Why Why 0.1 + 0.2 != 0.3 0.1 + 0.2 != 0.3 The Mysterious World of Floating Point Numbers The Mysterious World of Floating Point Numbers Benoit Jacquemont Benoit Jacquemont @bjacquemont @bjacquemont

$sum = 0.1 + 0.2; echo "Sum is ".$sum."\n"; if ($sum != 0.3) { echo "$sum is NOT EQUAL to 0.3\n"; } Sum is 0.3 0.3 is NOT EQUAL to 0.3 GO HOME GO HOME PHP PHP, ,

Let's try the same in JavaScript if (0.1 + 0.2 != 0.3) { console.log('0.1 + 0.2 is NOT EQUAL to 0.3'); } 0.1 + 0.2 is NOT EQUAL to 0.3

Let's continue with JavaScript... let sum = 0.1 + 0.2; console.log("Sum is " + sum); if (sum != 0.3) { console.log(sum + ' is NOT EQUAL to 0.3'); } Sum is 0.30000000000000004 0.30000000000000004 is NOT EQUAL to 0.3

Why in PHP ? $sum = 0.1 + 0.2; echo "Sum is ".$sum."\n"; if ($sum != 0.3) { echo "$sum is NOT EQUAL to 0.3\n"; } Sum is 0.3 0.3 is NOT EQUAL to 0.3

The precision con g option No impact on the precision of the computation! Round oating point numbers for display php.net/manual/en/ini.core.php#ini.precision

Increasing how many digits are displayed ini_set('precision', 17); // default 14 $sum = 0.1 + 0.2; echo "Sum is ".$sum."\n"; if ($sum != 0.3) { echo "$sum is NOT EQUAL to 0.3\n"; } Sum is 0.30000000000000004 0.30000000000000004 is NOT EQUAL to 0.3

0.1 + 0.2 == 0.30000000000000004 What is going on?

Our computers are Our computers are binary binary 0 0 or or 1 1 How to represent something in How to represent something in between? between?

Representing Real Numbers in Scienti c Notation 0.006458 6.458 x 10-3 6 . 458 x 10 -3 6: most signi cant digit, always non zero 458: other signi cant digits -3: exponent 3 parts, each representable by integers

Binary Implementation 0.0110111 1 . 10111 x 2 -10 Let's play with the exponent! Look ⇑, a oating point!

64 bits Floating Point Encoding sign 1 bit exponent 11 bits significand 52 bits

Binary Floating Point to Decimal Representation 000000000110000000000000000000000000000000000000000000000010101 1.10101 x 23 110.101 1 1 0 . 1 0 1 22= 4 21= 2 20= 1 2-1= ½ 2-2= ¼ 2-3= ⅛ 1.10101 x 22 = 4 + 2 + ½ + ⅛ = 6.625

Floating Point Numbers Solve Another Integers Limitation: Size

Largest Integer echo PHP_INT_MAX."\n"; 9223372036854775807 9.22 x 1018

Largest integer: 9.22 x 10 Largest integer: 9.22 x 1018 18 Earthworms mass: 7.6 x 10 Earthworms mass: 7.6 x 1012 12 kg kg

Largest integer: 9.22 x 10 Largest integer: 9.22 x 1018 18 Oceans mass: 1.42 x 10 Oceans mass: 1.42 x 1021 21 kg kg

Largest integer: 9.22 x 10 Largest integer: 9.22 x 1018 18 Earth mass: 5.42 x 10 Earth mass: 5.42 x 1024 24 kg kg

Largest integer: 9.22 x 10 Largest integer: 9.22 x 1018 18 Milky way mass: 2.98 x 10 Milky way mass: 2.98 x 1042 42 kg kg

Largest integer: 9.22 x 10 Largest integer: 9.22 x 1018 18 Observable Universe mass: Observable Universe mass: 1.5 x 10 1.5 x 1053 53 kg kg

Integers are very limited in size

Largest oating point number echo number_format(PHP_FLOAT_MAX, 0, "", "")."\n"; 1797693134862315708145274237317043567980705675258449965989174768031 5726078002853876058955863276687817154045895351438246423432132688946 4182768467546703537516986049910576551282076245490090389328944075868 5084551339423045832369032229481658085593321233482747978262041447231 68738177180919299881250404026184124858368 1.80 x 10308 290 orders of magnitude bigger than max integer

Floating point numbers are better than integers at representing very small and very large numbers

Float range is immensely wider than integers Integer: 1 to 9.2 x 1018 vs Float: 2.25 x 10-308 to 1.80 x 10308 But both use 64 bit storage...

Float can only represent a small sample of the numbers inside the boundaries

No content

No content

No content

No content

No content

No content

No content

No content

No content

Why 0.1 + 0.2 == 0.30000000000000004 ini_set('precision', 20); $op1 = 0.1; $op2 = 0.2; $expectedSum = 0.3; echo "Op #1: ".$op1."\n"; echo "Op #2: ".$op2."\n"; $actualSum = $op1 + $op2; echo "Actual Sum: ".$actualSum."\n"; echo "Expected Sum: ".$expectedSum."\n"; Op #1: 0.10000000000000000555 ← closest approximation of 0.1 Op #2: 0.2000000000000000111 ← closest approximation of 0.2 Actual Sum: 0.30000000000000004441 ← this is an approximation as well

Floating point numbers rule #1: Most oat numbers are approximations

How to have a consistent behavior between applications and hardware? 754 Standard So most computers and programming languages have the same behavior

Floating Point Numbers Dangers and Feats

No content

No content

No content

No content

Ariane 5 Maiden Flight 4 June 1996 Payload: 4 satellites of 1.2 T each. Total cost: $500 million

No content

Oops... Oops...

No content

No content

Same code as Ariane 4, but with 4x thrust... //64 bits float float horizontalVelocity = Inertial.getHorizontalVelocity(); // conversion to 16 bits integer (max 32767) int correction = computeCorrection((int) horizontalVelocity); // overflow! applyHorizontalCorrection(correction);

Lesson learned Converting oat to integer is highly risky Due to the oat wide range

Same in PHP $myFloat = 600000000000000000000; $myInt = (int) $myFloat; echo $myInt."\n"; -8742554432415203328 No error, and the result doesn't make sense...

The Gulf War & the The Gulf War & the Patriot Missile Event Patriot Missile Event

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

No content

// return the time in seconds in a float $timeBefore = getElapsedTimeFromStartup(); 1 2 3 sendRadarPingAndWaitForEcho(); 4 5 $timeAfter = getElapsedTimeFromStartup(); 6 7 $targetDistance = ($timeAfter - $timeBefore) * LIGHT_SPEED/2; 8 // return the time in seconds in a float $timeBefore = getElapsedTimeFromStartup(); sendRadarPingAndWaitForEcho(); 1 2 3 4 5 $timeAfter = getElapsedTimeFromStartup(); 6 7 $targetDistance = ($timeAfter - $timeBefore) * LIGHT_SPEED/2; 8 // return the time in seconds in a float $timeBefore = getElapsedTimeFromStartup(); sendRadarPingAndWaitForEcho(); $timeAfter = getElapsedTimeFromStartup(); 1 2 3 4 5 6 7 $targetDistance = ($timeAfter - $timeBefore) * LIGHT_SPEED/2; 8 // return the time in seconds in a float $timeBefore = getElapsedTimeFromStartup(); sendRadarPingAndWaitForEcho(); $timeAfter = getElapsedTimeFromStartup(); $targetDistance = ($timeAfter - $timeBefore) * LIGHT_SPEED/2; 1 2 3 4 5 6 7 8

Float rule #2 The bigger a oat, the less precise it becomes aka Loss of precision

// return the time in seconds in a float $timeBefore = getElapsedTimeFromStartup(); sendRadarPingAndWaitForEcho(); $timeAfter = getElapsedTimeFromStartup(); $targetDistance = ($timeAfter - $timeBefore) * LIGHT_SPEED/2; The longer the missile battery runs, the less precise getElapsedTimeFromStartup() becomes

Missile battery run time: 100 hours Inaccuray due to loss of precision: 0.3433s Scud missile speed: Mach 5 Patriot missile offset to its target: 589m

Lesson learned A oat number cannot be big and precise at the same time.

But Floating Point numbers can bring cool stuff

No content

Quake Quake Released in Feb. 1997 Released in Feb. 1997 Full real-time 3D Full real-time 3D rendering without 3D rendering without 3D accelerator accelerator Runs on a 60Mhz Runs on a 60Mhz Pentium Pentium

Lesson learned Floating Point Numbers calculations are crazy fast

What about us?

About me E-COMMERCE MOBILE APPLICATION PRINT CATALOG POINTS OF SALE ERP MEDIA SERVER SUPPLIERS PURCHASING DPT MARKETING DPT CSV FTP XML XLS SUPPLIERS PORTAL ENRIC H TRAN SLATE CONTROL Single Source of Truth for Product Information

No content

Storing real numbers in database mysql> CREATE TABLE my_float(f FLOAT); mysql> INSERT INTO my_float(f) VALUES(0.1); mysql> SELECT * FROM my_float; +------+ | f | +------+ | 0.1 | +------+ mysql> SELECT * FROM my_float WHERE f = 0.1; Empty set (0.00 sec) mysql> SELECT ROUND(f, 17) FROM my_float; +---------------------+ | ROUND(f, 17) | +---------------------+ | 0.10000000149011612 | +---------------------+ Loss of data due to approximation

Storing real numbers with Decimal mysql> CREATE TABLE my_decimal(d DECIMAL(10,5)); mysql> INSERT INTO my_decimal(d) VALUES(0.1); mysql> SELECT * FROM my_decimal; +---------+ | d | +---------+ | 0.10000 | +---------+ mysql> SELECT * FROM my_decimal WHERE d = 0.1; +---------+ | d | +---------+ | 0.10000 | +---------+ No approximation with xed point type

Transmitting real numbers ini_set('precision', 17); $myJsonFromAPI = '{"name": "foo", "price": 29.99}'; $myObject = json_decode($myJsonFromAPI, false); echo "Price:".$myObject->price."\n"; Price:29.989999999999998 Loss of data due to approximation Use strings to avoid approximation

When oat are shining Non-exhaustive list

Scienti c Calculations Scienti c Calculations

Sensors Sensors

Statistics Statistics

Machine Learning Machine Learning

When avoiding oat... ...and how. Non-exhaustive list

Money, payment & billing $price = 49.99; $coupon = 20.25; $card = 29.74; 1 2 3 4 echo "Price: $price. Using coupon $coupon.\n"; 5 $remaining = $price - $coupon; 6 7 echo "$remaining still to pay. Using card $card.\n"; 8 $remaining = $remaining - $card; 9 10 if ($remaining == 0) { 11 echo "Thank you for your payment!\n"; 12 } else { 13 echo "Paiement not finished: $remaining to pay\n"; 14 } 15 echo "Price: $price. Using coupon $coupon.\n"; $remaining = $price - $coupon; $price = 49.99; 1 $coupon = 20.25; 2 $card = 29.74; 3 4 5 6 7 echo "$remaining still to pay. Using card $card.\n"; 8 $remaining = $remaining - $card; 9 10 if ($remaining == 0) { 11 echo "Thank you for your payment!\n"; 12 } else { 13 echo "Paiement not finished: $remaining to pay\n"; 14 } 15 echo "$remaining still to pay. Using card $card.\n"; $remaining = $remaining - $card; $price = 49.99; 1 $coupon = 20.25; 2 $card = 29.74; 3 4 echo "Price: $price. Using coupon $coupon.\n"; 5 $remaining = $price - $coupon; 6 7 8 9 10 if ($remaining == 0) { 11 echo "Thank you for your payment!\n"; 12 } else { 13 echo "Paiement not finished: $remaining to pay\n"; 14 } 15 $price = 49.99; $coupon = 20.25; $card = 29.74; echo "Price: $price. Using coupon $coupon.\n"; $remaining = $price - $coupon; echo "$remaining still to pay. Using card $card.\n"; $remaining = $remaining - $card; if ($remaining == 0) { echo "Thank you for your payment!\n"; } else { echo "Paiement not finished: $remaining to pay\n"; } 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Price: 49.99. Using coupon 20.25. 1 29.74 still to pay. Using card 29.74. 2 Paiement not finished: 3.5527136788005E-15 to pay 3 Price: 49.99. Using coupon 20.25. 29.74 still to pay. Using card 29.74. 1 2 Paiement not finished: 3.5527136788005E-15 to pay 3 Price: 49.99. Using coupon 20.25. 29.74 still to pay. Using card 29.74. Paiement not finished: 3.5527136788005E-15 to pay 1 2 3

Money, payment & billing ini_set('precision', 17); //Only change! $price = 49.99; $coupon = 20.25; $card = 29.74; echo "Price: $price. Using coupon $coupon.\n"; $remaining = $price - $coupon; echo "$remaining still to pay. Using card $card.\n"; $remaining = $remaining - $card; if ($remaining == 0) { echo "Thank you for your payment!\n"; } else { echo "Paiement not finished: $remaining to pay\n"; } Price: 49.990000000000002. Using coupon 20.25. 29.740000000000002 still to pay. Using card 29.739999999999998. Paiement not finished: 3.5527136788005009E-15 to pay

Using Floating Point Numbers with money: It seems to work... ... until it doesn't.

Money is a discrete entity. No approximation in money!

Use cents and integers $price = 4999; $coupon = 2025; $card = 2974; function f(int $amount): string { return substr($amount, 0, -2).'.'.substr($amount, -2); } echo "Price:".f($price).". Using coupon ".f($coupon).".\n"; $remaining = $price - $coupon; echo f($remaining).' still to pay. Using card '.f($card)."\n"; $remaining = $remaining - $card; if ($remaining == 0) { echo "Thank you for your payment!\n"; } else { echo "Payement not finished.\nRemaining amount to pay:".f($remaining)."\n"; } Price:49.99. Using coupon 20.25. 29.74 still to pay. Using card 29.74 Thank you for your payment!

All amount in cents

Other alternatives to Floating point numbers

Strings ini_set('precision', 20); $myRealNumber = "0.1"; echo $myRealNumber."\n"; 0.1 Simple No computation No comparison

BCMath Arbitrary Precision Mathematics bcscale(2); $sum = bcadd("0.1", "0.2"); // strings!! if (bccomp($sum, "0.3")) { echo "$sum is NOT EQUAL to 0.3\n"; } else { echo "$sum is EQUAL to 0.3\n"; } 0.30 IS EQUAL to 0.3 Computation Comparison Needs an extension

php decimal Arbitrary-Precision Decimal Arithmetic For PHP 7 use Decimal\Decimal; $op1 = new Decimal("0.1"); $op2 = new Decimal("0.2"); $sum = $op1->add($op2); if (!$sum->equals("0.3")) { echo "$sum is NOT EQUAL to 0.3\n"; } else { echo "$sum is EQUAL to 0.3\n"; } 0.3 is EQUAL to 0.3 Nice object interface Not easily available php-decimal.io

Floating Point Numbers Fear them Respect them Love them Thank you! joind.in/talk/1d52e