Lattices for Communication Engineers

Lattices for Communication Engineers Jean-Claude Belfiore Télécom ParisTech CNRS, LTCI UMR 5141 February, 22 2011 Nanyang Technological University - SPMS

Part I Introduction

Signal Space The transmission problem Link between signal space and transmitted analog signal through an orthogonal basis of signals 3 / 41

Signal Space The transmission problem Link between signal space and transmitted analog signal through an orthogonal basis of signals Standard serial transmission Transmitted signal is x(t) = x k h(t kt) k where x k are the transmitted complex symbols and {h(t kt)} k is a family of orthogonal signals (h is a Nyquist root). OFDM transmission Transmitted signal is N/2 x(t) = k q= N/2 x k,q h(t kt)e i 2πk N+1 ft where x k,q are the transmitted complex symbols and { } h(t kt)e i N+1 2πk ft is a doubly indexed family of k,q orthogonal signals (for instance, h(t) = rect T (t) with f = 1 T ). 3 / 41

Signal Space Complex symbols and Signal Space We define vector x = (x 1,x 2,...,x m) as a vector living in a m dimensional complex space or a n dimensional real space (n = 2m). Complex symbols used in practice are QAM symbols, components of vector x. We need to introduce coding structure the QAM symbols. xk 64 QAM Figure: Symbol from a 64 QAM 4 / 41

Lattices Definition and properties Definition A Euclidean Z lattice is a discrete additive subgroup with rank p, p n of the Euclidean space R n. We restrict to the case p = n in the sequel. 5 / 41

Lattices Definition and properties Definition A Euclidean Z lattice is a discrete additive subgroup with rank p, p n of the Euclidean space R n. We restrict to the case p = n in the sequel. A lattice Λ is a Z module generated by vectors v 1,v 2,...,v n of R n. An element v of Λ can be written as : v = a 1 v 1 + a 2 v 2 +... + a n v n, a 1,a 2,...,a n Z 5 / 41

Lattices Lattices : Generator matrix The set of vectors v 1,v 2,...,v n is a lattice basis. Definition Matrix M whose columns are vectors v 1,v 2,...,v n is a generator matrix of the lattice denoted Λ M. Each vector x = (x 1,x 2,...,x n) in Λ M, can be written as, x = M z where z = (z 1,z 2,...,z n) Z n. Lattice Λ M may be seen as the result of a linear transform applied to lattice Z n (cubic lattice). 6 / 41

Lattices Lattices : Elementary Properties (I) Let Q M n (R), such that Q Q = I n be an isometry. The two lattices Λ M and Λ Q M are said equivalent. Lattice Λ Q M is a rotated version of Λ M if detq = 1. If T M n (Z) and dett ±1, then lattice Λ M T is a sublattice of Λ M. We will often consider sublattices of Z n. 7 / 41

Lattices Lattices : Elementary Properties (II) The generator matrix M describes the lattice Λ M, but it is not unique. All matrices M T with T M n (Z) and dett = ±1 are generator matrices of Λ M. T is called a unimodular matrix. G = M M is the Gram matrix of the lattice. M is also a generator matrix of the dual of Λ M. We define then gemetric parameters. Definitions The fundamental parallelotope of Λ M is the region, P = { x R n x = a 1 v 1 + a 2 v 2 +... + a n v n, 0 a i < 1, i = 1...n } The fundamental volume is the volume of the fundamental parallelotope. It is denoted vol ( Λ M ). The fundamental volume of the lattice is vol ( Λ M ) = det(m), which is det(g) either. 8 / 41

Lattices Lattices : Elementary Properties (III) Definition The Voronoï cell of a point u belonging to the lattice Λ is the region V Λ (u) = { x R n x u x y, y Λ } All Voronoï cells of a lattice are translated versions of the Voronoï cell of the zero point. This cell is called Voronoï cell of the lattice. The fundamental volume of a lattice is equal to the volume of its Voronoï cell. 9 / 41

Lattices The Z 2 -lattice v2 v1 Z 2 lattice (v1, v2) Lattice Point Lattice Basis Fundamental Parallelotope Voronoi region A QAM constellation is a finite part of Z 2. 10 / 41

Lattices The A 2 lattice v 2 v 1 The A 2 lattice (v 1, v 2) Lattice point Lattice basis Fundamental parallelotope Voronoi region An HEX constellation is a finite part of A 2, the hexagonal lattice. 11 / 41

Lattices Construction A (binary) Construction A for a Z lattice Let q be an integer. Then, Z/qZ is a finite field if q is a prime and a finite ring otherwise. For a linear code C of length n defined on Z/qZ, lattice Λ is given by Λ = qz n + C ( qz n ) + x. x C Construction of D 4 D 4 is obtained as D 4 = 2Z 4 + (4,3) F2 where (4,3) F2 is a binary parity-check code. 12 / 41

Lattices Construction A (quaternary) Construction A of the Leech lattice The Leech lattice can be obtained as 2Λ 24 = 2Z 24 + (24,12) Z4 where (24,12) Z4 is the quaternary self-dual code obtained by extending the quaternary cyclic Golay code over Z 4. Other constructions Construction A can be generalized. Constructions B or D for instance. But one can show that all these constructions are equivalent to construction A with a suitable alphabet (for the code). 13 / 41

Lattices Construction D: Barnes-Wall A family of lattices of dimension 2 m+1,m 2 can be constructed by construction D. Barnes-Wall Lattices Constructed as Z[i] lattices, BW m = (1 + i) m m 1 Z[i] 2m + (1 + i) r RM(m,r) r=0 where RM(m,r) is a Reed-Müller code (binary) of length n = 2 m, dimension k = ( r m ) l=0 l and minimum Hamming distance d = 2 m r. BW m is a Z lattice of dimension 2 m+1. Another construction of E 8 We have 2E 8 = (1 + i) 2 Z[i] 4 + (1 + i)(4,3,2) + (4,1,4) which can also be considered as a construction A on the ring R = F 2 + u F 2,u 2 = 0 by using the linear code of generator matrix G = 1 1 1 1 0 u 0 u 0 0 u u. 14 / 41

Part II Coding for the Gaussian Channel

Separation: Coding Gain and Shaping Gain What are Lattice Codes? An example Toy example: the 4 QAM A code with 4 codewords Figure: The 4 codewords are in red. Structure is Z 2/ 2Z 2. 16 / 41

Separation: Coding Gain and Shaping Gain What are Lattice Codes? An example Toy example: the 4 QAM A code with 4 codewords Figure: The 4 codewords are in red. Structure is Z 2/ 2Z 2. Centers of the squares are shifted points of a sublattice. 16 / 41

Separation: Coding Gain and Shaping Gain What are Lattice Codes? The general case Take a lattice Λ c and a sublattice Λ s Λ c of finite index M. Each point x Λ c + c can be written as x = x s + x q + c where x s Λ s and x q is the representative of x in Λ c /Λ s, of smallest Euclidean norm. c is a constant vector which ensures that the overall finite constellation has zero mean. Lattice Codes Lattice codes are the representatives of the quotient group Λ c /Λ s, with smallest Euclidean norm, shifted so that the overall constellation has zero mean. 17 / 41

Separation: Coding Gain and Shaping Gain Coding: Minimum distance of Λ c The Coding Lattice Λ c We want to characterize the performance of Λ c. Suppose that Λ s is a scaled version of Z n (separation). On the Gaussian channel, error probability is dominated by minimum pairwise error probability ( ) ( ) x t minx,t C max P (x t) = max Q x,t C x,t C 2 x t = Q N 0 2 N 0 where Q(x) is the error function ˆ + 1 Q(x) = e u2 2 du x 2π and N 0 is the power spectrum density of the noise. 18 / 41

Separation: Coding Gain and Shaping Gain Energetic considerations Communication engineers express error probability as a function of E b N 0 where E b is the required energy to transmit one bit and N 0 is the power spectrum density of the noise. Compare lattice codes (cubic shaping) with uncoded QAM with same spectral efficiency (same number of points) αz n with a carefully chosen α. Dominant term of the error probability is ( ) minx,t C x t Q 2 d = Q m min 2 Eb N 0 E s N 0 m being the spectral efficiency and E s the energy per symbol. Compare d2 min of the lattice code with Es the one of αz n. 19 / 41

Separation: Coding Gain and Shaping Gain Coding Gain: Examples Dimension 4 The checkerboard lattice D 4 has generator matrix M D4 = with det ( M D4 ) = 2 and d 2 min = 2. Coding gain is 1 1 0 0 1 1 0 0 0 1 1 0 0 0 1 1 γ c (D 4 ) = d2 min vol(d 4 ) 1 2 = 2 2 = 2. 20 / 41

Separation: Coding Gain and Shaping Gain Coding Gain: Examples Dimension 8 The Gosset lattice E 8 has generator matrix M E8 = with det ( M E8 ) = 1 and d 2 min = 2. Coding gain is 2 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 1/2 1/2 1/2 1/2 1/2 1/2 1/2 1/2 γ c (E 8 ) = d2 min vol(e 8 ) 1 4 = 2. 20 / 41

Separation: Coding Gain and Shaping Gain Normalized Second Order Moment Energy Performance of Λ s is related to the energy minimization of the lattice code. All points of the lattice code are in the Voronoï region of 0 of Λ s. For high rates, we assume points of Λ c uniformly distributed in the Voronoï region, so the energy per dimension of the lattice code becomes E = 1 ( n E x 2) = 1 ˆ 1 n V Λs (0) vol(λ s ) x 2 dx Normalized Second Order Moment The parameter ( 1 V G (Λ) = Λ (0) x 2 ) dx vol(λ) n 2 n vol(λ) is called the normalized second order moment of the lattice. It has to be minimized. 21 / 41

Separation: Coding Gain and Shaping Gain Coding Gain and Shaping Gain Dominant term of the Error Probability The error probability of a lattice code using Λ c as the coding lattice and Λ s as the shaping lattice is dominated by the term Q ( ) 3mE b γ c (Λ c ) γ s (Λ s ) N 0 22 / 41

Part III Nested Lattices and the Secrecy Gain

The Gaussian Wiretap Channel N0 A B N1 E Figure: The Gaussian Wiretap Channel model 24 / 41

Encoder Design The problem of Wiretap is a problem of labelling transmitted symbols with data bits 25 / 41

Encoder Design The problem of Wiretap is a problem of labelling transmitted symbols with data bits +2 mod (4) Channel We suppose the alphabet Z 4 and a channel Alice Eve that outputs y = x + 2 with probability 1/2 and x with same probability. The symbol error probability is 1/2. Symbol to Bits Labelling x = 2b 1 + b 0 Bit b 1 experiences error probability 1/2 while bit b 0 experiences error probability 0. 25 / 41

Uniform Noise Assume that Alice Eve channel is corrupted by an additive uniform noise Label points with data + pseudo random bits Transmitted point Figure: Constellation corrupted by uniform noise

Uniform Noise Assume that Alice Eve channel is corrupted by an additive uniform noise Label points with pseudo random bits Transmitted point Figure: Points can be decoded error free: label with pseudo-random symbols

Uniform Noise Assume that Alice Eve channel is corrupted by an additive uniform noise Label points with data Transmitted point Figure: Points are not distinguishable: label with data

Uniform Noise Label points with data Transmitted point Label points with pseudo random bits Transmitted point Figure: Constellation corrupted by uniform noise 26 / 41

Uniform Noise Error Probability Pseudo-random symbols are perfectly decoded by Eve when data error probability will be high. unfortunately not valid for Gaussian noise. Label points with data Transmitted point Label points with pseudo random bits Transmitted point Figure: Constellation corrupted by uniform noise 26 / 41

Coset Coding with Integers Label points with data + pseudo random bits Transmitted point Figure: Constellation corrupted by uniform noise 27 / 41

Coset Coding with Integers Example Suppose that points x are in Z. Euclidean division x = 3q + r q carries the pseudo-random symbols while r carries the data or pseudo-random symbols label points in 3Z while data label elements of Z/3Z. Label points with data + pseudo random bits Transmitted point Figure: Constellation corrupted by uniform noise 27 / 41

Lattice Coset Coding Gaussian noise is not bounded: it needs a n dimensional approach (then let n for sphere hardening). 1 dimensional n dimensional Transmitted lattice Z Fine lattice Λ b Pseudo-random symbols mz Z Coarse lattice Λ e Λ b Data Z/mZ Cosets Λ b /Λ e Table: From the example to the general scheme 28 / 41

Lattice Coset Coding Gaussian noise is not bounded: it needs a n dimensional approach (then let n for sphere hardening). points of Λ b Voronoi region of Λe Figure: Example of coset coding 28 / 41

Part IV The Secrecy Gain

The secrecy gain Eve s Probability of Correct Decision (data) 30 / 41

The secrecy gain Eve s Probability of Correct Decision (data) Can Eve decode the data? Figure: Eve correctly decodes when finding another coset representative 30 / 41

The secrecy gain Eve s Probability of Correct Decision (data) Can Eve decode the data? Eve s Probability of correct decision P c,e ( ) 1 n Vol ( ) Λ b e r 2 2N 1 2πN1 r Λ e ( ) 1 n Vol ( ( ) 1 Λ b ΘΛe 2πN1 2πN 1 ) where Figure: Eve correctly decodes when finding another coset representative Θ Λ (y) = q x 2, q = e πy, y > 0 x Λ is the theta series of Λ. 30 / 41

The secrecy gain Secrecy function Definition Let Λ be a n dimensional lattice with volume λ n. Its secrecy function is defined as, Ξ Λ (y) Θ λz n (y) Θ Λ (y) ϑ n 3 (e ) π λy = Θ Λ (y) where ϑ 3 (q) = + n= q n2 (Jacobi theta function) and y > 0. Examples E8 y 1.30 1.25 1.20 1.15 1.10 1.05 24 y 4.0 3.5 3.0 2.5 2.0 1.5 1.00 6 4 2 0 2 4 6 y db 1.0 6 4 2 0 2 4 6 y db Figure: Secrecy functions of E 8 and Λ 24 31 / 41

The secrecy gain Secrecy Gain Definition The strong secrecy gain of a lattice Λ is χ s Λ supξ Λ (y) y>0

The secrecy gain Secrecy Gain Definition The strong secrecy gain of a lattice Λ is χ s Λ supξ Λ (y) y>0 Definition For a lattice Λ equivalent to its dual and of determinant d (Λ) (determinant of the Gram matrix), we define the weak secrecy gain, χ Λ Ξ Λ (d (Λ) n 1 ) A lattice equivalent to its dual has a theta series with a multiplicative symmetry point at d (Λ) n 1 (Poisson-Jacobi s formula), Ξ Λ (d (Λ) n 1 ) y = Ξ Λ d (Λ) n 1 y 32 / 41

The secrecy gain First conjecture Conjecture If Λ is a lattice equivalent to its dual, then the strong and the weak secrecy gains coincide. Corollary The strong secrecy gain of a unimodular lattice Λ is χ s Λ Ξ Λ(1) Calculation of E 8 secrecy gain From E 8 theta series, 1 Ξ E8 (1) = = 1 ( 2 ϑ2 (e π ) 8 + ϑ 3 (e π ) 8 + ϑ 4 (e π ) 8) ϑ 3 (e π ) 8 3 (since ϑ ( 2 e π ) 4 ϑ 3 (e π ) = ϑ ( 4 e π ) ϑ 3 (e π ) = 4 1 ) 2 so we get χ E8 = Ξ E8 (1) = 4 3. 33 / 41

The secrecy gain Asymptotic behavior for unimodular lattices Want to study the behavior of even unimodular lattices when n. Question How does the optimal secrecy gain behaves when n? 34 / 41

The secrecy gain Asymptotic behavior for unimodular lattices Want to study the behavior of even unimodular lattices when n. Question How does the optimal secrecy gain behaves when n? First answer Apply the Siegel-Weil formula, where Θ Λ (q) ( Λ Ω Aut(Λ) = M n E k q 2) n M n = 1 Λ Ω Aut(Λ) n and E k is the Eisenstein series with weight k = n 2. Ω n is the set of all inequivalent n dimensional, even unimodular lattices. We get Θ n,opt ( e π ) E k ( e 2π) 34 / 41

The secrecy gain Asymptotic behavior (II) Maximal Secrecy gain For a given dimension n, multiple of 8, there exists an even unimodular lattice whose secrecy gain is ( χ n ϑn 3 e π ) ( E k e 2π ) 1 π 1 n 4 ( ) 1.086n 2 Γ 34 2 Behavior of Eisenstein Series We have ( E k e 2π) = 1 + 2k + m k 1 Bk m=1 e 2πm 1 B k being the Bernoulli numbers. For k a multiple of 4, then E k ( e 2π ) fastly converges to 2 (k ). 35 / 41

The secrecy gain Asymptotic behavior (II) Maximal Secrecy gain For a given dimension n, multiple of 8, there exists an even unimodular lattice whose secrecy gain is Bound from Siegel-Weil Formula vs. Extremal lattices ( χ n ϑn 3 e π ) ( E k e 2π ) 1 π 1 n 4 ( ) 1.086n 2 Γ 34 2 Behavior of Eisenstein Series We have ( E k e 2π) = 1 + 2k + m k 1 Bk m=1 e 2πm 1 B k being the Bernoulli numbers. For k a multiple of 4, then E k ( e 2π ) fastly converges to 2 (k ). Figure: Lower bound of the minimal secrecy gain as a function of n from Siegel-Weil formula. 35 / 41

Part V Wireless Channels - Other Lattices

Block Fading Channel Design Criterion Assume a wireless communication system transmitting on q subcarriers sufficiently spaced and during n channel uses. Assume Rayleigh fadings and 2 codewords X and T such that x 11 x 12 x 1n x 21 x 22 x 2n X =....... x q1 x q2 x qn Pairwise Error Probability Error probability will be dominated by T = ( ) max P (X T) q = max xi t 2 Γ n i X,T X,T i=1 4 t 11 t 12 t 1n t 21 t 22 t 2n....... t q1 t q2 t qn where Γ is the average signal to noise ratio and x i (resp. t i ) is the i th row of X (resp. T). This equality is valid if, for any i, x i t i. Hence, one has to find a code which maximizes min X,T µ(x,t) = min X,T q xi t 2 i i=1. 37 / 41

Block Fading Channel Lattice formulation over number fields Control the product in µ(x,t) The product q xi t 2 q i=1 i which becomes µ(x ) = xi 2 i=1 by linearity can be controled by introducing the algebraic norm in a well-chosen algebraic Galois extension K of degree q. Let ( σ 1,σ 2,...,σ q ) be the Galois group of K. Use the canonical embedding so that X = σ 1 (x 1 ) σ 1 (x 2 ) σ 1 (x n) σ 2 (x 1 ) σ 2 (x 2 ) σ 2 (x n)....... σ q (x 1 ) σ q (x 2 ) σ q (x n) 38 / 41

Block Fading Channel Construction of O K lattices A construction A for O K lattices where O K is the ring of integers of K can be given where a O K lattice is a O K module. 39 / 41

Block Fading Channel Construction of O K lattices A construction A for O K lattices where O K is the ring of integers of K can be given where a O K lattice is a O K module. Construction A (binary) over O K Take, for instance q = 2, K = Q ( 2 ). So, we have O K = Z[ 2]. Consider the principal ideal I = 2 Z[ 2]. As N (I ) = 2, then Z[ 2] / I = F 2. So, we can construct O K lattices in that way, where C is a binary linear code of length n. Λ = I n + C Construction A (quaternary) over O K Here, take q = 2, K = Q ( 5 ). So, we have O K = Z[φ] where φ = 1+ 5 2. Consider the ideal I = 2 Z[ φ]. As N (I ) = 4 and I is prime, then Z[ φ] / I = F 4. So, we have where C is a linear code of length n over F 4. Λ = I n + C 39 / 41

Perspectives Perspectives O lattices where O is a maximal order of some division algebra for the MIMO case Nested lattices for other applications in which 2 or more data streams must be constructed Han and Kobayashi Wyner-Ziv... Nested exotic lattices on other Dedekind domains? 40 / 41

Thank You!!