# Approximation $\log_2(x)$

Can anyone share an easy way to approximate $\log_2(x)$, given $x$ is between $0$ and 1?

I'm trying to solve this using an old fashioned calculator (i.e. no logs)

Thanks!

EDIT: I realize that I stepped a bit ahead. The x comes in the form of a fraction, e.g. 3/8, which is indeed between 0 and 1, but could also be written as log2(3) - log2(8). I am hoping there is a quick way to approximate this calculation to let's say 2 decimals

• Close to $0$ a logarithm diverges to $-\infty$... How close to zero do you need to go? A fast (and quite rough) approximation is a decimal logarithm is minus a number of zeros in front of your number (including the one before the decimal point), and the binary logarithm is 10/3 times the decimal logarithm (as $2^{10} = 1024$ is quite close to $10^3=1000$). But it is rough: counting zeros gives result $\pm 1$, multiplied by 10/3 makes max error about $3$... – CiaPan Mar 21 '16 at 9:38
• @CiaPan Please see my edited question. Not that close to 0 like 0.0001. Solving something in the lines of log(3/7) would be the goal. – Dimebag Mar 21 '16 at 9:57

First normalize the value to the range $[1,2)$, multiplying by $2$ as long as necessary (the number of multiplies will form the integer part of the logarithm).

Then use the formula

$$\log_2\left(\frac{1+t}{1-t}\right)=\frac2{\ln(2)}\left(t+\frac{t^3}3+\frac{t^5}5\cdots\right)$$ evaluating for

$$t=\frac{x-1}{x+1}$$which will be in range $[0,\dfrac13)$. It will converge reasonably quickly, about one correct decimal per term.

Another option is to keep a tabulated list of constants such as $1.5,1.25,1.125\cdots$ together with their logarithms, and use

$$x>C_i\to \log_2(x)=\log_2\left(\frac x{C_i}\right)+\log(C_i).$$

At the same time as you divide $x$ by the constants (but keeping $x>1$), you accumulate the logarithms of these constants. When $x\approx1$, you have it. You can choose the set of constants that suits you best.

• How would this fare on very small numbers? would the "about one correct decimal per term" apply after the first significant digit or in general? – whn Sep 25 '19 at 16:00
• @opa: one decimal per term is the worst case. – Yves Daoust Sep 25 '19 at 20:36

Using that $\log_2(x)=\mbox{ln}(x)/\mbox{ln}(2)$, now you can use the Taylor expansion: \begin{equation} \mbox{ln}(x)=\mbox{ln}\left((x-1)+1\right)=\sum_{k=1}^{\infty}\frac{(-1)^{k-1}}{k}(x-1)^k \end{equation} The same expansion can be used for determining the $\mbox{ln}(2)$. The first terms of the expansion are \begin{equation} \log_2(x)=\frac{1}{\mbox{ln}(2)}\left[(x-1)-(x-1)^2/2+...\right] \end{equation} This expansion is working better around $x=1$. These kind of expansion are commonly used in calculators or programming languages to compute analytic functions.

EDIT: Another possibility to compute the natural logarithm is to use the generalized continued fraction \begin{equation} \log((x-1)+1)=\cfrac{2(x-1)}{(x-1)+2-\cfrac{(x-1)^2}{3((x-1)+2)-\cfrac{4(x-1)^2}{5((x-1)+2)-\cfrac{9(x-1)^2}{7((x-1)+2)-\cdots}}}} \end{equation} I think this option will be converging faster than the Taylor expansion

As a final approximation, I would recommend you also to have a look at the Padé expansion of the logarithm. Having a fast look to the old questions in this website I found Approximating Logs and Antilogs by hand, where user153012 gives next Padé approximation to compute the logarithm $\phi_3(x)\leq\mbox{ln}(x)\leq\psi_3(x)$ where the lower bound is $$\phi_3(x)=\frac{x(60+60x+11x^2)}{3(20+30x+12x^2+x^3)},$$ and the upper $$\psi_3(x)=\frac{x(30+21x+x^2)}{3(10+12x+3x^2)}.$$

• The problem with this one is that it is not fast, i.e., doing by hand, you need to go up to (I guess) k=10 or 15 (up to k=5 is not enough) to get a good answer of let's say 2 decimals. – Dimebag Mar 21 '16 at 9:55
• Please check my edited question. I am looking for something as easy as approximating powers of e using the solution I found here: math.stackexchange.com/questions/71357/approximation-of-e-x – Dimebag Mar 21 '16 at 10:02
• You can also use a continued fraction to compute the logarithm, included in the answer (I have no experience with these continued fraction method, but apparently they are faster than Taylor series). If not, a Padé expansion of the logarithm should work fine – seoanes Mar 21 '16 at 10:13

When $x$ is in the range $]0.5, 1[$ you can easily find a binary representation of the $\log_2(x)$ using this algorithm:

1. Start writing $-0.$ (with the dot as the result will be fractional) and evaluate $z=1/x$.
2. if $z^2>2$ append $1$ and let $z=z/2$, else append $0$, to the result.
3. go to step 2 until you have enough digits (or you have $z=1$)

Calculate $3.3$ binary digit for each decimal to stop the algorithm.

You will find a number like $-0.0110101$ that can easily be converted in decimal as $-0.4140625$ weighting each digit as $1 \over 2^n$.

Following is an example of the calculations for $\log_2(0.75)\approx -0.415$.

1.333333333 0           -0.
1.333333333 1.777777778   0  0
1.777777778 3.160493827   1 -0.25
1.580246914 2.497180308   1 -0.125
1.248590154 1.558977373   0  0
1.558977373 2.430410448   1 -0.03125
1.215205224 1.476723736   0  0
1.476723736 2.180712994   1 -0.0078125
-0.4140625


If your input involves just multiplication or division of small natural numbers and you don't need accuracy exceeding 4 decimal numbers, then logarithmic tables could be the simplest solution.

See for example

For much bigger or smaller numbers you may apply standard reduction: $$\log_2 (x\cdot 2^n) = \log_2 x + n$$ $$\log_2 (x/ 2^n) = \log_2 x - n$$

with some pre-selected powers, say
$2^3 = 8$, $2^5 = 32$, $2^8=256$, $2^{10} = 1024$, $2^{15}=32768$, $2^{20} = 1048576$...

You may also search the Web for some online $\log_2$ calculator, like this one.

$$\log(2)=\sum_{n=1}^{\infty}\frac{1}{n2^n}$$

$$(1/2)+(1/(2\cdot2^2))+(1/(3\cdot2^3))+(1/(4\cdot2^4))+(1/(5\cdot2^5))+(1/(6\cdot2^6))+(1/(7\cdot2^7))+(1/(8\cdot2^8))+(1/(9\cdot2^9))+(1/(10\cdot2^{10}))=0.6930...$$

Even by hand this is giving a simple procedure of finding $\log(2)$ in base $2$