DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 1

Delay-Doppler Channel Estimation in Almost

Linear Complexity
Alexander Fish, Shamgar Gurevich, Ronny Hadani, Akbar Sayeed, and Oded Schwartz

To Solomon Golomb for the occasion of his 80 birthday mazel tov

Abstract—A fundamental task in wireless communication is

channel estimation: Compute the channel parameters a signal
undergoes while traveling from a transmitter to a receiver. In
the case of delay-Doppler channel, i.e., a signal undergoes only
delay and Doppler shifts, a widely used method to compute
the delay-Doppler parameters is the matched filter algorithm.
It uses a pseudo-random sequence of length N; and, in case
of non-trivial relative velocity between transmitter and receiver,
its computational complexity is O(N 2 log N ). In this paper we
introduce a novel approach of designing sequences that allow
faster channel estimation. Using group representation techniques
we construct sequences, which enable us to introduce a new
algorithm, called the flag method, that significantly improves the
matched filter algorithm. The flag method finds m delay-Doppler Fig. 1. Three paths scenario.
parameters in O(mN log N ) operations. We discuss applications
of the flag method to GPS, and radar systems.
Index Terms—Channel estimation, time-frequency shift prob- vector space H is endowed with the inner product
lem, fast matched filter, flag method, sequence design, P
Heisenberg–Weil sequences, fast moving users, high-frequency hf1 ; f2 i = f1 [n]f2 [n];
communication, radar, GPS. n2ZN

for f1 ; f2 2 H, and referred to as the Hilbert space of (digital)

p
2 i
sequences. Finally, we define e(t) = e N t ; where i = 1:
I. I NTRODUCTION
FUNDAMENTAL building block in many wireless
A communication protocols is channel estimation: learning
the channel parameters a signal undergoes while traveling
A. Channel Model
Let us start with the derivation of the discrete channel model
that we will consider throughout this paper. We follow closely
from a transmitter to a receiver [17]. In this paper we de-
the works [13], [14], [15], [16]. The transmitter sends—see
velop an efficient algorithm1 for delay-Doppler (also called
Figure 1 for ilustration—an analog signal SA (t); t 2 R; of
time-frequency) channel estimation. Our algorithm provides a
(two-sided) bandwidth W . While the actual signal is modu-
striking improvement over current methods in the presence of
lated onto a carrier frequency fc W; we consider a widely
high relative velocity between a transmitter and a receiver.
used complex baseband model for the multipath channel. In
The latter scenario occurs in GPS, radar systems, mobile
addition, we make the sparsity assumption on the finiteness of
communication of fast moving users, and very high frequency
the number of signal propagation paths. The complex baseband
(GHz) communication.
analog received signal is (see Equation (14) in [14]) given by
Throughout this paper we denote by H the vector space
m
X
of complex valued functions on the set of integers ZN =
f0; 1; :::; N 1g equipped with addition and multiplication RA (t) = k exp(2 ifk t) SA (t tk ) + W(t); (I-A.1)
k=1
modulo N: We assume that N is an odd prime number. The
where m denotes the number of propagation paths, k 2 C
A. Fish is with School of Mathematics and Statistics, University of Sydney, is the path coefficient, fk 2 R is the Doppler shift, and tk 2
Sydney, NSW 2006, Australia. Email: [email protected]. R+ is the path delay associated with the k-th path, and W
S. Gurevich is with the Department of Mathematics, University of Wiscon-
sin, Madison, WI 53706, USA. Email: [email protected]. denotes
Pm a random white noise. We assume the normalization
2
R. Hadani is with the Department of Mathematics, University of Texas, k=1 j kj 1. The Doppler shift depends on the relative
Austin, TX 78712, USA. Email: [email protected]. speed between the transmitter and the receiver along the path,
A. Sayeed is with the Department of Electrical and Computer Engi-
neering, University of Wisconsin, Madison, WI 53706, USA. Email: ak- and the delay encodes the distance between the transmitter and
[email protected]. receiver along the path. The parameter m will be called also
O. Schwartz is with the Department of Electrical Engineering and Com- the sparsity of the channel. We will call
puter Science, University of California, Berkeley, CA 94720, USA. Email:
[email protected].
1 Announcement of this result appeared in [3]. ( k ; tk ; fk ); k = 1; :::; m; (I-A.2)
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 2

B. Channel Estimation Problem

Consider sequences S; R 2 H, where R is given by the
following formula:
m
X
R[n] = k e(! k n) S[n k ]+W[n]; n 2 ZN ; (I-B.1)
k=1
Pm 2
with k 2 C, k=1 j k j 1; k ; ! k 2 ZN ; and W 2 H
denotes a random white noise. For the rest of the paper we
assume that all the coordinates of the sequence W are indepen-
dent, identically distributed random variables of expectation
zero. In analogy with the physical channel model described by
Equation (I-A.1), we will call k ; k ; ! k ; k = 1; :::; m; path
Fig. 2. Illustration of the process allowing the discrete channel model. coefficients, path delays, and Doppler shifts, respectively. The
objective is:
the channel parameters, and the main objective of channel Problem I-B.1 (Channel Estimation): Design S 2 H, and
estimation is to obtain them. We describe now a process— an effective method of extracting the channel parameters
see Figure 2 for illustration—allowing to reduce this task to ( k ; k ; ! k ), k = 1; :::; m; from S and R satisfying (I-B.1).
a problem for sequences. We start with a sequence S 2 H;
and transmit the analog signal Granting the solution of Problem I-B.1, we can compute the
M
X1 channel parameters (I-A.2) using Proposition I-A.1.
SA (t) = S[n mod(N )] sinc(W t n);
Remark I-B.2 (Frequency Resolution): From Proposition
n=0
I-A.1, follows that the resolution of the frequency shifts
where M N; and sinc(t) = sin( t)= t: To specify recognized by our digital method equals to W=N:
the transmission interval, we denote by Ts the time spread
of the channel, i.e., Ts = maxftk g, and we define K =
C. The GPS Problem
dW Ts e ;where dxe is the ceiling function. We assume K N;
and we take M = N + K. Then the transmission time We would like to discuss an important example of channel
of SA (t) is from t = 0 to t = M=W: At the receiver, estimation. A client on the earth surface wants to know his/her
the discrete system representation is obtained by sampling geographical location. The Global Positioning System (GPS) is
RA (t); satisfying (I-A.1), with sampling interval t = 1=W built to fulfill this task. Satellites send to earth their location—
starting from time2 K=W: As a result, we obtain the following see Figure 3 for illustration. For simplicity, the location of
sequence R 2 H:
R[n] = RA ((K + n)=W ) ; (I-A.3)
for n = 0; :::; N 1: By a direct calculation we obtain the
following:
1
Proposition I-A.1: Assume that tk 2 W Z; and fk 2 W N Z;
k = 1; :::; m: Then the sequence R given by (I-A.3) satisfies
m
X
R[n] = k e(! k n) S[K + n k] + W[n]; n 2 ZN ;
k=1

where k = k exp( 2 ifk K=W ); k = tk W , and ! k =

N fk =W:

Remark I-A.2: Even if an actual delay tk or Doppler shift

fk do not exactly lie on the lattice, they can be well approx-
imated by a few discrete delay-Doppler shifts on the lattice
(see [16], Section II-A).
Fig. 3. Satellites communicate location in GPS.
In the next section we formulate the mathematical problem
that we will solve in this paper, suggesting, according to
a satellite is modeled by a bit b 2 f 1g: The satellite
Proposition I-A.1, a method to compute the channel parame-
transmits—for example using the scheme proposed in Section
ters (I-A.2).
I-A—to the earth its sequence S 2 H of norm one multiplied
2 We start to sample at time K=W in order to sense all the terms in Equation by its location b: We assume, for simplicity, that the sequence
(I-A.1) at the receiver. travels through only one path (see Equation (1) in [1]). Hence,
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 3

by the sampling procedure described in Section I-A, the client In order to extract the time-frequency shift ( 0 ; ! 0 ); it is
receives the sequence R 2 H of the form “standard"6 (see [5], [7], [8], [10], [17], [18], [19]) to use
pseudo-random sequence S 2 H of norm one. In this case
R[n] = b 0 e(! 0 n) S[n 0] +W[n]; n 2 ZN ; (I-C.1)
M(S; S)[ 0; !
p ! 0 ] = 1 for ( ; !) = ( 0 ; ! 0 ), and
where 0 2 C, j 0 j 1; 0 ; ! 0 2 ZN , and W 2 H is a bounded by C= N ; C > 0; if ( ; !) 6= ( 0 ; ! 0 ). Hence, with
random white noise. Using 0 we can compute the distance probability going to one, as N goes to infinity, we have
from the satellite to the client3 , assuming a line of sight
1 + "N , if ( ; !) = ( 0 ; ! 0 );
between them. The problem of GPS can be formulated as M(R; S)[ ; !] = (I-E.2)
N, if ( ; !) 6= ( 0 ; ! 0 );
follows:
p p
where j"N j 1= N , and j N j (C + 1)= N :
Problem I-C.1 (GPS): Design S 2 H, and an effective
method of extracting (b; 0 ) from S and R satisfying (I-C.1).

In practice, the satellite transmits S = S1 + bS2 ; where

S1 ; S2 are almost orthogonal in some appropriate sense. Then
( 0 ; 0 ; ! 0 ); and (b 0 ; 0 ; ! 0 ) are computed using S1 ; and
S2 ; respectively, concluding with the derivation of the bit b.

D. The Time-Frequency Shift Problem

To suggest a solution to Problem I-C.1, and subsequently
to Problem I-B.1, we consider a simpler variant. Suppose the
transmitter and the receiver sequences S; R 2 H are related
by
R[n] = e(! 0 n) S[n 0] + W[n]; n 2 ZN ; (I-D.1)
Fig. 4. jM(R; S)j with pseudo-random S, and ( 0 ; !0 ) = (50; 50):
where ( 0 ; ! 0 ) 2 ZN ZN , and W 2 H denotes a random
white noise. The pair ( 0 ; ! 0 ) is called the time-frequency Identity (I-E.2)—see Figure 4 for a demonstration—
shift, and the vector space V = ZN ZN is called the time- suggests the following “entry-by-entry" solution to TFS prob-
frequency plane. We would like to solve the following: lem: Compute the matrix M(R; S); and choose ( 0 ; ! 0 )
for which M(R; S)[ 0 ; ! 0 ] 1: However, this solution of
Problem I-D.1 (Time-Frequency Shift (TFS)): Design TFS problem is expensive in terms of arithmetic complexity,
S 2 H, and an effective method of extracting the time- i.e., the number of multiplication and addition operations is
frequency shift ( 0 ; ! 0 ) from S and R satisfying (I-D.1). O(N 3 ): One can do better using a “line-by-line" computation.
This is due to the following observation:

E. The Matched Filter Algorithm Remark I-E.2 (FFT): The restriction of the matrix
A classical solution to Problem I-D.1, is the matched filter M(R; S) to any line7 (not necessarily through the origin)
algorithm [5], [7], [8], [10], [17], [18], [19]. We define the in the time-frequency plane V; is a certain convolution—for
following matched filter matrix4 of R and S: details see Section V—that can be computed, using the fast
Fourier transform8 (FFT), in O(N log N ) operations.
M(R; S)[ ; !] = hR[n]; e(!n) S[n ]i , ( ; !) 2 V: As a consequence of Remark I-E.2, one can solve TFS
For R and S satisfying (I-D.1), the law of the iterated problem in O(N 2 log N ) operations.
logarithm implies that, with probability going to one, as N
goes to infinity, we have
F. The Fast Matched Filter Problem
M(R; S)[ ; !] (I-E.1)
To the best of our knowledge, the “line-by-line" computa-
= 0 M(S; S)[ 0; !
! 0 ] + "N ; tion is also the fastest known method [11]. If N is large this
p p
where 0 = e(! 0 ! 0 ), j"N j 2 log log N = N SN R; may not suffice. For example, in applications to GPS [1], as
with SN R denotes the signal-to-noise ratio5 : in Problem I-C.1 above, we have N 1000: This leads to the
Remark I-E.1 (Noise Assumption): For the rest of the pa- following:
per we assume,
p for simplicity, that SN R 2 log log N; i.e.,
6 For example in spread-spectrum communication systems.
j"N j 1= N in (I-E.1): 7 In this paper, by a line through the origin we mean all scalar multiples
L = fau; a 2 ZN g of a fixed non-zero vector u 2 V: In addition, by a line
3 Since we work modulo N , the distance can be found modulo N c, where
W we mean a subset of V of the form L + v, where L is a fixed line through
W is the bandwidth, and c is the speed of light. the origin, and v 2 V is a fixed vector.
4 The matched filter matrix is called ambiguity function in radar theory. 8 The Rader algorithm [12] provides implementation of the FFT for se-
5 We define SN R = hS; Si =V ar(W). quences of prime length.
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 4

The Flag Algorithm

Input. The line L V , and SL ; R as in (I-D.1).

Output. The time-frequency shift ( 0 ; ! 0 ).

Step 1. Choose a line L? transversal to L.

Step 2. Compute M(R; SL ) on L? p , and find ( ; !)

such that jjM(R; SL )[ ; !]j 1=2j 5= N , i.e., ( ; !) on
the shifted line L + ( 0 ; ! 0 ).

Step 3. Compute M(R; SL ) on L + ( 0 ; ! 0 ), pand find

( ; !) such that jM(R; SL )[ ; !] 1j 1= N , i.e.,
Fig. 5. jM(R; SL )j for a flag SL with L = f(0; !)g; and ( 0 ; !0 ) =
(50; 50): ( 0 ; ! 0 ) = ( ; !).

Problem I-F.1 (Fast Matched Filter): Solve the TFS prob-

lem in almost linear complexity.

Note that computing one entry in M(R; S) already takes

O(N ) operations.

G. The Flag Method

We introduce the flag method to propose a solution to
solve the TFS problem with algorithm which reduces the
computational complexity from O(N 2 log N ) to O(N log N )
operations. The idea is to use a new set of sequences, which
enable, first, to find a line on which the time-frequency shift
Fig. 6. Diagram of flag algorithm.
is located, and, then, to search on the line to find the time-
frequency shift. With each line L through the origin in V; we
associate sequence SL 2 H of norm one, that we call flag. H. Solution to the GPS and Channel Estimation Problems
Since there are N + 1 lines through the origin in V , we obtain
Let L V be a line through the origin.
N + 1 different flag sequences. Each flag sequence satisfies—
see Figure 5 for illustration—the following “flag property"9 : Definition I-H.1 (Genericity): We say that the points
For a sequence R given by (I-D.1) with S = SL ; we have ( k ; ! k ) 2 V; k = 1; :::; m; are L-generic if no two of them
with probability going to one, as N goes to infinity, lie on a shift of L, i.e., on L + v, for some v 2 V:
M(R; SL )[ ; !] (I-G.1) Looking back to Problem I-B.1, we see that, under generic-
8 ity assumption, the flag method provides a fast computation,
< 1 + "N , if ( ; !) = ( 0 ; ! 0 );
= 1
+ N , in j j if ( ; !) 2 L0 r ( 0 ; ! 0 ); in O(mN log N ) operations, of the channel parameters of
: 2 channel with sparsity m. In particular, it calculates the GPS
N, if ( ; !) 2 V r L0 ;
p p parameters—see Problem I-C.1—in O(N log N ) operations.
where j"N j 1= N ; j N j 5= N ; j j denotes absolute Indeed, Identity (I-G.1), together with bilinearity of inner
value, and L0 is the shifted line L + ( 0 ; ! 0 ). In addi- product, implies that
tion10 , the flag sequences will satisfy the following “almost
k M(R; SL )[ k ; ! k ]; k = 1; :::; m;
orthogonality" property: Ifp L; M are two different lines, then
jM(SL ; SM )[ ; !]j 3= N ; for every ( ; !) 2 V: n where R is the sequence (I-B.1), with S = SL ; assuming that
As a consequence of Equation (I-G.1), we can apply the ( k ; ! k )’s are L-generic. So we can adjust the flag algorithm
Flag Algorithm, as described below, to solve the matched as follows:
filter problem, with probability going to one as N goes to Compute M(R; SL ) on L? . Find all ( ; !)’s such that
infinity. The complexity of the flag algorithm—see Figure 6 jM(R; SL )[ ; !]j is sufficiently large, i.e., find all the
for a demonstration—is O(N log N ); using FFT. This com- shifted lines L + ( k ; ! k )’s:
pletes our solution of Problem I-F.1—The Fast Matched Filter
Compute M(R; SL ) on each line L + ( k ; ! k ); and find
Problem.
( ; !) such that jM(R; SL )[ ; !]j is maximal on that line,
9 In linear algebra, a pair (` ; L) consisting of a line L
0 V; and a point i.e., ( ; !) = ( k ; ! k ) and k M(R; SL )[ k ; ! k ]:
`0 2 L, is called a flag.
10 This is important in various real-world applications, e.g., in GPS, radar, Figure 7 provides a visual illustration for the matched filter
and CDMA communication. matrix in three paths scenario. This completes our solutions
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 5

Fig. 8. Radar transmits wave and recieves echo.

Fig. 7. jM(R; SL )j ; for L = f(0; !)g; and ( k; k ; !k ) =

( p1 ; 50k; 50k); k = 1; 2; 3:
3
In Section III: You can see the design of the Heisenberg–
Weil flag sequences, using the Heisenberg–Weil opera-
of Problem I-B.1—The Channel Estimation Problem, and of tors, and diagonalization techniques of commuting op-
Problem I-C.1—The GPS Problem. erators. In addition, the investigation of the correlation
properties of the flag sequences is done in this section.
I. Applications to Radar These properties are formulated in Theorem III-C.1,
which guarantees applicability of the Heisenberg-Weil
The model of radar works as follows [10]. A radar
sequences to the flag method.
transmits—see Figure 8 for illustration—a sequence S 2 H
In Section IV: You can get explicit formulas for large
which bounces back from m targets. Using the sampling
collection of the Heisenberg–Weil flag sequences. In par-
procedure described in Section I-A, the radar receives as an
ticular, these formulas enable to generate the sequences
echo the following sequence R 2 H:
using low complexity algorithm.
m
X In Section V: You can find the formulas that suggest fast
R[n] = k e(! k n) S[n k] + W[n]; n 2 ZN ; computation of the matched filter matrix on any line in
k=1
the time-frequency plane. These formulas are of crucial
Pm 2
where k 2 C, k=1 j k j 1; ! k 2 ZN encodes the radial importance for the effectiveness of the flag method.
velocity of target k with respect to the radar, k 2 ZN encodes In Section VI: You can find needed proofs and justifica-
the distance between target k and the radar, and W is a random tions for all the claims and formulas that appear in the
white noise. body of the paper.
In order to determine the distances to the targets, and their
relative radial velocities with respect to the radar, we need to
II. T HE H EISENBERG AND W EIL O PERATORS
solve the following:
Problem I-I.1 (Radar): Having R and S, compute the pa- The flag sequences (see Section I-G) are defined, con-
rameters ( k ; ! k ); k = 1; :::; m: structed and analyzed using two special classes of operators
This is a particular case of the channel estimation problem. that act on the Hilbert space of sequences. The first class
Under the genericity assumption, the flag method solves it in consists of the Heisenberg operators and is a generalization
O(mN log N ) operations. of the time-shift and frequency-shift operators. The second
class consists of the Weil operators and is a generalization of
Remark I-I.2 (Velocity Resolution): From Remark I-B.2, the discrete Fourier transform. In this section we recall the
follows that larger N used by the radar implies better res- definitions and explicit formulas of these operators.
olution of recognized radial velocities of the targets.

J. What you can find in this paper A. The Heisenberg Operators

In the Section I: You can read about the derivation The Heisenberg operators are the unitary transformations
of the discrete delay-Doppler channel model, and the that act on the Hilbert space of sequences by
flag method for effective channel estimation. In addition,
concrete applications to GPS, and radar are discussed. ( ; !) : H ! H; ; ! 2 ZN ;
In Section II: You can find the definition and explicit [ ( ; !)f ][n] = e( 2 1 !) e(!n) f [n ];
formulas for the Heisenberg and Weil operators. These (II-A.1)
operators are our basic tool in the development of the flag where f 2 H, n 2 ZN ; and for the rest of this paper we use
method, in general, and the flag sequences, in particular. 2 1 to denote N2+1 which is the inverse of 2 modulo N:
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 6

B. The Weil Operators 1) Formulas for Weil Operators: It is important for our
Consider the discrete Fourier transform study to have the following explicit formulas [4], [6] for the
8 Weil operators:
< DF T : H ! H;
NP1 Fourier. We have
: [DF T (f )][n] = p1N e( nk) f [k];
k=0 0 1 N 1
f [n] = i 2 DF T (f )[n]; (II-B.4)
for every f 2 H, n 2 ZN : It is easy to check that the DF T 1 0
satisfies the following N 2 identities: Chirp. We have
DF T ( ; !) = (!; ) DF T; ; ! 2 ZN ; (II-B.1) 1 0 1
f [n] = e(2 cn2 )f [n]; (II-B.5)
where ( ; !) are the Heisenberg operators, and denotes c 1
composition of transformations. In [20] Weil found a large Scaling. We have
family of operators, which includes the DF T: His operators
satisfy identities analogous to (II-B.1). In more details, con-
a 0 a 1
sider the collection of matrices f [n] = f [a n]; (II-B.6)
0 a 1 N
a b
SL2 (ZN ) = ; a; b; c; d 2 ZN ; and ad bc = 1 :
c d for every f 2 H, 0 6= a; c; n 2 ZN ; where Na is the Legendre
symbol which is equal to 1 if a is a square modulo N; and
Note that G = SL2 (ZN ) is a group with respect to the
1 otherwise.
operation of matrix multiplication (see [2] for the notion of a
group). It is called the special linear group of order two over The group G admits the Bruhat decomposition
ZN : Each element G = U A [ U wU A;
a b
g= 2 G; where U G denotes the unipotent subgroup
c d
acts on the time-frequency plane V = ZN ZN via the change 1 0
U= ; c 2 ZN ;
of coordinates c 1
( ; !) 7! g( ; !) = (a + b!; c + d!): and A G denotes the diagonal subgroup
For g 2 G, let (g) be a linear operator on H which is a a 0
solution of the following system of N 2 linear equations: A= 1 ; 0 6= a 2 ZN ; (II-B.7)
0 a
g : (g) ( ; !) = (g( ; !)) (g);
; ! 2 ZN ; and w is the Weyl element. This means that every element
(II-B.2) g 2 G can be written in the form
Denote by Sol( g ) the space of all solutions to System
(II-B.2). For example for g = us or g = u0 wu00 s0
0 1 where u; u0 ; u00 2 U; s; s0 2 A; and w is the Weyl element.
w= ;
1 0 Hence, because is homomorphism, i.e., (gh) = (g) (h)
which is called the Weyl element, we have by (II-B.1) that for every g; h 2 G; we deduce that formulas (II-B.4), (II-B.5),
DF T 2 Sol( w ): Using results, from group representation and (II-B.6), extend to describe all the Weil operators.
theory, known as Stone–von Neumann (S-vN) theorem and
Schur’s lemma, one can show (see Section 2.3 in [6]) that III. S EQUENCE D ESIGN : H EISENBERG –W EIL F LAGS
dim Sol( g ) = 1, for every g 2 G. In fact there exists a
special set of solutions. This is the content of the following The flag sequences, that play the main role in the flag
result [20]: method, are of a special type. We define them as a sum
of a pseudorandom sequence and a structural sequence. The
Theorem II-B.1 (Weil operators): There exists a unique design of these sequences is done using group representation
collection of solutions f (g) 2 Sol( g ); g 2 Gg; which theory. The pseudorandom sequences are designed [7], [8],
are unitary operators, and satisfy the homomorphism condition [19] using the Weil representation operators (II-B.3), and will
(gh) = (g) (h); for every g; h 2 G: be called Weil (spike) sequences11 . The structural sequences
Denote by U (H) the collection of all unitary operators on are designed [9], [10] using the Heisenberg representation
the Hilbert space H of sequences. Theorem II-B.1 establishes operators (II-A.1), and will be called Heisenberg (line) se-
the map quences. Finally, the flag sequences are defined as a sum of
: G ! U (H); (II-B.3) Heisenberg sequence, and a Weil sequence, and will be called
Heisenberg–Weil flag sequences.
which is called the Weil representation [20]. We will call each
11 For the purpose of the Flag method, other pseudorandom signals may
(g); g 2 G; a Weil operator.
work.
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 7

A. The Heisenberg (Lines) Sequences B. The Weil (Spikes) Sequences

The operators (II-A.1) obey the Heisenberg commutation We follow closely the works [7], [8]. The group G =
relations SL2 (ZN ) is non-commutative, but contains a special class of
maximal commutative subgroups called tori [7]. There are two
( ; !) ( 0 ; ! 0 ) = e(! 0
!0 ) ( 0 ; !0 )
( ; !):
types of tori in G, split and non-split. A subgroup T G is
(III-A.1)
called split torus if there exists g 2 G such that
The expression ! 0 ! 0 vanishes if ( ; !), ( 0 ; ! 0 ) are on the
same line through the origin. Hence, for a given line through 1 1
T = gAg = fgsg ; s 2 Ag;
the origin L V = ZN ZN ; we have a commutative
collection of unitary operators where A is the subgroup (II-B.7) of diagonal matrices in G: A
subgroup T G is called non-split torus if there exists g 2 G
(l) : H ! H; l 2 L: (III-A.2) such that
The simultaneous diagonalization theorem from linear algebra T = gKg 1 ;
implies the existence of orthonormal basis BL for H; consist- where
ing of common eigensequences for all the operators (III-A.2).
1 0
Moreover, in our specific case there exists an explicit basis K = g 2 G; g t g g = g , g = ;
0
(see IV-A for explicit formulas) BL = ffL; g; parametrized
by characters12 of L. We will call the sequences fL; with 2 ZN a fixed non-square13 .
Heisenberg sequences, and they satisfy Example III-B.1: If N 3 is divisible by 4, then = 1
(l)fL; = (l)fL; ; for every l 2 L: is a non-square in ZN : In this case, an example of a non-split
torus is the group
The following theorem—see Figure 9 for a demonstration of
Property 1—describes their correlations properties: K = fg 2 SL2 (ZN ); g t g = Ig = SO2 (ZN );
Theorem III-A.1: The Heisenberg sequences satisfy the fol- of orthogonal matrices with determinant equal to one. This
lowing properties: group is also called the special orthogonal group.
1) Line. For every line L V , and every fL 2 BL , we
have Claim III-B.2: There are (N + 1)N=2 split tori, and (N
1)N=2 non-split tori in G:
1, if ( ; !) 2 L;
jM(fL ; fL )[ ; !]j =
0; if ( ; !) 2
= L; For a proof of Claim III-B.2 see Section VI-B.
and moreover, M(fL ; fL )[0; 0] = 1: For a given torus T G; we have by II-B.3 a commutative
2) Almost-orthogonality. For every two lines L1 6= L2 collection of diagonalizable Weil operators
V , and every fL1 2 BL1 , fL2 2 BL2 ; we have
p (g) : H ! H, g 2 T: (III-B.1)
jM(fL1 ; fL2 )[ ; !]j = 1= N ;
The simultaneous diagonalization theorem from linear algebra
for every ( ; !) 2 V: implies the existence of orthonormal basis BT for H; consist-
ing of common eigensequences for all the operators (III-B.1).
Theorem III-A.1 can be deduced from general results ob- Moreover, in our specific case there exists [7], [8] an explicit
tained in [9], [10]. However, for the sake of completeness we basis (see Section IV-B for explicit formulas in the case T is
supply a direct proof in Section VI-A. a split torus) BT = f'T; g; parametrized by characters14
of T . The sequences 'T; satisfy
(g)'T; = (g)'T; ; for every g 2 T: (III-B.2)
Remark III-B.3: There is a small abuse of notation in
(III-B.2). The torus T admits a unique non-trivial character
q —called the quadratic character—which takes the values
q (g) 2 f 1g; g 2 T: The dimension of the space H q of
sequences 'T; q , which satisfy (g)'T; q = q (g)'T; q is
equal to 2 or 0, if T is a split or non-split torus, respectively
[7], [8].
Let us denote by
ST = BT r H q ;
Fig. 9. jM(fL ; fL )j for L = f( ; )g:
13 An element y 2 Z
N is called square (non-square) if there exists (does
not exist) x 2 ZN such that y = x2 :
12 A functions : L ! C = C 0; is called character if it satisfies 14 A functions : T ! C ; is called character if it satisfies (gg 0 ) =
(l + l0 ) = (l) (l0 ), for every l; l0 2 L: (g) (g 0 ); for every g; g 0 2 T:
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 8

the set of sequences in BT ; which are not associated with the

quadratic character. We will call them Weil sequences. The
following theorem [7], [8]—see Figure 10 for illustration of
Property 1—describes their correlations properties:

Theorem III-B.4: The Weil sequences satisfy the following

properties:
1) Spike. For every torus T G, and every 'T 2 ST , we
have
1;
p if ( ; !) = (0; 0);
jM('T ; 'T )[ ; !]j =
2= N ; if ( ; !) 6= (0; 0);
and moreover, M('T ; 'T )[0; 0] = 1:
2) Almost-orthogonality. For every two tori T1 ; T2 G, Fig. 11. jM(SL ; SL )j for Heisenberg–Weil flag with L = f( ; )g:
and every 'T1 2 ST1 , 'T2 2 ST2 ; with 'T1 6= 'T2 ; we
have
p
4=pN ; if T1 6= T2 ; 2) Almost-orthogonality. For every two lines L1 6= L2
M('T1 ; 'T2 )[ ; !] V , tori T1 ; T2 G; and every two flags SLj =
2= N ; if T1 = T2 ; p
fLj + 'Tj = 2; with fLj 2 BLj , 'Tj 2 STj , j =
for every ( ; !) 2 V:
1; 2; 'T1 6= 'T2 ; we have for every ( ; !) 2 V
p
9=(2pN ); if T1 6= T2 ;
jM(SL1 ; SL2 )[ ; !]j
7=(2 N ); if T1 = T2 :
For a proof of Theorem III-C.1 see Section VI-C.
Remark III-C.2: As a consequence of Theorem III-C.1 we
obtain families of N + 1 almost-orthogonal flag sequences
which can be used for solving the TFS and GPS problems
in O(N log N ) operations, and channel estimation, and radar
problems in O(mN log N ) operations for channel of sparsity
m (see details in Section I).
This completes our design of the Heisenberg–Weil flag
sequences.
a 0
Fig. 10. M('T ; 'T ) for T = f 1 ; 0 6= a 2 ZN g:
0 a
IV. F ORMULAS FOR H EISENBERG –W EIL S EQUENCES
In order to implement the flag method it is important to
have explicit formulas for the Heisenberg and Weil sequences,
C. The Heisenberg–Weil Sequences which in particular enable one to generate them with an
We define the Heisenberg–Weil sequences. These are pse- efficient algorithm. In this section we supply such effective de-
quences in H, which are of the form SL = (fL + 'T ) = 2, scription for all Heisenberg sequences, and for Weil sequences
where fL and 'T are Heisenberg and Weil sequences, respec- associated with split tori.
tively. The following theorem—see Figure 11 for illustration
of Property 1—is the main technical result of this paper, and A. Formulas for Heisenberg Sequences
it describes their correlations properties:
First we parametrize the lines in the time-frequency plane,
Theorem III-C.1: The Heisenberg–Weil sequences satisfy and then we provide explicit formulas for the orthonormal
the properties bases of sequences associated with the lines.
1) Flag. For every line through the origin L pV , torus 1) Parametrization of Lines: The N + 1 lines in the time-
T G, and every flag SL = (fL + 'T ) = 2; with frequency plane V = ZN ZN can be described in terms of
fL 2 BL ; 'T 2 ST , we have their slopes. We have
8
< 1 + N , if ( ; !) = (0; 0); Lines with finite slope. These are the lines of the form
jM(SL ; SL )[ ; !]j = 1=2 + "N , if ( ; !) 2 L r (0; 0); Lc =spanf(1; c)g; c 2 ZN .
:
"N , if ( ; !) 2 V r L;
Line with infinite slope. This is the line
p p
where j N j 2= N ; and j"N j 3= N ; and moreover, L1 =spanf(0; 1)g:
M(SL ; SL )[0; 0] = 1 + N :
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 9

2) Formulas: Using the above parametrization, we obtain A function : ZN ! C is called multiplicative character
if (xy) = (x) (y) for every x; y 2 ZN : A way to write
Formulas for Heisenberg sequences associated with lines formulas for such functions is the following. Choose 2 C
of finite slope. For c 2 ZN we have the orthonormal basis which satisfies N 1 = 1; i.e., 2 N 1 = f exp N2 i1 k ;
1 1 k = 0; :::; N 2g, and define a multiplicative character by
BLc = ffc;b [n] = p e(2 cn2 + bn) ; b 2 ZN g;
N logr (n)
(IV-A.1) (n) = ; n 2 ZN : (IV-B.2)
of Heisenberg sequences associated with the line Lc :
Running over all the N 1 possible such ’s, we obtain
Formulas for Heisenberg sequences associated with the all the multiplicative characters of ZN : We are ready to
line of infinite slope. We have the orthonormal basis write, in terms of the parametrization (IV-B.1), the concrete
eigensequences associated with each of the tori. We obtain
BL1 = f b ; b 2 ZN g; (IV-A.2)
Formulas for Weil sequences associated with the diagonal
of Heisenberg sequences associated with the line L1 ; torus. For the diagonal torus A we have the set of Weil
where the b ’s denote the Dirac delta functions, b [n] = 1 sequences
if n = b, and = 0 otherwise. n o
The validity of Formula (IV-A.2) is immediate from De- SA = ' ; 1 6= 2 N 1 ;
finition (II-A.1). For a derivation of Formulas (IV-A.1), see
where ' 2 H is the sequence defined by
Section VI-D.
p 1 (n) if n 6= 0;
' [n] = N 1 (IV-B.3)
0 if n = 0;
B. Formulas for the Weil Sequences
where is the character defined by (IV-B.2).
We describe explicit formulas for the Weil sequences as-
sociated with split tori [5], [7], [8]. First we parametrize the
Formulas for Weil sequences associated with the torus
split tori in G = SL2 (ZN ), and then we write the explicit
Tuc ; for unipotent uc 2 G. For the torus Tuc associated
expressions for the orthonormal bases of sequences associated
with the unipotent element
with these tori.
1 0
1) Parametrization of Split Tori: Recall (see Section III-B) uc = ; c 2 ZN ;
c 1
that a split torus T G is a subgroup of the form T = Tg ;
g 2 G; with we have the set of Weil sequences
n o
Tg = gAg 1 ;
STuc = ' uc ; 1 6= 2 N 1 ;
where A G is the subgroup of all diagonal matrices (II-B.7).
where ' uc 2 H is the sequence defined by
We denote by T = fTg ; g 2 Gg the set of all split tori in
G: A direct computation shows that the collection of all Tg ’s ' uc [n] = e(2 1
cn2 )' [n]; (IV-B.4)
with
1 b for every n 2 ZN ; and ' is the sequence given by
g= ; b; c 2 ZN ; (IV-B.1)
c 1 + bc (IV-B.3).
exhausts the set T . Moreover, in (IV-B.1) the torus Tg can be Formulas for Weil sequences associated with other tori
written also as Tg0 ; for g 6= g 0 ; only if b 6= 0 and Tg ; g 2 G. For k 2 ZN ; and f 2 H, we define
(hk f ) [n] = f [kn]; and (mk f ) [n] = e(2 1 kn2 )f [n]: In
1 b 0 b
g0 = 1 : addition, for 0 6= y 2 ZN we denote by Ny the Legendre
c 1 + bc b 0
symbol of y; which is equal 1, or 1; if x is a square, or
not, respectively. Then, for the torus Tg associated with
2) Formulas: In order to provide the explicit formulas we the element
need to develop some basic facts and notations from the theory 1 b
of multiplicative characters. Consider the group ZN of all non- g= ; b; c 2 ZN ; b 6= 0; (IV-B.5)
c 1 + bc
zero elements in ZN ; with multiplication modulo N: A basic
we have the set of Weil sequences
fact about this group is that it is cyclic, i.e., there exists an n o
element—called generator (sometime called primitive root)— STg = ' g ; 1 6= 2 ;
N 1
r 2 ZN such that
where ' g 2 H denotes the sequence
ZN = f1; r; r2 ; :::; rN 2
g:
We fix, for the rest of this section, a generator r 2 ZN ; and ' g [n] = Cb m 1+bc DF T mb hb ' [n]; (IV-B.6)
b

we define the discrete logarithm map logr : ZN ! ZN 1 by with ' the sequence given by (IV-B.3), and Cb =
N 1
d b
logr (n) = d, if n = r : i 2
N :
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 10

The fact that Formula (IV-B.3) defines a set of Weil se- V. C OMPUTING THE M ATCHED F ILTER ON A L INE
quences is immediate from Identity (II-B.6). For a derivation Implementing the flag method, we need to compute in
of Weil sequences with Formulas (IV-B.4) and (IV-B.6), see O(N log N ) operations the restriction of the matched filter
Section VI-E. matrix to any given line in the time-frequency plane (see
Remark I-E.2). In this section we provide an algorithm for this
task. The upshot is—see Figure 12 for illustration of the case
C. Examples of Explicit Flag Sequences
of the diagonal line—that the restriction of the matched filter
We fix N = 37; and note that r = 2 is a generator for matrix to a line is a certain convolution that can be computed
Z37 , i.e., Z37 = f2d mod(37); d = 0; 1; :::; 35g: We give two fast using FFT. For a; b 2 ZN ; ' 2 H, we define
examples. 1
ma;b (')[n] = e(2 an2 + bn)'[n]; and ' [n] = '[ n]:
1) Flag associated with the time line and the diagonal (V-.1)
torus : We show how to use Formulas (IV-A.1), and (IV-B.3), In addition, for sequences '; 2 H, we denote by ' 2H
to obtain explicit flag sequence their convolution
p X
SL0 [n] = f0;1 [n] + ' 1 [n] = 2; n 2 Z37 ; (' )[ ] = '[n] [ n]; 2 ZN : (V-.2)
n2ZN
associated with the line L0 , and the diagonal torus A:
We consider two cases:
Heisenberg sequence associated with the time line. We
take in (IV-A.1), c = 0; b = 1; and obtain the Heisenberg 1) Formula on lines with finite slope and their shifts. For
sequence c 2 ZN consider the line Lc = f (1; c) ; 2 ZN g; and
1 for a fixed ! 2 ZN the shifted line L0c = Lc + (0; !):
f0;1 [n] = p exp(2 in=37); n 2 Z37 : On L0c we have
37
M('; )[ (1; c) + (0; !)] (V-.3)
Weil sequence associated with the diagonal torus. We
choose 36 3 = exp( 236i 18) = 1: We have then = m c; ! (') mc;0 ( )_ [ ];
d
the character 1 of Z37 , given by 1 (2 mod(37)) = where denotes the complex conjugate of the sequence
d
( 1) , d = 0; 1; :::; 35: Hence, using formula (IV-B.3)
we obtain the Weil sequence ' 1 [n]; n 2 Z37 ; given by :

p1 ( 1)log2 (n) if n 6= 0; 2) Formula on the line with infinite slope and its shifts.
' [n] = 36 Consider the line L1 = f!(0; 1) ; ! 2 ZN g; and for a
1
0 if n = 0:
fixed 2 ZN the shifted line L01 = L1 + ( ; 0): On
L01 we have
2) Flag associated with the diagonal line and a non- p
diagonal torus: We show how to use Formulas (IV-A.1), and M('; )[!(0; 1) + ( ; 0)] = N DF T (' )[!];
(IV-B.6), to obtain explicit flag sequence (V-.4)
p where [n] = [n ]::
SL1 [n] = f1;0 [n] + 'g i [n] = 2; n 2 Z37 ;
The validity of Formula (V-.4) is immediate from the
1
associated with the line L1 , and the torus Tg = gAg , with definition of the matched filter. For a verification of Formula
g given by (IV-B.5), with c = 0 and b = 1: (V-.3) see Section VI-F.
Heisenberg sequence associated with the diagonal line.
We take in (IV-A.1), c = 1; b = 0; and obtain the
Heisenberg sequence
1
f1;0 [n] = p exp( in2 =37); n 2 Z37 :
37

Weil sequence associated with the torus Tg . We choose

36 3 = exp( 236i 9) = i: We have then the character i
of Z37 , given by i (2d mod(37)) = id , d = 0; 1; :::; 35:
Hence, using formula (IV-B.6) we obtain the Weil se-
quence

' gi [n] = m1 DF T m1 ' i

[n]; h i
Fig. 12. M('; )[ ; ] = m 1;0 (') m1;0 ( )_ [ ] on L1 :
where ' i [n] = p1 ilog2 (n) if n 6= 0; and ' i [0] = 0:
36
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 11

VI. P ROOFS where in the second equality we use the unitarity of (g); in
the third equality we use Identity (II-B.2), and finally in the
A. Proof of Theorem III-A.1 last equality we use Step 1. This completes the proof of the
almost orthogonality property, and of Theorem III-A.1.
We will use two lemmas. First, let L V be a line, and for a
character : L ! C ; and vector v 2 V , define the character 3) Proof of Lemma VI-A.1 : For l 2 L we have
v : L ! C ; by v (l) = e( (v; l)) (l); l 2 L; where
(l)[ (v)fL ] = e( (v; l)) (v) (l)fL
: V V ! ZN is the symplectic form [( ; !); ( 0 ; ! 0 )] =
! 0 ! 0 : We have = e( (v; l)) (l)[ (v)fL ];
Lemma VI-A.1: Suppose fL 2 H is a -eigensequence for where in the first equality we use Identity (III-A.1). This
L; i.e., (l)fL = (l)fL ; for every l 2 L: Then the sequence completes the Proof of Lemma VI-A.1.
(v)fL is v -eigensequence for L:
4) Proof of Lemma VI-A.2 : For l 2 L we have
For the second Lemma, let L; M V be two lines, and (gl)fM = (gl) (g)fL
g 2 G = SL2 (ZN ) such that M = gL = fgl ; l 2 Lg: For a
character : L ! C ; define the character g : M ! C ; by = (g) (l)fL
g
(m) = (g 1 m), for every m 2 M: We have = (l) (g)fL
g
Lemma VI-A.2: Suppose fL is a -eigensequence for L; = (gl)fM ;
i.e., (l)fL = (l)fL ; for every l 2 L: Then the sequence where the second equality is by Identity (II-B.2): This com-
fM = (g)fL is g -eigensequence for M: pletes the proof of Lemma VI-A.2.
We verify Lemmas VI-A.1, and VI-A.2, after the proof of
the line, and almost-orthogonality properties. B. Proof of Claim III-B.2
1) Proof of Line Property: Let fL 2 BL be a - We use standard facts on G = SL2 (ZN ) and its toral
eigensequence. For v 2 V we have subgroups. Denote by Ts and Tns ; the collection of all
split, and non-split tori, respectively. The group G acts, by
jM(fL ; fL )[v]j = jhfL ; (v)fL ij conjugation, transitively, on both Ts and Tns : For a torus
1; if v 2 L; T G its stabilizer with respect to this action is its normalizer
=
0; if v 2
= L; subgroup N (T ) = fg 2 G; gT g 1 = T ): Hence, we
where in the first equality we use the definition of M, and in have #Ts = #G=#N (A); and #Tns = #G=#N (K):
the second we use Lemma VI-A.1. This completes the proof A direct calculation shows that #G = (N 2 1)N , and
#N (A) = 2(N 1), #N (K) = 2(N + 1): Hence,
of the line property.
2) Proof of Almost-Orthogonality Property: Consider the #Ts = (N + 1)N=2; and #Tns = (N 1)N=2:
time and frequency lines, L0 = f( ; 0)g; and L1 = f(0; !)g; This completes the proof of Claim III-B.2.
respectively. Recall that (see Section IV-A2) BL0 = ffa ; a 2
ZN g, where fa [n] = p1N e(an); n 2 ZN ; and BL1 = fhb ; b 2
C. Proof of Theorem III-C.1
ZN g, where hb = b . Hence, for every a; b 2 ZN we have p
p 1) Flag Property: Let SL = (fL + 'T ) = 2: We have
jM(fa ; hb )[v]j = 1= N ; v 2 V:
M(SL ; SL ) = [M(fL ; fL ) + M(fL ; 'T )
This implies
+M('T ; fL ) + M('T ; 'T )]=2:
Step 1. The almost-orthogonality holds for every fL0 2
BL0 ; and fL1 2 BL1 : We will show that
p
Next, let L; M V be any two distinct lines in V , and let jM('T ; fL )[ ; !]j 2= N ; ; ! 2 ZN : (VI-C.1)
fL 2 BL ; fM 2 BM :
Step 2. The almost-orthogonality holds for fL and fM : Noting that M(fL ; 'T )[ ; !] = M('T ; fL )[ ; !] we
Indeed, it is easy to see that there exists g 2 G = SL2 (ZN ) obtain from (VI-C.1) also the same bound for M(fL ; 'T ):
such that gL = L0 , and gM = L1 . From Lemma VI-A.2 Having this, using Theorems III-A.1 and III-B.4 we can
and the unitarity of (g) we have that up to unitary scalars deduce the Flag Property. So assume (l)fL = (l)fL for
fL0 = (g)fL 2 BL0 ; and fL1 = (g)fM 2 BL1 . Hence, l 2 L: By Lemma VI-A.1, it is enough to bound the inner
we obtain for every v 2 V product p
jh'T ; fL ij 2= N : (VI-C.2)
jM(fL ; fM )[v]j = jhfL ; (v)fM ij
We proceed in two steps.
= jh (g)fL ; (g) (v)fM ij
Step 1. The bound (VI-C.1) holds for L1 : Indeed, then
= jh (g)fL ; (gv) (g)fM ij
fL1 = b for some b 2 ZN ; hence
= jM(fL0 ; fL1 )[gv]j
p jh'T ; fL1 ij = j'T [b]j sup j'T [n]j:
= 1= N ; n2ZN
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 12

In [8] it was shown that for every Weil sequence 'T we have
p 1 0
sup j'T [n]j 2= N : uc = ; c 2 ZN ;
n2ZN c 1
Step 2. The bound (VI-C.1) holds for every line L: We
maps BL0 to the orthonormal basis BLc = ffc;b = (uc )f0;b ;
will use the following lemma. Consider a torus T G; and
b 2 ZN g of common eigensequences for the operators
an element g 2 G: Then we can define a new torus Tg =
( (1; c)); 2 ZN : Hence, using Formula (II-B.5) we derive
gT g 1 = fghg 1 ; h 2 T g: For a character : T ! C ;
our desired basis
we can associate a character g : Tg ! C , by g (ghg 1 ) =
(h); for every h 2 T: We have 1 1
BLc = ffc;b [n] = p e(2 cn2 + bn); b 2 ZN g:
Lemma VI-C.1: Suppose 'T is a -eigensequence for T; N
i.e., (h)'T = (h)'T ; for every h 2 T: Then the sequence
'Tg = (g)'T is g -eigensequence for Tg :
E. Derivation of Formulas (IV-B.4), and (IV-B.6)
For a proof of Lemma VI-C.1, see Section VI-C2. For a character : A ! C and an element g 2 G; define
the character g : Tg ! C ; by g (ghg 1 ) = (h), for every
Now we can verify Step 2. Indeed, given a line through the h 2 A: Using Lemma VI-C.1, we deduce that for g 2 G the
origin L V; there exists g 2 G such that gL = L1 : In set
particular, by Lemma VI-A.2 we obtain that fL1 = (g)fL is
up to a unitary scalar in BL1 : In addition, by Lemma VI-C.1 STg = f' g = (g)' ; is character of A, 6= q g;
we know that 'Tg = (g)'T is up to a unitary scalar in BTg :
Finally, we have is a set of Weil sequences associated with Tg : Specializing
to the characters = ; 1 6= 2 N 1 ; of A; and the
h'T ; fL i = h (g)'T ; (g)fL i
associated sequence ' 2 SA given by (IV-B.3), we can
= 'T g ; fL1 ; proceed to derive the formulas.
where the first equality is by the unitarity of (g). Hence, by
Step 1, we obtain the desired bound also in this case. 1) Derivation of Formula (IV-B.4): For the unipotent
element
1 0
2) Proof of Lemma VI-C.1 : For h 2 T we have uc = ; c 2 ZN ;
c 1
1 1
(ghg )'Tg = (ghg ) (g)'T
we have
= (g) (h)'T
h i
= (h) (g)'T ' uc [n] = (uc )' [n]
g 1
= (ghg )'Tg ; = e(2 1
cn2 )' [n] ;
where the second equality is because is homomorphism (see
Theorem II-B.1). This completes our proof of Lemma VI-C.1, for every n 2 ZN , where the second equality is by Formula
and of the Flag Property. (II-B.5). This completes our verification of Formula (IV-B.4).
p
3) Almost Orthogonality : Let SLj = fLj + 'T j = 2; 2) Derivation of Formula (IV-B.6): For the element
j = 1; 2; as in the assumptions. We have
1 b
M(SL1 ; SL2 ) = [M(fL1 ; fL2 ) + M(fL1 ; 'T2 ) g= ; b; c 2 ZN ; b 6= 0;
c 1 + bc
+M('T1 ; fL2 ) + M('T1 ; 'T2 )]=2:
its Bruhat decomposition is
The result now follows from Theorem III-A.1, Theorem
III-B.4, and the bound (VI-C.1). This completes our proof of 1 b 1 0 0 1 1 0 b 1
0
the Almost Orthogonality Property, and of Theorem III-C.1. = 1+bc :
c 1 + bc b 1 1 0 b 1 0 b
(VI-E.1)
D. Derivation of Formula IV-A.1 This implies that for n 2 ZN ; we have
h i
We have for the line L0 , i.e., for the operators ( ; 0); 2 ' g [n] = (g)' [n]
ZN ; the following orthonormal basis of eigensequences:
= Cb m 1+bc DF T mb hb ' [n];
1 b
BL0 = ff0;b [n] = p e(bn) ; b 2 ZN g:
N where, in the second equality we use identity (VI-E.1), the fact
Let us derive formulas for basis parametrized by a line with that is homomorphism, and the Formulas (II-B.4), (II-B.5),
finite slope. From Lemma VI-A.2, we know that the Weil (II-B.6). This completes our verification of Formula (IV-B.6).
operator (uc ) associated with the unipotent element
DELAY-DOPPLER CHANNEL ESTIMATION WITH ALMOST LINEAR COMPLEXITY — BY FISH, GUREVICH, HADANI, SAYEED, AND SCHWARTZ 13

F. Verification of Formula (V-.3) [7] Gurevich S., Hadani R., and Sochen N., The finite harmonic oscillator
and its associated sequences. PNAS, July 22, 2008 vol. 105 no. 29 9869–
We verify Formula (V-.3) for the matched filter M('; ); 9873.
'; 2 H, restricted to a line with finite slope. We define [8] Gurevich S., Hadani R., and Sochen N., The finite harmonic oscillator
M ('; )[ ; !] = h'; ( ; !) i, where ( ; !) are the and its applications to sequences, communication and radar . IEEE
Transactions on Information Theory, vol. 54, no. 9, September 2008.
Heisenberg operators (II-A.1). We note that [9] Howe R., Nice error bases, mutually unbiased bases, induced represen-
1
tations, the Heisenberg group and finite geometries. Indag. Math. (N.S.)
M ('; )[ ; !] = e(2 !)M('; )[ ; !]: (VI-F.1) 16 (2005), no. 3–4, 553–583.
[10] Howard S. D., Calderbank, R., and Moran W., The finite Heisenberg–
The element Weyl groups in radar and communications. EURASIP J. Appl. Signal
1 0 Process (2006).
u c = 2 G; [11] O’Toole J.M., Mesbah M., and Boashash B., Accurate and efficient im-
c 1
plementation of the time–frequency matched filter. IET Signal Process.,
satisfies 2010, Vol. 4, Iss. 4, pp. 428–437.
u c (1; c) = (1; 0); [12] Rader C. M., Discrete Fourier transforms when the number of data
(VI-F.2) samples is prime. Proc. IEEE 56, 1107–1108 (1968).
u c (0; !) = (0; !): [13] Rickard S., Balan R., Poor H.V., and Verdu S., Canonical Time-
Frequency, time-scale and frequency-scale representations of time-
For a fixed ! 2 ZN we compute the matched filter on L0c = varying channels. Communications in Information and Systems, vol. 5,
Lc + (0; !) = f (1; c) + (0; !); 2 ZN g: We obtain no. 1, pp. 197-226, 2005.
[14] Sayeed A., Sparse and Multipath Wireless Channels: Modeling and
M ('; )[ (1; c) + (0; !)] Implications. ASAP (2006).
[15] Sayeed, A.M., and Aazhang B., Joint Multipath-Doppler Diversity in
= h'; [ (1; c) + (0; !)] i Mobile Wireless Communications. IEEE Transactions on Communica-
= h (u tions, pp. 123-132, January 1999.
c )'; (u c) [ (1; c) + (0; !)] i
[16] Sayeed A., and Sivanadyan T., Wireless Communication and Sensing
= h (u c )'; ( ; !) (u c) i in Multipath Environments Using Multiantenna Transceivers. Handbook
1 on Array Processing and Sensor Networks, S. Haykin and K.J.R. Liu,
= m c;0 ('); e( 2 ! + !n)m c;0 ( )[n ] Eds, Wiley 2010.
1 [17] Tse D., and Viswanath P., Fundamentals of Wireless Communication.
= e(2 !) m c; ! (') mc;0 ( )_ [ ]; Cambridge University Press (2005).
[18] Verdu S., Multiuser Detection, Cambridge University Press (1998).
where, the second equality is by the unitarity of ; the third [19] Wang Z., and Gong G., New Sequences Design From Weil Represen-
equality is by Identities (II-B.2), (VI-F.2), the forth equality tation With Low Two-Dimensional Correlation in Both Time and Phase
is by Formula (II-B.5) and the definition (V-.1), and the last Shifts. IEEE Transactions on Information Theory, vol. 57, no. 7, July
2011.
equality is by definition (V-.2) of : Using Identity (VI-F.1), [20] Weil A., Sur certains groupes d’operateurs unitaires. Acta Math. 111,
we obtain Formula (V-.3). 143-211 (1964).

Acknowledgement. Warm thanks to J. Bernstein for his

support and encouragement in interdisciplinary research. We
are grateful to J. Lima, and A. Sahai, for sharing with us their
thoughts, and ideas on many aspects of signal processing and
wireless communication. The project described in this paper
was initiated by a question of M. Goresky and A. Klapper
during the conference SETA2008; we thank them very much.
We appreciate the support and encouragement of N. Boston,
R. Calderbank, S. Golomb, G. Gong, O. Holtz, R. Howe, P.
Sarnak, N. Sochen, D. Tse, and A. Weinstein. Finally, we
thank the referees for many important remarks. The research
reported in this paper was partially supported by NSF Grants
DMS-1101660, and DMS-1101698 - "The Heisenberg–Weil
Symmetries, their Geometrization and Applications".

R EFERENCES
[1] Agarwal N. et. al., Algorithms for GPS Operation Indoors and Down-
town. GPS Solutions 6, 149-160, 2002.
[2] Artin M., Algebra. Prentice Hall, Inc., Englewood Cliffs, NJ (1991)
[3] Fish A., Gurevich S., Hadani R., Sayeed A., and Schwartz O., Delay-
Doppler Channel Estimation with Almost Linear Complexity. Proceed-
ings of ISIT 2012 - IEEE International Symposium on Information
Theory, Cambridge, MA, USA, July 1–6, 2012.
[4] Gerardin P., Weil representations associated to finite fields. J. Algebra
46 (1977), 54–101.
[5] Golomb, S.W., and Gong G., Signal design for good correlation. For
wireless communication, cryptography, and radar. Cambridge University
Press, Cambridge (2005).
[6] Gurevich S., Hadani R., and Howe R., Quadratic Reciprocity and the
Sign of Gauss Sum via the Finite Weil Representation . International
Mathematics Research Notices, Vol. 2010, No. 19, pp. 3729–3745.