5

For a change, I will ask a derivative question. Please consider the Generalized W-Lambert/Product Logarithm function $\text W_k(z)$. Let’s see what happens when we try to differentiate with respect to the branch cut subscripted variable k using nth derivative notation. Note that partial derivative symbols are harder to type out:

Even though the function’s branch cuts are discrete, they can still use the discrete derivative. For a more generalized approach, it could be possible to just differentiate by ignoring the constraint of $k\in\Bbb Z$ and adding it back when done to see what happens. If the generalizations are the same regardless of the bolded link’s definitions, then we can ignore the constraint entirely.

$$\frac{d}{dk}\text W_k(z)=\text W^{(1,0)}_k(z), \frac{d}{dz} \text W_k^{(0,1)}(z)=\frac{\text W_k(z)}{x(\text W_k(z)+1)}=\frac{1}{e^{\text W_k(z)}+x} $$

Almost all special functions on Wolfram Mathworld has a closed form derivative, but the change of the branches of the Product Logarithm function are difficult to find. Here is an integral representation from the bolded link. Here assumes $k\in\Bbb Z, z\ne -\frac1e ,0$, but let’s see what happens if we differentiate with respect to k using the Special Case of the Leibniz Rule:

$$\frac{d}{dk}\text W_k(z) =\frac{d}{dk}\left(1+(\ln(z)+2\pi i k-1)e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}\right)=(1-\ln(z))\frac{1}{\pi}e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i)}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right]+\frac d{dk}(2\pi i k)e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}+2\pi i k\frac d{dk} e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt} $$

Now let’s finish the derivation by resubstitution and algebra with the next step already done:

$$ \frac{1-\ln(z)}{\pi}e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi)}\right)dt}\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i)}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right]+2\pi i e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}- k e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i)}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right]= \left(\frac{1-\ln(z)}{\pi}-k\right)e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt}\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right] +2\pi i e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi}\right)dt} = e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi)}\right)dt}\left[ \left(\frac{1-\ln(z)}{\pi}-k\right)\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i)}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right] +2\pi i \right]$$

Here is a partially closed form:

$$\text W_k(z) =1+(\ln(z)+2\pi i k-1)e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi)}\right)dt} \implies e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi)}\right)dt}= \frac{\text W_k(z) -1}{\ln(z)+2\pi i k-1}\implies e^{\frac{i}{2\pi}\int\limits_0^\infty\frac{1}{t+1}\ln\left(1-\frac{2\pi}{\ln(z)-\ln(t)+t+(2k+1)i\pi)}\right)dt}\left[ \left(\frac{1-\ln(z)}{\pi}-k\right)\left[\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2ik-2\pi +i)}-\int_0^\infty\frac{dt}{(t+1)(\ln(z)-\ln(t)+t+2i k+i)}\right] +2\pi i \right] = \frac{2\pi \text W_k(z) -1}{\ln(z)+2\pi i k-1}\left[ \left(\frac{1-\ln(z)}{\pi}-k\right)\int_0^\infty \frac{dt}{(t+1)(\ln(z)-\ln(t)+t+(2k+1)i)(\ln(z)-\ln(t)+t+(2k+1)i-2\pi)}+ i \right]=\frac d{dk} \text W_k(z), k\text{ is any constant}$$

Note that the restrictions on the branch cut k can be adjusted.

Here is a sum representation using the bolded link using the Stirling Numbers of the first kind $S^{(x)}_y$:

$$\frac{d}{dk}\text W_k(z)= \frac{d}{dk}\left((2i\pi k+\ln(z))-\ln(2i\pi k+\ln(z))-\sum_{p=0}^\infty\frac{(-1)^p}{(2i\pi k +\ln(z))^p}\sum_{j=1}^p\frac{S^{(p-j+1)}_p \ln^j(2i\pi k+\ln(z))}{j!} \right) , z\to 0,\infty=2\pi i+\frac{2i\pi}{\ln(z)+2i\pi k}-2 i \pi\sum_{p=0}^\infty \sum_{j=1}^p \frac{(-1)^p S^{(p-j+1)}_p\ \ln^{j-1}(2 i k \pi + \ln(z)) (j - p \ln(2 i k π + \ln(z))) }{(i (2 k \pi - i \ln(z)))^{p+1} j!}=? $$

There area few other representations, but only one should work. Also, the branches themselves have a distinct pattern. Please also see details in the Elementary properties, branches and range section. Note there could be typos. What is a closed form of derivative of the Lambert-W function with respect to the function’s branches? I have given a few possible forms here. Please correct me and give me feedback!

Тyma Gaidash
  • 12,081
  • $k$ is an integer, so how do you differentiate with respect to $k$? – WimC Sep 25 '21 at 08:38
  • @WimC You can still differentiate using the provided definitions, but just assume an integer constraint after differentiation, like the gamma functuon. If possible, maybe you can ignore the constraint which could “generalize” the function. It may change the definition, but will be for recreational mathematics. How does that sound? Thanks. – Тyma Gaidash Sep 25 '21 at 11:46
  • This doesn't make any sense. The branch cuts have no continuity across them as they are inherently discrete and discontinuous. They simply index the different "parts" of a multi-valued function so one can treat the function as single-valued. At best you could ask how $W_k(z)$ changes at each branch cut: $W_{k+1}(z) - W_k(z)$ but this too is not continuous. Anything you do to make k continuous is going to be somewhat arbitrary. Surely you can find a homeomorphism that works it will again be arbitrary. – Gupta Sep 25 '21 at 18:38
  • 1
    $W_{k+\frac{1}{2}}(z)$ is something "in between" $W_{k}(z)$ and $W_{k+1}(z)$. You could define $W_{\alpha}(z) = (1 - \left{\alpha\right}) W_{\left[\alpha\right]}(z) + \left{\alpha\right} W_{\left[\alpha\right]+1}(z)$ and clearly this is a generalization and just a linear interpolation. This generalization, though, may be completely useless. – Gupta Sep 25 '21 at 18:42
  • 1
    I guess that if $f(z)$ has a single branch point at $z=z_0$ then one could consider $2\pi \mathrm i (z-z_0) f’(z)$ to be the “branch derivative”? For $W$ the branch point is $z_0=-e^{-1}$. – WimC Sep 25 '21 at 18:44
  • Thanks for the healthy discussion. Would it be possible to define the branch cut as this integral representation, but assuming $k\in\Bbb R$ or $k\in\Bbb C$ instead of $k\in\Bbb Z$? There area few other representations I used in the question, but one could work. Also, the branches themselves have a distinct pattern. – Тyma Gaidash Sep 25 '21 at 18:45
  • You can define it to be that, sure, but it has to be consistent. Clearly with the representation you can extend $k$ to be complex. Since these are branch cuts your extension has to actually conform correctly. You would have to prove that the integral is continuous for $k$(it looks like it is but remember that the log function itself has branch cuts) and that there exists a $i \in \mathbb{Z}$ and $w$ that corresponds to $k$ and $z$. Branch cuts are generally somewhat arbitrary but what is important is that they are well-behaved with respect to the function and the other branches. – Gupta Sep 25 '21 at 18:52
  • I will leave a bit more time to see if others have ideas on generalization or using discrete calculus. Maybe there will be a new idea we haven’t thought of yet. Thanks again. – Тyma Gaidash Sep 25 '21 at 18:58
  • I define it as $$\text W_k(z)=\text W_k(e^{2\pi i k} z)$$ from the answer, but please tell me if there is a better way and I may accept your answer. – Тyma Gaidash Sep 28 '21 at 03:28

1 Answers1

1

Just for fun, the Wright Omega function can also have the interesting identity that:

$$ω(z)\mathop=^\text{def} \text W_{\left \lceil\frac{\text{Im}(z)}{2\pi}-\frac12\right\rceil}\left(e^z\right)\implies \text W_k(z)=ω(\ln(z)+2i\pi k)$$

which can only be used for the discrete version because it causes the branch cut to be discrete requiring discrete calculus for first solution, ignoring the differential equation for the function in the link:

Here is a discrete calculus solution using the difference quotient. The integer nature of the branch cuts creates the smallest possible value of $|h|=1$ just to have 2 variations:

$$\frac{\Delta \text W_k(z)}{\Delta k}= \frac{\Delta }{\Delta k}\text W_k(z)= \lim_{h\to 0}\frac{\text W_{k+h}(z)-\text W_k(z)}{h} \mathop=^\text{discrete} \pm \text W_{k\pm1}(z)\mp \text W_k(z) =\text W_{k+1}(z)- \text W_k(z), \text W_k(z)- \text W_{k-1}(z) $$

Here is a more general equation using Wolfram Functions $1.32.27.3.1$ and logarithmic differentiation:

$$\ln\left(\text W_k(z)\right)=\ln(z)-\text W_k(z)+2i\pi k\implies \frac{d}{dk} \ln\left(\text W_k(z)\right)=\frac{d}{dk}\left(\ln(z)-\text W_k(z)+2i\pi k \right)\implies \frac{\text W^{(1,0)}_k(z)}{\text W_k(z)}=0-\text W^{(1,0)}_k(z) +2\pi i=2\pi i -\text W^{(1,0)}_k(z) \implies 2\pi i= \frac{\text W^{(1,0)}_k(z)}{\text W_k(z)} +\text W^{(1,0)}_k(z) = \text W^{(1,0)}_k(z)\left(\frac1{\text W_k(z)}+1\right)\implies \text W^{(1,0)}_k(z)=\frac{d}{dk} \text W_k(z)=\frac{2i\pi}{\frac1{\text W_k(z)}+1}=\frac{2i\pi \text W_k(z)}{\text W_k(z)+1}$$

Similarly we can solve for $\text W_k(z)$:

$$\ln\left(\text W_k(z)\right)=\ln(z)-\text W_k(z)+2i\pi k\implies \text W_k(z)= \text W_b\left(e^{2i\pi k}z\right) = \text W\left(e^{2i\pi k}z\right)$$

where $b$ just adds more branches of the function.

This means that $\text W_k(z)$ satisfies the following first order nonlinear ordinary differential equation. I assume that the constant of integration $c$ is $z$:

$$y’(k)=\frac{2i\pi\, y(k)}{y(k)+1}\implies y(k)= \text W_k(z) = \text W\left(e^{c+2i\pi k}\right)=\text W_k\left(ce^{2\pi i k}\right) =\text W_k\left(ze^{2\pi ik}\right) $$

I am still unsure about the differential equation result generalization.

Look at the $\frac{d\,w(k,z)}{dk}$ result for proof. Please correct me and give me feedback!

Тyma Gaidash
  • 12,081
  • I still have do not know how to generalize $\text W_k(z)$ to all $k$ or at least rational k or complex $k$, but I do have a derivative definition which I cannot think of how to use. – Тyma Gaidash Sep 28 '21 at 03:21
  • Does the series expansions not work? – Rounak Sarkar Oct 08 '21 at 17:35
  • @RounakSarkar The series expansions work, but they may not agree with “non integer” branches, and they were already in the question. They are also not a closed form. If you can prove, very optional, that various series expansions/ integral representations give the same result for non integer branch cuts, then I will consider it. For now the definition is the one in the answer as a derivative. – Тyma Gaidash Oct 08 '21 at 18:40
  • I would say that non-integer branches aren't defined because by definition they must be integers, so in order for us to extend it to any number, we need to define it in terms of something else, and defining it in terms of the series expansion seems natural and useful to me. – Rounak Sarkar Oct 09 '21 at 04:33
  • @RounakSarkar This is my analytic continuation definition which I used to implicitly derive the derivative. If I can assume a derivative for this, then it should imply some extensions. I am not sure if two different series expansions would agree on the same non integer branch cut values. – Тyma Gaidash Oct 09 '21 at 13:07