In this note we identify the condition that tells whether a given basic feasible solution (BFS)

can be declared optimal. The LP is represented in the standard as: Maximize 𝑐̅ ⋅ 𝑥̅ s.t. 𝐴𝑥̅ = 𝑏̅
and 𝑥̅ ≥ 0, where 𝐴 is 𝑚 × 𝑛 matrix with 𝑚 ≤ 𝑛 and full-rank, i.e., 𝑟𝑎𝑛𝑘(𝐴) = 𝑚. Let 𝑥̅
denote a BFS. Without loss of generality, let us assume that the first 𝑚 components of 𝑥̅ are
the basic variables, i.e., 𝑥̅ = (𝑥1 , 𝑥2 , … , 𝑥𝑚 , 0, … ,0)𝑇 with 𝐴1̅ , 𝐴̅2 , … , 𝐴̅𝑚 being linearly
independent. Then 𝑥1 𝐴̅1 + 𝑥2 𝐴̅2 + ⋯ + 𝑥𝑚 𝐴̅𝑚 = 𝑏̅ and 𝑥1 , 𝑥2 , … , 𝑥𝑚 ≥ 0. Let 𝑧0 denote the
objective function value at 𝑥̅ , i.e., 𝑧0 = 𝑐̅ ⋅ 𝑥̅ = 𝑐1 𝑥1 + 𝑐2 𝑥2 + ⋯ + 𝑐𝑚 𝑥𝑚 . Let us define
𝑥1,𝑗 , 𝑥2,𝑗 , … , 𝑥𝑚,𝑗 and 𝑧𝑗 for 𝑗 = 1,2, … , 𝑛 as follows:
(1) 𝑥1,𝑗 𝐴1̅ + 𝑥2,𝑗 𝐴̅2 + ⋯ + 𝑥𝑚,𝑗 𝐴̅𝑘 = 𝐴𝑗̅
(2) 𝑧𝑗 = 𝑐1 𝑥1,𝑗 + 𝑐2 𝑥2,𝑗 + ⋯ + 𝑐𝑚 𝑥𝑚,𝑗
Observe that (𝑥1 − 𝑥1,𝑗 )𝐴1̅ + (𝑥2 − 𝑥2,𝑗 )𝐴̅2 + ⋯ + (𝑥𝑚 − 𝑥𝑚,𝑗 )𝐴̅𝑚 + 𝐴𝑗̅ = (𝑥1 𝐴1̅ + 𝑥2 𝐴̅2 +
⋯ + 𝑥𝑚 𝐴̅𝑚 ) − (𝑥1,𝑗 𝐴1̅ + 𝑥2,𝑗 𝐴̅2 + ⋯ + 𝑥𝑚,𝑗 𝐴̅𝑚 ) + 𝐴𝑗̅ = 𝑏̅ for all 𝑗. Then 𝑥1,𝑗 , 𝑥2,𝑗 , … , 𝑥𝑚,𝑗
can be viewed as the amounts by which 𝑥1 , 𝑥2 , … , 𝑥𝑚 must be decreased in order to increase
𝑥𝑗 by one unit while satisfying the equations 𝐴𝑥̅ = 𝑏̅. The decrease in 𝑥1 , 𝑥2 , … , 𝑥𝑚 decreases
the objective function by 𝑧𝑗 , and the increase in 𝑥𝑗 increases it by 𝑐𝑗 . If the net effect 𝑐𝑗 − 𝑧𝑗 is
positive, then we can expect to improve the solution by increasing 𝑥𝑗 . However, if 𝑐𝑗 ≤ 𝑧𝑗 ∀𝑗,
then intuition suggests that the current BFS is optimal.

Result: For a given BFS, if 𝑐𝑗 ≤ 𝑧𝑗 for all 𝑗, then the BFS is optimal.
Proof: Let 𝑥̅ denote the BFS with all the above properties. Let 𝑦̅ = (𝑦1 , 𝑦2 , … , 𝑦𝑛 ) denote an
arbitrary feasible solution, i.e., 𝑦1 , 𝑦2 , … , 𝑦𝑛 ≥ 0 and 𝑦1 𝐴1̅ + 𝑦2 𝐴̅2 + ⋯ + 𝑦𝑛 𝐴̅𝑛 = 𝑏̅ . Then
𝑧 = 𝑐1 𝑦1 + 𝑐2 𝑦2 + ⋯ + 𝑐𝑛 𝑦𝑛 is the objective function value at 𝑦̅. If show that 𝑧0 ≥ 𝑧, then
the proof is complete. Using (1), we can rewrite 𝑦1 𝐴1̅ + 𝑦2 𝐴̅2 + ⋯ + 𝑦𝑛 𝐴̅𝑛 = 𝑏̅ as:
𝑚 𝑚 𝑚

𝑦1 (∑ 𝑥𝑖,1 𝐴̅𝑖 ) + 𝑦2 (∑ 𝑥𝑖,2 𝐴̅𝑖 ) + ⋯ + 𝑦𝑛 (∑ 𝑥𝑖,𝑛 𝐴̅𝑖 ) = 𝑏̅

𝑖=1 𝑖=1 𝑖=1
For 𝑖 = 1,2, … , 𝑚, if we take the 𝑖-th term out from each of the above sums and put them
together, we get (𝑦1 𝑥𝑖,1 + 𝑦2 𝑥𝑖,2 + ⋯ + 𝑦𝑛 𝑥𝑖,𝑛 )𝐴̅𝑖 . Then we have
𝑛 𝑛 𝑛

(∑ 𝑦𝑗 𝑥1,𝑗 ) 𝐴1̅ + (∑ 𝑦𝑗 𝑥2,𝑗 ) 𝐴̅2 + ⋯ + (∑ 𝑦𝑗 𝑥𝑚,𝑗 ) 𝐴̅𝑚 = 𝑏̅

𝑗=1 𝑗=1 𝑗=1

Since 𝐴1̅ , 𝐴̅2 , … , 𝐴̅𝑚 are linearly independent, the equations 𝑥1 𝐴1̅ + 𝑥2 𝐴̅2 + ⋯ + 𝑥𝑚 𝐴̅𝑚 = 𝑏̅
have unique solution. Hence, 𝑥𝑖 = ∑𝑛𝑗=1 𝑦𝑗 𝑥𝑖,𝑗 for 𝑖 = 1,2, … , 𝑚. Then
𝑛 𝑛 𝑛

𝑧0 = (∑ 𝑦𝑗 𝑥1,𝑗 ) 𝑐1 + (∑ 𝑦𝑗 𝑥2,𝑗 ) 𝑐2 + ⋯ + (∑ 𝑦𝑗 𝑥𝑚,𝑗 ) 𝑐𝑚

𝑗=1 𝑗=1 𝑗=1
We can rewrite the above expression by taking the 𝑗-th term, for 𝑗 = 1,2, … , 𝑛, out from each
of the above sums and then put them together to obtain
𝑚 𝑚 𝑚

𝑧0 = 𝑦1 (∑ 𝑥𝑖,1 𝑐𝑖 ) + 𝑦2 (∑ 𝑥𝑖,2 𝑐𝑖 ) + ⋯ + 𝑦𝑛 (∑ 𝑥𝑖,𝑛 𝑐𝑖 )

𝑖=1 𝑖=1 𝑖=1
Then by (2), 𝑧0 = 𝑦1 𝑧1 + 𝑦2 𝑧2 + ⋯ + 𝑦𝑛 𝑧𝑛 . Since 𝑐𝑗 ≤ 𝑧𝑗 for all 𝑗, then 𝑧0 ≥ 𝑦1 𝑐1 + 𝑦2 𝑐2 +
⋯ + 𝑦𝑛 𝑐𝑛 = 𝑧, as required. This completes the proof.

The above result gives sufficient condition for optimality of a BFS. It does not tell what
happens when 𝑐𝑗 > 𝑧𝑗 for at least one 𝑗 ∈ {1,2, … , 𝑛}. Intuition may suggest that the BFS is
sub-optimal, but it may not be true. Let 𝑥̅ denote the BFS with all previously mentioned
properties. If we can construct a feasible solution with a better objective function value, than
𝑥̅ is sub-optimal. First note that 𝑐𝑗 = 𝑧𝑗 for all 𝑗 ≤ 𝑚. This is because if 𝑗 ≤ 𝑚, 𝑥𝑖,𝑗 = 0 for
𝑖 ≠ 𝑗 and 𝑥𝑗,𝑗 = 1 solve (1), and then 𝑧𝑗 = 𝑐𝑗 by (2). So 𝑗 with 𝑐𝑗 > 𝑧𝑗 must be from {𝑚 +
1, 𝑚 + 2, … , 𝑛}. Without loss of generality, let us assume that 𝑐𝑚+1 > 𝑧𝑚+1.
For an arbitrary 𝜃 ≥ 0, let us define 𝑥𝑖′ = 𝑥𝑖 − 𝜃𝑥𝑖,𝑚+1 for 𝑖 = 1,2, … , 𝑚. Then 𝑥1′ 𝐴1̅ +
𝑥2′ 𝐴̅2 + ⋯ + 𝑥𝑚
′ ̅
𝐴𝑚 + 𝜃𝐴̅𝑚+1 = (𝑥1 𝐴1̅ + 𝑥2 𝐴̅2 + ⋯ + 𝑥𝑚 𝐴̅𝑚 ) − 𝜃(𝑥1,𝑚+1 𝐴1̅ + 𝑥2,𝑚+1 𝐴̅2 +
⋯ + 𝑥𝑚,𝑚+1 𝐴̅𝑚 ) + 𝜃𝐴̅𝑚+1 = 𝑏̅ . Then 𝑥̅ ′ = (𝑥1′ , 𝑥2′ , … , 𝑥𝑚
′
, 𝜃, 0, … ,0) satisfies 𝐴𝑥̅ ′ = 𝑏̅. Now
if we can ensure that 𝑥𝑖′ ≥ 0 for all 𝑖 ≤ 𝑚, then 𝑥̅ ′ is a feasible solution. Since 𝑥̅ is a BFS,
𝑥𝑖 ≥ 0 ∀𝑖 ≤ 𝑚. Then 𝑥𝑖,𝑚+1 ≤ 0 ⇒ 𝑥𝑖′ ≥ 0, as required. However, if 𝑥𝑖,𝑚+1 > 0, we need
𝜃 ≤ 𝑥𝑖 ⁄𝑥𝑖,𝑚+1 for ensuring 𝑥𝑖′ ≥ 0. There can be multiple 𝑖 with positive 𝑥𝑖,𝑚+1 . Then there
are multiple such restrictions on 𝜃. Overall we need 𝜃 ≤ 𝜃0 = min{𝑥𝑖 ⁄𝑥𝑖,𝑚+1 : 𝑥𝑖,𝑚+1 >
0, 𝑖 = 1,2, … , 𝑚} in order to ensure 𝑥𝑖′ ≥ 0 ∀𝑖 ≤ 𝑚. Note that if 𝑥𝑖,𝑚+1 ≤ 0 ∀𝑖 ≤ 𝑚, then 𝜃0
is undefined and there is no upper bound for 𝜃.
Let us check the objective function value of the new feasible solution. 𝑧 ′ = 𝑐 ⋅ 𝑥̅ ′ = 𝑐1 𝑥1′ +
𝑐2 𝑥2′ + ⋯ + 𝑐𝑚 𝑥𝑚 ′
+ 𝑐𝑚+1 𝜃 = (𝑐1 𝑥1 + 𝑐2 𝑥2 + ⋯ + 𝑐𝑚 𝑥𝑚 ) − 𝜃(𝑐1 𝑥1,𝑚+1 + 𝑐2 𝑥2,𝑚+1 + ⋯ +
𝑐𝑚 𝑥𝑚,𝑚+1 ) + 𝜃𝑐𝑚+1 = 𝑧0 + 𝜃(𝑐𝑚+1 − 𝑧𝑚+1 ). Then 𝜃 > 0 ⇒ 𝑧 ′ > 𝑧0 , i.e., if we can choose
a strictly positive 𝜃 for constructing 𝑥̅ ′ , then the current BFS is sub-optimal. If 𝜃0 > 0 or 𝜃0
is undefined, then we can choose a strictly positive 𝜃. If 𝜃0 = 0, which happens when 𝑥𝑖 = 0
and 𝑥𝑖,𝑚+1 > 0 for some 𝑖 = 1,2, … , 𝑚, then we cannot say that the BFS is sub-optimal. We
also cannot say that it is optimal. Then the nature of 𝑥̅ remains inconclusive.
If 𝑥𝑖 > 0 ∀𝑖 ≤ 𝑚, i.e., if all the basic variables are strictly positive, then 𝑥̅ is called a non-
degenerate BFS, otherwise it’s called a degenerate BFS. If 𝑥̅ is non-degenerate, then 𝜃0 =
min{𝑥𝑖 ⁄𝑥𝑖,𝑚+1 : 𝑥𝑖,𝑚+1 > 0, 𝑖 = 1,2, … , 𝑚} > 0 whenever 𝜃0 exists. Then we can always find
𝑥̅ ′ with 𝑧 ′ > 𝑧0 . Hence, for a non-degenerate BFS if 𝑐𝑗 > 𝑧𝑗 for some 𝑗, then the BFS is sub-
optimal. In the case of degenerate BFS, 𝜃0 may be zero, and we cannot conclude anything.
Then the safe option is to proceed with further calculations. Hence, if we find a BFS with
𝑐𝑗 ≤ 𝑧𝑗 for all 𝑗, then we have found optimal solution and we stop, otherwise we continue to
explore new BFS with a hope to stop soon.