压缩映射
若映射 T : X → X 以一个固定比例缩小所有距离,则称其为压缩映射:存在常数 k 满足 0 ≤ k < 1,使得对所有 x, y 都有 d(T(x), T(y)) ≤ k · d(x, y)。常数 k 严格小于 1——施加 T 会把任意两点拉得更近,至少缩小 1 − k 的比例。不动点是满足 T(p) = p 的点 p,即映射不移动它。下面的定理断言:在完备空间上,压缩映射有且仅有一个不动点,且你只需从任意一点出发迭代即可找到它。
巴拿赫不动点定理
下面是完整的陈述与证明。策略是:从任意点出发,迭代,证明轨道是柯西的(这里压缩因子与几何级数登场),援引完备性得到极限,再验证该极限是不动的且唯一。
BANACH FIXED-POINT THEOREM.
Let (X, d) be a NON-EMPTY COMPLETE metric space and
T : X -> X a contraction with constant k, 0 <= k < 1.
Then T has exactly one fixed point p, and for any start x_0
the iterates x_{n+1} = T(x_n) converge to p.
PROOF.
Step 1 (consecutive gaps shrink geometrically).
Let D = d(x_0, x_1). Applying the contraction repeatedly,
d(x_n, x_{n+1}) = d(T(x_{n-1}), T(x_n)) <= k * d(x_{n-1}, x_n),
so by induction d(x_n, x_{n+1}) <= k^n * D.
Step 2 (the orbit is Cauchy).
For m > n, chain the triangle inequality across the gaps:
d(x_n, x_m) <= d(x_n,x_{n+1}) + ... + d(x_{m-1}, x_m)
<= (k^n + k^{n+1} + ... + k^{m-1}) * D
<= k^n * (1 + k + k^2 + ...) * D
= k^n * D / (1 - k). [geometric series, k<1]
Since k^n -> 0, the right side -> 0 as n -> infinity,
so for every epsilon > 0 we can force d(x_n, x_m) < epsilon.
Hence (x_n) is Cauchy.
Step 3 (a limit exists).
X is complete, so x_n -> p for some p in X.
Step 4 (p is fixed).
T is a contraction, hence continuous, so T(x_n) -> T(p).
But T(x_n) = x_{n+1} -> p as well.
Limits are unique, so T(p) = p.
Step 5 (uniqueness).
Suppose T(p) = p and T(q) = q. Then
d(p, q) = d(T(p), T(q)) <= k * d(p, q).
So (1 - k) * d(p, q) <= 0. Since 1 - k > 0 and d >= 0,
we get d(p, q) = 0, i.e. p = q. QED稠密集与可分性
我们以另一种风味的结构收尾。若集合 D 的闭包就是整个 X,则称 D 是 X 的稠密子集——等价地,X 中每个点都是 D 中点的极限,因此无论多小的球都与 D 相交。有理数在 R 中稠密:任何实数都可被分数任意逼近。若一个空间具有可数的稠密子集,则称其可分——一个可数的脚手架,从它出发取极限即可到达每个点。
可分性是一种“小”的条件,现实空间满足它,而病态空间违反它。轴 R 是可分的,以有理数为见证;R^n 亦然,以坐标为有理数的点为见证。可分性正是使我们能够逼近与计算的东西:一个可数稠密集是一张你可以遍历、存储、并输入算法的列表。许多后续定理都默默地假定了它。