Padé Approximants

A Taylor series approximates a function near a point by matching its value and as many derivatives as you like: $f(x) \approx c_0 + c_1 x + c_2 x^2 + \cdots + c_n x^n$. It is the workhorse of calculus, and for gentle functions it is superb. But a polynomial has a structural limit: it is finite everywhere. It can never shoot off to infinity at some particular point, because polynomials just don't do that.

Many real functions do shoot off to infinity — they have poles, places where the function genuinely diverges, like $1/(2-x)$ at $x=2$. A truncated Taylor series has no way to represent that. It just keeps being a smooth, bounded curve, drifting further and further from the truth as you approach the singularity.

The French mathematician Henri Padé, working in the 1890s under Charles Hermite, formalized a fix: instead of a polynomial, build a ratio of two polynomials — $P(x)/Q(x)$ — chosen so that its own Taylor expansion matches $f(x)$'s known coefficients exactly, term for term. Nothing about the input changes: you still only use $c_0, c_1, c_2, \dots$. What changes is the shape you pour that information into. And a ratio, unlike a sum, is allowed to have a zero in its denominator — a pole of its very own, positioned wherever the algebra demands.

Both curves below are built from the exact same five numbers: the first five Taylor coefficients of $f(x) = -\ln(1-x)$, a function with a genuine singularity at $x=1$. One curve sums them as a degree-4 polynomial. The other divides two degree-2 polynomials built from those same numbers — a Padé[2/2] approximant.

<p class="hint">{{hint_para}}</p>
<div class="plotwrap">
  <canvas id="plot" width="640" height="320"></canvas>
</div>
<div class="controls">
  <label for="xslider">{{slider_label}}</label>
  <input type="range" id="xslider" min="0" max="99" value="30" step="1" aria-label="{{slider_label}}">
  <span id="xval" class="xval">x = 0.30</span>
</div>
<div class="readout" id="readout"></div>
<div class="btns">
  <button id="reset" type="button" class="ghost">{{btn_reset}}</button>
</div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; }
.hint { font-size: .9rem; color: #444; margin: 0 0 .7rem; line-height: 1.45; }
.plotwrap { background: #f7f9fb; border: 1px solid #dbe3ea; border-radius: 10px; padding: .5rem; }
canvas { width: 100%; height: auto; display: block; }
.controls { display: flex; align-items: center; gap: .6rem; margin: .7rem 0 .3rem; flex-wrap: wrap; }
.controls label { font-size: .85rem; font-weight: 600; color: #1d3557; }
input[type="range"] { flex: 1 1 200px; accent-color: #1d3557; }
.xval { font: 700 14px ui-monospace, monospace; color: #1d3557; min-width: 5.5em; }
.readout { font-size: .88rem; line-height: 1.6; margin: .3rem 0 .6rem; }
.readout .row { display: flex; justify-content: space-between; gap: .8rem; padding: .15rem 0; border-bottom: 1px dashed #e1e6ea; }
.readout .row:last-child { border-bottom: none; }
.readout .swatch { display: inline-block; width: .7em; height: .7em; border-radius: 2px; margin-right: .4em; }
.sw-exact { background: #444; }
.sw-taylor { background: #e63946; }
.sw-pade { background: #0a7d33; }
.readout .err { font-family: ui-monospace, monospace; }
.btns { display: flex; gap: .5rem; flex-wrap: wrap; }
button { font: 600 14px system-ui, sans-serif; padding: .45rem .9rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }

// Code not found

Drag the slider toward $x=1$ and watch the gap open up. Near the origin both curves are nearly indistinguishable from the true function — that's expected, they were built to match it there. But as $x$ approaches the pole, the plain polynomial falls further and further behind while the rational function, which is allowed to grow without bound, keeps tracking the real curve.

Building a Padé approximant sounds like it should require knowing where the pole is in advance. It doesn't — that's the elegant part.

• Setup. Write the target ratio as $P(x)/Q(x)$, with $P$ of degree $L$, $Q$ of degree $M$, and normalize $Q(0)=1$. The pair is called the Padé[L/M] approximant.
• Match the series. Requiring $P(x) = Q(x) \cdot \left(c_0 + c_1 x + c_2 x^2 + \cdots\right)$ order by order up to $x^{L+M}$ gives $L+M+1$ linear equations in the unknown coefficients of $P$ and $Q$.
• Solve two small systems. The equations for the $Q$ coefficients decouple from $P$ — they form an $M \times M$ linear system by themselves. Solve it (ordinary Gaussian elimination), then the $P$ coefficients fall out by direct substitution.
• No search, no guessing. There is no trial and error over candidate pole locations. The denominator's roots — the approximant's poles — emerge automatically from the linear system, and if $f$ genuinely has a pole nearby, $Q$'s smallest root usually lands close to it.
• The cost is modest. Solving an $M \times M$ linear system is roughly $O(M^3)$ with plain elimination — utterly routine, the same order of work as numerically stable linear-algebra codes already do millions of times a day.

The catch is honesty, not hardness: a Padé approximant is only as good as the Taylor coefficients it started from, and if $f$ has more than one competing singularity nearby, a low-order approximant may lock onto the wrong one or produce a spurious pole that isn't in $f$ at all.

"I only trust the first few terms of my series, but I need an answer far outside where it converges" is a strikingly common problem, and Padé approximants are a standard answer:

• Perturbation theory in physics. Quantum field theory and fluid dynamics often produce power series that converge for absurdly small couplings. Padé approximants (and their generalizations) extend usable predictions well past that radius.
• Special-function libraries. Software that evaluates $\exp$, $\log$, trigonometric or Bessel functions on real hardware frequently uses rational (Padé-style) approximations internally, because they need fewer terms than a Taylor series for the same accuracy.
• Continued fractions. Padé approximants and continued fractions are two views of the same underlying structure — each convergent of a continued fraction is itself a Padé approximant of the function it represents.
• Control theory and signal processing. Approximating a time delay $e^{-sT}$ by a low-order rational function (a Padé delay model) is standard practice for keeping control-system simulations both accurate and simple.

Anywhere a truncated power series is the only local information available, but the function is suspected to have structure — a pole, a branch cut — a Padé approximant is often the cheapest way to see further.

Nothing about the raw information changes between a Taylor series and a Padé approximant — both start from the exact same list of derivatives at a point. What changes is the container: a sum of powers versus a ratio of two such sums. That one structural choice is the difference between a curve that must stay smooth forever and a curve that is free to blow up exactly where the real function does.

It's a small, general lesson disguised as a numerical trick: when a model keeps failing in the same predictable way, sometimes the fix isn't more data or more terms — it's giving the model a shape that is actually capable of representing what you're asking it to describe.