5.11. SADDLE POINTS∗ 135
This is because minx Hε (x,y∗)≡ Hε (g(y∗) ,y∗) so
Hε (g((1− t)y∗+ ty) ,y∗)≥ Hε (g(y∗) ,y∗)
Then subtracting the first term on the right, one gets
tHε (g(y∗) ,y∗)≥ tHε (g((1− t)y∗+ ty) ,y)
and cancelling the t,
Hε (g(y∗) ,y∗)≥ Hε (g((1− t)y∗+ ty) ,y)
Now apply Lemma 5.11.3 and let t→ 0+ . This along with lower semicontinuity yields
Hε (g(y∗) ,y∗)≥ lim inft→0+
Hε (g((1− t)y∗+ ty) ,y) = Hε (g(y∗) ,y) (5.12)
Hence for every x,y
Hε (x,y∗)≥ Hε (g(y∗) ,y∗)≥ Hε (g(y∗) ,y)
Thusmin
xHε (x,y∗)≥ Hε (g(y∗) ,y∗)≥max
yHε (g(y∗) ,y)
and so
maxy∈B
minx∈A
Hε (x,y) ≥ minx
Hε (x,y∗)≥ Hε (g(y∗) ,y∗)
≥ maxy
Hε (g(y∗) ,y)≥minx∈A
maxy∈B
Hε (x,y)
Thus, letting C ≡max{∥x∥ : x ∈ A}
εC2 +maxy∈B
minx∈A
H (x,y)≥minx∈A
maxy∈B
H (x,y)
Since ε is arbitrary, it follows that
maxy∈B
minx∈A
H (x,y)≥minx∈A
maxy∈B
H (x,y)
This proves the first part because it was shown above in 5.9 that
minx∈A
maxy∈B
H (x,y)≥maxy∈B
minx∈A
H (x,y)
Now consider 5.8 about the existence of a “saddle point” given the equality of minmaxand maxmin. Let
α = maxy∈B
minx∈A
H (x,y) = minx∈A
maxy∈B
H (x,y)
Then fromy→min
x∈AH (x,y) and x→max
y∈BH (x,y)
being upper semicontinuous and lower semicontinuous respectively, there exist y0 and x0such that
α = minx∈A
H (x,y0) =
.
maxy∈B
minimum of u.s.cminx∈A
H (x,y) = minx∈A
maximum of l.s.c.maxy∈B
H (x,y) = maxy∈B
H (x0,y)