9.15. Directional Derivatives#

This section deals with the subject of directional derivatives for convex functions.

9.15.1. Convex Real Functions#

For real functions \(f : \RR \to \RR\), several useful results are available.

Remark 9.13 (Domain of a convex real function)

The domain of a convex real function is an interval.

Let \(f: \RR \to \RR\) be convex. Then \(\dom f\) is convex. But every convex subset of real line is an interval.

  1. Now let \(I = \dom f\) be an interval.

  2. We say \(a = \inf I\) as the left end point of \(I\).

  3. We say \(b = \sup I\) as the right end point of \(I\).

  4. \(a\) and \(b\) may or may not belong to \(I\).

  5. If both \(a\) and \(b\) belong to \(I\), then \(I\) is a closed interval.

  6. If neither \(a\) nor \(b\) belongs to \(I\), then \(I\) is an open interval.

9.15.1.1. Characterization#

Theorem 9.199 (Characterization of real convex functions)

Let \(f: \RR \to \RR\) be a real function with \(\dom f = I\) which is an interval (closed or open or semi-open). Let \(a\) and \(b\) be the left and right endpoints of the interval \(I\).

The following are equivalent.

  1. \(f\) is convex over \(I\).

  2. For every \(x_1, x_2, x_3 \in I\) with \(x_1 < x_2 < x_3\),

    \[ \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1}. \]
  3. For every \(x_1, x_2, x_3 \in I\) with \(x_1 < x_2 < x_3\),

    \[ \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_2)}{x_3 - x_2}. \]
  4. For every \(x_1, x_2, x_3 \in I\) with \(x_1 < x_2 < x_3\),

    \[ \frac{f(x_3) - f(x_1)}{x_3 - x_1} \leq \frac{f(x_3) - f(x_2)}{x_3 - x_2}. \]

Proof. (1) \(\implies\) (2) Assume that \(f\) is convex.

  1. Let

    \[ \alpha = \frac{x_3 - x_2}{x_3 - x_1}, \beta = \frac{x_2 - x_1}{x_3 - x_1}. \]
  2. Then, \(\alpha + \beta = 1\) and \(\alpha , \beta \in (0,1)\).

  3. Also, verify that

    \[ x_2 = \alpha x_1 + \beta x_3. \]
  4. Thus,

    \[\begin{split} & f(x_2) \leq \alpha f(x_1) + \beta f(x_3) \\ & \iff f(x_2) \leq \frac{x_3 - x_2}{x_3 - x_1} f(x_1) + \frac{x_2 - x_1}{x_3 - x_1} f(x_3)\\ & \iff (x_3 - x_1) f(x_2) \leq (x_3 - x_2) f(x_1) + (x_2 - x_1) f(x_3)\\ & \iff (x_3 - x_1) f(x_2) \leq ((x_3 - x_1) - (x_2 - x_1)) f(x_1) + (x_2 - x_1) f(x_3)\\ & \iff (x_3 - x_1) (f(x_2) - f(x_1)) \leq (x_2 - x_1) (f(x_3) - f(x_1))\\ & \iff \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1}. \end{split}\]

(2) \(\implies\) (1)

  1. Let \(x_1, x_3 \in I\) and \(t \in (0, 1)\).

  2. WLOG, assume that \(x_1 < x_3\).

  3. Let \(\alpha = t\) and \(\beta = (1-t)\).

  4. Let \(x_2 = \alpha x_1 + \beta x_3\).

  5. Then, \(x_1 < x_2 < x_3\).

  6. From the hypothesis, we have

    \[ \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1}. \]
  7. Using the previous argument backwards, this implies

    \[ f(x_2) \leq \alpha f(x_1) + \beta f(x_3) = t f(x_1) + (1-t) f(x_3). \]
  8. Thus, \(f\) is convex.

(2) \(\iff\) (3)

  1. Pick any \(x_1, x_2, x_3 \in I\) with \(x_1 < x_2 < x_3\).

  2. By hypothesis (2)

    \[\begin{split} & \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1} \\ \iff & (x_3 - x_1) (f(x_2) - f(x_1)) \leq (x_2 - x_1) (f(x_3) - f(x_1))\\ \iff & (x_3 - x_1) f(x_2) \leq ((x_3 - x_1) - (x_2 - x_1)) f(x_1) + (x_2 - x_1) f(x_3)\\ \iff & ((x_3 - x_2) + (x_2 - x_1) ) f(x_2) \leq ((x_3 - x_2) f(x_1) + (x_2 - x_1) f(x_3)\\ \iff & (x_3 - x_2) (f(x_2) - f(x_1)) \leq (x_2 - x_1) (f(x_3) - f(x_2))\\ \iff & \frac{f(x_2) - f(x_1)}{x_2 - x_1 } \leq \frac{f(x_3) - f(x_2)}{x_3 - x_2}. \end{split}\]

(2) \(\iff\) (4)

  1. Pick any \(x_1, x_2, x_3 \in I\) with \(x_1 < x_2 < x_3\).

  2. By hypothesis (2)

    \[\begin{split} & \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1} \\ \iff & (x_3 - x_1) (f(x_2) - f(x_1)) \leq (x_2 - x_1) (f(x_3) - f(x_1))\\ \iff & (x_3 - x_1) f(x_2) \leq (x_3 - x_2) f(x_1) + ((x_2 - x_3) + (x_3 - x_1)) f(x_3)\\ \iff & (x_3 - x_1) (f(x_2) - f(x_3)) \leq (x_3 - x_2) (f(x_1) - f(x_3)) \\ \iff & (x_3 - x_1) (f(x_3) - f(x_2)) \geq (x_3 - x_2) (f(x_3) - f(x_1)) \\ \iff & (x_3 - x_2) (f(x_3) - f(x_1)) \leq (x_3 - x_1) (f(x_3) - f(x_2))\\ \iff & \frac{f(x_3) - f(x_1)}{x_3 - x_1} \leq \frac{f(x_3) - f(x_2)} {x_3 - x_2}. \end{split}\]

9.15.1.2. One Sided Derivatives#

Recall from Definition 2.83 that if \(f\) is defined over \([a,b)\), then the right hand derivative is defined as:

\[ f'_+(a) = \lim_{x \to a^+} \frac{f(x) - f(a)}{x - a} = \lim_{h \downarrow 0} \frac{f(a + h) - f(a)}{h} \]

if the limit exists. Similarly, if \(f\) is defined over \((c,a]\), then the left hand derivative is defined as:

\[ f'_-(a) = \lim_{x \to a^-} \frac{f(x) - f(a)}{x - a} = \lim_{h \uparrow 0} \frac{f(a + h) - f(a)}{h} = \lim_{r \downarrow 0} \frac{f(a) - f(a - r)}{r} \]

if the limit exists.

We introduce two helper functions

\[ s_+(x, h) = \frac{f(x + h) - f(x)}{h} \]

and

\[ s_-(x, h) = \frac{f(x) - f(x-h)}{h} \]

where \(h > 0\).

Then

\[ f'_+(x) = \lim_{h \downarrow 0} s_+(x, h) \]

and

\[ f'_-(x) = \lim_{h \downarrow 0} s_-(x, h). \]

An interesting property of convex functions is that the one sided derivatives always exist. On the real line, there are only two directions to move; left and right. The one sided derivatives play the role of directional derivatives on the real line.

Lemma 9.4 (Monotonicity of \(s_+\) and \(s_-\) for convex functions)

Let \(f: \RR \to \RR\) with \(\dom f = I\) be convex. Let \(a\) and \(b\) be the left and right endpoints of the interval \(I\). Then \(s_+\) and \(s_-\) as functions of \(h\) are monotonic.

  1. \(s_+(x, h)\) is a nondecreasing function of \(h\).

  2. \(s_-(x, h)\) is a nonincreasing function of \(h\).

Proof. Monotonicity of \(s_+\).

  1. Let \(x \in I\) such that \(x \neq b\).

  2. Let \(h_1, h_2 > 0\) such that \(h_1 < h_2\) and \(x + h_2 \in I\).

  3. Consider the three points \(x < x + h_1 < x + h_2\).

  4. Then

    \[ \frac{f(x+h_1) - f(x)}{h_1} \leq \frac{f(x+h_2) - f(x)}{h_2}. \]
  5. Hence \(s_+(x, h_1) \leq s_+(x, h_2)\).

  6. Hence \(s_+\) is a nondecreasing function of \(h\).

The argument for monotonicity of \(s_-\) is similar.

Observation 9.6 (One sided derivatives as infimum/supremum)

Due to the monotonicity of \(s_+\) over \(h\), we have

\[ f'_+(x) = \lim_{h \downarrow 0} s_+(x, h) = \inf_{h > 0}s_+(x, h). \]

Similarly, due to the monotonicity of \(s_+\) over \(h\), we have

\[ f'_-(x) = \lim_{h \downarrow 0} s_-(x, h) = \sup_{h > 0}s_-(x, h). \]

Theorem 9.200 (Real convex functions and one sided derivatives)

Let \(f: \RR \to \RR\) with \(\dom f = I\) be convex. Let \(a\) and \(b\) be the left and right endpoints of the interval \(I\). Then, for every \(x \in \interior I = (a,b)\), the left hand derivative \(f'_-(x)\) and the right hand derivative \(f'_+(x)\) exist.

If \(a \in I\), then the right hand derivative \(f'_+(a)\) exists. If \(b \in I\), then the left hand derivative \(f'_-(b)\) exists.

If \(a \in I\), we define \(f'_-(a) = -\infty\). If \(b \in I\), we define \(f'_+(b) = \infty\).

Proof. We are given that \(f\) is convex over \((a,b)\).

  1. Let \(x \in \interior I\).

  2. Then there exists \(r > 0\) such that \((x - r, x + r) \subseteq I\).

  3. For \(h > 0\), define

    \[ F(h) = \frac{f(x + h) - f(x)}{h}. \]
  4. Let \(0 < h_1 < h_2\) such that \(h_2 < r\).

  5. Let \(x_1 = x\), \(x_2 = x + h_1\), \(x_3 = x + h_2\).

  6. Since \(f\) is convex, hence by Theorem 9.199

    \[ \frac{f(x_2) - f(x_1)}{x_2 - x_1} \leq \frac{f(x_3) - f(x_1)}{x_3 - x_1}. \]
  7. But that means

\[ \frac{f(x + h_1) - f(x)}{h_1} \leq \frac{f(x + h_2) - f(x)}{h_2}. \]
  1. Thus, whenever \(h_1 < h_2\) (up to \(h_2 < r\)), \(F(h_1) \leq F(h_2)\).

  2. Thus, \(F\) is a nondecreasing (monotone) function of \(h\) in some interval \((0, \delta)\) where \(\delta < r\).

  3. Then, \(f'_+(x) = \lim_{h \downarrow 0} F(h)\) exists.

A similar argument shows that \(f'_-(x)\) also exists. Similar arguments apply for the one sided derivatives at the end points.

9.15.1.3. Continuity#

Theorem 9.201 (Convex real function is continuous)

Let \(f: \RR \to \RR\) be a real convex function with \(\dom f = I\). Let \(a\) and \(b\) be the left and right endpoints of the interval \(I\). Then,

  1. \(f\) is continuous at every \(x \in (a,b)\).

  2. If \(a \in I\), then \(f\) is continuous from the right at \(a\).

  3. If \(b \in I\), then \(f\) is continuous from the left at \(b\).

In other words, \(f\) is continuous on \(I\).

Proof. We proceed as follows.

  1. Let \(x \in (a,b)\).

  2. By Theorem 9.200, the one sided derivatives \(f'_+(x)\) and \(f'_-(x)\) exist.

  3. Then, by limit arithmetic

    \[ \lim_{h \to 0^-} (f(x + h) - f(x)) = \left ( \lim_{h \to 0^-} \frac{f(x + h) - f(x)}{h} \right ) \left ( \lim_{h \to 0^-} h \right) = 0. \]
  4. Similarly,

    \[ \lim_{h \downarrow 0} (f(x + h) - f(x)) = \left ( \lim_{h \downarrow 0} \frac{f(x + h) - f(x)}{h} \right ) \left ( \lim_{h \downarrow 0} h \right) = 0. \]
  5. Thus,

    \[ \lim_{h \to 0^-} (f(x + h) - f(x)) = \lim_{h \downarrow 0} (f(x + h) - f(x)) = 0. \]
  6. Thus, \(f\) is continuous at \(x\).

  7. Since \(x\) was arbitrary, hence \(f\) is continuous on \((a,b)\).

Now consider the case where \(a \in I\).

  1. By Theorem 9.200, the one sided derivative \(f'_+(a)\) exists.

  2. Then, by limit arithmetic

    \[ \lim_{h \to 0^-} (f(a + h) - f(a)) = \left ( \lim_{h \to 0^-} \frac{f(a + h) - f(a)}{h} \right ) \left ( \lim_{h \to 0^-} h \right) = 0. \]
  3. Hence \(f\) is continuous from the right at \(a\).

A similar argument holds for continuity from the left at \(b\).

9.15.1.4. Properties of One Sides Derivatives#

Theorem 9.202 (Properties of one-sided derivatives)

Let \(f: \RR \to \RR\) be a real convex function with \(\dom f = I\). Let \(a\) and \(b\) be the left and right endpoints of the interval \(I\).

  1. We have \(f'_-(x) \leq f'_+(x)\) for every \(x \in I\).

  2. If \(x \in \interior I\) then both \(f'_-(x)\) and \(f'_+(x)\) are finite.

  3. If \(x, z \in I\) and \(x < z\), then \(f'_+(x) \leq f'_-(z)\).

  4. The functions \(f'_-, f'_+ : \RR \to \ERL\) are nondecreasing over \(I\).

  5. The function \(f'_+\) is right-continuous at every interior point of \(I\). If \(a \in I\) then \(f'_+\) is right-continuous at \(a\).

  6. The function \(f'_-\) is left-continuous at every interior point of \(I\). If \(b \in I\) then \(f'_-\) is left-continuous at \(b\).

  7. The function \(f'_+\) is upper-semicontinuous at every \(x \in I\).

  8. The function \(f'_-\) is lower-semicontinuous at every \(x \in I\).

Proof. (1)

  1. If \(a \in I\), then by convention \(f'_-(a) = -\infty\). Hence \(f'_-(a) \leq f'_+(a)\).

  2. If \(b \in I\), then by convention \(f'_+(b) = \infty\). Hence \(f'_-(b) \leq f'_+(b)\).

  3. Now let \(x \in \interior I\).

  4. Then there is \(r > 0\) such that \((x-r, x+r) \in I\).

  5. Pick any \(h > 0\) such that \(h < r\).

  6. Then, using the three points \(x - h, x, x + h\), we have

    \[ \frac{f(x) - f(x - h)}{h} \leq \frac{f(x + h) - f(x)}{h} \]

    due to Theorem 9.199.

  7. Taking the limit \(h \downarrow 0\), we see that

    \[ f'_-(x) \leq f'_+(x) \]

    holds true for every \(x \in \interior I\).

(2)

  1. Let \(x \in \interior I\).

  2. Let \(h > 0\) such that \((x - h, x + h) \subseteq I\).

  3. Then we have

    \[ f'_+(x) \leq \frac{f(x+h) - f(x)}{h} < \infty. \]
  4. Similarly, we have

    \[ - \infty < \frac{f(x) - f(x -h)}{h} \leq f'_-(x). \]
  5. By (1), we have

    \[ -\infty < f'_-(x) \leq f'_+(x) < \infty. \]
  6. Hence both are finite at interior points of \(I\).

(3)

  1. Let \(y = \frac{x + z}{2}\).

  2. Due to Theorem 9.199, we have

    \[ \frac{f(y) - f(x)}{y - x} \leq \frac{f(z) - f(y)}{z - y}. \]
  3. We also have

    \[ f'_+(x) \leq \frac{f(y) - f(x)}{y - x} \text{ and } \frac{f(z) - f(y)}{z - y} \leq f'_-(z). \]
  4. Combining, we get \(f'_+(x) \leq f'_-(z)\).

(4)

  1. Let \(x, z \in I\) such that \(x < z\).

  2. From (3), we have \(f'_+(x) \leq f'_-(z)\).

  3. From (1), we have \(f'_-(z) \leq f'_+(z)\).

  4. Combining, we have \(f'_+(x) \leq f'_+(z)\).

  5. Hence \(f'_+\) is nondecreasing.

  6. Similarly, \(f'_-(x) \leq f'_+(x) \leq f'_-(z)\).

  7. Hence \(f'_-\) is nondecreasing.

(5)

  1. Pick any \(x \in I\) such that \(x \neq b\) (if \(b \in I\)).

  2. Then \(x < b\).

  3. We can pick \(h > 0\) and \(r > 0\) such that \(x + h + r < b\).

  4. Then \(f'_+(x + h) \leq \frac{f(x + h + r) - f(x + h)}{r}\).

  5. We established in Theorem 9.201 that \(f\) is continuous.

  6. Taking the limit \(h \downarrow 0\), we obtain

    \[ \lim_{h \downarrow 0} f'_+(x + h) \leq \frac{f(x + r) - f(x)}{r}. \]

    This is valid since \(f\) is continuous.

  7. Now taking the limit \(r \downarrow 0\) on the R.H.S., we obtain

    \[ \lim_{h \downarrow 0} f'_+(x + h) \leq f'_+(x). \]
  8. Since \(f'_+\) is nondecreasing by claim (4), hence

    \[ f'_+(x) \leq \lim_{h \downarrow 0} f'_+(x + h). \]
  9. Together, we must have

    \[ f'_+(x) = \lim_{h \downarrow 0} f'_+(x + h). \]
  10. Hence \(f'_+\) is right continuous at \(x\).

(6)

  1. An argument similar to (5) shows that \(f'_-\) is left continuous at every \(x \in I\) except for \(x=a\) (if \(a \in I\)).

(7) Upper semicontinuity of \(f'_+\)

  1. We need to show that for every \(\epsilon > 0\) there exists \(r > 0\) such that

    \[ f'_+(y) < f'_+(x) + \epsilon \text{ for every } y \in (x-r, x+r) \cap I. \]
  2. Pick some \(\epsilon > 0\).

  3. Consider any \(x \in \interior I\).

  4. By (6) \(f'_+\) is right continuous at \(x\).

  5. Hence there exists \(r_1 > 0\) such that for every \(y \in [x, x+r_1)\), we have

    \[ |f'_+(y) - f'_+(x)| < \epsilon. \]
  6. By (4), \(f'_+\) is nondecreasing. Hence for every \(y \in [x, x+r_1)\)

    \[ |f'_+(y) - f'_+(x)| = f'_+(y) - f'_+(x). \]
  7. Hence for every \(y \in [x, x+r_1)\), we have

    \[ f'_+(y) - f'_+(x) < \epsilon \iff f'_+(y) < f'_+(x) + \epsilon. \]
  8. Now, let \(r = \min(x - a, r_1)\).

  9. By monotonicity of \(f'_+\), for every \(y \in (x - r, x)\)

    \[ f'_+(y) \leq f'_+(x). \]
  10. Hence for every \(y \in (x, -r, x)\)

    \[ f'_+(y) < f'_+(x) + \epsilon. \]
  11. Combining the two, for every \(y \in (x - r, x + r)\), we have

    \[ f'_+(y) < f'_+(x) + \epsilon. \]
  12. Hence \(f'_+\) is u.s.c. at \(x\).

  13. Now, if \(a \in I\) then let \(x = a\).

  14. By right continuity of \(f'_+\) at \(a\), there exists \(r > 0\) such that for every \(y \in [a, a + r)\), we have

    \[ f'_+(y) < f'_+(a) + \epsilon. \]
  15. Also \((a - r, a + r) \cap I = [a, a + r)\).

  16. Hence \(f'_+\) is u.s.c. at \(a\).

  17. If \(b \in I\), then by convention \(f'_+(b) = \infty\).

  18. Hence \(f'_+\) is u.s.c. at \(b\).

(8) Lower semicontinuity of \(f'_-\)

  1. The argument is similar to (7).

9.15.2. Directional Derivatives of Proper Functions#

Definition 9.70 (Directional derivative)

Let \(f : \VV \to \RERL\) be a proper function with \(S = \dom f\). Let \(\bx \in \interior S\). The directional derivative at \(\bx\) in the direction \(\bd \in \VV\) is defined by

\[ f'(\bx;\bd) \triangleq \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha \bd) - f(\bx)}{\alpha} \]

provided the limit exists.

We say that \(f\) is directionally differentiable at \(\bx\) if it is directionally differentiable in every direction at \(\bx\).

The directional derivative is a scalar quantity (\(\in \RR\)) if it is defined (i.e., the limit exists). When we say that

\[ f'(\bx;\bd) = \lim_{t \downarrow 0} \frac{f(\bx + t \bd) - f(\bx)}{t}, \]

we mean that \(f\) is defined over a set \(\{ \bv \ST \bv = \bx + t \bd, 0 < t < t_{\max} \}\) and for every \(\epsilon > 0\), there exists \(\delta > 0\) such that

\[ \left | \frac{f(\bx + t \bd) - f(\bx)}{t} - f'(\bx;\bd) \right | < \epsilon \text{ whenever } 0 < t < \delta. \]

Since \(\bx \in \interior S\), hence there exists \(r > 0\) such that \(B(\bx, r) \subseteq S\).

With \(\bv \in B(\bx, r)\), we need \(\| t \bd \| < r\). Thus, a \(t_{\max} = \frac{r}{\| \bd \|}\) is a suitable range of allowed values for \(t\). Accordingly, \(0 < \delta < t_{\max}\) can be chosen.

Remark 9.14 (Directional derivative for zero vector)

If \(\bd = \bzero\) then, \(f'(\bx; \bd) = 0\).

We can see this from the fact that

\[ f'(\bx;\bzero) = \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha \bzero) - f(\bx)}{\alpha} = 0. \]

A useful result is for computing the directional derivative of a function which is the pointwise maximum of a finite number of proper functions.

We recall from Theorem 3.22 that the interior of a finite intersection of sets is the intersection of their interiors. This is useful in identifying the interior of the domain for a pointwise maximum of a finite set of functions.

9.15.3. Differentiability#

9.15.3.1. Differentiability of Proper Functions#

Definition 9.71 (Differentiability of proper functions)

Let \(f : \VV \to \RERL\) be a proper function with \(S = \dom f\). Let \(\bx \in \interior S\). \(f\) is said to be differentiable at \(\bx\) if there exists \(\bg \in \VV^*\) such that:

(9.5)#\[\underset{\bh \to \bzero}{\lim} \frac{f(\bx + \bh) - f(\bx) - \langle \bh, \bg \rangle}{\| \bh \|} = 0.\]

The unique vector \(\bg\) satisfying this condition is called the gradient of \(f\) at \(\bx\) and is denoted by \(\nabla f(\bx)\).

If \(f\) is differentiable at some \(\bx \in \interior S\), then there is a simple formula to connect the gradient and the directional derivatives.

9.15.3.2. Gradient and Directional Derivatives#

Theorem 9.203 (Gradient and directional derivatives)

Let \(f : \VV \to \RERL\) be a proper function with \(S = \dom f\). Let \(\bx \in \interior S\). Assume that \(f\) is differentiable at \(\bx\). Then, for any \(\bd \in \VV\),

\[ f'(\bx; \bd) = \langle \bd, \nabla f(\bx) \rangle. \]

In other words, the directional derivative is the projection of the gradient in the specified direction.

Proof. For \(\bd = \bzero\), the equality is obvious. We shall consider the case where \(\bd \neq \bzero\).

  1. Since \(f\) is differentiable at \(\bx\), hence

\[ \underset{\bh \to \bzero}{\lim} \frac{f(\bx + \bh) - f(\bx) - \langle \bh, \nabla f(\bx) \rangle}{\| \bh \|} = 0. \]
  1. In particular, if we take the limit of \(\bh\) along the direction of \(\bd\) as \(t \bd\) where \(t > 0\) and \(t \to 0^+\), then

    \[ \underset{t \to 0^+}{\lim} \frac{f(\bx + t \bd) - f(\bx) - \langle t \bd, \nabla f(\bx) \rangle}{\| t \bd \|} = 0. \]
  2. Splitting the terms, we get

    \[ \underset{t \to 0^+}{\lim} \frac{f(\bx + t \bd) - f(\bx)}{\| t \bd \|} - \underset{t \to 0^+}{\lim} \frac{\langle \bd, \nabla f(\bx) \rangle}{\| \bd \|} = 0. \]
  3. Multiplying with \(\| \bd \|\) and simplifying, we get:

    \[ \underset{t \to 0^+}{\lim} \frac{f(\bx + t \bd) - f(\bx)}{t} - \underset{t \to 0^+}{\lim} \langle \bd, \nabla f(\bx) \rangle = 0. \]
  4. Thus,

    \[ f'(\bx; \bd ) = \underset{t \to 0^+}{\lim} \frac{f(\bx + t \bd) - f(\bx)}{t} = \langle \bd, \nabla f(\bx) \rangle. \]

9.15.3.3. Gradient in \(\RR^n\)#

Remark 9.15 (Gradient in \(\RR^n\))

It is imperative to compare the definition of gradients in this section with Definition 5.1 (differentiability of functions from \(\RR^n\) to \(\RR^m\)) and the notion of the gradient as defined in Definition 5.4.

To better develop our understanding of gradients, let us examine the gradient in the Euclidean space \(\RR^n\). The standard basis is given by \(\BBB = \{\be_1, \dots, \be_n \}\) which are the coordinate unit vectors. The standard inner product is given by the dot product

\[ \langle \bx, \by \rangle = \by^T \bx = \bx^T \by. \]

A vector \(\bx \in \RR^n\) is written as

\[ \bx = \sum_{i=1}^n x_i \be_i. \]

The individual coordinates are obtained via

\[ x_i = \langle \be_i, \bx \rangle = \langle \bx, \be_i \rangle = \bx^T \be_i \Forall i \in [1,\dots, n]. \]

Let \(f: \RR^n \to \RERL\) be a proper function. Let \(S = \dom f\). Let \(\bx \in \interior S\). Assume that \(f\) is differentiable at \(\bx\). Let \(\bg = \nabla f (\bx)\). Let

\[ \bg = \sum_i^n g_i \be_i. \]

Following the notation in Observation 5.1, the derivative of \(f\) at \(\bx\), denoted by \(Df(\bx)\) is given by

\[ \underset{\bh \to \bzero}{\lim} \frac{\| f(\bx + \bh) - f(\bx) - Df(\bx) \bh \|_2}{\| \bh \|_2} = 0. \]

We don’t have to check for \(\bx + \bh \in \dom f\) as \(f\) is a proper function with a value of \(\infty\) at points outside its effective domain.

Compare this with (9.5). For \(f : \RR^n \to \RR\), \(Df(\bx)\) is a row vector. If we let \(\tilde{\bg} = Df(\bx)^T\), then

\[ Df(\bx) \bh = \tilde{\bg}^T \bh = \langle \bh, \tilde{\bg} \rangle. \]

Then, the definition of \(\tilde{\bg}\) in the limit above is exactly the same as \(\bg\) in (9.5). Thus, \(\tilde{\bg} = \bg\). We can see that our definition of gradient coincides with the definition in Definition 5.4 for \(\RR^n\) with the dot product as standard inner product.

Now consider the components of \(\bg\).

\[ g_i = \langle \be_i, \bg \rangle = \langle \be_i, \nabla f (\bx) \rangle. \]

By Theorem 9.203, the directional derivative in the direction \(\be_i\) is given by

\[ f'(\bx; \be_i) = \langle \be_i, \nabla f(\bx) \rangle = \langle \be_i, \bg \rangle = g_i. \]

Thus,

\[ \frac{\partial f(\bx)}{\partial x_i} = \langle \be_i, \nabla f(\bx) \rangle. \]

The partial derivatives of \(f\) at \(\bx\) along the standard basis vectors are identical to the directional derivatives of \(f\).

\[\begin{split} \nabla f(\bx) = \begin{bmatrix} \frac{\partial f(\bx)}{\partial x_1}\\ \vdots\\ \frac{\partial f(\bx)}{\partial x_n} \end{bmatrix} = \begin{bmatrix} \langle \be_1, \nabla f(\bx) \rangle\\ \vdots\\ \langle \be_n, \nabla f(\bx) \rangle \end{bmatrix}. \end{split}\]

Then, for an arbitrary direction \(\bd = \sum_{i=1}^n \be_i\), the directional derivative becomes

\[ f'(\bx; \bd) = \langle \bd, \nabla f(\bx) \rangle = \nabla f(\bx)^T \bd = \sum_{i=1}^n \frac{\partial f(\bx)}{\partial x_i} d_i. \]

Recall from Definition 9.70, that the directional derivative is independent on the choice of the inner product. This is also clear from the expression \(\sum_{i=1}^n \frac{\partial f(\bx)}{\partial x_i} d_i\) as the partial derivatives are independent of the choice of the inner product.

However, this means that the gradient itself must depend on the choice of inner product. If \(\langle \cdot, \cdot \rangle_a\) and \(\langle \cdot, \cdot \rangle_b\) are two different inner products defined on \(\RR^n\), then the gradients of \(f\) at \(\bx\) w.r.t. the two inner products, denoted by \(\nabla_a f(\bx)\) and \(\nabla_b f(\bx)\) must satisfy the relationship

\[ f'(\bx; \bd) = \langle \bd, \nabla_a f(\bx) \rangle_a = \langle \bd, \nabla_b f(\bx) \rangle_b \Forall \bd \in \VV. \]

In the following, we shall assume that \(\nabla f(\bx)\) denotes the gradient w.r.t. the dot product.

Consider the inner product given by

\[ \langle \bx, \by \rangle_H = \bx^T \bH \by \]

where \(\bH \in \RR^{n \times n}\) is a symmetric positive definite matrix.

Then,

\[\begin{split} (\nabla_H f(\bx))_i &= \nabla_H f(\bx)^T \be_i & \text{coordinate in standard basis}\\ &= \nabla_H f(\bx)^T (\bH \bH^{-1}) \be_i & \text{$\bH$ is invertible}\\ &= \nabla_H f(\bx)^T \bH (\bH^{-1}\be_i) & \\ &= \langle \bH^{-1}\be_i, \nabla_H f(\bx) \rangle_H & \text{by definition of this inner product} \\ &= f'(\bx; \bH^{-1}\be_i) & \text{directional derivative w.r.t. this inner product} \\ &= \nabla f(\bx)^T \bH^{-1}\be_i & \text{directional derivative w.r.t. dot product}\\ &= (\bH^{-1} \nabla f(\bx))^T \be_i & \text{$\bH$ is symmetric}. \end{split}\]

Thus,

\[ \nabla_H f(\bx) = \bH^{-1} \nabla f(\bx). \]

Thus, the gradient w.r.t. the inner product \(\langle \cdot, \cdot \rangle_H\) is the scaled version of the standard gradient where the scaling factor is \(\bH^{-1}\).

9.15.3.4. Gradient in \(\RR^{m \times n}\)#

Remark 9.16 (Gradient in \(\RR^{m \times n}\))

We next look at the vector space of real matrices. The standard basis is a family of unit matrices \(\{ \bE_{i j} \}_{1 \leq i \leq m, 1 \leq j \leq n}\) where \(\bE_{i j}\) has the \((i,j)\)-th entry as 1 and other entries as 0.

The standard inner product is given by

\[ \langle \bX , \bY \rangle = \Trace(\bY^T \bX ) \Forall \bX, \bY \in \RR^{m \times n}. \]

Let \(f : \RR^{m \times n} \to \RR\) be a proper function. Let \(S = \dom f\). Let \(\bX \in \interior S\). Assume that \(f\) is differentiable at \(\bX\).

The gradient is given by

\[ \partial f(\bX) = \left ( \frac{\partial f(\bX)}{\partial x_{i j}} \right )_{i j }. \]

The directional derivative for some direction \(\bD \in \RR^{m \times n}\) is given by

\[ f(\bX ; \bD) = \langle \bD, \partial f(\bX) \rangle = \Trace(\partial f(\bX)^T \bD). \]

Consider the inner product given by

\[ \langle \bX, \bY \rangle_H = \Trace(\bX^T \bH \bY) \]

where \(\bH \in \RR^{m \times m}\) is a symmetric positive definite matrix.

Then,

\[\begin{split} (\nabla_H f(\bX))_{i j} &= \Trace(\nabla_H f(\bX)^T \bE_{i j}) & \text{coordinate in standard basis}\\ &= \Trace(\nabla_H f(\bX)^T (\bH \bH^{-1}) \bE_{i j}) & \text{$\bH$ is invertible}\\ &= \Trace(\nabla_H f(\bX)^T \bH (\bH^{-1}\bE_{i j}) ) & \\ &= \langle \bH^{-1}\bE_{i j}, \nabla_H f(\bX) \rangle_H & \text{by definition of this inner product} \\ &= f'(\bX; \bH^{-1}\bE_{i j}) & \text{directional derivative w.r.t. this inner product} \\ &= \Trace(\nabla f(\bX)^T \bH^{-1}\bE_{i j}) & \text{directional derivative w.r.t. standard inner product}\\ &= (\bH^{-1} \nabla f(\bX))^T \bE_{i j} & \text{$\bH$ is symmetric}. \end{split}\]

Thus,

\[ \nabla_H f(\bX) = \bH^{-1} \nabla f(\bX). \]

9.15.4. Proper Convex Functions#

9.15.4.1. Existence of Directional Derivatives#

An important property of directional derivatives is that if \(f\) is a proper convex function then \(f\) is directionally differentiable at every \(\bx \in \interior \dom f\).

Theorem 9.204 (Existence of directional derivatives for convex functions.)

Let \(f: \VV \to \RERL\) be a proper convex function with \(S = \dom f\). Let \(\bx \in \interior S\). Then, for any \(\bd \in \VV\), the directional derivative \(f'(\bx; \bd)\) exists.

Proof. This is a consequence of the directional differentiability of the scalar convex functions.

  1. Define the convex function \(F : \RR \to \RR\) as

    \[ F(t) = f(\bx + t \bd). \]
  2. Let \(I = \dom F\).

  3. Then \(I\) is an interval of values for which \(\bx + t \by \in S\).

  4. Since \(\bx \in \interior S\), hence \(t=0 \in \interior I\).

  5. We now note that

    \[ f'(\bx; \bd) = \lim_{t \downarrow 0} \frac{f(\bx + t \bd) - f(\bx)}{t} = \lim_{t \downarrow 0} \frac{F(t) - F(0)}{t} = F'_+ (0). \]

    It is the right hand derivative of \(F\) at \(t=0\).

  6. By Theorem 9.200, \(F'_+(0)\) exists.

  7. Hence \(f'(\bx; \bd)\) exists for every \(\bx \in \interior S\) and every \(\bd \in \VV\).

Observation 9.7 (Relation between the directional derivatives in opposite directions)

We can see that

\[ f'(\bx; -\bd) = \lim_{t \downarrow 0} \frac{f(\bx - t \bd) - f(\bx)}{t} = \lim_{t \downarrow 0} \frac{F(-t) - F(0)}{t} = - \lim_{r \uparrow 0} \frac{F(r) - F(0)}{r} = - F'_- (0). \]

Hence

\[ F'_- (0) = - f'(\bx; -\bd). \]

By Theorem 9.202

\[ F'_- (0) \leq F'_+(0). \]

Hence

\[ - f'(\bx; -\bd) \leq f'(\bx; \bd) \Forall \bd \in \VV. \]

Observation 9.8 (Directional derivative as infimum)

Let \(f: \VV \to \RERL\) be a proper convex function with \(S = \dom f\). Let \(\bx \in \interior S\). Then, for any \(\bd \in \VV\),

\[ f'(\bx; \bd) = \inf_{t > 0} \frac{f(\bx + t \bd) - f(\bx)}{t}. \]

This follows from the fact that \(F(t) = f(\bx + t \bd)\) is convex and due to Observation 9.6,

\[ f'(\bx; \bd) = F'_+(0) = \inf_{t > 0} \frac{F(t) - F(0)}{t} = \inf_{t > 0} \frac{f(\bx + t \bd) - f(\bx)}{t}. \]

9.15.4.2. Upper Semicontinuity#

The next result generalizes the upper semicontinuity property of the right hand derivatives of real convex functions.

Theorem 9.205

Let \(f: \VV \to \RERL\) be a proper convex function with \(\dom f = S\). Assume that \(S\) is an open subset of \(\VV\). Let \(\{ f_k \}\) be a sequence of proper convex functions \(f_k : \VV \to \RERL\) with \(\dom f_k = S\) with the property that

\[ \lim_{k \to \infty} f_k(\bx_k) = f(\bx) \]

holds true for every \(\bx \in S\) and every sequence \(\{ \bx_k \}\) of \(S\) that converges to \(\bx\). Then for any \(\bx \in S\) and any direction \(\bd \in \VV\) and any sequences \(\{ \bx_k \}\) of \(S\) and \(\{ \bd_k \}\) of \(\VV\) converging to \(\bx\) and \(\bd\) respectively, we have

\[ \limsup_{k \to \infty} f'_k (\bx_k; \bd_k ) \leq f'(\bx; \bd). \]

Furthermore if \(f\) is differentiable over \(S\), then \(f\) is also continuously differentiable over \(S\).

Proof. Limit superior

  1. Choose any \(\epsilon > 0\).

  2. By definition of the directional derivative, there exists a \(t > 0\) such that

    \[ \frac{f(\bx + t \bd) - f(\bx)}{t} < f'(\bx; \bd) + \epsilon. \]
  3. Due to Observation 9.8, for every \(k\) and every \(t > 0\), we have

    \[ f'_k(\bx_k; \bd_k) \leq \frac{f_k(\bx_k + t \bd_k) - f(\bx_k)}{t}. \]
  4. Now

    \[ \lim_{k \to \infty} \frac{f_k(\bx_k + t \bd_k) - f(\bx_k)}{t} = \frac{f(\bx + t \bd) - f(\bx)}{t}. \]
  5. Hence for sufficiently large \(k\), we have

    \[ f'_k(\bx_k; \bd_k) \leq \frac{f_k(\bx_k + t \bd_k) - f(\bx_k)}{t} < f'(\bx; \bd) + \epsilon. \]
  6. By taking the limit superior on the L.H.S. as \(k \to \infty\), we have

    \[ \limsup_{k \to \infty} f'_k (\bx_k; \bd_k ) \leq f'(\bx; \bd) + \epsilon. \]
  7. Since this is valid for every \(\epsilon > 0\), hence we must have

    \[ \limsup_{k \to \infty} f'_k (\bx_k; \bd_k ) \leq f'(\bx; \bd) \]

    as desired.

Continuous differentiability

  1. We are given that \(f\) is differentiable over \(S\).

  2. Then \(f\) is also continuous over \(S\).

  3. Let \(\bx \in S\).

  4. Let \(\{ \bx_k \}\) be a sequence of \(S\) converging to \(\bx\).

  5. Let \(\bd \in \VV\) be any nonzero direction.

  6. Due to Theorem 9.203, for every \(k\),

    \[ f'(\bx_k; \bd) = \langle \bd, \nabla f(\bx_k) \rangle. \]
  7. Hence

    \[\begin{split} \limsup_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle &= \limsup_{k \to \infty} f'(\bx_k; \bd)\\ &\leq f'(\bx; \bd) \\ &= \langle \bd, \nabla f(\bx) \rangle. \end{split}\]
  8. By replacing \(\bd\) with \(-\bd\) in the previous argument, we have

    \[\begin{split} - \liminf_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle &= \limsup_{k \to \infty} \langle -\bd, \nabla f(\bx_k) \rangle\\ &= \limsup_{k \to \infty} f'(\bx_k; -\bd)\\ &\leq f'(\bx; -\bd)\\ &= -\langle \bd, \nabla f(\bx) \rangle. \end{split}\]
  9. Hence

    \[ \liminf_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle \geq \langle \bd, \nabla f(\bx) \rangle. \]
  10. Thus we have

    \[ \limsup_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle \leq \liminf_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle. \]
  11. But then this must be an equality. Hence

    \[ \lim_{k \to \infty} \langle \bd, \nabla f(\bx_k) \rangle = \langle \bd, \nabla f(\bx) \rangle. \]
  12. Since this is valid for every nonzero direction \(\bd \in \VV\), hence we must have

    \[ \lim_{k \to \infty} \nabla f(\bx_k) = \nabla f(\bx). \]
  13. Hence \(\nabla f\) is continuous at every \(\bx \in S\).

  14. Hence \(f\) is continuously differentiable at every \(\bx \in S\).

9.15.4.3. Directional Derivatives Map#

The existence of directional derivatives in all directions allows us to consider a mapping from a direction \(\bd \in \VV\) to the directional derivative of \(f\) in this direction at \(\bx\). We can define a directional derivative map parameterized by \(\bx \in S\) as:

\[ \bg_x(\bd) \triangleq f'(\bx; \bd) = \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha \bd) - f(\bx)}{\alpha}. \]

We shall refer to such maps by \(\bd \mapsto f'(\bx; \bd)\).

Theorem 9.206 (Convexity and homogeneity of \(\bd \mapsto f'(\bx; \bd)\))

Let \(f: \VV \to \RERL\) be a proper convex function with \(S = \dom f\). Let \(\bx \in \interior S\). Then, the function \(\bd \mapsto f'(\bx; \bd)\) is convex and nonnegative homogeneous.

Nonnegative homogeneity: For any \(t \geq 0\) and \(\bd \in \VV\),

\[ f'(\bx; t \bd) = t f'(\bx; \bd). \]

Proof. Convexity

  1. Let \(\bd_1, \bd_2 \in \VV\) and \(t \in (0, 1)\).

  2. Let \(\bd = t \bd_1 + (1-t) \bd_2\).

  3. Then,

    \[\begin{split} f'(\bx; \bd) &= f'(\bx; t \bd_1 + (1-t) \bd_2)\\ &= \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha [t \bd_1 + (1-t) \bd_2]) - f(\bx)}{\alpha}\\ &= \lim_{\alpha \downarrow 0} \frac{f(t \bx + \alpha t \bd_1 + (1-t) \bx + \alpha (1-t) \bd_2) - f(\bx)}{\alpha}\\ &= \lim_{\alpha \downarrow 0} \frac{f(t (\bx + \alpha \bd_1) + (1-t) (\bx + \alpha \bd_2)) - f(\bx)}{\alpha}\\ &\leq \lim_{\alpha \downarrow 0} \frac{t f(\bx + \alpha \bd_1) + (1-t) f(\bx + \alpha \bd_2) - t f(\bx) - (1-t)f(\bx)}{\alpha}\\ &= t \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha \bd_1) - f(\bx)}{\alpha} + (1-t) \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha \bd_2) - f(\bx)}{\alpha}\\ &= t f'(\bx; \bd_1) + (1-t) f'(\bx; \bd_2). \end{split}\]

    We used the convexity property of \(f\) in this derivation.

  4. Thus, \(f'(\bx; \bd)\) is convex.

Nonnegative homogeneity

  1. For \(t=0\),

    \[ f'(\bx, 0 \bd) = f'(\bx, \bzero) = 0 = 0 f'(\bx; \bd). \]

    Thus, the homogeneity property is trivial for \(t=0\).

  2. Now consider \(t > 0\).

  3. Then,

    \[\begin{split} f'(\bx; t \bd) &= \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha t \bd) - f(\bx)}{\alpha}\\ &= t \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha t \bd) - f(\bx)}{\alpha t}\\ &= t f'(\bx; \bd). \end{split}\]
  4. Thus, \(f'(\bx; \bd)\) is nonnegative homogeneous.

9.15.4.4. As Linear Underestimator#

Directional derivatives are a linear underestimator for convex functions.

Theorem 9.207 (Directional derivative as linear underestimator)

Let \(f: \VV \to \RERL\) be a proper convex function with \(S = \dom f\). Let \(\bx \in \interior S\). Then, for every \(\by \in S\)

\[ f(\by) \geq f(\bx) + f'(\bx; \by - \bx). \]

Proof. Note that

\[\begin{split} f'(\bx; \by - \bx) &= \lim_{\alpha \downarrow 0} \frac{f(\bx + \alpha (\by - \bx)) - f(\bx)}{\alpha}\\ &= \lim_{\alpha \downarrow 0} \frac{f((1-\alpha) \bx + \alpha \by) - f(\bx)}{\alpha}\\ &\leq \lim_{\alpha \downarrow 0} \frac{(1-\alpha)f(\bx) + \alpha f(\by) - f(\bx)}{\alpha}\\ &= \lim_{\alpha \downarrow 0} \frac{\alpha (f(\by) - f(\bx))}{\alpha}\\ &= \lim_{\alpha \downarrow 0}(f(\by) - f(\bx)) = f(\by) - f(\bx). \end{split}\]

Thus,

\[ f(\by) \geq f(\bx) + f'(\bx; \by - \bx). \]

9.15.5. Pointwise Maximum of Finite Set of Functions#

9.15.5.1. Directional Derivative#

Theorem 9.208 (Directional derivative of a maximum of functions)

Let \(f_1, f_2, \dots, f_m : \VV \to \RERL\) be proper functions. Let \(f : \VV \to \RERL\) be defined as

\[ f(\bx) = \max\{f_1(\bx), \dots, f_m(\bx) \} \]

with \(\dom f = \bigcap_{i=1}^m \dom f_i\).

Let \(\bx \in \interior \dom f = \bigcap_{i=1}^m \interior \dom f_i\) and \(\bd \in \VV\). Assume that \(f'(\bx; \bd)\) exists for every \(i \in 1,\dots,m\).

Let \(I(\bx) = \{i \in 1,\dots,m \ST f_i(\bx) = f(\bx) \}\) be the set of indices of functions whose value at \(\bx\) equals \(f(\bx)\). Then,

\[ f'(\bx; \bd) = \underset{i \in I(\bx)}{\max} f'(\bx; \bd). \]

In other words, the directional derivative of a pointwise maximum of functions equals the maximum of directional directives of functions which attain the pointwise maximum at a specific point.

Proof. The key idea here is that for computing the directional derivative \(f'(\bx; \bd)\), only those functions are relevant for which \(f_i(\bx) = f(\bx)\). We need to show this first.

  1. Since \(\bx \in \interior \dom f\), there exists \(B(\bx, r)\) such that \(f\) and \(f_i\) are all defined over this open ball.

  2. Let \(s = \frac{r}{\| \bd \|}\).

  3. For every \(i \in 1,\dots,m\), let \(g_i : \RR \to \RR\) be defined as

    \[ g_i(t) = f_i(\bx + t \bd) \]

    with \(\dom g_i = [0, s)\). \(\| s \bd \| = r\). Thus, \(\bx + t \bd \in B(\bx, r)\). Hence, \(g_i\) are well defined.

  4. Then,

    \[\begin{split} \lim_{t \to 0+} g_i(t) &= \lim_{t \to 0+} f_i(\bx + t \bd) \\ &= \lim_{t \to 0+} [(f_i(\bx + t \bd) - f_i(\bx)) + f_i(\bx)]\\ &= \lim_{t \to 0+}\left [ t \frac{f_i(\bx + t \bd) - f_i(\bx)}{t} + f_i(\bx) \right] \\ &= 0 \cdot f'_i(\bx; \bd) + f_i(\bx)\\ &= f_i(\bx) = g_i(0). \end{split}\]

    We used the fact that \(f'_i(\bx; \bd)\) exists for every \(f_i\).

  5. Thus, \(g_i\) is continuous from the right at \(t=0\) for every \(i \in 1,\dots, m\).

  6. Let \(i \in I(\bx)\) and \(j \notin I(\bx)\).

  7. Then, \(f_i(\bx) > f_j (\bx)\). Alternatively \(g_i(0) > g_j(0)\).

  8. Since \(g_i, g_j\) are continuous from the right, hence there exists \(\epsilon_{i j} > 0\) such that \(g_i(t) > g_j(t)\) for every \(t \in [0, \epsilon_{i j}]\).

  9. Minimizing \(\epsilon_{ij}\) over all pairs of \(i \in I(\bx)\) and \(j \notin I(\bx)\), there exists \(\epsilon > 0\) such that for any \(i \in I(\bx)\) and \(j \notin I(\bx)\),

    \[ f_i(\bx + t \bd) = g_i(t) > g_j(t) = f_j(\bx + t \bd) \Forall t \in [0, \epsilon]. \]

We can now compute the directional derivative.

  1. For every \(t \in [0, \epsilon]\),

    \[ f(\bx + t \bd) = \underset{i=1,\dots,m}{\max} f_i(\bx + t \bd) = \underset{i \in I(\bx)}{\max} f_i(\bx + t \bd). \]
  2. Consequently, for any \(t \in (0, \epsilon]\)

    \[\begin{split} \frac{f(\bx + t \bd) - f(\bx)}{t} &= \frac{\underset{i \in I(\bx)}{\max} f_i(\bx + t \bd) - f(\bx)}{t}\\ &= \frac{\underset{i \in I(\bx)}{\max} (f_i(\bx + t \bd) - f_i(\bx))}{t}\\ &= \underset{i \in I(\bx)}{\max} \frac{f_i(\bx + t \bd) - f_i(\bx)}{t}. \end{split}\]

    We used the fact that \(f_i(\bx) = f(\bx)\) for every \(i \in I(\bx)\).

  3. Taking the limit \(t \downarrow 0\),

    \[\begin{split} f'(\bx; \bd) &= \lim_{t \downarrow 0} \frac{f(\bx + t \bd) - f(\bx)}{t}\\ &= \lim_{t \downarrow 0} \underset{i \in I(\bx)}{\max} \frac{f_i(\bx + t \bd) - f_i(\bx)}{t}\\ &= \underset{i \in I(\bx)}{\max} \lim_{t \downarrow 0} \frac{f_i(\bx + t \bd) - f_i(\bx)}{t}\\ &= \underset{i \in I(\bx)}{\max} f'_i(\bx; \bd). \end{split}\]

9.15.5.2. Finite Set of Convex Functions Case#

Theorem 9.209 (Directional derivative of pointwise maximum of convex functions)

Let \(f_1, f_2, \dots, f_m : \VV \to \RERL\) be proper convex functions. Let \(f : \VV \to \RERL\) be defined as

\[ f(\bx) = \max\{f_1(\bx), \dots, f_m(\bx) \} \]

with \(\dom f = \bigcap_{i=1}^m \dom f_i\).

Let \(\bx \in \interior \dom f\) and \(\bd \in \VV\). Then,

\[ f'(\bx; \bd) = \underset{i \in I(\bx)}{\max} f'(\bx; \bd). \]

where \(I(\bx) = \{i \in 1,\dots,m \ST f_i(\bx) = f(\bx) \}\).

Proof. Since \(f_i\) are proper convex, hence their pointwise maximum \(f\) is proper convex.

By Theorem 9.204, the directional derivatives \(f'(\bx; \bd)\) and \(f'_i(\bx; \bd)\) for \(i=1,\dots,m\) exist.

By Theorem 9.208,

\[ f'(\bx; \bd) = \underset{i \in I(\bx)}{\max} f'(\bx; \bd). \]

where \(I(\bx) = \{i \in 1,\dots,m \ST f_i(\bx) = f(\bx) \}\).