### Prologue

SICP reaches a point, in Chapter 3, where for significant parts of it you’re not doing any coding. It has exercises, but they’re all about thinking about the concepts, not doing anything with a computer. It has you do substitutions to see what expressions result. It has you make diagrams that focus in on particular systemic aspects of processes. It also gets into operational models, talking about simulating logic gates, how concurrent processing can work (expressed in hypothetical Scheme code). It’s all conceptual. Some of it was good, I thought. The practice of doing substitutions manually helps you really get what your Scheme functions are doing, rather than guessing. The rest didn’t feel that engaging. One could be forgiven for thinking that the book is getting dry at this point.

It gets into some coding again in Section 3.3, where it covers building data structures. It gets more interesting with an architecture called “streams” in Section 3.5.

One thing I will note is that the only way I was able to get the code in Section 3.5 to work in Racket was to go into “Lazy Scheme.” I don’t remember what language setting I used for the prior chapters, maybe R5RS, or “Pretty Big.” Lazy Scheme does lazy evaluation on Scheme code. One can be tempted to think that this makes using the supporting structures for streams covered in this section pointless, because the underlying language is doing the delayed evaluation that this section implements in code. Anyway, Lazy Scheme doesn’t interfere with anything in this section. It all works. It just makes the underlying structure for streams redundant. For the sake of familiarity with the code this section discusses (which I think helps in preserving one’s sanity), I think it’s best to play along and use its code.

Another thing I’ll note is this section makes extensive use of knowledge derived from calculus. Some other parts of this book do that, too, but it’s emphasized here. It helps to have that background.

I reached a stopping point in SICP, here, 5 years ago, because of a few things. One was I became inspired to pursue a history project on the research and development that led to the computer technology we use today. Another is I’d had a conversation with Alan Kay that inspired me to look more deeply at the STEPS project at Viewpoints Research, and try to make something of my own out of that. The third was Exercise 3.61 in SICP. It was a problem that really stumped me. So I gave up, and looked up an answer for it on the internet, in a vain attempt to help me understand it. The answer didn’t help me. It worked when I tried the code, but I found it too confusing to understand why it produced correct results. The experience was really disappointing, and disheartening. Looking it up was a mistake. I wished that I could forget I’d seen the answer, so I could go back to working on trying to figure it out, but I couldn’t. I’d seen it. I worked on a few more exercises after that, but then I dropped SICP. I continued working on my history research, and I got into exploring some fundamental computing concepts on processors, language parsing, and the value of understanding a processor as a computing/programming model.

I tried an idea that Kay told me about years earlier, of taking something big, studying it, and trying to improve on it. That took me in some interesting directions, but I hit a wall in my skill, which I’m now trying to get around. I figured I’d continue where I left off in SICP. One way this diversion helped me is I basically forgot the answers for the stuff I did. So, I was able to come back to the problem, using a clean slate, almost fresh. I had some faded memories of it, which didn’t help. I just had to say to myself “forget it,” and focus on the math, and the streams architecture. That’s what finally helped me solve 3.60, which then made 3.61 straightforward. That was amazing.

### A note about Exercise 3.59

It was a bit difficult to know at first why the streams (cosine-series and sine-series) were coming out the way they were for this exercise. The cosine-series is what’s called an even series, because its powers are even (0, 2, 4, etc.). The sine-series is what’s called an odd series, because its powers are odd (1, 3, 5, etc.). However, when you’re processing the streams, they just compute a general model of series, with *all* of the terms, regardless of whether the series you’re processing has values in each of the terms or not. So, cosine-series comes out (starting at position 0) as: [1, 0, -1/2, 0, 1/24, …], since the stream is computing *a _{0}x^{0}* +

*a*+

_{1}x^{1}*a*…, where

_{2}x^{2}*a*is each term’s coefficient, and some of the terms are negative, in this case. The coefficients of the terms that don’t apply to the series come out as 0. With sine-series, it comes out (starting at position 0) as: [0, 1, 0, -1/6, 0, 1/120, …].

_{i}What’s really interesting is that exercises 3.59, 3.61, and 3.62 are pretty straightforward. Like with some of the prior exercises, all you have to do is translate the math into stream terms (and it’s a pretty direct translation from one to the other), and it works! You’re programming in mathland! I discovered this in the earlier exercises in this section, and it amazed me how expressive this is. I could pretty much write code as if I was writing out the math. I could think in terms of the math, not so much the logistics of implementing computational logic to make the math work. At the same time, this felt disconcerting to me, because when things went wrong, I wanted to know why, computationally, and I found that working with streams, it was difficult to conceptualize what was going on. I felt as though I was just supposed to trust that the math I expressed worked, just as it should. I realize now that’s what I should have done. It was starting to feel like I was playing with magic. I didn’t like that. It was difficult for me to trust that the math was actually working, and if it wasn’t, that it was because I was either not understanding the math, and/or not understanding the streams architecture, not because there was something malfunctioning underneath it all. I really wrestled with that, and I think it was for the good, because now that I can see how elegant it all is, it looks so beautiful!

### Exercise 3.60

This exercise gave me a lot of headaches. When I first worked on it 5 years ago, I kind of got correct results with what I wrote, but not quite. I ended up looking up someone else’s solution to it, which worked. It was kind of close to what I had. I finally figured out what I did wrong when I worked on it again recently: I needed to research how to multiply power series, but in a computationally efficient manner. It turns out you need to use a method called the Cauchy product, but there’s a twist, because of the way the streams architecture works. A lot of math sources I looked up use the Cauchy product method, anyway (including a source that covers power series multiplication that I cite below), but the reason it needs to be used is it automatically collects all like terms as each term of the product is produced. The problem you need to work around is that you can’t go backwards through a stream, except by using stream-ref and indexes, and I’ve gotten the sense by going through these exercises that when it comes to doing math problems, you’re generally not supposed to be doing that, though there’s an example later in SICP where they talk about a method Euler devised for accelerating computation of power series where they use stream-ref to go backwards and forwards inside a stream. I think that’s an exception.

Here are a couple sources that I found helpful when trying to work this out:

Formal Power Series, from Wikipedia. Pay particular attention to the section on “Operations on Formal Power Series.” It gives succinct descriptions on multiplying power series, inverting them (though I’d only pay attention to the mathematical expression of this in Exercise 3.61), and dividing them (which you’ll need for Exercise 3.62).

**Edit** 6/4/2018: I originally had a video here, which I thought gave a good illustration of what the Cauchy product accomplishes. It also provided a nice visual aide that pointed me toward a solution to this problem in SICP. Unfortunately, the video has disappeared from YouTube. So, I’ll substitute a presentation of the Cauchy product that I used as a tool for figuring out how to do this exercise. This will not give you the answer for how to do this exercise, but it’s on the way to the answer I got.

Let *A* and *B* be power series that we’re going to multiply. We will use *a _{i}* to represent the coefficients for

*A*, and

*b*to represent the coefficients for

_{i}*B:*

*A* = *a _{0}x^{0} + a_{1}x^{1} + a_{2}x^{2} + …*

*B* = *b _{0}x^{0} + b_{1}x^{1} + b_{2}x^{2} + …*

Multiplying these series creates a new series, *C*, made by creating a sum of products from terms in *A* and *B*:

*A x B = C*

*C _{0} = a_{0}x^{0}b_{0}x^{0}*

*C _{1} = a_{1}x^{1}b_{0}x^{0} + a_{0}x^{0}b_{1}x^{1}*

*C _{2} = a_{2}x^{2}b_{0}x^{0} + a_{1}x^{1}b_{1}x^{1} + a_{0}x^{0}b_{2}x^{2}*

*C _{3} = a_{3}x^{3}b_{0}x^{0} + a_{2}x^{2}b_{1}x^{1} + a_{1}x^{1}b_{2}x^{2} + a_{0}x^{0}b_{3}x^{3}*

*C _{4} = …*

*A x B = C _{0} + C_{1} + C_{2} + C_{3} + C_{4} + …*

You’ll note that, with the exception of the very first term, as we progress to computing each term, we start with the *i _{th}* term in series

*A*, and go down to the

*0*term, but with series

_{th}*B*, we start with the

*0*term, and go upward to the

_{th}*i*term.

_{th}To reiterate, you can’t use the stream architecture in SICP to compute the product for this exercise using the exact method I’ve outlined above (when I thought about trying to do that, it get extremely convoluted, and it’s not worth pursuing), but there is a way to compute the product series that allows you to go through both streams from left to right.

What helped me find a succinct way to implement mul-series was to just focus on the coefficients:

a_{0}b_{0}

a_{1}b_{0} + a_{0}b_{1}

a_{2}b_{0} + a_{1}b_{1} + a_{0}b_{2}

a_{3}b_{0} + a_{2}b_{1} + a_{1}b_{2} + a_{0}b_{3}

…

A couple hints I’ll give are:

- There is a way to do the Cauchy product without using mutable state in variables, nor is it necessary to do anything particularly elaborate. You can do it just using the stream architecture.
- Think about how addition is commutative, as you look at the above pattern of computation.

The solution is a little counterintuitive, but once you find it, it’s pretty neat.

It is *crucial* that you get the solution for this exercise right, though, because the next two exercises (3.61 and 3.62) will give you no end of headaches if you don’t. Exercise 3.61 builds on what you create for this one, and 3.62 builds on what you create for 3.61.

Exercise 3.60 says you can try out mul-series using sin(x) and cos(x), squaring both, and then adding the product series together. This should give you a result of 1. How is “1” represented in streams, though? Well, it makes sense that it will be in the constant-term position (the zeroth position) in the stream. That’s where a scalar value would be in a power series.

There is a related concept to working with power series called a unit series. A unit series is just a list of the coefficients in a power series. If you’ve done the previous exercises where SICP says you’re working with power series, this is what you’re really working with (though SICP doesn’t mention this). It’s why in Exercise 3.61 SICP has you writing a function called “invert-unit-series”.

The unit series equivalent for the scalar value 1 is [1, 0, 0, 0, 0, …].

### A note about Exercise 3.62

The exercise talks about using your div-series function to compute tan-series. Here was a good source for finding out how to do that:

The unit series for tan-series should come out as: [0, 1, 0, 1/3, 0, 2/15, 0, 17/315, …]