Cauchy Integral Theorem

kuniga.me > NP-Incompleteness > Cauchy Integral Theorem

Cauchy Integral Theorem

26 Apr 2024

portrait of Cauchy from Wikipedia


This is our third post in the series with my notes on complex integration, corresponding to Chapter 4 in Ahlfors’ Complex Analysis.

The Cauchy integral theorem provides conditions under which the integral over a closed curve is zero.

The previous posts from the series:

  1. Complex Integration
  2. Path-Independent Line Integrals

Recap

In the previous post [3] we ended with the following Corollary 2 stating that:

The complex line integral $\int_\gamma f(z)dz$, defined in $\Omega$, depends only on the endpoints of $\gamma$ if and only if $f$ is the derivative of some holomorphic function in $\Omega$.

Another corollary is the following:

Corollary 1. Let $f(z)$ be a function defined in $\Omega$. Then

\[\int_\gamma f(z)dz = 0\]

If and only if $f$ is the derivative of some holomorphic function $F$ in $\Omega$.

The general idea of Cauchy’s theorem that we’ll cover in this post is that we only need $f$ itself to be holomorphic in $\Omega$, for special types of the region $\Omega$.

A result we haven’t proved yet says that the derivative of a holomorphic function is itself holomorphic, but not all holomorphic functions are derivatives of of a holomorphic function. So Cauchy’s theorem is a stronger result.

In a rectangle

We first consider the case where $\Omega$ is a rectangle $R$ defined by

\[a \le x \le b, c \le y \le d\]

The curve we’ll use is the border of $R$, denoted by $\partial R$, with a clock-wise orientation as depicted in Figure 1. In other words, it’s the segments $(a, c) \rightarrow (b, c)$, $(b, c) \rightarrow (b, d)$, $(b, d) \rightarrow (a, d)$ and $(a, d) \rightarrow (a, c)$.

See caption.
Figure 1: Reference rectangle with clockwise orientation

Theorem 1. If the function $f(z)$ is holomorphic in $R$, then

\[\int_{\partial R} f(z) dz = 0\] In [1], the author defines the notation $$\eta(R) = \int_{\partial R} f(z) dz$$ for any given $R$ and when a rectangle is partitioned into four coungruent rectangles (Figure 1.1), we can refer to them as $R^{(1)}, R^{(2)}, R^{(3)}, R^{(4)}$.
See caption.
Figure 1.1: Partition into congruent subrectangles, recursively.
We then have that $$(1.1) \quad \eta(R) = \eta(R^{(1)}) + \eta(R^{(2)}) + \eta(R^{(3)}) + \eta(R^{(4)})$$ The idea is that for a segment shared between two sub-recangles they have opposite directions, so their integral cancels out. The only segments left uncancelled are those belonging to the original rectangle. By the triangle inequality, we thus have: $$(1.2) \quad \abs{\eta(R)} \le \abs{\eta(R^{(1)})} + \abs{\eta(R^{(2)})} + \abs{\eta(R^{(3)})} + \abs{\eta(R^{(4)})}$$ We choose the subrectangle with the largest modulus and call it $R_1$. For $k = 1, 2, 3, 4$ we have: $$\abs{\eta(R_1)} \ge \abs{\eta(R^{(k)})}$$ So that $$4 \abs{\eta(R_1)} \ge \abs{\eta(R^{(1)})} + \abs{\eta(R^{(2)})} + \abs{\eta(R^{(3)})} + \abs{\eta(R^{(4)})} \ge \abs{\eta(R)}$$ Or $$\abs{\eta(R_1)} \ge \frac{1}{4} \abs{\eta(R)}$$ If we repeat this process recursively for $R_n$ and obtain $R_{n+1}$, we'll get more generally that: $$\abs{\eta(R_n)} \ge \frac{1}{4} \abs{\eta(R_{n-1})}$$ Which leads us to $$(1.3) \quad \abs{\eta(R_n)} \ge 4^{-n} \abs{\eta(R)}$$ As $n$ grows, the subrectangle $R_n$ converges to a point which we'll call $z^*$ (shown in Figure 1.1). Since $f(z)$ is holomorphic in $R$ by hypothesis, the limit $$f'(z^*) = \lim_{z \rightarrow z^*} \frac{f(z) - f(z^*)}{z - z^*}$$ exists, that is, for every $\epsilon$, there exists $\delta > 0$ such that $\abs{z - z^*} \lt \delta$ such that: $$\abs{f'(z^*) - \frac{f(z) - f(z^*)}{z - z^*}} \lt \epsilon$$ Or $$(1.4) \quad \abs{f(z) - f(z^*) + f'(z*) (z - z*)} \lt \epsilon \abs{z - z*}$$ Now consider the open circle $\abs{z - z^*} \lt \delta$ (i.e. the set of points for which $(1.4)$ holds). We can choose $n$ sufficiently large we can guarantee that $R_n$ is inside that open circle and that all $z$ in $\partial R_n$ satisfy $(1.4)$.

So far we have $\abs{z - z^*} \lt \delta$, but we can find a tigher upper bound. The maximum distance between two points in a rectangle is its diagonal. Let $\Delta$ be the diagonal of $R$. Every time we pick a subrectangle the diagonal is halved, so the diagonal of $R_n$ is $2^{-n} \Delta$. Thus if $z$ and $z^*$ are in $R_n$, $\abs{a - z^*} \le 2^{-n} \Delta$. We have then: $$(1.5) \quad \abs{f(z) - f(z^*) + f'(z*) (z - z*)} \lt \epsilon \abs{z - z*} \le \epsilon 2^{-n} \Delta$$ In [2] we saw in the last example that for any curve $\gamma$, $$\int_\gamma (z - a)^{n}dz = 0$$ For $n \gt 0$. This let's us conclude the following: $$(1.6) \quad \int_{\partial R_n} (z - z^*) dz = 0$$ By replacing $a$ with $z^*$, $\gamma$ with $\partial R$, and setting $n = 1$. Similarly, if we do it for $n = 0$: $$(1.7) \quad \int_{\partial R_n} dz = 0$$ Since $-f(z*)$ exists, we can multiply it by $(1.7)$ and still get a 0: $$(1.8) \quad -f(z*) \int_{\partial R_n} dz = 0$$ Similarly, $f'(z*)$ exists and we can multiple it by $(1.6)$ and still get a 0: $$(1.9) \quad f'(z*) \int_{\partial R_n} (z - z^*) dz = 0$$ Since $(1.8)$ and $(1.9)$ are zero we can add them to $\eta(R_n)$ and obtain [4]: $$\eta(R_n) = \int_{\partial R_n} f(z)dz = \int_{\partial R_n} f(z)dz - f(z*) \int_{\partial R_n} dz + \int_f'(z*) {\partial R_n} (z - z^*) dz$$ Moving them under one integral (we can also move $f(z*)$ and $f'(z*)$ inside since they're constants with respect to $z \in R_n$): $$\eta(R_n) = \int_{\partial R_n} f(z) - f(z*) + f'(z*) (z - z^*) dz$$ All this trickery so that we get to the form of the inequality $(1.5)$. However, that inequality is with respect to the modulus, so we can use Theorem 2 from [2] which states: $$\abs{\int_\gamma{f(z)dz}} \le \int_\gamma \abs{f(z)} \abs{dz}$$ In our case: $$\abs{\eta(R_n)} \le \int_{\partial R_n} \abs{f(z) - f(z*) + f'(z*) (z - z^*)} \abs{dz}$$ Now we can leverage the inequality $(1.5)$ and obtain: $$\abs{\eta(R_n)} \lt \int_{\partial R_n} \epsilon 2^{-n} \Delta \abs{dz} = \epsilon 2^{-n} \Delta \int_{\partial R} \abs{dz}$$ Where $\int_{\partial R} \abs{dz}$ is the perimeter of $R_n$. Let $P$ be the perimeter of $R$. Every time we pick a subrectangle the perimeter halves, so the perimeter of $R_n$ is $2^{-n} R$, and we have $$\abs{\eta(R_n)} \lt 4^{-n} \epsilon \Delta P$$ Replacing this in $(1.3)$ gives us, $$4^{-n} \abs{\eta(R)} \le \abs{\eta(R_n)} \lt 4^{-n} \epsilon \Delta P$$ So finally $$\abs{\eta(R)} \lt \epsilon \Delta P$$ $\Delta P$ is fixed for any given $R$ and we're free to choose $\epsilon \gt 0$, which leads to $\abs{\eta(R)} = 0$, implying $$\int_{\partial R} f(z) dz = 0$$

The proof is very clever but I don’t have a good intuition on why it works. Anyway, we can generalize the theorem a bit by allowing points in $R$ for which $f(z)$ isn’t holomorphic:

Theorem 2. Let $f(z)$ be holomorphic in $R’$, obtained from the rectangle $R$ by removing a finite set of points $\xi_j$. Then if

\[\lim_{z \rightarrow \xi_j} (z - \xi_j) f(z) = 0\]

for all $j$, then

\[(1) \quad \int_{\partial R} f(z) dz = 0\] First, we use again the fact that if we subdivide a rectangle into subrectangles and $(1)$ holds for each them, it also holds for the original rectangle, which was used to show $(1.1)$ in Theorem 1. We can then partition the rectangle in such a way that subrectangle has at most one $\xi_j$.

This lets us reduce the theorem to the case where we remove exactly one point $\xi$ from $R$ (the case with zero points is Theorem 1). So we just need to prove $(1)$ holds for such rectangle which we'll denote $R'$.

Using the same argument, we can subdivide a rectangle $R'$ containing $\xi$ such that the subrectangle containing $\xi$, denoted by $R_0$, is:
  • A square of size $L$
  • $\xi$ lies on its center
  • L is infinitesimally small
  • Satisfying: $$(2.1) \quad \int_{\partial R'} f(z)dz = \int_{\partial R_0} f(z)dz$$
See Figure 2.1 for an illustration.
See caption.
Figure 2.1: Partition of a rectangle such that the inner rectangle $R_0$ is a square and $\xi$ lies in the center.
For any given $\epsilon \gt 0$ and $z \in \partial R_0$ we can make: $$(2.2) \quad \frac{\epsilon}{\abs{z - \xi}}$$ arbitrarily large. That's because $\abs{z - \xi}$ is bounded by the diagonal of $R_0$ and $\epsilon$ is constant, so shrinking $R_0$ grows $(2.2)$. This lets us use this value as an upper bound for $f(z)$: $$(2.3) \quad \abs{f(z)} \lt \frac{\epsilon}{\abs{z - \xi}}$$ Applying Theorem 2 from [2] to $(2.1)$, we have: $$\abs{\int_{\partial R'} f(z)dz} = \abs{\int_{\partial R_0} f(z)dz} = \int_{\partial R_0} \abs{f(z)}\abs{dz}$$ Using $(2.3)$: $$(2.4) \quad \abs{\int_{\partial R'} f(z)dz} \lt \epsilon \int_{\partial R_0} \frac{\abs{dz}}{\abs{z - \xi}}$$ Since we chose $R_0$ to be a square of size $L$ with $\xi$ at is center, we have the lower bound $\abs{z - \xi} \ge L/2$, or that $$\frac{1}{z - \xi} \le \frac{2}{L}$$ Using these in $(2.4)$: $$\abs{\int_{\partial R'} f(z)dz} \lt \frac{2 \epsilon}{L} \int_{\partial R_0} \abs{dz}$$ We have that $\int_{\partial R_0} \abs{dz}$ is the perimeter of $R_0$, that is, $4L$, so: $$\abs{\int_{\partial R'} f(z)dz} \lt 8\epsilon$$ Since this is true for any $\epsilon \gt 0$, it must be that: $$\abs{\int_{\partial R'} f(z)dz} = 0$$

So we’re saying that if $f(z)$ is not holomorphic at specific points in $R$ but it tends to 0 there, integrating over its boundary is still yields 0.

In a disk

We now consider the case where $\Omega$ is the open circle $\abs{z - z_0} \lt \rho$, which we’ll denote by $\Delta$. We have the following result:

Theorem 3. If $f(z)$ is holomorphic in $\Delta$, then

\[\int_{\gamma} f(z) dz = 0\]

for any closed curve $\gamma$ in $\Delta$.

Let $(x_0, y_0)$ be the center of $\Delta$ and $(x', y')$ some point in $\Delta$. Let $\sigma$ be the curve formed by the segments $(x_0, y_0) \rightarrow (x', y_0)$ and then $(x', y_0) \rightarrow (x', y')$ as shown in Figure 3.1:
See caption.
Figure 3.1: Circle with points $(x_0, y_0)$ and $(x, y)$. The blue path corresponds to $\sigma$, the red path to $\overline{\sigma}$.

We'll keep $(x_0, y_0)$ fixed and assume $(x', y')$ is variable. We can then define a function of $x'$ and $y'$: $$F(x', y') = \int_{\sigma} f(z) dz$$ Note that $\sigma$ is implicitly a function of $x'$ and $y'$ since its endpoint is $(x', y')$. And as discussed in [3] (section Complex line integral), we can also write as: $$F(x', y') = \int_{\sigma} f(z) dx + i \int_{\sigma} f(z) dy$$ We can split each integral into the horizontal and the vertical segments. For the first integral, on the second segment $x$ doesn't vary, so the integral is 0. For the second integral, on the first segment $y$ doesn't vary, so the integral is 0. We can simplify to: $$(3.1) \quad F(x', y') = \int_{x_0}^{x'} f(x, y_0) dx + i \int_{y_0}^{y'} f(x', y) dy$$ Suppose we want to compute $\frac{\partial F}{\partial y}(x', y')$. By definition we have: $$(3.2) \quad \frac{\partial F}{\partial y}(x', y') = \lim_{h \rightarrow 0} \frac{F(x', y' + h) - F(x', y')}{h}$$ Replacing by the definition of $F$: $$= \lim_{h \rightarrow 0} \frac{\paren{\int_{x_0}^{x'} f(x, y_0) dx + i \int_{y_0}^{y' + h} f(x', y) dy} - \paren{\int_{x_0}^{x'} f(x, y_0) dx + i \int_{y_0}^{y'} f(x', y) dy}}{h}$$ The integral on $x$ doesn't depend on $y$, so it cancels out, leaving us with: $$= i \lim_{h \rightarrow 0} \frac{\int_{y_0}^{y' + h} f(x', y) dy - \int_{y_0}^{y'} f(x', y) dy}{h}$$ If we define $H(y')$ as $\int_{y_0}^{y'} f(x', y) dy$, then we have: $$= i \lim_{h \rightarrow 0} \frac{H(y' + h) - H(y')}{h}$$ The limit is the definition of $\frac{dH}{dy}(y')$, and we have that $\frac{dH}{dy}(y') = f(x', y')$ from the fundamental theorem of calculus. Thus $$(3.3) \quad \frac{\partial F}{\partial y}(x', y') = i f(x', y')$$ Now consider the curve $\overline{\sigma}$ composed of the segments $(x', y') \rightarrow (x_0, y')$ and $(x_0, y') \rightarrow (x_0, y_0)$ (the red path in Figure 3.1). We have that $\sigma + \overline{\sigma}$ form a oriented rectangle, so we have, from Theorem 1: $$\int_{\sigma} f(z) dz + \int_{\overline{\sigma}} f(z) dz = 0$$ Or that $$\int_{\sigma} f(z) dz = \int_{-\overline{\sigma}} f(z) dz = F(x', y')$$ Where $-\overline{\sigma}$ is the reverse of $\overline{\sigma}$, meaning the segments $(x_0, y_0) \rightarrow (x_0, y')$ and $(x_0, y') \rightarrow (x', y')$. Using a very similar argument we used for $\sigma$, we'll conclude that: $$(3.4) \quad \frac{\partial F}{\partial x}(x', y') = f(x', y')$$ Equations $(3.3)$ and $(3.4)$ tells us that $F$ satisfy the Cauchy-Riemman equations and is hence holomorphic in $\Delta$ (there's a corner case to be considered, see Corner case at the end) and that $f$ is the derivative of $F$. We can use Corollary 1 to conclude: $$\int_{\gamma} f(z) dz = 0$$ for any closed curve $\gamma$ in $\Delta$.

Corner case. What if the point $(x, y)$ lies on the same axis as $(x_0, y_0)$, say $(x, y_0)$? In this case, our curve degenerates into the segment $(x_0, y_0) \rightarrow (x, y_0)$. But let's pretend we still have the vertical segment, but that the endpoints coincide with $(x, y_0)$. How does $\frac{\partial F}{\partial y}(x', y')$ look like now? We can still follow the same reasoning and conclude that $$\frac{\partial F}{\partial y}(x', y_0) = i \lim_{h \rightarrow 0} \frac{H(y_0 + h) - H(y_0)}{h}$$ with $\frac{dH}{dy}(y_0) = f(x', y_0)$ even though $H(y_0) = 0$. So $$\frac{\partial F}{\partial y}(x', y_0) = i f(x', y_0)$$ Note that we can't form a rectangle anymore with $\overline{\sigma}$, but it's ok: for computing $\frac{\partial F}{\partial x}(x', y')$ nothing changed, since it only depends on the horizontal segment which is $\sigma$ now. Generalizing, for the case where $(x', y') = (x_0, y_0)$ we can still pretend we have 2 degenerate segments.

One question that came to mind when trying to deal with these corner cases: can't we simply choose a different starting point for these cases? I believe the answer to be no. Then we wouldn't be able to treat $x_0$ and $y_0$ as constants and they would be a function of $x$ and $y$.

Note that it's fine for the curve $\sigma$ to be a function of $x$ and $y$ (as it is in fact) since we don't make any assumption about its constancy in $(3.1)$.

In the same way we generalized Theorem 1. to allow for points in the rectangle where we allow $f$ to be non-holomorphic, we can generalize Theorem 3.

Theorem 4. Let $f(z)$ be holomorphic in the region $\Delta’$ obtained by omitting a finite number of points $\xi_j$ from the open disk $\Delta$. If $f(z)$ is such that

\[\lim_{z \rightarrow \xi_j} (z - \xi_j) f(z) = 0\]

then

\[\int_{\gamma} f(z) dz = 0\]

for any closed curve $\gamma$ in $\Delta’$.

As in the proof of Theorem 3, we can use two segments connecting the center $(x_0, y_0)$ to a point $(x', y')$ as long as the segments don't contain any of the $\xi_j$. In this case we use the exact same arguments, except that the rectangle might contain $\xi_j$ inside it and we have to use Theorem 2 instead of Theorem 1 to conclude: $$\int_{\sigma} f(z) dz + \int_{\overline{\sigma}} f(z) dz = 0$$ However, if any of the segments do go through one, such as in Figure 4.1, we can use 3 segments to avoid the missing point.
See caption.
Figure 4.1: Curve to avoid going over a missing point $\xi$.
We can find $x_1$ in between $x_0$ and $x'$ such that the line $x = x_1$ contains no $\xi_j$, and similarly for $y_1$ between $y_0$ and $y$. For $\sigma$ we use the segments: $(x_0, y_0) \rightarrow (x_0, y_1)$, $(x_0, y_1) \rightarrow (x', y_1)$, and $(x', y_1) \rightarrow (x', y')$. For $\overline{\sigma}$ we use $(x', y') \rightarrow (x_1, y')$, $(x_1, y') \rightarrow (x_1, y_0)$ and $(x_1, y_0) \rightarrow (x_0, y_0)$, depicted in Figure 4.1.

The first difference is that equation $(3.1)$ will now look like: $$F(x', y') = \int_{x_0}^{x'} f(x, y_1) dx + i \paren{\int_{y_0}^{y_1} f(x_0, y) dy + \int_{y_1}^{y'} f(x', y) dy}$$ To determine $\frac{\partial F}{\partial y}(x', y')$, we'll need to compute $F(x', y' + h)$, with $h \rightarrow 0$. We can use the exact same curve as we did for $F(x', y')$ except that the last segment will now go to $y' + h$, that is: $$F(x', y') = \int_{x_0}^{x'} f(x, y_1) dx + i \paren{\int_{y_0}^{y_1} f(x_0, y) dy + \int_{y_1}^{y' + h} f(x', y) dy}$$ By considering the limit $(3.2)$: $$\frac{\partial F}{\partial y}(x', y') = \lim_{h \rightarrow 0} \frac{F(x', y' + h) - F(x', y')}{h}$$ We'll arrive at the same conclusion that $$\frac{\partial F}{\partial y}(x', y') = i f(x', y')$$ If we call the blue segments $\sigma$ and the red ones $\overline{\sigma}$, we'll still conclude that $$(4.1) \quad F(x', y') = \int_{\sigma} f(z) dz = \int_{-\overline{\sigma}} f(z) dz$$ Because the 2 rectangles cancel out. Computing $\partial F/\partial x (x', y')$ is trickier however. That's because when we consider the point $(x' \pm h, y')$, the curve utilized is the 2-segment one because the path to $(x' \pm h, y')$ doesn't contain a $\xi_j$. This is illustrated in Figure 4.2 (note that we inverted the directions due to $(4.1)$).
See caption.
Figure 4.2: When we increment $x'$ by a small amount, the curve changes drastically.

So when we consider the difference $F(x' + h, y') - F(x', y')$ we need to be careful. The trick is to use rectangles to reduce the differences. Taking the example for Figure 4.2, we label the relevant segments.
See caption.
Figure 4.3: Labeling segments from Figure 4.2.
Let's define: $$S(\sigma) = \int_{s} f(z) dz$$ For each of the segments $s = A, B, \cdots$. We have that $F(x', y') = S(F) + S(E) + S(C)$ and $F(x' + h, y') = S(1) + S(B) + S(C) + S(D)$. We also know that $A, B, -E, -F$ form a directed rectangle and from Theorem 2 we have $$(4.2) \quad S(1) + S(B) - S(E) - S(F) = 0$$ Now, if we want to compute $F(x' + h, y') - F(x', y')$ we have: $$= S(1) + S(B) + S(C) + S(D) - (S(F) + S(E) + S(C)) \\ = S(1) + S(B) + S(D) - S(F) - S(E)$$ From $(4.2)$ we know $S(1) + S(B) = S(E) + S(F)$, so we can cancel some terms out and obtain: $$= S(D)$$ $D$ is a horizontal segment from $x'$ to $x' + h$, so we have: $$F(x' + h, y') - F(x', y') = \int_{x'}^{x' + h} f(x, y') dx$$ If we take $\lim_{h \rightarrow 0}$, we'll conclude, using arguments similar to those for $\partial F/\partial y$, that: $$\frac{\partial F}{\partial x} (x', y') = f(x', y')$$ The conclusion from this exercise is that even when infinitesimal changes cause a dramatic change in the chosen curve, we're still able to obtain a difference that is an infinitesimal segment.

Conclusion

It took me a long time to figure out the proof of Theorem 4. In [1], Ahlfors provides almost no details besides Figure 4.1 included in the proof of Theorem 4.

On my first read of the book I thought I had understood the proof but once I tried to plug it into the definition of derivative as a limit, I realized I didn’t understand it properly.

References