I toss off a mention in this essay, about its book publication. By the time it appeared I was thinking whether I could assemble these A-to-Z’s, or a set of them, into a book. I haven’t had the energy to put that together but it still seems viable.
Mr Wu, author of the Singapore Maths Tuition blog, asked me to explain a technical term today. I thought that would be a fun, quick essay. I don’t learn very fast, do I?
A note on style. I make reference here to “Big-O” and “Little-O”, capitalizing and hyphenating them. This is to give them visual presence as a name. In casual discussion they’re just read, or said, as the two words or word-and-a-letter. Often the Big- or Little- gets dropped and we just talk about O. An O, without further context, in my experience means Big-O.
The part of me that wants smooth consistency in prose urges me to write “Little-o”, as the thing described is represented with a lowercase ‘o’. But Little-o sounds like a midway game or an Eyerly Aircraft Company amusement park ride. And I never achieve consistency in my prose anyway. Maybe for the book publication. Until I’m convinced another is better, though, “Little-O” it is.
Big-O and Little-O Notation.
When I first went to college I had a campus post office box. I knew my box number. I also knew the length of the sluggish line for the combination lock code. The lock was a dial, lettered A through J. Being a young STEM-class idiot I thought, boy, would it actually be quicker to pick the lock than wait for the line? A three-letter combination, of ten options? That’s 1,000 possibilities. If I could try five a minute that’s, at worst, three hours 20 minutes. Combination might be anywhere in that set; I might get lucky. I could expect to spend 80 minutes picking my lock.
I decided to wait in line instead, and good that I did. I was unaware lock settings might not be a letter, like ‘A’. It could be the midway point between adjacent letters, like ‘AB’. That meant there were eight times as many combinations as I estimated, and I could expect to spend over ten hours. Even the slow line was faster than that. It transpired that my combination had two of these midway letters.
But that’s a little demonstration of algorithmic complexity. Also in cracking passwords by trial-and-error. Doubling the set of possible combination codes octuples the time it takes to break into the set. Making the combination longer would also work; each extra letter would multiply the cracking time by twenty. So you understand why your password should include “special characters” like punctuation, but most of all should be long.
We’re often interested in how long to expect a task to take. Sometimes we’re interested in the typical time it takes. Often we’re interested in the longest it could ever take. If we have a deterministic algorithm, we can say. We can count how many steps it takes. Sometimes this is easy. If we want to add two two-digit numbers together we know: it will be, at most, three single-digit additions plus, maybe, writing down a carry. (To add 98 and 37 is adding 8 + 7 to get 15, to add 9 + 3 to get 12, and to take the carry from the 15, so, 1 + 12 to get 13, so we have 135.) We can get a good quarrel going about what “a single step” is. We can argue whether that carry into the hundreds column is really one more addition. But we can agree that there is some smallest bit of arithmetic work, and proceed from that.
For any algorithm we have something that describes how big a thing we’re working on. It’s often ‘n’. If we need more than one variable to describe how big it is, ‘m’ gets called up next. If we’re estimating how long it takes to work on a number, ‘n’ is the number of digits in the number. If we’re thinking about a square matrix, ‘n’ is the number of rows and columns. If it’s a not-square matrix, then ‘n’ is the number of rows and ‘m’ the number of columns. Or vice-versa; it’s your matrix. If we’re looking for an item in a list, ‘n’ is the number of items in the list. If we’re looking to evaluate a polynomial, ‘n’ is the order of the polynomial.
In normal circumstances we don’t work out how many steps some operation does take. It’s more useful to know that multiplying these two long numbers would take about 900 steps than that it would need only 816. And so this gives us an asymptotic estimate. We get an estimate of how much longer cracking the combination lock will take if there’s more letters to pick from. This allowing that some poor soul will get the combination A-B-C.
There are a couple ways to describe how long this will take. The more common is the Big-O. This is just the letter, like you find between N and P. Since that’s easy, many have taken to using a fancy, vaguely cursive O, one that looks like . I agree it looks nice. Particularly, though, we write , where f is some function. In practice, we’ll see functions like or or . Usually something simple like that. It can be tricky. There’s a scheme for multiplying large numbers together that’s . What you will not see is something like , or or such. This comes to what we mean by the Big-O.
It’ll be convenient for me to have a name for the actual number of steps the algorithm takes. Let me call the function describing that g(n). Then g(n) is if once n gets big enough, g(n) is always less than C times f(n). Here c is some constant number. Could be 1. Could be 1,000,000. Could be 0.00001. Doesn’t matter; it’s some positive number.
There’s some neat tricks to play here. For example, the function ‘‘ is . It’s also and and . The function ‘‘ is also and those later terms, but it is not . And you can see why is right out.
There is also a Little-O notation. It, too, is an upper bound on the function. But it is a stricter bound, setting tighter restrictions on what g(n) is like. You ask how it is the stricter bound gets the minuscule letter. That is a fine question. I think it’s a quirk of history. Both symbols come to us through number theory. Big-O was developed first, published in 1894 by Paul Bachmann. Little-O was published in 1909 by Edmund Landau. Yes, the one with the short Hilbert-like list of number theory problems. In 1914 G H Hardy and John Edensor Littlewood would work on another measure and they used Ω to express it. (If you see the letter used for Big-O and Little-O as the Greek omicron, then you see why a related concept got called omega.)
What makes the Little-O measure different is its sternness. g(n) is if, for every positive number C, whenever n is large enough g(n) is less than or equal to C times f(n). I know that sounds almost the same. Here’s why it’s not.
If g(n) is , then you can go ahead and pick a C and find that, eventually, . If g(n) is , then I, trying to sabotage you, can go ahead and pick a C, trying my best to spoil your bounds. But I will fail. Even if I pick, like a C of one millionth of a billionth of a trillionth, eventually f(n) will be so big that . I can’t find a C small enough that f(n) doesn’t eventually outgrow it, and outgrow g(n).
This implies some odd-looking stuff. Like, that the function n is not . But the function n is at least , and and those other fun variations. Being Little-O compels you to be Big-O. Big-O is not compelled to be Little-O, although it can happen.
These definitions, for Big-O and Little-O, I’ve laid out from algorithmic complexity. It’s implicitly about functions defined on the counting numbers. But there’s no reason I have to limit the ideas to that. I could define similar ideas for a function g(x), with domain the real numbers, and come up with an idea of being on the order of f(x).
We make some adjustments to this. The important one is that, with algorithmic complexity, we assumed g(n) had to be a positive number. What would it even mean for something to take minus four steps to complete? But a regular old function might be zero or negative or change between negative and positive. So we look at the absolute value of g(x). Is there some value of C so that, when x is big enough, the absolute value of g(x) stays less than C times f(x)? If it does, then g(x) is . Is it the case that for every positive number C it’s true that g(x) is less than C times f(x), once x is big enough? Then g(x) is .
Fine, but why bother defining this?
A compelling answer is that it gives us a way to describe how different a function is from an approximation to that function. We are always looking for approximations to functions because most functions are hard. We have a small set of functions we like to work with. Polynomials are great numerically. Exponentials and trig functions are great analytically. That’s about all the functions that are easy to work with. Big-O notation particularly lets us estimate how bad an error we make using the approximation.
For example, the Runge-Kutta method numerically approximates solutions to ordinary differential equations. It does this by taking the information we have about the function at some point x to approximate its value at a point x + h. ‘h’ is some number. The difference between the actual answer and the Runge-Kutta approximation is . We use this knowledge to make sure our error is tolerable. Also, we don’t usually care what the function is at x + h. It’s just what we can calculate. What we want is the function at some point a fair bit away from x, call it x + L. So we use our approximate knowledge of conditions at x + h to approximate the function at x + 2h. And use x + 2h to tell us about x + 3h, and from that x + 4h and so on, until we get to x + L. We’d like to have as few of these uninteresting intermediate points as we can, so look for as big an h as is safe.
That context may be the more common one. We see it, particularly, in Taylor Series and other polynomial approximations. For example, the sine of a number is approximately:
This has consequences. It tells us, for example, that if x is about 0.1, this approximation is probably pretty good. So it is: the sine of 0.1 (radians) is about 0.0998334166468282 and that’s exactly what five terms here gives us. But it also warns that if x is about 10, this approximation may be gibberish. And so it is: the sine of 10.0 is about -0.5440 and the polynomial is about 1448.27.
The connotation in using Big-O notation here is that we look for small h’s, and for to be a tiny number. It seems odd to use the same notation with a large independent variable and with a small one. The concept carries over, though, and helps us talk efficiently about this different problem.
Thank you for reading.