Floating-point representation in computers

Basic structure

A typical floating-point value has three components: sign, significand (also called mantissa) and exponent. The significand carries the significant digits of the number; the exponent scales the significand by a power of the base so the radix point can "float" to represent very large or very small values. Hardware and software commonly normalize the significand so its leading digit is nonzero when possible. For background on how the shifting point works see radix point.

Standards and development

The behavior of floating-point arithmetic has been standardized to ensure portability between systems. The most widely used specification is the IEEE 754 family, which defines formats, rounding modes and special values such as infinity and NaN (not-a-number). The standard emerged from work by computer architects and numerical analysts to make arithmetic predictable across different computer platforms and languages.

Common formats and uses

• Single precision (32-bit) and double precision (64-bit) are the most common formats; they differ in exponent and significand widths, affecting range and precision.
• Floating-point is essential in scientific computing, graphics, simulations, signal processing and machine learning where a wide dynamic range is needed.
• Many systems also support extended formats, decimal floating-point for exact decimal fractions, and arbitrary-precision libraries for high-accuracy needs.

Limitations and notable behaviors

Because many decimal fractions have no exact binary representation, floating-point arithmetic incurs rounding error. Other important phenomena include overflow and underflow when values exceed representable bounds, subnormal numbers that fill the gap around zero, and special values: infinities and NaNs. Comparison of floating-point values must account for rounding and representation error; exact equality tests often yield surprising results. The exponent is central to range and scaling; see exponent discussions for details.

Practical advice: choose an appropriate precision for your problem, prefer algorithms designed for numerical stability, avoid equality checks on raw floats, and use decimal or fixed-point arithmetic for exact currency computations. For background on how computers represent numbers in binary hardware, consult resources about binary systems and the role of the radix point.