451.67 From One Division, and Always a Shade Too High

The anchor is 45 squared, the shortfall is 1500, and one division by 900 lands on 1355/3 = 451.6667 against a true 451.66359. The estimate overshoots by exactly h squared over four a squared, which is 25/9 in the square here, so the error has a known sign as well as a known size. The article carries the bracket that names 452 as the nearest integer, one Newton step to nine figures, and what happens when the anchor is chosen too far away.

Lambdia
August 22, 2026 · 9 min read

The square root of 204,000, in your head, to five figures. No calculator, and about ten seconds.

451.67, from one anchor and one division. The true value is 451.66359, so the estimate is high by three thousandths, which is seven parts in a million. The method is short enough to say in a sentence and it comes with a proof that it always errs in the same direction.

Start from a square you already know

Everyone knows $45^2 = 2025$ . Shift the decimal point and you know $450^2 = 202{,}500$ , which is close to the target and is the whole reason 204,000 is a tractable number to be handed. The shortfall is the first thing to write down:

204{,}000 - 450^2 = 204{,}000 - 202{,}500 = 1{,}500

(1)

Stopping here and answering 450 is the actual trap, and it is a real answer rather than a slow route: 450 squared misses the target by fifteen hundred. The anchor is not the estimate, it is the place the estimate is measured from.

How much a step is worth near 450

You need to know what one extra unit buys in the square. The gap between consecutive squares is an identity, not a measurement:

(a+1)^2 - a^2 = 2a + 1, \qquad 451^2 - 450^2 = 901

(2)

Call it nine hundred, because dividing 1,500 by 900 is something a person can do in their head and dividing by 901 is not. Then share the shortfall out at that rate:

450 + \frac{1{,}500}{900} = 450 + \frac{5}{3} = \frac{1355}{3} = 451.6\overline{6}

(3)

Rounded, 451.67. The rounding of 901 down to 900 is not a sloppy step that happens to work: 900 is exactly $2a$ , and $2a$ rather than $2a+1$ is the correct coefficient for the linear approximation. The two readings coincide, which is the small piece of luck that makes this problem clean.

The linear correction

For $h$ small relative to $a^2$ , $\sqrt{a^2 + h} \approx a + h/(2a)$ . It is the first-order Taylor expansion of the square root at $a^2$ , and the derivative $1/(2a)$ is the reciprocal of the same $2a$ that appears in the gap between consecutive squares. That is not a coincidence: the gap is a discrete derivative.

The estimate is always too high, by a known amount

A one-directional error is much better than a small one, because it can be corrected. Square the estimate and compare:

\left(a + \frac{h}{2a}\right)^{2} = a^2 + h + \frac{h^2}{4a^2} \;>\; a^2 + h

(4)

The leftover term $h^2/(4a^2)$ is a square, so it is never negative and the estimate always overshoots. At these numbers it is $1500^2/(4 \cdot 450^2) = 25/9 = 2.78$ in the square, and 2.78 in the square is roughly $2.78/900 = 0.0031$ in the root, which is exactly the error you actually get. The geometry behind it is that the square root is concave, so its tangent line lies above the curve everywhere.

Fig. 1 — The tangent at 450 sits above the curve, so the one-division estimate can only be too big. The gap at 204 000 is three thousandths.

Concretely: 0.003075 of absolute error, and $0.003075/451.66 = 6.81 \times 10^{-6}$ of relative error. Seven parts in a million, from arithmetic that takes one division by 900.

One more line buys nine figures

If five figures are not enough, the same estimate feeds the oldest iteration there is. Average a guess with the number divided by that guess, and the relative error of the new guess is roughly the square of the old one:

x_{k+1} = \frac{1}{2}\left(x_k + \frac{N}{x_k}\right), \qquad x_0 = \frac{1355}{3} \;\Rightarrow\; x_1 = 451.6635916\ldots

(5)

That is correct to nine figures against the true 451.66359162544859, an improvement from $3 \times 10^{-3}$ to $10^{-8}$ in one step, which is the quadratic convergence of Newton's method doing what it does. The division is no longer head-sized, so this is a pencil step rather than a mental one, but it is the same method the Babylonians used and it needs nothing beyond a division and a halving.

Which integer is nearest, and how you would know

The estimate does not tell you the nearest integer, and it is worth checking rather than assuming. The target sits strictly between two consecutive squares:

451^2 = 203{,}401 \;<\; 204{,}000 \;<\; 204{,}304 = 452^2

(6)

So the root is between 451 and 452, and 204,000 sits 599 above the lower square and only 304 below the upper one. It is past the midpoint of the interval, so the nearest integer is 452, not 451. Also, since neither bound is met with equality, 204,000 is not a perfect square.

Fig. 2 — The bracket. The target lands two thirds of the way along the gap, which puts the root closer to 452.

When one division is not enough

The accuracy in (4) is governed by $h^2/(4a^2)$ , so the method degrades as the target moves away from the anchor. What matters is the ratio $h/a^2$ , which here is $1500/202500 = 0.0074$ . Squaring something that small is why the error is invisible at five figures.

Move the anchor closer and the estimate improves at once. Anchor at 451 instead, so $a^2 = 203{,}401$ , $h = 599$ and $2a = 902$ , and (3) returns $451 + 599/902 = 451.66408$ , high by only 0.00049. Six times better than the 450 version, for a division nobody can do in their head. That tradeoff is the whole engineering of a mental method: the anchor has to be a square you know, and the divisor has to be a number you can divide by.

Push the anchor far away and the estimate becomes useless in the ordinary way. Estimate $\sqrt{2}$ from $a = 1$ and $h = 1$ , and (3) returns 1.5 against 1.41421, an error of six percent, because $h/a^2 = 1$ . The practical rule is to choose the anchor so that the shortfall is a small fraction of the anchor's square, which is exactly why the first move is to find the nearest square you know rather than to start dividing.

The one thing the method never does is undershoot, at any distance. That is (4), and it holds for every $a > 0$ and every $h > -a^2$ , which was also swept over eight hundred targets without a single exception.

Sources and further reading

The family of hand methods this belongs to — Methods of computing square roots
The approximation in the definition box — Linear approximation
The iteration in equation (5) — Newton's method
Why the tangent lies above the curve — Concave function

None of the figures above came from a floating-point square root. The estimate is the exact fraction $1355/3$ and the overshoot is the exact fraction $25/9$ ; the true root was obtained twice over, once by integer bisection on $204000 \times 10^{12}$ , which returns 451,663,591 with no floating point at all, and once at fifty decimal digits to bound the error at +0.00307504.

Comments · 0

Be the first to comment.