Logarithm of a BigDecimal

https://stackoverflow.com/questions/739532

09-09-2019
|

Question

How can I calculate the logarithm of a BigDecimal? Does anyone know of any algorithms I can use?

My googling so far has come up with the (useless) idea of just converting to a double and using Math.log.

I will provide the precision of the answer required.

edit: any base will do. If it's easier in base x, I'll do that.

Solution

Java Number Cruncher: The Java Programmer's Guide to Numerical Computing provides a solution using Newton's Method. Source code from the book is available here. The following has been taken from chapter 12.5 Big Decmial Functions (p330 & p331):

/**
 * Compute the natural logarithm of x to a given scale, x > 0.
 */
public static BigDecimal ln(BigDecimal x, int scale)
{
    // Check that x > 0.
    if (x.signum() <= 0) {
        throw new IllegalArgumentException("x <= 0");
    }

    // The number of digits to the left of the decimal point.
    int magnitude = x.toString().length() - x.scale() - 1;

    if (magnitude < 3) {
        return lnNewton(x, scale);
    }

    // Compute magnitude*ln(x^(1/magnitude)).
    else {

        // x^(1/magnitude)
        BigDecimal root = intRoot(x, magnitude, scale);

        // ln(x^(1/magnitude))
        BigDecimal lnRoot = lnNewton(root, scale);

        // magnitude*ln(x^(1/magnitude))
        return BigDecimal.valueOf(magnitude).multiply(lnRoot)
                    .setScale(scale, BigDecimal.ROUND_HALF_EVEN);
    }
}

/**
 * Compute the natural logarithm of x to a given scale, x > 0.
 * Use Newton's algorithm.
 */
private static BigDecimal lnNewton(BigDecimal x, int scale)
{
    int        sp1 = scale + 1;
    BigDecimal n   = x;
    BigDecimal term;

    // Convergence tolerance = 5*(10^-(scale+1))
    BigDecimal tolerance = BigDecimal.valueOf(5)
                                        .movePointLeft(sp1);

    // Loop until the approximations converge
    // (two successive approximations are within the tolerance).
    do {

        // e^x
        BigDecimal eToX = exp(x, sp1);

        // (e^x - n)/e^x
        term = eToX.subtract(n)
                    .divide(eToX, sp1, BigDecimal.ROUND_DOWN);

        // x - (e^x - n)/e^x
        x = x.subtract(term);

        Thread.yield();
    } while (term.compareTo(tolerance) > 0);

    return x.setScale(scale, BigDecimal.ROUND_HALF_EVEN);
}

/**
 * Compute the integral root of x to a given scale, x >= 0.
 * Use Newton's algorithm.
 * @param x the value of x
 * @param index the integral root value
 * @param scale the desired scale of the result
 * @return the result value
 */
public static BigDecimal intRoot(BigDecimal x, long index,
                                 int scale)
{
    // Check that x >= 0.
    if (x.signum() < 0) {
        throw new IllegalArgumentException("x < 0");
    }

    int        sp1 = scale + 1;
    BigDecimal n   = x;
    BigDecimal i   = BigDecimal.valueOf(index);
    BigDecimal im1 = BigDecimal.valueOf(index-1);
    BigDecimal tolerance = BigDecimal.valueOf(5)
                                        .movePointLeft(sp1);
    BigDecimal xPrev;

    // The initial approximation is x/index.
    x = x.divide(i, scale, BigDecimal.ROUND_HALF_EVEN);

    // Loop until the approximations converge
    // (two successive approximations are equal after rounding).
    do {
        // x^(index-1)
        BigDecimal xToIm1 = intPower(x, index-1, sp1);

        // x^index
        BigDecimal xToI =
                x.multiply(xToIm1)
                    .setScale(sp1, BigDecimal.ROUND_HALF_EVEN);

        // n + (index-1)*(x^index)
        BigDecimal numerator =
                n.add(im1.multiply(xToI))
                    .setScale(sp1, BigDecimal.ROUND_HALF_EVEN);

        // (index*(x^(index-1))
        BigDecimal denominator =
                i.multiply(xToIm1)
                    .setScale(sp1, BigDecimal.ROUND_HALF_EVEN);

        // x = (n + (index-1)*(x^index)) / (index*(x^(index-1)))
        xPrev = x;
        x = numerator
                .divide(denominator, sp1, BigDecimal.ROUND_DOWN);

        Thread.yield();
    } while (x.subtract(xPrev).abs().compareTo(tolerance) > 0);

    return x;
}

/**
 * Compute e^x to a given scale.
 * Break x into its whole and fraction parts and
 * compute (e^(1 + fraction/whole))^whole using Taylor's formula.
 * @param x the value of x
 * @param scale the desired scale of the result
 * @return the result value
 */
public static BigDecimal exp(BigDecimal x, int scale)
{
    // e^0 = 1
    if (x.signum() == 0) {
        return BigDecimal.valueOf(1);
    }

    // If x is negative, return 1/(e^-x).
    else if (x.signum() == -1) {
        return BigDecimal.valueOf(1)
                    .divide(exp(x.negate(), scale), scale,
                            BigDecimal.ROUND_HALF_EVEN);
    }

    // Compute the whole part of x.
    BigDecimal xWhole = x.setScale(0, BigDecimal.ROUND_DOWN);

    // If there isn't a whole part, compute and return e^x.
    if (xWhole.signum() == 0) return expTaylor(x, scale);

    // Compute the fraction part of x.
    BigDecimal xFraction = x.subtract(xWhole);

    // z = 1 + fraction/whole
    BigDecimal z = BigDecimal.valueOf(1)
                        .add(xFraction.divide(
                                xWhole, scale,
                                BigDecimal.ROUND_HALF_EVEN));

    // t = e^z
    BigDecimal t = expTaylor(z, scale);

    BigDecimal maxLong = BigDecimal.valueOf(Long.MAX_VALUE);
    BigDecimal result  = BigDecimal.valueOf(1);

    // Compute and return t^whole using intPower().
    // If whole > Long.MAX_VALUE, then first compute products
    // of e^Long.MAX_VALUE.
    while (xWhole.compareTo(maxLong) >= 0) {
        result = result.multiply(
                            intPower(t, Long.MAX_VALUE, scale))
                    .setScale(scale, BigDecimal.ROUND_HALF_EVEN);
        xWhole = xWhole.subtract(maxLong);

        Thread.yield();
    }
    return result.multiply(intPower(t, xWhole.longValue(), scale))
                    .setScale(scale, BigDecimal.ROUND_HALF_EVEN);
}

OTHER TIPS

A hacky little algorithm that works great for large numbers uses the relation log(AB) = log(A) + log(B). Here's how to do it in base 10 (which you can trivially convert to any other logarithm base):

Count the number of decimal digits in the answer. That's the integral part of your logarithm, plus one. Example: floor(log10(123456)) + 1 is 6, since 123456 has 6 digits.
You can stop here if all you need is the integer part of the logarithm: just subtract 1 from the result of step 1.
To get the fractional part of the logarithm, divide the number by 10^(number of digits), then compute the log of that using math.log10() (or whatever; use a simple series approximation if nothing else is available), and add it to the integer part. Example: to get the fractional part of log10(123456), compute math.log10(0.123456) = -0.908..., and add it to the result of step 1: 6 + -0.908 = 5.092, which is log10(123456). Note that you're basically just tacking on a decimal point to the front of the large number; there is probably a nice way to optimize this in your use case, and for really big numbers you don't even need to bother with grabbing all of the digits -- log10(0.123) is a great approximation to log10(0.123456789).

This one is super fast, because:

No toString()
No BigInteger math (Newton's/Continued fraction)
Not even instantiating a new BigInteger
Only uses a fixed number of very fast operations

One call takes about 20 microseconds (about 50k calls per second)

But:

Only works for BigInteger

Workaround for BigDecimal (not tested for speed):

Shift the decimal point until the value is > 2^53
Use toBigInteger() (uses one div internally)

This algorithm makes use of the fact that the log can be calculated as the sum of the exponent and the log of the mantissa. eg:

12345 has 5 digits, so the base 10 log is between 4 and 5. log(12345) = 4 + log(1.2345) = 4.09149... (base 10 log)

This function calculates base 2 log because finding the number of occupied bits is trivial.

public double log(BigInteger val)
{
    // Get the minimum number of bits necessary to hold this value.
    int n = val.bitLength();

    // Calculate the double-precision fraction of this number; as if the
    // binary point was left of the most significant '1' bit.
    // (Get the most significant 53 bits and divide by 2^53)
    long mask = 1L << 52; // mantissa is 53 bits (including hidden bit)
    long mantissa = 0;
    int j = 0;
    for (int i = 1; i < 54; i++)
    {
        j = n - i;
        if (j < 0) break;

        if (val.testBit(j)) mantissa |= mask;
        mask >>>= 1;
    }
    // Round up if next bit is 1.
    if (j > 0 && val.testBit(j - 1)) mantissa++;

    double f = mantissa / (double)(1L << 52);

    // Add the logarithm to the number of bits, and subtract 1 because the
    // number of bits is always higher than necessary for a number
    // (ie. log2(val)<n for every val).
    return (n - 1 + Math.log(f) * 1.44269504088896340735992468100189213742664595415298D);
    // Magic number converts from base e to base 2 before adding. For other
    // bases, correct the result, NOT this number!
}

You could decompose it using

log(a * 10^b) = log(a) + b * log(10)

Basically b+1 is going to be the number of digits in the number, and a will be a value between 0 and 1 which you could compute the logarithm of by using regular double arithmetic.

Or there are mathematical tricks you can use - for instance, logarithms of numbers close to 1 can be computed by a series expansion

ln(x + 1) = x - x^2/2 + x^3/3 - x^4/4 + ...

Depending on what kind of number you're trying to take the logarithm of, there may be something like this you can use.

EDIT: To get the logarithm in base 10, you can divide the natural logarithm by ln(10), or similarly for any other base.

This is what I've come up with:

//http://everything2.com/index.pl?node_id=946812        
public BigDecimal log10(BigDecimal b, int dp)
{
    final int NUM_OF_DIGITS = dp+2; // need to add one to get the right number of dp
                                    //  and then add one again to get the next number
                                    //  so I can round it correctly.

    MathContext mc = new MathContext(NUM_OF_DIGITS, RoundingMode.HALF_EVEN);

    //special conditions:
    // log(-x) -> exception
    // log(1) == 0 exactly;
    // log of a number lessthan one = -log(1/x)
    if(b.signum() <= 0)
        throw new ArithmeticException("log of a negative number! (or zero)");
    else if(b.compareTo(BigDecimal.ONE) == 0)
        return BigDecimal.ZERO;
    else if(b.compareTo(BigDecimal.ONE) < 0)
        return (log10((BigDecimal.ONE).divide(b,mc),dp)).negate();

    StringBuffer sb = new StringBuffer();
    //number of digits on the left of the decimal point
    int leftDigits = b.precision() - b.scale();

    //so, the first digits of the log10 are:
    sb.append(leftDigits - 1).append(".");

    //this is the algorithm outlined in the webpage
    int n = 0;
    while(n < NUM_OF_DIGITS)
    {
        b = (b.movePointLeft(leftDigits - 1)).pow(10, mc);
        leftDigits = b.precision() - b.scale();
        sb.append(leftDigits - 1);
        n++;
    }

    BigDecimal ans = new BigDecimal(sb.toString());

    //Round the number to the correct number of decimal places.
    ans = ans.round(new MathContext(ans.precision() - ans.scale() + dp, RoundingMode.HALF_EVEN));
    return ans;
}

A Java implementation of Meower68 pseudcode which I tested with a few numbers:

public static BigDecimal log(int base_int, BigDecimal x) {
        BigDecimal result = BigDecimal.ZERO;

        BigDecimal input = new BigDecimal(x.toString());
        int decimalPlaces = 100;
        int scale = input.precision() + decimalPlaces;

        int maxite = 10000;
        int ite = 0;
        BigDecimal maxError_BigDecimal = new BigDecimal(BigInteger.ONE,decimalPlaces + 1);
        System.out.println("maxError_BigDecimal " + maxError_BigDecimal);
        System.out.println("scale " + scale);

        RoundingMode a_RoundingMode = RoundingMode.UP;

        BigDecimal two_BigDecimal = new BigDecimal("2");
        BigDecimal base_BigDecimal = new BigDecimal(base_int);

        while (input.compareTo(base_BigDecimal) == 1) {
            result = result.add(BigDecimal.ONE);
            input = input.divide(base_BigDecimal, scale, a_RoundingMode);
        }

        BigDecimal fraction = new BigDecimal("0.5");
        input = input.multiply(input);
        BigDecimal resultplusfraction = result.add(fraction);
        while (((resultplusfraction).compareTo(result) == 1)
                && (input.compareTo(BigDecimal.ONE) == 1)) {
            if (input.compareTo(base_BigDecimal) == 1) {
                input = input.divide(base_BigDecimal, scale, a_RoundingMode);
                result = result.add(fraction);
            }
            input = input.multiply(input);
            fraction = fraction.divide(two_BigDecimal, scale, a_RoundingMode);
            resultplusfraction = result.add(fraction);
            if (fraction.abs().compareTo(maxError_BigDecimal) == -1){
                break;
            }
            if (maxite == ite){
                break;
            }
            ite ++;
        }

        MathContext a_MathContext = new MathContext(((decimalPlaces - 1) + (result.precision() - result.scale())),RoundingMode.HALF_UP);
        BigDecimal roundedResult = result.round(a_MathContext);
        BigDecimal strippedRoundedResult = roundedResult.stripTrailingZeros();
        //return result;
        //return result.round(a_MathContext);
        return strippedRoundedResult;
    }

If all you need is to find the powers of 10 in the number you can use:

public int calculatePowersOf10(BigDecimal value)
{
    return value.round(new MathContext(1)).scale() * -1;
}

Pseudocode algorithm for doing a logarithm.

Assuming we want log_n of x

double fraction, input;
int base;
double result;

result = 0;
base = n;
input = x;

while (input > base){
  result++;
  input /= base;
}
fraction = 1/2;
input *= input;   

while (((result + fraction) > result) && (input > 1)){
  if (input > base){
    input /= base;
    result += fraction;
  }
  input *= input;
  fraction /= 2.0;
 }

The big while loop may seem a bit confusing.

On each pass, you can either square your input or you can take the square root of your base; either way, you must divide your fraction by 2. I find squaring the input, and leaving the base alone, to be more accurate.

If the input goes to 1, we're through. The log of 1, for any base, is 0, which means we don't need to add any more.

if (result + fraction) is not greater than result, then we've hit the limits of precision for our numbering system. We can stop.

Obviously, if you're working with a system which has arbitrarily many digits of precision, you will want to put something else in there to limit the loop.

I was searching for this exact thing and eventually went with a continued fraction approach. The continued fraction can be found at here or here

Code:

import java.math.BigDecimal;
import java.math.MathContext;

public static long ITER = 1000;
public static MathContext context = new MathContext( 100 );
public static BigDecimal ln(BigDecimal x) {
    if (x.equals(BigDecimal.ONE)) {
        return BigDecimal.ZERO;
    }

    x = x.subtract(BigDecimal.ONE);
    BigDecimal ret = new BigDecimal(ITER + 1);
    for (long i = ITER; i >= 0; i--) {
    BigDecimal N = new BigDecimal(i / 2 + 1).pow(2);
        N = N.multiply(x, context);
        ret = N.divide(ret, context);

        N = new BigDecimal(i + 1);
        ret = ret.add(N, context);

    }

    ret = x.divide(ret, context);
    return ret;
}

Old question, but I actually think this answer is preferable. It has good precision and supports arguments of practically any size.

private static final double LOG10 = Math.log(10.0);

/**
 * Computes the natural logarithm of a BigDecimal 
 * 
 * @param val Argument: a positive BigDecimal
 * @return Natural logarithm, as in Math.log()
 */
public static double logBigDecimal(BigDecimal val) {
    return logBigInteger(val.unscaledValue()) + val.scale() * Math.log(10.0);
}

private static final double LOG2 = Math.log(2.0);

/**
 * Computes the natural logarithm of a BigInteger. Works for really big
 * integers (practically unlimited)
 * 
 * @param val Argument, positive integer
 * @return Natural logarithm, as in <tt>Math.log()</tt>
 */
public static double logBigInteger(BigInteger val) {
    int blex = val.bitLength() - 1022; // any value in 60..1023 is ok
    if (blex > 0)
        val = val.shiftRight(blex);
    double res = Math.log(val.doubleValue());
    return blex > 0 ? res + blex * LOG2 : res;
}

The core logic (logBigInteger method) is copied from this other answer of mine.

I created a function for BigInteger but it can be easily modified for BigDecimal. Decomposing the log and using some properties of the log is what I do but I get only double precision. But it works for any base. :)

public double BigIntLog(BigInteger bi, double base) {
    // Convert the BigInteger to BigDecimal
    BigDecimal bd = new BigDecimal(bi);
    // Calculate the exponent 10^exp
    BigDecimal diviser = new BigDecimal(10);
    diviser = diviser.pow(bi.toString().length()-1);
    // Convert the BigDecimal from Integer to a decimal value
    bd = bd.divide(diviser);
    // Convert the BigDecimal to double
    double bd_dbl = bd.doubleValue();
    // return the log value
    return (Math.log10(bd_dbl)+bi.toString().length()-1)/Math.log10(base);
}

Licensed under: CC-BY-SA with attribution

Not affiliated with StackOverflow