Why This Book Actually Shows Up In My Shelf

Most numerical analysis textbooks either spend too much time proving convergence theorems or skim through the algorithms without explaining why they break. Kendall Atkinson's version lands somewhere in between. I keep it around because when I'm debugging a solver that stopped making sense after the seventh iteration, flipping to the chapter on roundoff error tends to point at the problem faster than starting from scratch. The Dover edition is cheap enough that you don't feel guilty about marking it up. The paper is thin but legible. The second edition adds material on splines and eigenvalue problems that the first one glossed over. If you grab the first edition for a dollar at a used bookstore, skip it. The corrections and expanded chapters matter, especially if you're working with anything involving polynomial approximation.

Introduction To Numerical Analysis Second Edition Dover Books On Mathematics

The structure follows a standard progression: floating point arithmetic, interpolation, numerical differentiation and integration, solving linear systems, least squares, eigenvalue computation, and ordinary differential equations. What makes it workable as a reference is the order of the chapters. Atkinson introduces error analysis before most methods, which is actually how you should read it. Most people flip straight to Gaussian elimination and then wonder why their results look wrong when the condition number is high. I keep this book open on my desk when I'm setting up finite difference schemes for partial differential equations. The treatment of stability criteria in the ODE chapters is concise without being dismissive. You get the theory you need to pick a time step that won't blow up, not a full proof of Lax equivalence. That's usually enough for practical work.

What It Covers Well

The floating point chapter is where most resources mess up. They tell you about precision limits but don't explain catastrophic cancellation in a way that sticks. Atkinson walks through subtraction of nearly equal numbers with actual examples. I remember running into an issue once where a quadratic formula implementation was giving garbage results for certain parameter ranges. Reading through that section showed me exactly why. The workaround wasn't reformulating the equation; it was using an alternative form for the smaller root. The book explains both without making you feel stupid for not knowing. Interpolation gets the right amount of attention. Splines are covered properly, not just as an afterthought. B-splines show up in the later chapters with enough detail to actually use them. If you've ever tried to build a curve that doesn't oscillate into oblivion between data points, you'll appreciate the treatment. The linear algebra sections assume you know basic matrix operations but don't assume you've taken a numerical linear algebra course. That's a reasonable middle ground. Gaussian elimination, LU factorization, iterative methods, and conditioning all get explained with worked examples. The sparse matrix discussion is brief but points you toward the right literature if you need more.

Where It Falls Short