Documentation |
MATLAB^{®} represents floating-point numbers in either double-precision or single-precision format. The default is double precision, but you can make any number single precision with a simple conversion function.
MATLAB constructs the double-precision (or double) data type according to IEEE^{®} Standard 754 for double precision. Any value stored as a double requires 64 bits, formatted as shown in the table below:
Bits | Usage |
---|---|
63 | Sign (0 = positive, 1 = negative) |
62 to 52 | Exponent, biased by 1023 |
51 to 0 | Fraction f of the number 1.f |
MATLAB constructs the single-precision (or single) data type according to IEEE Standard 754 for single precision. Any value stored as a single requires 32 bits, formatted as shown in the table below:
Bits | Usage |
---|---|
31 | Sign (0 = positive, 1 = negative) |
30 to 23 | Exponent, biased by 127 |
22 to 0 | Fraction f of the number 1.f |
Because MATLAB stores numbers of type single using 32 bits, they require less memory than numbers of type double, which use 64 bits. However, because they are stored with fewer bits, numbers of type single are represented to less precision than numbers of type double.
Use double-precision to store values greater than approximately 3.4 x 10^{38} or less than approximately -3.4 x 10^{38}. For numbers that lie between these two limits, you can use either double- or single-precision, but single requires less memory.
Because the default numeric type for MATLAB is double, you can create a double with a simple assignment statement:
x = 25.783;
The whos function shows that MATLAB has created a 1-by-1 array of type double for the value you just stored in x:
whos x Name Size Bytes Class x 1x1 8 double
Use isfloat if you just want to verify that x is a floating-point number. This function returns logical 1 (true) if the input is a floating-point number, and logical 0 (false) otherwise:
isfloat(x) ans = 1
You can convert other numeric data, characters or strings, and logical data to double precision using the MATLAB function, double. This example converts a signed integer to double-precision floating point:
y = int64(-589324077574); % Create a 64-bit integer x = double(y) % Convert to double x = -5.8932e+11
Because MATLAB stores numeric data as a double by default, you need to use the single conversion function to create a single-precision number:
x = single(25.783);
The whos function returns the attributes of variable x in a structure. The bytes field of this structure shows that when x is stored as a single, it requires just 4 bytes compared with the 8 bytes to store it as a double:
xAttrib = whos('x'); xAttrib.bytes ans = 4
You can convert other numeric data, characters or strings, and logical data to single precision using the single function. This example converts a signed integer to single-precision floating point:
y = int64(-589324077574); % Create a 64-bit integer x = single(y) % Convert to single x = -5.8932e+11
This section describes which classes you can use in arithmetic operations with floating-point numbers.
You can perform basic arithmetic operations with double and any of the following other classes. When one or more operands is an integer (scalar or array), the double operand must be a scalar. The result is of type double, except where noted otherwise:
single — The result is of type single
double
int* or uint* — The result has the same data type as the integer operand
char
logical
This example performs arithmetic on data of types char and double. The result is of type double:
c = 'uppercase' - 32; class(c) ans = double char(c) ans = UPPERCASE
You can perform basic arithmetic operations with single and any of the following other classes. The result is always single:
single
double
char
logical
In this example, 7.5 defaults to type double, and the result is of type single:
x = single([1.32 3.47 5.28]) .* 7.5; class(x) ans = single
For the double and single classes, there is a largest and smallest number that you can represent with that type.
The MATLAB functions realmax and realmin return the maximum and minimum values that you can represent with the double data type:
str = 'The range for double is:\n\t%g to %g and\n\t %g to %g'; sprintf(str, -realmax, -realmin, realmin, realmax) ans = The range for double is: -1.79769e+308 to -2.22507e-308 and 2.22507e-308 to 1.79769e+308
Numbers larger than realmax or smaller than -realmax are assigned the values of positive and negative infinity, respectively:
realmax + .0001e+308 ans = Inf -realmax - .0001e+308 ans = -Inf
The MATLAB functions realmax and realmin, when called with the argument 'single', return the maximum and minimum values that you can represent with the single data type:
str = 'The range for single is:\n\t%g to %g and\n\t %g to %g'; sprintf(str, -realmax('single'), -realmin('single'), ... realmin('single'), realmax('single')) ans = The range for single is: -3.40282e+38 to -1.17549e-38 and 1.17549e-38 to 3.40282e+38
Numbers larger than realmax('single') or smaller than -realmax('single') are assigned the values of positive and negative infinity, respectively:
realmax('single') + .0001e+038 ans = Inf -realmax('single') - .0001e+038 ans = -Inf
If the result of a floating-point arithmetic computation is not as precise as you had expected, it is likely caused by the limitations of your computer's hardware. Probably, your result was a little less exact because the hardware had insufficient bits to represent the result with perfect accuracy; therefore, it truncated the resulting value.
Because there are only a finite number of double-precision numbers, you cannot represent all numbers in double-precision storage. On any computer, there is a small gap between each double-precision number and the next larger double-precision number. You can determine the size of this gap, which limits the precision of your results, using the eps function. For example, to find the distance between 5 and the next larger double-precision number, enter
format long eps(5) ans = 8.881784197001252e-16
This tells you that there are no double-precision numbers between 5 and 5 + eps(5). If a double-precision computation returns the answer 5, the result is only accurate to within eps(5).
The value of eps(x) depends on x. This example shows that, as x gets larger, so does eps(x):
eps(50) ans = 7.105427357601002e-15
If you enter eps with no input argument, MATLAB returns the value of eps(1), the distance from 1 to the next larger double-precision number.
Similarly, there are gaps between any two single-precision numbers. If x has type single, eps(x) returns the distance between x and the next larger single-precision number. For example,
x = single(5); eps(x)
returns
ans = 4.7684e-07
Note that this result is larger than eps(5). Because there are fewer single-precision numbers than double-precision numbers, the gaps between the single-precision numbers are larger than the gaps between double-precision numbers. This means that results in single-precision arithmetic are less precise than in double-precision arithmetic.
For a number x of type double, eps(single(x)) gives you an upper bound for the amount that x is rounded when you convert it from double to single. For example, when you convert the double-precision number 3.14 to single, it is rounded by
double(single(3.14) - 3.14) ans = 1.0490e-07
The amount that 3.14 is rounded is less than
eps(single(3.14)) ans = 2.3842e-07
Almost all operations in MATLAB are performed in double-precision arithmetic conforming to the IEEE standard 754. Because computers only represent numbers to a finite precision (double precision calls for 52 mantissa bits), computations sometimes yield mathematically nonintuitive results. It is important to note that these results are not bugs in MATLAB.
Use the following examples to help you identify these cases:
The decimal number 4/3 is not exactly representable as a binary fraction. For this reason, the following calculation does not give zero, but rather reveals the quantity eps.
e = 1 - 3*(4/3 - 1) e = 2.2204e-16
Similarly, 0.1 is not exactly representable as a binary number. Thus, you get the following nonintuitive behavior:
a = 0.0; for i = 1:10 a = a + 0.1; end a == 1 ans = 0
Note that the order of operations can matter in the computation:
b = 1e-16 + 1 - 1e-16; c = 1e-16 - 1e-16 + 1; b == c ans = 0
There are gaps between floating-point numbers. As the numbers get larger, so do the gaps, as evidenced by:
(2^53 + 1) - 2^53 ans = 0
Since pi is not really pi, it is not surprising that sin(pi) is not exactly zero:
sin(pi) ans = 1.224646799147353e-16
When subtractions are performed with nearly equal operands, sometimes cancellation can occur unexpectedly. The following is an example of a cancellation caused by swamping (loss of precision that makes the addition insignificant).
sqrt(1e-16 + 1) - 1 ans = 0
Some functions in MATLAB, such as expm1 and log1p, may be used to compensate for the effects of catastrophic cancellation.
Round-off, cancellation, and other traits of floating-point arithmetic combine to produce startling computations when solving the problems of linear algebra. MATLAB warns that the following matrix A is ill-conditioned, and therefore the system Ax = b may be sensitive to small perturbations:
A = diag([2 eps]); b = [2; eps]; y = A\b; Warning: Matrix is close to singular or badly scaled. Results may be inaccurate. RCOND = 1.110223e-16.
These are only a few of the examples showing how IEEE floating-point arithmetic affects computations in MATLAB. Note that all computations performed in IEEE 754 arithmetic are affected, this includes applications written in C or FORTRAN, as well as MATLAB.
See Floating-Point Functions for a list of functions most commonly used with floating-point numbers in MATLAB.
The following references provide more information about floating-point arithmetic.
[1] Moler, Cleve, "Floating Points," MATLAB News and Notes, Fall, 1996. A PDF version is available on the MathWorks Web site at http://www.mathworks.com/company/newsletters/news_notes/pdf/Fall96Cleve.pdf
[2] Moler, Cleve, Numerical Computing with MATLAB, S.I.A.M. A PDF version is available on the MathWorks Web site at http://www.mathworks.com/moler/.