What Is Power Series

Replacing Functions With Polynomials

A power series is just an infinite polynomial. That's it. You're approximating a function as a sum of terms involving powers of x, each with its own coefficient. In practice, this means you can take something ugly like sin(x) or e^x and express it as a and get arbitrarily close to the true value by adding more terms. The standard form looks like this: the series is centered at a point a, and you're summing from n equals zero to infinity of c sub n times x minus a to the n. When the center is zero, which is the most common case, it collapses into what we call a Maclaurin series. I tend to use centers other than zero when I'm dealing with functions that behave badly near the origin, like logarithmic functions or rational expressions with poles.

What Is Power Series and Why It Matters in Practice

The real reason people use power series isn't because it looks elegant on paper. It's because calculators and computers don't natively know how to evaluate transcendental functions. They know how to add, multiply, and raise things to integer powers. A power series gives you a recipe: chop the function into polynomial pieces, evaluate those pieces, and you've got your answer to whatever precision you need. I've spent years working with numerical routines in embedded systems where floating-point performance was critically tight. One concrete example: I was implementing a thermal simulation for a battery management system, and the code needed to compute exponential decay terms repeatedly inside a loop running on a microcontroller with no FPU. I wrote out the Taylor expansion of e to the negative x as a fifth-order power series centered at zero. The whole thing ran in about 0.3 microseconds per evaluation instead of calling the system exp function, which took roughly 12 microseconds. That difference mattered because the simulation was processing tens of thousands of data points per second. The catch is that the series only works well near the center point. Move too far away and the higher-order terms blow up, and you need a ridiculous number of terms before convergence kicks in. For the exponential function the radius of convergence is technically infinite, but in practice floating-point overflow in the intermediate terms becomes your limiting factor. I learned that the hard way when someone tried to evaluate e raised to twelve using a straightforward Taylor series and the intermediate factorial terms exceeded the floating-point range before the final result even came close to being correct.

A better approach for large arguments is to use a half-angle reduction: split the argument in half repeatedly, evaluate the series on the small remainder, then square back up the required number of times. This keeps every term within a safe numerical range and cuts the number of significant terms needed from around fifteen down to four or five. It's a standard trick in mathematical libraries but something you won't find in most introductory courses. There are a few things about power series that beginners consistently get wrong, mostly because textbooks present them in a vacuum. The first is the distinction between radius of convergence and interval of convergence. The radius tells you how far from the center you can travel before the series stops converging, but it says nothing about what happens at the boundary. You have to check the endpoints individually, and they can behave completely differently. I once had a student spend two hours debugging a solution because the series converged conditionally at one endpoint and diverged at the other, and he'd assumed both endpoints behaved the same way based on the radius alone. The second thing is that not every smooth function is equal to its power series. There's a class of functions that are infinitely differentiable but whose Taylor series either converges to the wrong thing or doesn't converge at all in any useful sense. The classic example is the function that equals e to the negative one over x squared for nonzero x and zero at x equals zero. Every derivative at zero is zero, so the Taylor series is identically zero everywhere, which is clearly not the function itself. These edge cases matter in physics and engineering applications where someone might be working with approximate models that look smooth but aren't analytic.

Operationally, the main things you do with power series are substitution, differentiation, integration, and multiplication. Substitution means replacing x with some expression and checking that the new input stays within the radius of convergence. Differentiation and integration term by term is usually straightforward and preserves the radius, though convergence at the boundary can change. Multiplication comes in two forms: you can multiply two series together using the Cauchy product, or you can divide one series by another using long division techniques adapted for power series. I find myself doing a lot of series manipulations when solving differential equations, particularly when the coefficients are polynomials. You assume a solution of the form of a power series with unknown coefficients, plug it into the equation, shift the indices so everything lines up, and then match coefficients term by term. The algebra gets messy but the logic is mechanical. Once you've done it a dozen times the pattern becomes almost automatic. One practical limitation worth noting: power series expansions are local approximations. If you need global behavior, a single series centered at one point won't cut it. You might need multiple series in different regions, or you might be better served by a Padé approximant instead. Padé approximants are rational functions that match the same Taylor coefficients but generally have better global behavior and larger effective domains. I switched my battery management code from a sixth-order Taylor series to a Padé approximant of the same complexity and got an order of magnitude better accuracy across the full operating range of the controller. The derivation is slightly more involved but the payoff is significant.

If you're looking to get hands-on with this, the Wolfram MathWorld page on power series has solid derivations, and the Wikipedia entry on Taylor series gives you a good reference for standard expansions. For implementation details, the C math library source code in musl or glibc shows how professionals handle the edge cases I mentioned, and OpenLibm is a good open-source reference for portable implementations. The GNU Scientific Library also has a clean power series routine worth examining if you're building something of your own. Start by memorizing the five most common series: the exponential, sine, cosine, natural log, and geometric series. Everything else builds on those through substitution or algebraic manipulation. When you hit a function you don't have memorized, work out its derivatives at the center point, pattern-match the coefficients, and verify the radius of convergence. That process takes about twenty minutes the first time and gets down to two or three minutes once you've seen enough examples. I still pull out a notebook for genuinely unfamiliar functions, but the standard cases are muscle memory now.