Namespaces
Variants
Actions

Difference between revisions of "LLL basis reduction method"

From Encyclopedia of Mathematics
Jump to: navigation, search
m (MR/ZBL numbers added)
m (tex encoded by computer)
 
Line 1: Line 1:
 +
<!--
 +
l1101601.png
 +
$#A+1 = 65 n = 0
 +
$#C+1 = 65 : ~/encyclopedia/old_files/data/L110/L.1100160 LLL basis reduction method
 +
Automatically converted into TeX, above some diagnostics.
 +
Please remove this comment and the {{TEX|auto}} line below,
 +
if TeX found to be correct.
 +
-->
 +
 +
{{TEX|auto}}
 +
{{TEX|done}}
 +
 
Basis reduction is the process of transforming an arbitrary basis of a [[Lattice|lattice]] into a reduced basis, where a reduced basis consists of short vectors, or, equivalently, a basis that has a small orthogonality defect. Various notions of reduced basis have been proposed by Ch. Hermite [[#References|[a3]]], and by A. Korkin, G. Zolotarev and H. Minkowski in the language of quadratic forms (cf. also [[Geometry of numbers|Geometry of numbers]]; [[Quadratic form|Quadratic form]]). A.K. Lenstra, H.W. Lenstra and L. Lovász [[#References|[a5]]] introduced a notion of reduced lattice basis together with an efficient algorithm that transforms a given lattice basis into a reduced basis of the same lattice. This algorithm has found numerous applications in number theory and integer optimization, and various cryptographic schemes have been broken using this algorithm.
 
Basis reduction is the process of transforming an arbitrary basis of a [[Lattice|lattice]] into a reduced basis, where a reduced basis consists of short vectors, or, equivalently, a basis that has a small orthogonality defect. Various notions of reduced basis have been proposed by Ch. Hermite [[#References|[a3]]], and by A. Korkin, G. Zolotarev and H. Minkowski in the language of quadratic forms (cf. also [[Geometry of numbers|Geometry of numbers]]; [[Quadratic form|Quadratic form]]). A.K. Lenstra, H.W. Lenstra and L. Lovász [[#References|[a5]]] introduced a notion of reduced lattice basis together with an efficient algorithm that transforms a given lattice basis into a reduced basis of the same lattice. This algorithm has found numerous applications in number theory and integer optimization, and various cryptographic schemes have been broken using this algorithm.
  
Let <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101601.png" /> be the <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101602.png" />-dimensional real vector space with inner product <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101603.png" /> and the Euclidean norm <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101604.png" />. A lattice <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101605.png" /> is a discrete additive subgroup of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101606.png" /> (cf. [[Group|Group]]). Its rank or dimension is the dimension of the smallest linear subspace that contains <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101607.png" />. A lattice <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101608.png" /> of rank <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l1101609.png" /> has a basis, that is, a set of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016010.png" /> linearly independent vectors <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016011.png" />, so that <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016012.png" /> consists of all linear integer combinations of the basis vectors, <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016013.png" />.
+
Let $  \mathbf R  ^ {n} $
 +
be the $  n $-
 +
dimensional real vector space with inner product $  \langle  {\cdot, \cdot } \rangle $
 +
and the Euclidean norm $  \| y \| = \langle  {y,y } \rangle ^ { {1 / 2 } } $.  
 +
A lattice $  L \subset  \mathbf R  ^ {n} $
 +
is a discrete additive subgroup of $  \mathbf R  ^ {n} $(
 +
cf. [[Group|Group]]). Its rank or dimension is the dimension of the smallest linear subspace that contains $  L $.  
 +
A lattice $  L $
 +
of rank $  m $
 +
has a basis, that is, a set of $  m $
 +
linearly independent vectors $  b _ {1} \dots b _ {m} $,  
 +
so that $  L $
 +
consists of all linear integer combinations of the basis vectors, $  L = \{ {t _ {1} b _ {1} + \dots + t _ {m} b _ {m} } : {t _ {1} \dots t _ {m} \in \mathbf Z } \} $.
  
Let <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016014.png" /> be a basis of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016015.png" /> with basis matrix <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016016.png" /> consisting of the column vectors <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016017.png" />. Then all basis matrices of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016018.png" /> are of the form <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016019.png" />, where <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016020.png" /> is an <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016021.png" /> integer matrix with determinant <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016022.png" />.
+
Let $  b _ {1} \dots b _ {m} \in \mathbf R  ^ {n} $
 +
be a basis of $  L $
 +
with basis matrix $  B = [ b _ {1} \dots b _ {m} ] $
 +
consisting of the column vectors $  b _ {1} \dots b _ {m} $.  
 +
Then all basis matrices of $  L $
 +
are of the form $  {\overline{B}\; } = BM $,  
 +
where $  M $
 +
is an $  ( m \times m ) $
 +
integer matrix with determinant $  \pm  1 $.
  
With an ordered basis <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016023.png" /> one can associate the Gram–Schmidt orthogonalization <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016024.png" />, which can be computed, together with the Gram–Schmidt coefficients <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016025.png" />, by the recursion <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016026.png" />, <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016027.png" /> for <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016028.png" /> (cf. also [[Orthogonalization method|Orthogonalization method]]). A lattice basis <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016029.png" /> is LLL-reduced if
+
With an ordered basis $  b _ {1} \dots b _ {m} \in \mathbf R  ^ {n} $
 +
one can associate the Gram–Schmidt orthogonalization $  {\widehat{b}  } _ {1} \dots {\widehat{b}  } _ {m} \in \mathbf R  ^ {n} $,  
 +
which can be computed, together with the Gram–Schmidt coefficients $  \mu _ {i,j }  = { {\langle  {b _ {i} , {\widehat{b}  } _ {j} } \rangle } / {\langle  { {\widehat{b}  } _ {j} , {\widehat{b}  } _ {j} } \rangle } } $,  
 +
by the recursion $  {\widehat{b}  } _ {1} = b _ {1} $,
 +
$  {\widehat{b}  } _ {i} = b _ {i} - \sum _ {j = 1 }  ^ {i - 1 } \mu _ {i,j }  {\widehat{b}  } _ {j} $
 +
for $  i = 2 \dots m $(
 +
cf. also [[Orthogonalization method|Orthogonalization method]]). A lattice basis $  b _ {1} \dots b _ {m} $
 +
is LLL-reduced if
  
1) <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016030.png" /> for <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016031.png" />;
+
1) $  | {\mu _ {i,j }  } | \leq  {1 / 2 } $
 +
for $  1 \leq  j < i \leq  m $;
  
2) for <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016032.png" />,
+
2) for $  1 < i \leq  m $,
  
<table class="eq" style="width:100%;"> <tr><td valign="top" style="width:94%;text-align:center;"><img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016033.png" /></td> </tr></table>
+
$$
 +
{
 +
\frac{3}{4}
 +
} \left \| { {\widehat{b}  } _ {i - 1 }  } \right \|  ^ {2} \leq  \left \| { {\widehat{b}  } _ {i} + \mu _ {i,i - 1 }  {\widehat{b}  } _ {i + 1 }  } \right \|  ^ {2} .
 +
$$
  
For every LLL-reduced basis the length <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016034.png" /> approximates the <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016035.png" />th successive minimum <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016036.png" /> of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016037.png" />, where <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016038.png" /> is the smallest radius of a ball centred at the origin that contains <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016039.png" /> linearly independent lattice vectors. Always:
+
For every LLL-reduced basis the length $  \| {b _ {i} } \| $
 +
approximates the $  i $
 +
th successive minimum $  \lambda _ {i} $
 +
of $  L $,  
 +
where $  \lambda _ {i} ( L ) $
 +
is the smallest radius of a ball centred at the origin that contains $  i $
 +
linearly independent lattice vectors. Always:
  
<table class="eq" style="width:100%;"> <tr><td valign="top" style="width:94%;text-align:center;"><img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016040.png" /></td> </tr></table>
+
$$
 +
2 ^ {1 - i } \leq  \left \| { {\widehat{b}  } _ {i} } \right \|  ^ {2} \lambda _ {i} ( L ) ^ {- 2 } \leq  \left \| {b _ {i} } \right \|  ^ {2} \lambda _ {i} ( L ) ^ {- 2 } \leq  2 ^ {m - 1 } .
 +
$$
  
 
The novelty of LLL-reduction is a polynomial-time algorithm that transforms an arbitrary integer lattice basis into an LLL-reduced basis [[#References|[a5]]]. This algorithm has numerous applications.
 
The novelty of LLL-reduction is a polynomial-time algorithm that transforms an arbitrary integer lattice basis into an LLL-reduced basis [[#References|[a5]]]. This algorithm has numerous applications.
  
For example, polynomials with integer coefficients <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016041.png" /> can be factored in polynomial time into irreducible factors [[#References|[a5]]]. There is a polynomial-time algorithm which, given a rational point <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016042.png" /> and an integer lattice basis <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016043.png" />, finds a lattice point <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016044.png" /> that is nearest to <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016045.png" /> up to a factor <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016046.png" /> [[#References|[a1]]]. Linear inequalities with integer coefficients in <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016047.png" /> variables <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016048.png" /> can be solved over the integers in time <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016049.png" />, where <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016050.png" /> is the maximal bit length of the entries in the inequalities [[#References|[a4]]], [[#References|[a6]]]. In particular, <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016051.png" />-<img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016052.png" /> solutions of a single equation with integer coefficients can be found quite efficiently [[#References|[a2]]]. Univariate polynomial equations <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016053.png" /> with an arbitrary integer modulus <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016054.png" /> can be solved in polynomial time for small solutions, i.e., solutions <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016055.png" /> with <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016056.png" />.
+
For example, polynomials with integer coefficients $  c _ {0} + c _ {1} x + \dots + c _ {n} x  ^ {n} $
 +
can be factored in polynomial time into irreducible factors [[#References|[a5]]]. There is a polynomial-time algorithm which, given a rational point $  x \in \mathbf R  ^ {n} $
 +
and an integer lattice basis $  b _ {1} \dots b _ {m} $,  
 +
finds a lattice point $  b $
 +
that is nearest to $  x $
 +
up to a factor $  2 ^ { {n / 2 } } $[[#References|[a1]]]. Linear inequalities with integer coefficients in $  n $
 +
variables $  x _ {1} \dots x _ {n} $
 +
can be solved over the integers in time $  n ^ {O ( n ) } M ^ {O ( 1 ) } $,  
 +
where $  M $
 +
is the maximal bit length of the entries in the inequalities [[#References|[a4]]], [[#References|[a6]]]. In particular, 0 $-
 +
$  1 $
 +
solutions of a single equation with integer coefficients can be found quite efficiently [[#References|[a2]]]. Univariate polynomial equations $  p ( x ) = 0 ( { \mathop{\rm mod} } N ) $
 +
with an arbitrary integer modulus $  N $
 +
can be solved in polynomial time for small solutions, i.e., solutions $  x $
 +
with $  | x | \leq  O ( N ^ { {1 / 4 } } ) $.
  
Some applications of LLL-reduction are possible because LLL-reduction can be refined to block reduction so that every reduced basis contains a vector that is not longer than <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016057.png" /> times the shortest lattice vector, where <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016058.png" /> is fixed [[#References|[a8]]], [[#References|[a9]]]. Block reduction strongly reduces blocks <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016059.png" /> of the basis with fixed size <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016060.png" />. LLL-reduction is the case <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016061.png" />. The running time of block-reduction increases with <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016062.png" /> as <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016063.png" /> approaches <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016064.png" />. Shortest lattice vectors of lattices of dimension up to <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/l/l110/l110160/l11016065.png" /> can be found within a few hours running time. LLL-reduction can be extended to reduce a set of linearly dependent real vectors. One can derive from LLL-reduction a continued fraction algorithm for arbitrary dimension [[#References|[a4]]].
+
Some applications of LLL-reduction are possible because LLL-reduction can be refined to block reduction so that every reduced basis contains a vector that is not longer than $  ( 1 + \epsilon )  ^ {n} $
 +
times the shortest lattice vector, where $  \epsilon > 0 $
 +
is fixed [[#References|[a8]]], [[#References|[a9]]]. Block reduction strongly reduces blocks $  b _ {i} \dots b _ {i + \beta - 1 }  $
 +
of the basis with fixed size $  \beta $.  
 +
LLL-reduction is the case $  \beta = 2 $.  
 +
The running time of block-reduction increases with $  \beta $
 +
as $  \epsilon $
 +
approaches 0 $.  
 +
Shortest lattice vectors of lattices of dimension up to $  80 $
 +
can be found within a few hours running time. LLL-reduction can be extended to reduce a set of linearly dependent real vectors. One can derive from LLL-reduction a continued fraction algorithm for arbitrary dimension [[#References|[a4]]].
  
 
====References====
 
====References====
 
<table><TR><TD valign="top">[a1]</TD> <TD valign="top"> L. Babai, "On Lovász lattice reduction and the nearest lattice point problem" ''Combinatorica'' , '''6''' (1986) pp. 1–13 {{MR|0856638}} {{ZBL|0593.68030}} </TD></TR><TR><TD valign="top">[a2]</TD> <TD valign="top"> M.J. Coster, A. Joux, B.A. LaMacchia, A.M. Odlyzko, C.P. Schnorr, J. Stern, "An improved low-density subset sum algorithm" ''Comp. Complexity'' , '''2''' (1992) pp. 111–128 {{MR|1227795}} {{ZBL|0774.11075}} </TD></TR><TR><TD valign="top">[a3]</TD> <TD valign="top"> C. Hermite, "Extraits de lettres de M. Ch. Hermite á M. Jacobi sur différents objets de la théorie des nombres (2eme lettre du 6-8-1845)" ''J. Reine Angew. Math.'' , '''40''' (1850) pp. 279–290</TD></TR><TR><TD valign="top">[a4]</TD> <TD valign="top"> B. Just, "Generalizing the continued fraction algorithm to arbitrary dimensions" ''Siam J. Comput.'' , '''21''' (1992) pp. 909–926 {{MR|1181407}} {{ZBL|0763.11027}} </TD></TR><TR><TD valign="top">[a5]</TD> <TD valign="top"> R. Kannan, "Minkowski's convex body theorem and integer programming" ''Math. Oper. Res.'' , '''12''' (1987) pp. 415–440</TD></TR><TR><TD valign="top">[a6]</TD> <TD valign="top"> A.K. Lenstra, H.W. Lenstra, Jr., L. Lovász, "Factoring polynomials with rational coefficients" ''Math. Ann.'' , '''261''' (1982) pp. 515–534 {{MR|0682664}} {{ZBL|0488.12001}} {{ZBL|0477.68043}} </TD></TR><TR><TD valign="top">[a7]</TD> <TD valign="top"> H.W. Lenstra, Jr., "Integer programming with a fixed number of variables" ''Math. Oper. Res.'' , '''8''' (1983) pp. 538–548 {{MR|0727410}} {{ZBL|0524.90067}} </TD></TR><TR><TD valign="top">[a8]</TD> <TD valign="top"> L. Lovász, "An algorithmic theory of numbers, graphs and convexity" , SIAM (1986) {{MR|0861822}} {{ZBL|0606.68039}} </TD></TR><TR><TD valign="top">[a9]</TD> <TD valign="top"> C.P. Schnorr, "A hierarchy of polynomial time lattice basis reduction algorithms" ''Theor. Comput. Sci.'' , '''53''' (1987) pp. 201–224 {{MR|0918090}} {{ZBL|0642.10030}} </TD></TR><TR><TD valign="top">[a10]</TD> <TD valign="top"> C.P. Schnorr, "Block reduced lattice bases and successive minima" ''Combinatorics, Probab. Computing'' , '''3''' (1994) pp. 507–522 {{MR|1314072}} {{ZBL|0845.11025}} </TD></TR></table>
 
<table><TR><TD valign="top">[a1]</TD> <TD valign="top"> L. Babai, "On Lovász lattice reduction and the nearest lattice point problem" ''Combinatorica'' , '''6''' (1986) pp. 1–13 {{MR|0856638}} {{ZBL|0593.68030}} </TD></TR><TR><TD valign="top">[a2]</TD> <TD valign="top"> M.J. Coster, A. Joux, B.A. LaMacchia, A.M. Odlyzko, C.P. Schnorr, J. Stern, "An improved low-density subset sum algorithm" ''Comp. Complexity'' , '''2''' (1992) pp. 111–128 {{MR|1227795}} {{ZBL|0774.11075}} </TD></TR><TR><TD valign="top">[a3]</TD> <TD valign="top"> C. Hermite, "Extraits de lettres de M. Ch. Hermite á M. Jacobi sur différents objets de la théorie des nombres (2eme lettre du 6-8-1845)" ''J. Reine Angew. Math.'' , '''40''' (1850) pp. 279–290</TD></TR><TR><TD valign="top">[a4]</TD> <TD valign="top"> B. Just, "Generalizing the continued fraction algorithm to arbitrary dimensions" ''Siam J. Comput.'' , '''21''' (1992) pp. 909–926 {{MR|1181407}} {{ZBL|0763.11027}} </TD></TR><TR><TD valign="top">[a5]</TD> <TD valign="top"> R. Kannan, "Minkowski's convex body theorem and integer programming" ''Math. Oper. Res.'' , '''12''' (1987) pp. 415–440</TD></TR><TR><TD valign="top">[a6]</TD> <TD valign="top"> A.K. Lenstra, H.W. Lenstra, Jr., L. Lovász, "Factoring polynomials with rational coefficients" ''Math. Ann.'' , '''261''' (1982) pp. 515–534 {{MR|0682664}} {{ZBL|0488.12001}} {{ZBL|0477.68043}} </TD></TR><TR><TD valign="top">[a7]</TD> <TD valign="top"> H.W. Lenstra, Jr., "Integer programming with a fixed number of variables" ''Math. Oper. Res.'' , '''8''' (1983) pp. 538–548 {{MR|0727410}} {{ZBL|0524.90067}} </TD></TR><TR><TD valign="top">[a8]</TD> <TD valign="top"> L. Lovász, "An algorithmic theory of numbers, graphs and convexity" , SIAM (1986) {{MR|0861822}} {{ZBL|0606.68039}} </TD></TR><TR><TD valign="top">[a9]</TD> <TD valign="top"> C.P. Schnorr, "A hierarchy of polynomial time lattice basis reduction algorithms" ''Theor. Comput. Sci.'' , '''53''' (1987) pp. 201–224 {{MR|0918090}} {{ZBL|0642.10030}} </TD></TR><TR><TD valign="top">[a10]</TD> <TD valign="top"> C.P. Schnorr, "Block reduced lattice bases and successive minima" ''Combinatorics, Probab. Computing'' , '''3''' (1994) pp. 507–522 {{MR|1314072}} {{ZBL|0845.11025}} </TD></TR></table>

Latest revision as of 22:15, 5 June 2020


Basis reduction is the process of transforming an arbitrary basis of a lattice into a reduced basis, where a reduced basis consists of short vectors, or, equivalently, a basis that has a small orthogonality defect. Various notions of reduced basis have been proposed by Ch. Hermite [a3], and by A. Korkin, G. Zolotarev and H. Minkowski in the language of quadratic forms (cf. also Geometry of numbers; Quadratic form). A.K. Lenstra, H.W. Lenstra and L. Lovász [a5] introduced a notion of reduced lattice basis together with an efficient algorithm that transforms a given lattice basis into a reduced basis of the same lattice. This algorithm has found numerous applications in number theory and integer optimization, and various cryptographic schemes have been broken using this algorithm.

Let $ \mathbf R ^ {n} $ be the $ n $- dimensional real vector space with inner product $ \langle {\cdot, \cdot } \rangle $ and the Euclidean norm $ \| y \| = \langle {y,y } \rangle ^ { {1 / 2 } } $. A lattice $ L \subset \mathbf R ^ {n} $ is a discrete additive subgroup of $ \mathbf R ^ {n} $( cf. Group). Its rank or dimension is the dimension of the smallest linear subspace that contains $ L $. A lattice $ L $ of rank $ m $ has a basis, that is, a set of $ m $ linearly independent vectors $ b _ {1} \dots b _ {m} $, so that $ L $ consists of all linear integer combinations of the basis vectors, $ L = \{ {t _ {1} b _ {1} + \dots + t _ {m} b _ {m} } : {t _ {1} \dots t _ {m} \in \mathbf Z } \} $.

Let $ b _ {1} \dots b _ {m} \in \mathbf R ^ {n} $ be a basis of $ L $ with basis matrix $ B = [ b _ {1} \dots b _ {m} ] $ consisting of the column vectors $ b _ {1} \dots b _ {m} $. Then all basis matrices of $ L $ are of the form $ {\overline{B}\; } = BM $, where $ M $ is an $ ( m \times m ) $ integer matrix with determinant $ \pm 1 $.

With an ordered basis $ b _ {1} \dots b _ {m} \in \mathbf R ^ {n} $ one can associate the Gram–Schmidt orthogonalization $ {\widehat{b} } _ {1} \dots {\widehat{b} } _ {m} \in \mathbf R ^ {n} $, which can be computed, together with the Gram–Schmidt coefficients $ \mu _ {i,j } = { {\langle {b _ {i} , {\widehat{b} } _ {j} } \rangle } / {\langle { {\widehat{b} } _ {j} , {\widehat{b} } _ {j} } \rangle } } $, by the recursion $ {\widehat{b} } _ {1} = b _ {1} $, $ {\widehat{b} } _ {i} = b _ {i} - \sum _ {j = 1 } ^ {i - 1 } \mu _ {i,j } {\widehat{b} } _ {j} $ for $ i = 2 \dots m $( cf. also Orthogonalization method). A lattice basis $ b _ {1} \dots b _ {m} $ is LLL-reduced if

1) $ | {\mu _ {i,j } } | \leq {1 / 2 } $ for $ 1 \leq j < i \leq m $;

2) for $ 1 < i \leq m $,

$$ { \frac{3}{4} } \left \| { {\widehat{b} } _ {i - 1 } } \right \| ^ {2} \leq \left \| { {\widehat{b} } _ {i} + \mu _ {i,i - 1 } {\widehat{b} } _ {i + 1 } } \right \| ^ {2} . $$

For every LLL-reduced basis the length $ \| {b _ {i} } \| $ approximates the $ i $ th successive minimum $ \lambda _ {i} $ of $ L $, where $ \lambda _ {i} ( L ) $ is the smallest radius of a ball centred at the origin that contains $ i $ linearly independent lattice vectors. Always:

$$ 2 ^ {1 - i } \leq \left \| { {\widehat{b} } _ {i} } \right \| ^ {2} \lambda _ {i} ( L ) ^ {- 2 } \leq \left \| {b _ {i} } \right \| ^ {2} \lambda _ {i} ( L ) ^ {- 2 } \leq 2 ^ {m - 1 } . $$

The novelty of LLL-reduction is a polynomial-time algorithm that transforms an arbitrary integer lattice basis into an LLL-reduced basis [a5]. This algorithm has numerous applications.

For example, polynomials with integer coefficients $ c _ {0} + c _ {1} x + \dots + c _ {n} x ^ {n} $ can be factored in polynomial time into irreducible factors [a5]. There is a polynomial-time algorithm which, given a rational point $ x \in \mathbf R ^ {n} $ and an integer lattice basis $ b _ {1} \dots b _ {m} $, finds a lattice point $ b $ that is nearest to $ x $ up to a factor $ 2 ^ { {n / 2 } } $[a1]. Linear inequalities with integer coefficients in $ n $ variables $ x _ {1} \dots x _ {n} $ can be solved over the integers in time $ n ^ {O ( n ) } M ^ {O ( 1 ) } $, where $ M $ is the maximal bit length of the entries in the inequalities [a4], [a6]. In particular, $ 0 $- $ 1 $ solutions of a single equation with integer coefficients can be found quite efficiently [a2]. Univariate polynomial equations $ p ( x ) = 0 ( { \mathop{\rm mod} } N ) $ with an arbitrary integer modulus $ N $ can be solved in polynomial time for small solutions, i.e., solutions $ x $ with $ | x | \leq O ( N ^ { {1 / 4 } } ) $.

Some applications of LLL-reduction are possible because LLL-reduction can be refined to block reduction so that every reduced basis contains a vector that is not longer than $ ( 1 + \epsilon ) ^ {n} $ times the shortest lattice vector, where $ \epsilon > 0 $ is fixed [a8], [a9]. Block reduction strongly reduces blocks $ b _ {i} \dots b _ {i + \beta - 1 } $ of the basis with fixed size $ \beta $. LLL-reduction is the case $ \beta = 2 $. The running time of block-reduction increases with $ \beta $ as $ \epsilon $ approaches $ 0 $. Shortest lattice vectors of lattices of dimension up to $ 80 $ can be found within a few hours running time. LLL-reduction can be extended to reduce a set of linearly dependent real vectors. One can derive from LLL-reduction a continued fraction algorithm for arbitrary dimension [a4].

References

[a1] L. Babai, "On Lovász lattice reduction and the nearest lattice point problem" Combinatorica , 6 (1986) pp. 1–13 MR0856638 Zbl 0593.68030
[a2] M.J. Coster, A. Joux, B.A. LaMacchia, A.M. Odlyzko, C.P. Schnorr, J. Stern, "An improved low-density subset sum algorithm" Comp. Complexity , 2 (1992) pp. 111–128 MR1227795 Zbl 0774.11075
[a3] C. Hermite, "Extraits de lettres de M. Ch. Hermite á M. Jacobi sur différents objets de la théorie des nombres (2eme lettre du 6-8-1845)" J. Reine Angew. Math. , 40 (1850) pp. 279–290
[a4] B. Just, "Generalizing the continued fraction algorithm to arbitrary dimensions" Siam J. Comput. , 21 (1992) pp. 909–926 MR1181407 Zbl 0763.11027
[a5] R. Kannan, "Minkowski's convex body theorem and integer programming" Math. Oper. Res. , 12 (1987) pp. 415–440
[a6] A.K. Lenstra, H.W. Lenstra, Jr., L. Lovász, "Factoring polynomials with rational coefficients" Math. Ann. , 261 (1982) pp. 515–534 MR0682664 Zbl 0488.12001 Zbl 0477.68043
[a7] H.W. Lenstra, Jr., "Integer programming with a fixed number of variables" Math. Oper. Res. , 8 (1983) pp. 538–548 MR0727410 Zbl 0524.90067
[a8] L. Lovász, "An algorithmic theory of numbers, graphs and convexity" , SIAM (1986) MR0861822 Zbl 0606.68039
[a9] C.P. Schnorr, "A hierarchy of polynomial time lattice basis reduction algorithms" Theor. Comput. Sci. , 53 (1987) pp. 201–224 MR0918090 Zbl 0642.10030
[a10] C.P. Schnorr, "Block reduced lattice bases and successive minima" Combinatorics, Probab. Computing , 3 (1994) pp. 507–522 MR1314072 Zbl 0845.11025
How to Cite This Entry:
LLL basis reduction method. Encyclopedia of Mathematics. URL: http://encyclopediaofmath.org/index.php?title=LLL_basis_reduction_method&oldid=47549
This article was adapted from an original article by C.P. Schnorr (originator), which appeared in Encyclopedia of Mathematics - ISBN 1402006098. See original article