Related
I'm just curious, why in IEEE-754 any non zero float number divided by zero results in infinite value? It's a nonsense from the mathematical perspective. So I think that correct result for this operation is NaN.
Function f(x) = 1/x is not defined when x=0, if x is a real number. For example, function sqrt is not defined for any negative number and sqrt(-1.0f) if IEEE-754 produces a NaN value. But 1.0f/0 is Inf.
But for some reason this is not the case in IEEE-754. There must be a reason for this, maybe some optimization or compatibility reasons.
So what's the point?
It's a nonsense from the mathematical perspective.
Yes. No. Sort of.
The thing is: Floating-point numbers are approximations. You want to use a wide range of exponents and a limited number of digits and get results which are not completely wrong. :)
The idea behind IEEE-754 is that every operation could trigger "traps" which indicate possible problems. They are
Illegal (senseless operation like sqrt of negative number)
Overflow (too big)
Underflow (too small)
Division by zero (The thing you do not like)
Inexact (This operation may give you wrong results because you are losing precision)
Now many people like scientists and engineers do not want to be bothered with writing trap routines. So Kahan, the inventor of IEEE-754, decided that every operation should also return a sensible default value if no trap routines exist.
They are
NaN for illegal values
signed infinities for Overflow
signed zeroes for Underflow
NaN for indeterminate results (0/0) and infinities for (x/0 x != 0)
normal operation result for Inexact
The thing is that in 99% of all cases zeroes are caused by underflow and therefore in 99%
of all times Infinity is "correct" even if wrong from a mathematical perspective.
I'm not sure why you would believe this to be nonsense.
The simplistic definition of a / b, at least for non-zero b, is the unique number of bs that has to be subtracted from a before you get to zero.
Expanding that to the case where b can be zero, the number that has to be subtracted from any non-zero number to get to zero is indeed infinite, because you'll never get to zero.
Another way to look at it is to talk in terms of limits. As a positive number n approaches zero, the expression 1 / n approaches "infinity". You'll notice I've quoted that word because I'm a firm believer in not propagating the delusion that infinity is actually a concrete number :-)
NaN is reserved for situations where the number cannot be represented (even approximately) by any other value (including the infinities), it is considered distinct from all those other values.
For example, 0 / 0 (using our simplistic definition above) can have any amount of bs subtracted from a to reach 0. Hence the result is indeterminate - it could be 1, 7, 42, 3.14159 or any other value.
Similarly things like the square root of a negative number, which has no value in the real plane used by IEEE754 (you have to go to the complex plane for that), cannot be represented.
In mathematics, division by zero is undefined because zero has no sign, therefore two results are equally possible, and exclusive: negative infinity or positive infinity (but not both).
In (most) computing, 0.0 has a sign. Therefore we know what direction we are approaching from, and what sign infinity would have. This is especially true when 0.0 represents a non-zero value too small to be expressed by the system, as it frequently the case.
The only time NaN would be appropriate is if the system knows with certainty that the denominator is truly, exactly zero. And it can't unless there is a special way to designate that, which would add overhead.
NOTE:
I re-wrote this following a valuable comment from #Cubic.
I think the correct answer to this has to come from calculus and the notion of limits. Consider the limit of f(x)/g(x) as x->0 under the assumption that g(0) == 0. There are two broad cases that are interesting here:
If f(0) != 0, then the limit as x->0 is either plus or minus infinity, or it's undefined. If g(x) takes both signs in the neighborhood of x==0, then the limit is undefined (left and right limits don't agree). If g(x) has only one sign near 0, however, the limit will be defined and be either positive or negative infinity. More on this later.
If f(0) == 0 as well, then the limit can be anything, including positive infinity, negative infinity, a finite number, or undefined.
In the second case, generally speaking, you cannot say anything at all. Arguably, in the second case NaN is the only viable answer.
Now in the first case, why choose one particular sign when either is possible or it might be undefined? As a practical matter, it gives you more flexibility in cases where you do know something about the sign of the denominator, at relatively little cost in the cases where you don't. You may have a formula, for example, where you know analytically that g(x) >= 0 for all x, say, for example, g(x) = x*x. In that case the limit is defined and it's infinity with sign equal to the sign of f(0). You might want to take advantage of that as a convenience in your code. In other cases, where you don't know anything about the sign of g, you cannot generally take advantage of it, but the cost here is just that you need to trap for a few extra cases - positive and negative infinity - in addition to NaN if you want to fully error check your code. There is some price there, but it's not large compared to the flexibility gained in other cases.
Why worry about general functions when the question was about "simple division"? One common reason is that if you're computing your numerator and denominator through other arithmetic operations, you accumulate round-off errors. The presence of those errors can be abstracted into the general formula format shown above. For example f(x) = x + e, where x is the analytically correct, exact answer, e represents the error from round-off, and f(x) is the floating point number that you actually have on the machine at execution.
Prompted by a spot of earlier code golfing why would:
>NaN^0
[1] 1
It makes perfect sense for NA^0 to be 1 because NA is missing data, and any number raised to 0 will give 1, including -Inf and Inf. However NaN is supposed to represent not-a-number, so why would this be so? This is even more confusing/worrying when the help page for ?NaN states:
In R, basically all mathematical functions (including basic
Arithmetic), are supposed to work properly with +/- Inf and NaN as
input or output.
The basic rule should be that calls and relations with Infs really are
statements with a proper mathematical limit.
Computations involving NaN will return NaN or perhaps NA: which of
those two is not guaranteed and may depend on the R platform (since
compilers may re-order computations).
Is there a philosophical reason behind this, or is it just to do with how R represents these constants?
This is referenced in the help page referenced by ?'NaN'
"The IEC 60559 standard, also known as the ANSI/IEEE 754 Floating-Point Standard.
http://en.wikipedia.org/wiki/NaN."
And there you find this statement regarding what should create a NaN:
"There are three kinds of operations that can return NaN:[5]
Operations with a NaN as at least one operand.
It is probably is from the particular C compiler, as signified by the Note you referenced. This is what the GNU C documentation says:
http://www.gnu.org/software/libc/manual/html_node/Infinity-and-NaN.html
" NaN, on the other hand, infects any calculation that involves it. Unless the calculation would produce the same result no matter what real value replaced NaN, the result is NaN."
So it seems that the GNU-C people have a different standard in mind when writing their code. And the 2008 version of ANSI/IEEE 754 Floating-Point Standard is reported to make that suggestion:
http://en.wikipedia.org/wiki/NaN#Function_definition
The published standard is not free. So if you are have access rights or money you can look here:
http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=4610933
The answer can be summed up by "for historical reasons".
It seems that IEEE 754 introduced two different power functions - pow and powr, with the latter preserving NaN's in the OP case and also returning NaN for Inf^0, 0^0, 1^Inf, but eventually the latter was dropped as explained briefly here.
Conceptually, I'm in the NaN preserving camp, because I'm coming at the issue from viewpoint of limits, but from convenience point of view I expect current conventions are slightly easier to deal with, even if they don't make a lot of sense in some cases (e.g. sqrt(-1)^0 being equal to 1 while all operations are on real numbers makes little sense if any).
Yes, I'm late here, but as R Core member who was involved in this design, let me recall what I commented above. NaN preserving and NA preserving work "equivalently" in R, so if you agree that NA^0 should give 1, NaN^0 |-> 1 is a consequence.
Indeed (as others said) you should really read R's help pages and not C or
IEEE standards, to answer such questions,
and SimonO101 correctly cited
1 ^ y and y ^ 0 are 1, always
and I'm pretty sure that I was heavily involved (if not the author) of that.
Note that it is good, not bad, to be able to provide non-NaN answers, also in cases other programming languages do differently.
The consequence of such a rule is that more things work automatically correctly;
in the other case, the R programmer would have been urged to do more special casing herself.
Or put differently, a simple rule as the above (returning non-NaN in all cases) is a good rule, because it propagates continuity in a mathematical sense: lim_x f(x) = f(lim x).
We have had a few cases where it was clearly advantageous (i.e. did not need special casing, I'm repeating..) to adhere to the above "= 1" rule, rather than to propagate NaN. As I said further up, the sqrt(-1)^0 is also such an example, as 1 is the correct result as soon as you extend to the complex plane.
Here's one reasoning. From Goldberg:
In IEEE 754, NaNs are often represented as floating-point numbers with
the exponent e_max + 1 and nonzero significands.
So NaN is a floating-point number, though with a special meaning. Raising a number to the power zero sets its exponent to zero, therefore it will no longer be NaN.
Also note:
> 1^NaN
[1] 1
One is a number whose exponent is zero already.
Conceptually, the only problem with NaN^0 == 1 is that zero values can come about at least four different ways, but the IEEE format uses the same representation for three of them. The above formula equality sense for the most common case (which is one of the three), but not for the others.
BTW, the four cases I would recognize would be:
A literal zero
Unsigned zero: the difference between two numbers that are indistinguishable
Positive infinitesimal: The product or quotient of two numbers of matching sign, which is too small to be distinguished from zero.
Negative infinitesimal: The product or quotient of two numbers of opposite sign, which is too small to be distinguished from zero.
Some of these may be produced via other means (e.g. literal zero could be produced as the sum of two literal zeros; positive infinitesimal by the division of a very small number by a very large one, etc.).
If a floating-point recognized the above, it could usefully regard raising NaN to a literal zero as yielding one, and raising it to any other kind of zero as yielding NaN; such a rule would allow a constant result to be assumed in many cases where something that might be NaN would be raised to something the compiler could identify as a constant zero, without such assumption altering program semantics. Otherwise, I think the issue is that most code isn't going to care whether x^0 might would NaN if x is NaN, and there's not much point to having a compiler add code for conditions code isn't going to care about. Note that the issue isn't just the code to compute x^0, but for any computations based on that which would be constant if x^0 was.
If you look at the type of NaN, it is still a number, it's just not a specific number that can be represented by the numeric type.
EDIT:
For example, if you were to take 0/0. What is the result? If you tried to solve this equation on paper, you get stuck at the very first digit, how many zero's fit into another 0? You can put 0, you can put 1, you can put 8, they all fit into 0*x=0 but it's impossible to know which one the correct answer is. However, that does not mean the answer is no longer a number, it's just not a number that can be represented.
Regardless, any number, even a number that you can't represent, to the power of zero is still 1. If you break down some math x^8 * x^0 can be further simplified by x^(8+0) which equates to x^8, where did the x^0 go? It makes sense if x^0 = 1 because then the equation x^8 * 1 explains why x^0 just sort of disappears from existence.
If I run following line of code, I get DIVIDE BY ZERO error
1. System.out.println(5/0);
which is the expected behavior.
Now I run the below line of code
2. System.out.println(5/0F);
here there is no DIVIDE BY ZERO error, rather it shows INFINITY
In the first line I am dividing two integers and in the second two real numbers.
Why does dividing by zero for integers gives DIVIDE BY ZERO error while in the case of real numbers it gives INFINITY
I am sure it is not specific to any programming language.
(EDIT: The question has been changed a bit - it specifically referred to Java at one point.)
The integer types in Java don't have representations of infinity, "not a number" values etc - whereas IEEE-754 floating point types such as float and double do. It's as simple as that, really. It's not really a "real" vs "integer" difference - for example, BigDecimal represents real numbers too, but it doesn't have a representation of infinity either.
EDIT: Just to be clear, this is language/platform specific, in that you could create your own language/platform which worked differently. However, the underlying CPUs typically work the same way - so you'll find that many, many languages behave this way.
EDIT: In terms of motivation, bear in mind that for the infinity case in particular, there are ways of getting to infinity without dividing by zero - such as dividing by a very, very small floating point number. In the case of integers, there's obviously nothing between zero and one.
Also bear in mind that the cases in which integers (or decimal floating point types) are used typically don't need to concept of infinity, or "not a number" results - whereas in scientific applications (where float/double are more typically useful), "infinity" (or at least, "a number which is too large to sensibly represent") is still a potentially valid result.
This is specific to one programming language or a family of languages. Not all languages allow integers and floats to be used in the same expression. Not all languages have both types (for example, ECMAScript implementations like JavaScript have no notion of an integer type externally). Not all languages have syntax like this to convert values inline.
However, there is an intrinsic difference between integer arithmetic and floating-point arithmetic. In integer arithmetic, you must define that division by zero is an error, because there are no values to represent the result. In floating-point arithmetic, specifically that defined in IEEE-754, there are additional values (combinations of sign bit, exponent and mantissa) for the mathematical concept of infinity and meta-concepts like NaN (not a number).
So we can assume that the / operator in this programming language is generic, that it performs integer division if both operands are of the language's integer type; and that it performs floating-point division if at least one of the operands is of a float type of the language, whereas the other operands would be implicitly converted to that float type for the purpose of the operation.
In real-number math, division of a number by a number close to zero is equivalent to multiplying the first number by a number whose absolute is very large (x / (1 / y) = x * y). So it is reasonable that the result of dividing by zero should be (defined as) infinity as the precision of the floating-point value would be exceeded.
Implementation details were to be found in that programming language's specification.
It's clear that one shouldn't use floating precision when working with, say, monetary amounts since the variation in precision leads to inaccuracies when doing calculations with that amount.
That said, what are use cases when that is acceptable? And, what are the general principles one should have in mind when deciding?
Floating point numbers should be used for what they were designed for: computations where what you want is a fixed precision, and you only care that your answer is accurate to within a certain tolerance. If you need an exact answer in all cases, you're best using something else.
Here are three domains where you might use floating point:
Scientific Simulations
Science apps require a lot of number crunching, and often use sophisticated numerical methods to solve systems of differential equations. You're typically talking double-precision floating point here.
Games
Think of games as a simulation where it's ok to cheat. If the physics is "good enough" to seem real then it's ok for games, and you can make up in user experience what you're missing in terms of accuracy. Games usually use single-precision floating point.
Stats
Like science apps, statistical methods need a lot of floating point. A lot of the numerical methods are the same; the application domain is just different. You find a lot of statistics and monte carlo simulations in financial applications and in any field where you're analyzing a lot of survey data.
Floating point isn't trivial, and for most business applications you really don't need to know all these subtleties. You're fine just knowing that you can't represent some decimal numbers exactly in floating point, and that you should be sure to use some decimal type for prices and things like that.
If you really want to get into the details and understand all the tradeoffs and pitfalls, check out the classic What Every Programmer Should Know About Floating Point, or pick up a book on Numerical Analysis or Applied Numerical Linear Algebra if you're really adventurous.
I'm guessing you mean "floating point" here. The answer is, basically, any time the quantities involved are approximate, measured, rather than precise; any time the quantities involved are larger than can be conveniently represented precisely on the underlying machine; any time the need for computational speed overwhelms exact precision; and any time the appropriate precision can be maintained without other complexities.
For more details of this, you really need to read a numerical analysis book.
Short story is that if you need exact calculations, DO NOT USE floating point.
Don't use floating point numbers as loop indices: Don't get caught doing:
for ( d = 0.1; d < 1.0; d+=0.1)
{ /* Some Code... */ }
You will be surprised.
Don't use floating point numbers as keys to any sort of map because you can never count on equality behaving like you may expect.
Most real-world quantities are inexact, and typically we know their numeric properties with a lot less precision than a typical floating-point value. In almost all cases, the C types float and double are good enough.
It is necessary to know some of the pitfalls. For example, testing two floating-point numbers for equality is usually not what you want, since all it takes is a single bit of inaccuracy to make the comparison non-equal. tgamblin has provided some good references.
The usual exception is money, which is calculated exactly according to certain conventions that don't translate well to binary representations. Part of this is the constants used: you'll never see a pi% interest rate, or a 22/7% interest rate, but you might well see a 3.14% interest rate. In other words, the numbers used are typically expressed in exact decimal fractions, not all of which are exact binary fractions. Further, the rounding in calculations is governed by conventions that also don't translate well into binary. This makes it extremely difficult to precisely duplicate financial calculations with standard floating point, and therefore people use other methods for them.
It's appropriate to use floating point types when dealing with scientific or statistical calculations. These will invariably only have, say, 3-8 significant digits of accuracy.
As to whether to use single or double precision floating point types, this depends on your need for accuracy and how many significant digits you need. Typically though people just end up using doubles unless they have a good reason not to.
For example if you measure distance or weight or any physical quantity like that the number you come up with isn't exact: it has a certain number of significant digits based on the accuracy of your instruments and your measurements.
For calculations involving anything like this, floating point numbers are appropriate.
Also, if you're dealing with irrational numbers floating point types are appropriate (and really your only choice) eg linear algebra where you deal with square roots a lot.
Money is different because you typically need to be exact and every digit is significant.
I think you should ask the other way around: when should you not use floating point. For most numerical tasks, floating point is the preferred data type, as you can (almost) forget about overflow and other kind of problems typically encountered with integer types.
One way to look at floating point data type is that the precision is independent of the dynamic, that is whether the number is very small of very big (within an acceptable range of course), the number of meaningful digits is approximately the same.
One drawback is that floating point numbers have some surprising properties, like x == x can be False (if x is nan), they do not follow most mathematical rules (distributivity, that is x( y + z) != xy + xz). Depending on the values for z, y, and z, this can matters.
From Wikipedia:
Floating-point arithmetic is at its
best when it is simply being used to
measure real-world quantities over a
wide range of scales (such as the
orbital period of Io or the mass of
the proton), and at its worst when it
is expected to model the interactions
of quantities expressed as decimal
strings that are expected to be exact.
Floating point is fast but inexact. If that is an acceptable trade off, use floating point.
Most mathematicians agree that:
eπi + 1 = 0
However, most floating point implementations disagree. How well can we settle this dispute?
I'm keen to hear about different languages and implementations, and various methods to make the result as close to zero as possible. Be creative!
It's not that most floating point implementations disagree, it's just that they cannot get the accuracy necessary to get a 100% answer. And the correct answer is that they can't.
PI is an infinite series of digits that nobody has been able to denote by anything other than a symbolic representation, and e^X is the same, and thus the only way to get to 100% accuracy is to go symbolic.
Here's a short list of implementations and languages I've tried. It's sorted by closeness to zero:
Scheme: (+ 1 (make-polar 1 (atan 0 -1)))
⇒ 0.0+1.2246063538223773e-16i (Chez Scheme, MIT Scheme)
⇒ 0.0+1.22460635382238e-16i (Guile)
⇒ 0.0+1.22464679914735e-16i (Chicken with numbers egg)
⇒ 0.0+1.2246467991473532e-16i (MzScheme, SISC, Gauche, Gambit)
⇒ 0.0+1.2246467991473533e-16i (SCM)
Common Lisp: (1+ (exp (complex 0 pi)))
⇒ #C(0.0L0 -5.0165576136843360246L-20) (CLISP)
⇒ #C(0.0d0 1.2246063538223773d-16) (CMUCL)
⇒ #C(0.0d0 1.2246467991473532d-16) (SBCL)
Perl: use Math::Complex; Math::Complex->emake(1, pi) + 1
⇒ 1.22464679914735e-16i
Python: from cmath import exp, pi; exp(complex(0, pi)) + 1
⇒ 1.2246467991473532e-16j (CPython)
Ruby: require 'complex'; Complex::polar(1, Math::PI) + 1
⇒ Complex(0.0, 1.22464679914735e-16) (MRI)
⇒ Complex(0.0, 1.2246467991473532e-16) (JRuby)
R: complex(argument = pi) + 1
⇒ 0+1.224606353822377e-16i
Is it possible to settle this dispute?
My first thought is to look to a symbolic language, like Maple. I don't think that counts as floating point though.
In fact, how does one represent i (or j for the engineers) in a conventional programming language?
Perhaps a better example is sin(π) = 0? (Or have I missed the point again?)
I agree with Ryan, you would need to move to another number representation system. The solution is outside the realm of floating point math because you need pi to represented as an infinitely long decimal so any limited precision scheme just isn't going to work (at least not without employing some kind of fudge-factor to make up the lost precision).
Your question seems a little odd to me, as you seem to be suggesting that the Floating Point math is implemented by the language. That's generally not true, as the FP math is done using a floating point processor in hardware. But software or hardware, floating point will always be inaccurate. That's just how floats work.
If you need better precision you need to use a different number representation. Just like if you're doing integer math on numbers that don't fit in an int or long. Some languages have libraries for that built in (I know java has BigInteger and BigDecimal), but you'd have to explicitly use those libraries instead of native types, and the performance would be (sometimes significantly) worse than if you used floats.
#Ryan Fox In fact, how does one represent i (or j for the engineers) in a conventional programming language?
Native complex data types are far from unknown. Fortran had it by the mid-sixties, and the OP exhibits a variety of other languages that support them in hist followup.
And complex numbers can be added to other languages as libraries (with operator overloading they even look just like native types in the code).
But unless you provide a special case for this problem, the "non-agreement" is just an expression of imprecise machine arithmetic, no? It's like complaining that
float r = 2/3;
float s = 3*r;
float t = s - 2;
ends with (t != 0) (At least if you use an dumb enough compiler)...
I had looooong coffee chats with my best pal talking about Irrational numbers and the diference between other numbers. Well, both of us agree in this different point of view:
Irrational numbers are relations, as functions, in a way, what way? Well, think about "if you want a perfect circle, give me a perfect pi", but circles are diferent to the other figures (4 sides, 5, 6... 100, 200) but... How many more sides do you have, more like a circle it look like. If you followed me so far, connecting all this ideas here is the pi formula:
So, pi is a function, but one that never ends! because of the ∞ parameter, but I like to think that you can have "instance" of pi, if you change the ∞ parameter for a very big Int, you will have a very big pi instance.
Same with e, give me a huge parameter, I will give you a huge e.
Putting all the ideas together:
As we have memory limitations, the language and libs provide to us huge instance of irrational numbers, in this case, pi and e, as final result, you will have long aproach to get 0, like the examples provided by #Chris Jester-Young
In fact, how does one represent i (or j for the engineers) in a conventional programming language?
In a language that doesn't have a native representation, it is usually added using OOP to create a Complex class to represent i and j, with operator overloading to properly deal with operations involving other Complex numbers and or other number primitives native to the language.
Eg: Complex.java, C++ < complex >
Numerical Analysis teaches us that you can't rely on the precise value of small differences between large numbers.
This doesn't just affect the equation in question here, but can bring instability to everything from solving a near-singular set of simultaneous equations, through finding the zeros of polynomials, to evaluating log(~1) or exp(~0) (I have even seen special functions for evaluating log(x+1) and (exp(x)-1) to get round this).
I would encourage you not to think in terms of zeroing the difference -- you can't -- but rather in doing the associated calculations in such a way as to ensure the minimum error.
I'm sorry, it's 43 years since I had this drummed into me at uni, and even if I could remember the references, I'm sure there's better stuff around now. I suggest this as a starting point.
If that sounds a bit patronising, I apologise. My "Numerical Analysis 101" was part of my Chemistry course, as there wasn't much CS in those days. I don't really have a feel for the place/importance numerical analysis has in a modern CS course.
It's a limitation of our current floating point computational architectures. Floating point arithmetic is only an approximation of numeric poles like e or pi (or anything beyond the precision your bits allow). I really enjoy these numbers because they defy classification, and appear to have greater entropy(?) than even primes, which are a canonical series. A ratio defy's numerical representation, sometimes simple things like that can blow a person's mind (I love it).
Luckily entire languages and libraries can be dedicated to precision trigonometric functions by using notational concepts (similar to those described by Lasse V. Karlsen ).
Consider a library/language that describes concepts like e and pi in a form that a machine can understand. Does a machine have any notion of what a perfect circle is? Probably not, but we can create an object - circle that satisfies all the known features we attribute to it (constant radius, relationship of radius to circumference is 2*pi*r = C). An object like pi is only described by the aforementioned ratio. r & C can be numeric objects described by whatever precision you want to give them. e can be defined "as the e is the unique real number such that the value of the derivative (slope of the tangent line) of the function f(x) = ex at the point x = 0 is exactly 1" from wikipedia.
Fun question.