Bellman Equation#

Exercises from the Class Notes#

Exercise 2.2#

Prove that the set \(\mathcal{C}(X)\) of bounded continuous functions \(f:X\rightarrow \mathbb{R}\) equipped with the sup norm \(|| y || \equiv \max_{\{ x \in X \}}|y(x)|\) is a complete normed vector space.

Solution

See Example 3 in Chapter 2.11 of Luenberger, \(\textit{Optimization by Vector Space Methods}\).

Exercise 2.3#

Use an inductive argument similar to the one in the proof of Theorem 2.5. to establish that, when assumptions 2.1-2.4 hold, \(V\) is concave.

Solution

Pick \(x_{1},\ x_{2}\in X\) with \(x_{1}\neq x_{2}\). Let \(\theta\in\left(0,1\right)\), and \(x_{\theta} = \left(1-\theta\right)x_{1} + \theta x_{2}\). We define \(g^{\theta}(x_1,x_2)=(1-\theta) g(x_1)+\theta g(x_2)\). Since \(\Gamma\) is convex, we have \(g^{\theta}(x_1,x_2) \in\Gamma\left(x_{\theta}\right)\). By the optimality of the \(T\) operator, we have

\[\begin{align*} Tf\left(x_{\theta}\right) \geq F\left(x_{\theta}, g^{\theta}(x_1,x_2) \right) + \beta f\left(g^{\theta}(x_1,x_2) \right).\\ \end{align*}\]

Assuming that \(f\) is concave, we find that

\[\begin{align*} Tf\left(x_{\theta}\right) &\geq F\left(x_{\theta}, g^{\theta}(x_1,x_2) \right) + \beta f\left(g^{\theta}(x_1,x_2) \right)\\ & > \left(1-\theta\right)[F\left(x_{1}, g\left(x_{1}\right) \right) + \beta f\left(g\left(x_{1}\right)\right)] + \theta [F\left(x_{2}, g\left(x_{2}\right) \right) + \beta f\left(g\left(x_{2}\right)\right)]\\ & = \left(1-\theta\right)Tf\left(x_{1}\right) + \theta Tf\left(x_{2}\right).\\ \end{align*}\]

The second equality follows from the strict concavity of \(F\) with respect to its second argument, and the premise that \(f\) is also concave. Since \(V = \lim_{n\rightarrow \infty}T^nV_0\), and the space of concave functions is the closure of the space of strictly concave functions, assuming that \(V_0\) is concave, we can conclude that \(V\) is concave.

Additional Exercises#

Exercise 2.4#

  1. Show that the space of non-decreasing functions is the closure of the space of increasing functions.

  2. Show that the space of differentiable functions is not closed.

Solution
  1. We show that any non-decreasing function can be approximated arbitrarily closely by a sequence of strictly increasing function.

    Let \(f:[a,b]\rightarrow R\) be a non-decreasing function. We construct a sequence \(\{f_{n}\}\) such that \(f_{n}\left(x\right)=f\left(x\right) + \frac{x}{n}\), where \(n\in \mathbb{N}\).

    Each \(f_{n}\) is strictly increasing as for any \(x_{1}\) and \(x_{2}\), if \(x_{1}<x_{2}\), \(f_{n}\left(x_{1}\right)<f_{n}\left(x_{2}\right)\).

    As \(n\rightarrow +\infty\), we have \(\sup_{x\in [a,b]}\ |f_{n}\left(x\right) - f\left(x\right)| = \sup_{x\in [a,b]}\ \frac{x}{n} \leq \frac{b}{n}\). Therefore \(f_{n}\rightarrow f\) as \(n\rightarrow +\infty\), and we conclude that the space of non-decreasing functions is the closure of the space of strictly increasing functions.

  2. We show that a sequence of differentiable functions in a closed set converges uniformly to a non-differentiable function.

    Suppose \(f_{n}\left(x\right) = \sqrt{\left(x-\frac{1}{2}\right)^2 + \frac{1}{n^2}}\) for \(x\in [0,1]\). It can be easily shown that \(\lim_{n\rightarrow \infty}\ f\left(x\right) = |x - \frac{1}{2}|\), which is non-differentiable at \(x = \frac{1}{2}\).

Exercise 2.5#

Let \(S\) be the space of all continuous, strictly increasing functions on \([a,b]\), with \(||x,y||=\max_{a\leq t \leq b}|x(t)-y(t)|\).

  1. Show that \(S\) is not complete.

  2. Show that \(S\) is complete if “strictly increasing” is replaced with “nondecreasing”.

Solution
  1. We show that there exists a Cauchy sequence that does not converge to an element in S.

    Consider \(x_{n}^{t} = 1 + \frac{t}{n},\ t\in[a,b]\). Pick an arbitrary \(x_{m}\).

\[\begin{align*} \rho\left(x_{n}, x_{m}\right) &= \max_{a\leq t\leq b} |\frac{t}{n} - \frac{t}{m}|\\ & = \max_{a\leq t\leq b} |\frac{t\left(m-n\right)}{nm}|\\ & = |\frac{b\left(m-n\right)}{nm}| \end{align*}\]

It can be easily shown that \(\rho\left(x_{n}, x_{m}\right) \rightarrow 0\) as \(n,m\rightarrow \infty\). However, \(x_{n}\left(t\right)\rightarrow x\left(t\right) = 1\), which is a constant. This contradicts our premise that the limit is strictly increasing, thus showing that \(S\) is not complete.

  1. From Question 1 of Exercise 2.4, we know that the space of non-decreasing functions is the closure of the space of increasing functions, meaning that any non-decreasing function in \(S\) can be arbitrarily approximated by a sequence of strictly increasing functions. Therefore, any Cauchy sequence will converge to an element contained by \(S\), and such a space is complete.

Exercise 2.6#

Show that if \(T\) is a contraction on \(S\), then \(T\) is uniformly continuous on \(S\).

Solution

If \(T\) is a contraction mapping, then

\( \exists \beta\in\left(0,1\right), \forall x,y\in S,\ \rho\left(Tx, Ty\right) \leq \beta\rho\left(x,y\right) \).

If \(T\) is uniformly continuous, then

\( \forall \epsilon>0, \exists \delta>0, \forall x,y\in S,\ \rho\left(x,y\right)<\delta\ \Rightarrow\ \rho\left(Tx, Ty\right)<\epsilon\).

Now take \(\delta = \frac{\epsilon}{\beta}\), then \(\forall \epsilon>0\), if \(\rho\left(x,y\right)<\delta\),

\(\rho\left(Tx, Ty\right) \leq \beta\rho\left(x,y\right) < \beta \delta = \epsilon\).

Therefore, \(T\) is uniformly continuous on S.

Exercise 2.7#

Show that in a Banach space, i.e. a complete metric space, a subset is complete if and only if it is closed.

Solution

A complete subset is closed since every convergent (and hence Cauchy) sequence has a limit in the subset. A Cauchy sequence from a closed subset has a limit somewhere in the Banach space. By closure the limit must be in the subset.

More explicitly, Since \(S_0 ⊆ S\) is closed, any convergent sequence in \(S_0\) converges to a point in \(S_0\). Take the set of Cauchy sequences in \(S\). They all converge to points in \(S\) since \(S\) is complete. Take the subset of those sequences that belong to \(S_0\), then by the argument above they converge to a point in \(S_0\), so \(S_0\) is complete

Exercise 2.8#

  1. Show that a continuous function on a compact metric space is bounded and uniformly continuous.

  2. (Weierstrass’ extreme value theorem). If \(X\) is compact and \(f : X \rightarrow R\) is continuous, then \(f\) is bounded and attains its maximum and minimum values.

Solution
  1. (Boundedness) Suppose \(f\) is continuous on a compact metric space \(X\), but it is not bounded. Then \(\forall n\in \mathbb{N}, \exists x_{n}\in X\), such that \(f\left(x_{n}\right)>n\). Since \(X\) is compact, by Bolzano–Weierstrass theorem, there exists a convergent subsequence \(\{x_{n_{k}}\}_{k\in\mathbb N}\). Denote its limit by \(x\). Since \(f\) is continuous at \(x\), \(f\left(x_{n_{k}}\right)\) should converge to \(f\left(x\right)\). However \(\forall k\), \(f\left(x_{n_{k}}\right) > x_{n_{k}}\), then \(\left(x_{n_{k}}\right) \rightarrow +\infty\). This contradicts our premise, implying that a continuous function on a compact metric space is bounded.

    (Uniform Continuity) Suppose \(\exists \epsilon_{0}>0\), such that \(\forall \delta>0, \exists x_{n},y_{n}\in X\), such that \(\rho\left(x_{n}, y_{n}\right)<\delta\), but \(\rho\left(f\left(x_{n_{k}}\right),f\left(y_{n_{k}}\right)\right)\geq \epsilon_{0}\). Since \(X\) is compact, the sequence \(\{x_{n_{k}}\}\) has a convergent subsequence, denoted by \(x^{*}\). Then as \(k\rightarrow \infty\), \(\{x_{n_{k}}\}\rightarrow x^{*}\) and \(\{y_{n_{k}}\}\rightarrow x^{*}\). However, as \(f\left(x_{n_{k}}\right)\rightarrow f\left(x^{*}\right)\) and \(f\left(y_{n_{k}}\right)\rightarrow f\left(x^{*}\right)\), we have \(\rho\left(f\left(x_{n_{k}}\right), f\left(y_{n_{k}}\right)\right)\rightarrow 0\), as \(k\rightarrow +\infty\). This contradicts our premise, implying that a continuous function on a compact metric space is uniformly continuous.

  2. (Maximum) Suppose \(f\) is upper-bounded by \(M\) on a compact metric space \(X\), but it cannot attain \(M\). Then \(f\left(x\right)<M,\ \forall x\in X\). Let \(g\left(x\right) = \frac{1}{M-f\left(x\right)}\), then g is also bounded on \(X\). Then \(\exists K>0\) such that \(g\left(x\right)\leq K,\ \forall x\in X\) and \(\frac{1}{M-f\left(x\right)}\leq K\), \(f\left(x\right)\leq M-\frac{1}{k}\). \(M\) is no longer an upperbound now, thus \(f\) attains its maximum values.

    (Minimum) Suppose \(f\) is lower-bounded by \(N\) but cannot attain \(N\), and let \(g\left(x\right) = \frac{1}{N-f\left(x\right)}\). We can easily show that \(N\) is no longer a lower bound. Therefore \(f\) attains its minimum values.