙 𝑉𝑖𝑛𝑡,𝑚𝑖𝑛 ≤ 𝑉𝑖𝑛𝑡 ≤ 𝑉𝑖𝑛𝑡,𝑚𝑎𝑥 𝑙 𝑙 𝑙 𝑙 𝑔𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ) ≤ 𝑘𝑗 (𝑉𝑖𝑛𝑡 ) ≤ ℎ𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 ) for 𝑗 = 1 … 𝑚𝐸,𝑖𝑛𝑡 ∗ and such that 𝑉𝑒𝑥𝑡 satisfies ∗ ∗ 𝑁22 𝑉𝑒𝑥𝑡 = −𝑁21 𝑉𝑖𝑛𝑡 ∗ Problem (4) is the optimization problem of finding some 𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 such that ∗ ) = 𝑚𝑖𝑛{𝑓(𝑉𝑖𝑛𝑡 ); 𝑉𝑖𝑛𝑡 ∈ 𝐴𝑙 } 𝑓(𝑉𝑖𝑛𝑡 where for l=1,…,9 (corresponding to the nine stages of fruit development) 𝐴𝑙 = { 𝑙 𝑙 𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 ; 𝑁11 𝑉𝑖𝑛𝑡 = 0; 𝑉𝑖𝑛𝑡,𝑚𝑖𝑛 ≤ 𝑉𝑖𝑛𝑡 ≤ 𝑉𝑖𝑛𝑡,𝑚𝑎𝑥 ; 𝑙 𝑙 𝑙 𝑙 𝑔𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ) ≤ 𝑘𝑗 (𝑉𝑖𝑛𝑡 ) ≤ ℎ𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 ) 𝑗 = 1 … 𝑚𝐸,𝑖𝑛𝑡 } ⊂ ℝ𝑛𝑖𝑛𝑡 is the feasible set and 𝑓: ℝ𝑛𝑖𝑛𝑡 → ℝ is the objective function. As 𝐴𝑙 is a compact convex subset of ℝ𝑛𝑖𝑛𝑡 (see lemma 1) and f is a strictly convex and continuous function ∗ on ℝ𝑛𝑖𝑛𝑡 , from theorem 2, this strictly convex problem admits a unique solution 𝑉𝑖𝑛𝑡 (for each 𝑙 = 1. . .9). ∗ ) the augmented matrix obtained by appending the columns of 𝑁22 and Then we denote by (𝑁22 |−𝑁21 𝑉𝑖𝑛𝑡 ∗ ∗ )) = 𝑛𝑒𝑥𝑡 . So −𝑁21 𝑉𝑖𝑛𝑡 . We have 𝑟𝑎𝑛𝑘(𝑁22 ) = 𝑛𝑒𝑥𝑡 and we check numerically that 𝑟𝑎𝑛𝑘((𝑁22 |−𝑁21 𝑉𝑖𝑛𝑡 ∗ ∗ according to the Rouché–Capelli theorem, the linear system 𝑁22 𝑉𝑒𝑥𝑡 = −𝑁21 𝑉𝑖𝑛𝑡 admits a unique solution ∗ 𝑉𝑒𝑥𝑡 (for each 𝑙 = 1. . .9) given by ∗ 𝑡 𝑡 ∗ 𝑉𝑒𝑥𝑡 = −(𝑁22 𝑁22 )−1 𝑁22 𝑁21 𝑉𝑖𝑛𝑡 (5) Lemma 1: 𝐴𝑙 is a compact convex subset of ℝ𝑛𝑖𝑛𝑡 Proof : Let’s denote 𝑙 𝑙 𝐴1 𝑙 = {𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 ; 𝑉𝑖𝑛𝑡,𝑚𝑖𝑛 ≤ 𝑉𝑖𝑛𝑡 ≤ 𝑉𝑖𝑛𝑡,𝑚𝑎𝑥 } 𝐴2 𝑙 = {𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 ; 𝑁11 𝑉𝑖𝑛𝑡 = 0} 𝑙 𝑙 𝑙 𝑙 𝐴3 𝑙 = {𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 ; 𝑔𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 , 𝑉𝑒𝑥𝑡,𝑚𝑎𝑥 ) ≤ 𝑘𝑗 (𝑉𝑖𝑛𝑡 ) ≤ ℎ𝑗 (𝑉𝑒𝑥𝑡,𝑚𝑖𝑛 ) 𝑗 = 1 … 𝑚𝐸,𝑖𝑛𝑡 } 1. 𝐴𝑙 Is bounded because 𝐴𝑙 ⊂ 𝐴1𝑙 which is bounded. 2. ∀𝑖 = 1. .3, the subset 𝐴𝑖 𝑙 is the inverse image of a closed set and so 𝐴𝑖 𝑙 is closed. Then 𝐴𝑙 is closed as intersection of three closed subsets 𝐴𝑖 𝑙 . 3. ∀(𝑋, 𝑌) ∈ 𝐴𝑙 , it is easy to prove that ∀𝜆 ∈ [0,1] ∀(𝑋, 𝑌) ∈ 𝐴𝑙 we also have 𝜆𝑋 + (1 − 𝜆)𝑌 ∈ 𝐴𝑙 . 𝐴𝑙 is convex So 𝐴𝑙 is a compact convex subset. ∗ Theorem 2: The optimization problem of finding some 𝑉𝑖𝑛𝑡 ∈ ℝ𝑛𝑖𝑛𝑡 such that ∗ ∗ ) = 𝑚𝑖𝑛{𝑓(𝑉𝑖𝑛𝑡 ); 𝑉𝑖𝑛𝑡 ∈ 𝐴𝑙 } admits a unique solution 𝑉𝑖𝑛𝑡 𝑓(𝑉𝑖𝑛𝑡 (for each 𝑙 = 1. . .9). Proof: The function 𝑓 is continuous and is strictly convex that is 𝑓(𝜆𝑋 + (1 − 𝜆)𝑌) < 𝜆𝑓(𝑋) + (1 − 𝜆)𝑓(𝑌) for 𝑋 ≠ 𝑌 and ∀𝜆 ∈ ]0,1[ From the Weierstrass extreme values theorem, f continuous from 𝐴𝑙 ⊂ ℝ𝑛 , 𝐴𝑙 compact (closed and bounded) then the minimization problem 𝑚𝑖𝑛𝑓(𝑋) with 𝑋 ∈ 𝐴𝑙 admits an optimal solution with 𝑋 ∗ ∈ 𝐴𝑙 . The optimal solution is unique. Indeed if we assume that 𝑋 ≠ 𝑌 are two minima of f in 𝐴𝑙 Then we have 𝑓(𝑋) = 𝑓(𝑌) = 𝑚𝑖𝑛 𝑓 𝑋+𝑌 ) 2 But as f is strictly convex, 𝑓( < 𝑓(𝑋) 2 + 𝑓(𝑌) 2 = 𝑓(𝑋) which leads to a contradiction.