Search code examples
javabigdecimal

Rounding error from BigDecimal to byte array in Java

How should one properly handle a conversion rounding error with BigDecimal in Java:

BigDecimal -> byte[] -> BigDecimal

I have a custom datatype 32 bytes in length (yep, 32 bytes not 32 bits) and I need to decode the fractional part of BigDecimal into byte[].

I understand that I will lose some accuracy. Are there any established techniques to implement such a conversion?

NOTE: It is fixed point datatype of form MxN, where M % 8 == N % 8 == 0

Solution

  • Your fixed-point fractional part can be interpreted as the numerator, n, of a fraction n/(2256). I suggest, therefore, computing the BigDecimal value representing 1/(2256) (this is exactly representable as a BigDecimal) and storing a reference to it in a final static field.

    To convert to a byte[], then, use the two-arg version of BigDecimal.divideToIntegralValue() to divide the fractional part of your starting number by 1/(2256), using the MathContext argument to specify the rounding mode you want. Presumably you want either RoundingMode.HALF_EVEN or RoundingMode.HALF_UP. Then get the BigInteger unscaled value of the result (which should be numerically equal to the scaled value, since an integral value should have scale 0) via BigDecimal.unscaledValue(). BigInteger.toByteArray() will then give you a byte[] closely related to what you're after.*

    To go the other way, you can pretty much reverse the process. BigDecimal has a constructor that accepts a byte[] that, again, is very closely related to your representation. Using that constructor, convert your byte[] to a BigInteger, and thence to BigDecimal via the appropriate constructor. Multiply by that stored 1/(2256) value to get the fractional part you want.

    * The biggest trick here may involve twiddling signs appropriately. If your BigDecimals may be negative, then you probably want to first obtain their absolute values before converting to byte[]. More importantly, the byte[]s produced and consumed by BigInteger use a two's complement representation (i.e. with a sign bit), whereas I suppose you'll want an unsigned, pure binary representation. That mainly means that you'll need to allow for an extra bit -- and therefore a whole extra byte -- when you convert. Be also aware of byte order; check the docs of BigInteger for the byte order it uses, and adjust as appropriate.