Skip to content

A performant numpy extension for Galois fields and their applications

License

Notifications You must be signed in to change notification settings

TrendingTechnology/galois

 
 

Repository files navigation

Galois: A performant numpy extension for Galois fields and their applications

PyPI version Supported Versions Documentation Status Lint Test Codecov

The main idea of the galois package can be summarized as follows. The user creates a "Galois field array class" using GF = galois.GF(p**m). A Galois field array class GF is a subclass of np.ndarray and its constructor x = GF(array_like) mimics the call signature of np.array(). A Galois field array x is operated on like any other numpy array, but all arithmetic is performed in GF(p^m) not Z or R.

Internally, the Galois field arithmetic is implemented by replacing numpy ufuncs. The new ufuncs are written in python and then just-in-time compiled with numba. The ufuncs can be configured to use either lookup tables (for speed) or explicit calculation (for memory savings).

In addition to normal array arithmetic, galois also supports linear algebra (with np.linalg functions), polynomials over Galois fields (with galois.Poly), and forward error correction codes (with galois.BCH and galois.ReedSolomon).

Features

Roadmap

  • Number-theoretic transform, DFT over Galois fields
  • Elliptic curves over Galois fields
  • Group and ring arrays
  • GPU support

Documentation

The documentation for galois can be found at https://galois.readthedocs.io/en/stable/. It includes installation instructions, basic usage, tutorials, a development guide, an API reference, and release notes.

Installation

The latest version of galois can be installed from PyPI using pip.

python3 -m pip install galois

Versioning

This project uses semantic versioning. Releases are versioned major.minor.patch. Major releases introduce API-changing features. Minor releases add features and are backwards compatible with other releases in major.x.x. Patch releases fix bugs in a minor release and are backwards compatible with other releases in major.minor.x.

Releases before 1.0.0 are alpha and beta releases. Alpha releases are 0.0.alpha. There is no API compatibility guarantee for them. They can be thought of as 0.0.alpha-major. Beta releases are 0.beta.x and are API compatible. They can be thought of as 0.beta-major.beta-minor.

Basic Usage

Galois field arrays

Class construction

Galois field array classes are created using the galois.GF() class factory function.

>>> import numpy as np

>>> import galois

>>> GF256 = galois.GF(2**8)

>>> print(GF256)
<class 'np.ndarray over GF(2^8)'>

These classes are subclasses of galois.FieldArray (which itself subclasses np.ndarray) and are instances of galois.FieldClass.

>>> isinstance(GF256, galois.FieldClass)
True

>>> issubclass(GF256, galois.FieldArray)
True

>>> issubclass(GF256, np.ndarray)
True

A Galois field array class contains attributes relating to its Galois field and has methods to modify how the field is calculated or displayed. See the attributes and methods in galois.FieldClass.

# Summarizes some properties of the Galois field
>>> print(GF256.properties)
GF(2^8):
  characteristic: 2
  degree: 8
  order: 256
  irreducible_poly: Poly(x^8 + x^4 + x^3 + x^2 + 1, GF(2))
  is_primitive_poly: True
  primitive_element: GF(2, order=2^8)

# Access each attribute individually
>>> GF256.irreducible_poly
Poly(x^8 + x^4 + x^3 + x^2 + 1, GF(2))

The galois package even supports arbitrarily-large fields! This is accomplished by using numpy arrays with dtype=object and pure-python ufuncs. This comes at a performance penalty compared to smaller fields which use numpy integer dtypes (e.g., np.uint32) and have compiled ufuncs.

>>> GF = galois.GF(36893488147419103183); print(GF.properties)
GF(36893488147419103183):
  characteristic: 36893488147419103183
  degree: 1
  order: 36893488147419103183

>>> GF = galois.GF(2**100); print(GF.properties)
GF(2^100):
  characteristic: 2
  degree: 100
  order: 1267650600228229401496703205376
  irreducible_poly: Poly(x^100 + x^57 + x^56 + x^55 + x^52 + x^48 + x^47 + x^46 + x^45 + x^44 + x^43 + x^41 + x^37 + x^36 + x^35 + x^34 + x^31 + x^30 + x^27 + x^25 + x^24 + x^22 + x^20 + x^19 + x^16 + x^15 + x^11 + x^9 + x^8 + x^6 + x^5 + x^3 + 1, GF(2))
  is_primitive_poly: True
  primitive_element: GF(2, order=2^100)

Array creation

Galois field arrays can be created from existing numpy arrays.

# Represents an existing numpy array
>>> array = np.random.randint(0, GF256.order, 10, dtype=int); array
array([ 31, 254, 155, 154, 121, 185,  16, 246, 216, 244])

# Explicit Galois field array creation (a copy is performed)
>>> GF256(array)
GF([ 31, 254, 155, 154, 121, 185,  16, 246, 216, 244], order=2^8)

# Or view an existing numpy array as a Galois field array (no copy is performed)
>>> array.view(GF256)
GF([ 31, 254, 155, 154, 121, 185,  16, 246, 216, 244], order=2^8)

Or they can be created from "array-like" objects. These include strings representing a Galois field element as a polynomial over its prime subfield.

# Arrays can be specified as iterables of iterables
>>> GF256([[217, 130, 42], [74, 208, 113]])
GF([[217, 130,  42],
    [ 74, 208, 113]], order=2^8)

# You can mix-and-match polynomial strings and integers
>>> GF256(["x^6 + 1", 2, "1", "x^5 + x^4 + x"])
GF([65,  2,  1, 50], order=2^8)

# Single field elements are 0-dimensional arrays
>>> GF256("x^6 + x^4 + 1")
GF(81, order=2^8)

Galois field arrays also have constructor class methods for convenience. They include:

  • FieldArray.Zeros, FieldArray.Ones, FieldArray.Identity, FieldArray.Range, FieldArray.Random, FieldArray.Elements

Galois field elements can either be displayed using their integer representation, polynomial representation, or power representation. Calling FieldClass.display will change the element representation. If called as a context manager, the display mode will only be temporarily changed.

>>> x = GF256(["y**6 + 1", 0, 2, "1", "y**5 + y**4 + y"]); x
GF([65,  0,  2,  1, 50], order=2^8)

# Set the display mode to represent GF(2^8) field elements as polynomials over GF(2) with degree less than 8
>>> GF256.display("poly");

>>> x
GF([α^6 + 1, 0, α, 1, α^5 + α^4 + α], order=2^8)

# Temporarily set the display mode to represent GF(2^8) field elements as powers of the primitive element
>>> with GF256.display("power"):
...     print(x)

GF([α^191, 0, α, 1, α^194], order=2^8)

# Resets the display mode to the integer representation
>>> GF256.display();

Field arithmetic

Galois field arrays are treated like any other numpy array. Array arithmetic is performed using python operators or numpy functions.

In the list below, GF is a Galois field array class created by GF = galois.GF(p**m), x and y are GF arrays, and z is an integer np.ndarray. All arithmetic operations follow normal numpy broadcasting rules.

  • Addition: x + y == np.add(x, y)
  • Subtraction: x - y == np.subtract(x, y)
  • Multiplication: x * y == np.multiply(x, y)
  • Division: x / y == x // y == np.divide(x, y)
  • Scalar multiplication: x * z == np.multiply(x, z), e.g. x * 3 == x + x + x
  • Additive inverse: -x == np.negative(x)
  • Multiplicative inverse: GF(1) / x == np.reciprocal(x)
  • Exponentiation: x ** z == np.power(x, z), e.g. x ** 3 == x * x * x
  • Logarithm: np.log(x), e.g. GF.primitive_element ** np.log(x) == x
  • COMING SOON: Logarithm base b: GF.log(x, b), where b is any field element
  • Matrix multiplication: A @ B == np.matmul(A, B)
>>> x = GF256.Random((2,5)); x
GF([[166,  71, 214, 164, 228],
    [168, 202,  73,  54, 180]], order=2^8)

>>> y = GF256.Random(5); y
GF([ 25, 102, 131, 233, 188], order=2^8)

# y is broadcast over the last dimension of x
>>> x + y
GF([[191,  33,  85,  77,  88],
    [177, 172, 202, 223,   8]], order=2^8)

Linear algebra

The galois package intercepts relevant calls to numpy's linear algebra functions and performs the specified operation in GF(p^m) not in R. Some of these functions include:

  • np.dot, np.vdot, np.inner, np.outer, np.matmul, np.linalg.matrix_power
  • np.linalg.det, np.linalg.matrix_rank, np.trace
  • np.linalg.solve, np.linalg.inv
>>> A = GF256.Random((3,3)); A
GF([[151, 147, 229],
    [162, 192,  59],
    [ 52, 213,  37]], order=2^8)

>>> b = GF256.Random(3); b
GF([154, 193, 235], order=2^8)

>>> x = np.linalg.solve(A, b); x
GF([114, 170, 178], order=2^8)

>>> np.array_equal(A @ x, b)
True

Galois field arrays also contain matrix decomposition routines not included in numpy. These include:

  • FieldArray.row_reduce, FieldArray.lu_decompose, FieldArray.lup_decompose

Numpy ufunc methods

Galois field arrays support numpy ufunc methods. This allows the user to apply a ufunc in a unique way across the target array. The ufunc method signature is <ufunc>.<method>(*args, **kwargs). All arithmetic ufuncs are supported. Below is a list of their ufunc methods.

  • <method>: reduce, accumulate, reduceat, outer, at
>>> X = GF256.Random((2,5)); X
GF([[210,  67, 167, 137,  95],
    [104,  74, 178,  13, 142]], order=2^8)

>>> np.multiply.reduce(X, axis=0)
GF([ 63, 169, 209, 171, 161], order=2^8)
>>> x = GF256.Random(5); x
GF([210,  49,  66, 251, 148], order=2^8)

>>> y = GF256.Random(5); y
GF([  3, 123, 247, 144, 197], order=2^8)

>>> np.multiply.outer(x, y)
GF([[107, 245,  37, 192,  98],
    [ 83,  67, 183, 146, 140],
    [198,  93, 248, 206, 128],
    [ 16, 170, 178,  83,  68],
    [161,  89,  38, 116, 209]], order=2^8)

Polynomials over Galois fields

The galois package supports polynomials over Galois fields with the galois.Poly class. galois.Poly does not subclass np.ndarray but instead contains a FieldArray of coefficients as an attribute (implements the "has-a", not "is-a", architecture).

Polynomials can be created by specifying the polynomial coefficients as either a FieldArray or an "array-like" object with the field keyword argument.

>>> p = galois.Poly([172, 22, 0, 0, 225], field=GF256); p
Poly(172x^4 + 22x^3 + 225, GF(2^8))

>>> coeffs = GF256([33, 17, 0, 225]); coeffs
GF([ 33,  17,   0, 225], order=2^8)

>>> p = galois.Poly(coeffs, order="asc"); p
Poly(225x^3 + 17x + 33, GF(2^8))

Polynomials over Galois fields can also display the field elements as polynomials over their prime subfields. This can be quite confusing to read, so be warned!

>>> print(p)
Poly(225x^3 + 17x + 33, GF(2^8))

>>> with GF256.display("poly"):
...     print(p)

Poly((α^7 + α^6 + α^5 + 1)x^3 + (α^4 + 1)x + (α^5 + 1), GF(2^8))

Polynomials can also be created using a number of constructor class methods. They include:

  • Poly.Zero, Poly.One, Poly.Identity, Poly.Random, Poly.Integer, Poly.String, Poly.Degrees, Poly.Roots
# Construct a polynomial by specifying its roots
>>> q = galois.Poly.Roots([155, 37], field=GF256); q
Poly(x^2 + 190x + 71, GF(2^8))

>>> q.roots()
GF([ 37, 155], order=2^8)

Polynomial arithmetic is performed using python operators.

>>> p
Poly(225x^3 + 17x + 33, GF(2^8))

>>> q
Poly(x^2 + 190x + 71, GF(2^8))

>>> p + q
Poly(225x^3 + x^2 + 175x + 102, GF(2^8))

>>> divmod(p, q)
(Poly(225x + 57, GF(2^8)), Poly(56x + 104, GF(2^8)))

>>> p ** 2
Poly(171x^6 + 28x^2 + 117, GF(2^8))

Polynomials over Galois fields can be evaluated at scalars or arrays of field elements.

>>> p = galois.Poly.Degrees([4, 3, 0], [172, 22, 225], field=GF256); p
Poly(172x^4 + 22x^3 + 225, GF(2^8))

# Evaluate the polynomial at a single value
>>> p(1)
GF(91, order=2^8)

>>> x = GF256.Random((2,5)); x
GF([[212, 211, 244, 125,  75],
    [113, 139, 247, 223, 106]], order=2^8)

# Evaluate the polynomial at an array of values
>>> p(x)
GF([[158, 129,  28, 122, 186],
    [184, 132, 179,  49, 223]], order=2^8)

Polynomials can also be evaluated in superfields. For example, evaluating a Galois field’s irreducible polynomial at one of its elements.

# Notice the irreducible polynomial is over GF(2), not GF(2^8)
>>> p = GF256.irreducible_poly; p
Poly(x^8 + x^4 + x^3 + x^2 + 1, GF(2))

>>> GF256.is_primitive_poly
True

# Notice the primitive element is in GF(2^8)
>>> alpha = GF256.primitive_element; alpha
GF(2, order=2^8)

# Since p(x) is a primitive polynomial, alpha is one of its roots
>>> p(alpha, field=GF256)
GF(0, order=2^8)

BCH codes

See full documentation.

In [1]: import numpy as np

In [2]: import galois

In [3]: bch = galois.BCH(15, 7)

# Messages can be either vectors or matrices of np.ndarray or galois.FieldArray (galois.GF2 in this case)
In [4]: M = np.random.randint(0, 2, (5,bch.k)); M
Out[4]:
array([[1, 0, 0, 0, 1, 1, 1],
       [0, 1, 1, 1, 1, 1, 1],
       [1, 0, 0, 0, 0, 1, 0],
       [1, 1, 0, 0, 1, 1, 1],
       [0, 1, 1, 1, 1, 1, 1]])

In [5]: C = bch.encode(M); C
Out[5]:
array([[1, 0, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 0],
       [0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1],
       [1, 0, 0, 0, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1],
       [1, 1, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0],
       [0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1]])

# Corrupt the first bit in each codeword
In [6]: C[:,0] ^= 1; C
Out[6]:
array([[0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 0],
       [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1],
       [0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1],
       [0, 1, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0],
       [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1]])

In [7]: bch.decode(C)
Out[7]:
array([[1, 0, 0, 0, 1, 1, 1],
       [0, 1, 1, 1, 1, 1, 1],
       [1, 0, 0, 0, 0, 1, 0],
       [1, 1, 0, 0, 1, 1, 1],
       [0, 1, 1, 1, 1, 1, 1]])

Reed-Solomon codes

See full documentation.

In [1]: import numpy as np

In [2]: import galois

In [3]: rs = galois.ReedSolomon(15, 9)

In [4]: (rs.n, rs.k, rs.t)
Out[4]: (15, 9, 3)

In [5]: GF = rs.field; GF
Out[5]: <class 'numpy.ndarray over GF(2^4)'>

# Messages can be either vectors or matrices of np.ndarray or galois.FieldArray
In [6]: M = GF.Random((5,rs.k)); M
Out[6]:
GF([[ 0, 11, 13,  7,  9,  9,  3,  2, 12],
    [ 0,  8, 15, 10, 13,  2,  6,  2,  6],
    [ 1,  9, 13,  1, 13,  2,  6,  4, 12],
    [ 5, 14, 11, 10,  9, 15,  5,  0,  0],
    [ 6,  1,  4,  9,  9,  3, 14, 11, 13]], order=2^4)

In [7]: C = rs.encode(M); C
Out[7]:
GF([[ 0, 11, 13,  7,  9,  9,  3,  2, 12,  6,  3, 13,  0,  8,  4],
    [ 0,  8, 15, 10, 13,  2,  6,  2,  6,  1, 15,  8, 14,  0, 15],
    [ 1,  9, 13,  1, 13,  2,  6,  4, 12,  8, 11,  7,  1,  5, 13],
    [ 5, 14, 11, 10,  9, 15,  5,  0,  0,  1,  8, 13, 12, 13,  3],
    [ 6,  1,  4,  9,  9,  3, 14, 11, 13, 10,  0, 12,  3,  0,  1]],
   order=2^4)

# Corrupt the first symbol in each codeword
In [8]: C[:,0] += GF(13); C
Out[8]:
GF([[13, 11, 13,  7,  9,  9,  3,  2, 12,  6,  3, 13,  0,  8,  4],
    [13,  8, 15, 10, 13,  2,  6,  2,  6,  1, 15,  8, 14,  0, 15],
    [12,  9, 13,  1, 13,  2,  6,  4, 12,  8, 11,  7,  1,  5, 13],
    [ 8, 14, 11, 10,  9, 15,  5,  0,  0,  1,  8, 13, 12, 13,  3],
    [11,  1,  4,  9,  9,  3, 14, 11, 13, 10,  0, 12,  3,  0,  1]],
   order=2^4)

In [9]: rs.decode(C)
Out[9]:
GF([[ 0, 11, 13,  7,  9,  9,  3,  2, 12],
    [ 0,  8, 15, 10, 13,  2,  6,  2,  6],
    [ 1,  9, 13,  1, 13,  2,  6,  4, 12],
    [ 5, 14, 11, 10,  9, 15,  5,  0,  0],
    [ 6,  1,  4,  9,  9,  3, 14, 11, 13]], order=2^4)

Performance

To compare the performance of galois and native numpy, we'll use a prime field GF(p). This is because it is the simplest field. Namely, addition, subtraction, and multiplication are modulo p, which can be simply computed with numpy arrays (x + y) % p. For extension fields GF(p^m), the arithmetic is computed using polynomials over GF(p) and can't be so tersely expressed in numpy.

Lookup performance

For fields with order less than or equal to 2^20, galois uses lookup tables for efficiency. Here is an example of multiplying two arrays in GF(31) using native numpy and galois with ufunc_mode="jit-lookup".

In [1]: import numpy as np

In [2]: import galois

In [3]: GF = galois.GF(31)

In [4]: GF.ufunc_mode
Out[4]: 'jit-lookup'

In [5]: a = GF.Random(10_000, dtype=int)

In [6]: b = GF.Random(10_000, dtype=int)

In [7]: %timeit a * b
79.7 µs ± 1 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

In [8]: aa, bb = a.view(np.ndarray), b.view(np.ndarray)

# Equivalent calculation of a * b using native numpy implementation
In [9]: %timeit (aa * bb) % GF.order
96.6 µs ± 2.4 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

The galois ufunc runtime has a floor, however. This is due to a requirement to view the output array and convert its dtype with astype(). For example, for small array sizes numpy is faster than galois because it doesn't need to do these conversions.

In [4]: a = GF.Random(10, dtype=int)

In [5]: b = GF.Random(10, dtype=int)

In [6]: %timeit a * b
45.1 µs ± 1.82 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

In [7]: aa, bb = a.view(np.ndarray), b.view(np.ndarray)

# Equivalent calculation of a * b using native numpy implementation
In [8]: %timeit (aa * bb) % GF.order
1.52 µs ± 34.8 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

However, for large N galois is strictly faster than numpy.

In [10]: a = GF.Random(10_000_000, dtype=int)

In [11]: b = GF.Random(10_000_000, dtype=int)

In [12]: %timeit a * b
59.8 ms ± 1.64 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [13]: aa, bb = a.view(np.ndarray), b.view(np.ndarray)

# Equivalent calculation of a * b using native numpy implementation
In [14]: %timeit (aa * bb) % GF.order
129 ms ± 8.01 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

Calculation performance

For fields with order greater than 2^20, galois will use explicit arithmetic calculation rather than lookup tables. Even in these cases, galois is faster than numpy!

Here is an example multiplying two arrays in GF(2097169) using numpy and galois with ufunc_mode="jit-calculate".

In [1]: import numpy as np

In [2]: import galois

In [3]: GF = galois.GF(2097169)

In [4]: GF.ufunc_mode
Out[4]: 'jit-calculate'

In [5]: a = GF.Random(10_000, dtype=int)

In [6]: b = GF.Random(10_000, dtype=int)

In [7]: %timeit a * b
68.2 µs ± 2.09 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

In [8]: aa, bb = a.view(np.ndarray), b.view(np.ndarray)

# Equivalent calculation of a * b using native numpy implementation
In [9]: %timeit (aa * bb) % GF.order
93.4 µs ± 2.12 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

And again, the runtime comparison with numpy improves with large N because the time of viewing and type converting the output is small compared to the computation time. galois achieves better performance than numpy because the multiplication and modulo operations are compiled together into one ufunc rather than two.

In [10]: a = GF.Random(10_000_000, dtype=int)

In [11]: b = GF.Random(10_000_000, dtype=int)

In [12]: %timeit a * b
51.2 ms ± 1.08 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [13]: aa, bb = a.view(np.ndarray), b.view(np.ndarray)

# Equivalent calculation of a * b using native numpy implementation
In [14]: %timeit (aa * bb) % GF.order
111 ms ± 1.48 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

Linear algebra performance

Linear algebra over Galois fields is highly optimized. For prime fields GF(p), the performance is comparable to the native numpy implementation (using BLAS/LAPACK).

In [1]: import numpy as np

In [2]: import galois

In [3]: GF = galois.GF(31)

In [4]: A = GF.Random((100,100), dtype=int)

In [5]: B = GF.Random((100,100), dtype=int)

In [6]: %timeit A @ B
720 µs ± 5.36 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [7]: AA, BB = A.view(np.ndarray), B.view(np.ndarray)

# Equivalent calculation of A @ B using the native numpy implementation
In [8]: %timeit (AA @ BB) % GF.order
777 µs ± 4.6 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

For extension fields GF(p^m), the performance of galois is close to native numpy linear algebra (about 10x slower). However, for extension fields, each multiplication operation is equivalently a convolution (polynomial multiplication) of two m-length arrays and polynomial remainder division with the irreducible polynomial. So it's not an apples-to-apples comparison.

Below is a comparison of galois computing the correct matrix multiplication over GF(2^8) and numpy computing a normal integer matrix multiplication (which is not the correct result!). This comparison is just for a performance reference.

In [1]: import numpy as np

In [2]: import galois

In [3]: GF = galois.GF(2**8)

In [4]: A = GF.Random((100,100), dtype=int)

In [5]: B = GF.Random((100,100), dtype=int)

In [6]: %timeit A @ B
7.13 ms ± 114 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [7]: AA, BB = A.view(np.ndarray), B.view(np.ndarray)

# Native numpy matrix multiplication, which doesn't produce the correct result!!
In [8]: %timeit AA @ BB
651 µs ± 12.4 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Acknowledgements

This package heavily relies on Numba and its just-in-time compiler for performance. We use Frank Luebeck's compilation of Conway polynomials for computing primitive polynomials for extension fields. We utilize SageMath for generating test vectors.

About

A performant numpy extension for Galois fields and their applications

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 100.0%