equal-norm tight frames with erasures
TRANSCRIPT
Uniform Tight Frames with Erasures
Peter G� Casazza
Department of Mathematics� University of Missouri�Columbia
Columbia� MO �����
E�mail� pete�math�missouri�edu
and
Jelena Kova�cevi�c
Mathematics of Communication Research� Bell Labs� Lucent Technologies�
��� Mountain Avenue� Murray Hill� NJ ���� USA
E�mail� jelena�bell�labs�com
Uniform tight frames have been shown to be useful for robust data trans�
mission� The losses in the network are modeled as erasures of transmitted
frame coe�cients� We give the �rst systematic study of the general class
of uniform tight frames and their properties� We search for e�cient con�
structions of such frames� We show that the only uniform tight frames with
the group structure and one or two generators are the generalized harmonic
frames� Finally� we give a complete classi�cation of frames in terms of their
robustness to erasures�
�� INTRODUCTION
Frames are redundant sets of vectors in a Hilbert space which yield one natural
representation for each vector in the space� but which may have in�nitely many
di�erent representations for a given vector ��� �� � � �� �� �� � � �� ��� ����
Frames have been used in signal processing because of their resilience to additive
noise � ��� resilience to quantization � ��� as well as their numerical stability of
reconstruction � ��� and greater freedom to capture signal characteristics ��� ��� Re�
cently� several new applications for �uniform tight� frames have been developed�
The �rst� developed by Goyal� Kova�cevi�c and Vetterli � �� ��� ��� ��� uses the
redundancy of a frame to mitigate the e�ect of losses in packet�based communica�
tion systems� Modern communication networks transport packets of data from a
Some of the results in this paper were reported at the SPIE Conf� on Wavelet Appl� in Signaland Image Proc� �San Diego� CA� July ������ The �rst author was supported by NSF DMS�������
�
� CASAZZA AND KOVACEVI�C
source to a recipient� These packets are sequences of information bits of a certain
length surrounded by error�control� addressing� and timing information that assure
that the packet is delivered without errors� This is accomplished by not delivering
the packet if it contains errors� Failures here are due primarily to bu�er over�ows
at intermediate nodes in the network� So to most users� the behavior of a packet
network is not characterized by random loss� but by unpredictable transport time�
This is due to a protocol� invisible to the user� that retransmits lost packets� Re�
transmission of packets takes much longer than the original transmission� In many
applications� retransmission of lost packets is not feasible and the potential for large
delay is unacceptable�
If a lost packet is independent of the other transmitted data� then the information
is truly lost to the receiver� However� if there are dependencies between transmitted
packets� one could have partial or complete recovery despite losses� This leads us
naturally to use frames for encoding� The question� however� is� What are the best
frames for this purpose� With an additive noise model for quantization� in ����� the
authors show that a uniform frame minimizes mean�squared error if and only if it
is tight� So it is this class of frames � the uniform normalized tight frames � which
we seek to identify and study�
Another recent important application of uniform normalized tight frames is in
multiple�antenna code design � �� Much theoretical work has been done to show
that communication systems which employ multiple antennas can have very high
channel capacities � �� � �� These methods rely on the assumption that the receiver
knows the complex�valued Rayleigh fading coe�cients� To remove this assumption�
in � �� new classes of unitary space�time signals are proposed� If we have N trans�
mitting antennas and we transmit in blocks of M time samples �over which the
fading coe�cients are approximately constant�� then a constellation of K unitary
space�time signals is a �weighted bypM� collection of M � N complex matrices
f�kg for which ��k�k � I � The nth column of any �k contains the signal trans�
mitted on antenna n as a function of time� The only structure required in general
is the time�orthogonality of the signals�
Originally it was believed that designing such constellations was a too cumber�
some and di�cult optimization problem for practice� However� in � ��� it was shown
that constellations arising in a �systematic� fashion can be done with relatively
little e�ort� Systematic here means that we need to design high�rate space�time
constellations with low encoding and decoding complexity� It is known that full
transmitter diversity �that is� where the constellation is a set of unitary matrices
whose di�erences have nonzero determinant� is a desirable property for good per�
formance� In a tour�de�force� in � �� the authors used �xed�point�free groups and
their representations to design high�rate constellations with full diversity� More�
over� they classi�ed all full�diversity constellations that form a group� for all rates
and numbers of transmitting antennas�
For these applications� and a host of other applications in signal processing� it has
become important that we understand the class of uniform normalized tight frames�
In this paper� we make the �rst systematic study of such frames� In Section � we
review basic notions on frames� In particular� we state the Naimark�s Theorem
� � which has been rediscovered several times in recent years � � �� although it
has been used for several decades in operator theory� We give examples of uniform
UNIFORM TIGHT FRAMES WITH ERASURES �
normalized tight frames such as harmonic and Gabor frames� For harmonic frames�
we de�ne a more general class � general harmonic frames �GHF� � and study when
harmonic frames are equivalent to each other and general harmonic frames� In
Proposition ���� we give a simple equivalence condition on general harmonic frames
which states that the inner product between any two frame vectors �i and �j in a
GHF is equal to the inner product between �i�� and �j��� Section � concentrates
on uniform tight frames �UTF�� We start with a review in Section �� and proceed
with several ways of classifying UTFs including providing a correspondence between
subspaces of the original space and the UTFs� obtaining UTFs as alternate dual
frames for a given frame as well as �nding UTFs through frames equivalent to them�
In Section �� we shift our attention to UTFs with group structure� We show that
the UNTFs generated by the family fUk��gM��k�� �where U is unitary and �� �
H � are precisely the GHFs� We then extend our discussion to UNTFs generated
by fUkV j��g and higher numbers of generators� Finally� Section � introduceserasures modeled as losses of transform coe�cients hf� �ii� where f is the signalto be transmitted and f�igi�I is a set of frames vectors corresponding to erasedtransform coe�cients� We give a complete classi�cation of frames with respect
to their robustness to erasures� We study when we can obtain frames robust to
a certain number of erasures as a projection from another frame with a di�erent
number of erasures�
�� FRAME REVIEW
A set of vectors � � f�igi�I in a Hilbert space H � is called a frame if
� � Akxk� �Xi�I
jhx� �iij� � Bkxk� � ��� x �� �� � �
where I is the index set and the constants A�B are called frame bounds� Although
many of our results hold in more general settings� in this paper� we concentrate
mostly on the N �dimensional real or complex Hilbert spaces RN and CN �which
we denote HN � with the usual Euclidean inner product� When results generalize
to the in�nite�dimensional setting� we will point it out�
When A � B the frame is tight �TF�� If A � B � � the frame is normalized tight
�NTF�� A frame is uniform �UF� if all its elements have the same norm c�� k�ik � c�
For a uniform tight frame �UTF�� the frame bound A gives the redundancy ratio�
A UNTF of norm� vectors is an orthonormal basis �ONB��
The analysis frame operator� F maps the Hilbert space H into ���I�
�Fx�i � h�i� xi� ���
�Often� uniform is used to mean c � � here� however� we use the more general de�nition��F is sometimes called just frame operator� Here� we use analysis frame operator for F � synthesis
frame operator for F � and frame operator for F �F �
� CASAZZA AND KOVACEVI�C
for i � I � When H � HN � the analysis frame operator is an M �N matrix whose
rows are the transposed frame vectors ��i �
F �
�B�
���� � � � ���N��� � � �
���
��M� � � � ��MN
�CA � ���
We say that two frames f�igi�I and f�igi�I for H are equivalent if there is an
invertible operator L on H for which L�i � �i for all i � I � and they are unitarily
equivalent if L can be chosen to be a unitary operator� A direct calculation shows
that fS�����ig is a normalized tight frame for any frame f�ig� Thus� every frameis equivalent to a normalized tight frame� If f�igi�I is a frame with frame boundsA�B� and if P is an orthogonal projection on H � then by � � we have that fP�igi�Iis a frame for PH with frame bounds A and B� In particular� if f�igi�I is anormalized tight frame for H �for example� if it is an orthonormal basis for H ��
then fP�igi�I is a normalized tight frame for PH �The following theorem tells us that every normalized tight frame can be realized
as a projection of an orthonormal basis from a larger space� It serves as a converse
to the observation above that orthogonal projections of normalized tight frames
produce normalized tight frames for their span�
Theorem ��� �Naimark � �� Han Larson � ��� � A family f�igi�I in a Hilbert
space H is a normalized tight frame for H if and only if there is a larger Hilbert space
H � K and an orthonormal basis feigi�I for K so that the orthogonal projection P
of K onto H satis�es� Pei � �i� for all i � I�
Let us now go through certain important frame notions� Using the analysis frame
operator F � � � can be rewritten as
AI � F �F � BI� ���
We call S � F �F the frame operator� It follows that S is invertible �Lemma �����
in � ���� and furthermore
B��I � S�� � A��I� ���
It follows that f�igi�I is a normalized tight frame if and only if F is an isometry�
Also� S is a positive self�adjoint invertible operator on H and S � AI if and only
if the frame is tight� In �nite dimensions� the canonical dual frame of � is a frame
de�ned as !� � f !�igMi�� � fS���igMi��� where
!�i � S���i� ��
�This theorem has been rediscovered by several people in recent years� The �rst author �rstheard it from I� Daubechies in the mid����s� Han and Larson rediscovered it in ���� they came upwith the idea that a frame could be obtained by compressing a basis in a larger space and thatthe process is reversible �the statement in this paper is due to Han and Larson�� Finally� it waspointed out to the �rst author by E� Soljanin ���� that this is� in fact� the Naimark�s theorem�which has been widely known in operator theory and has been used in quantum theory�
UNIFORM TIGHT FRAMES WITH ERASURES �
for k � � � � � �M � Now�
f �Xi�Ihf� S���ii�i� for all f � H�
So the canonical dual frame can be used to reconstruct the elements of H from
the frame� However� there may be other sequences in H which give reconstruction�
This formula points out both the strengths and weaknesses of frames� First� we
see that every element f � H has at least one natural series representation in
terms of the frame elements� Also� this element may have in�nitely many other
representations� However� in order to �nd this natural representation of f � we
need to invert the frame operator� which may be di�cult or even impossible in
practice� The best frames then are clearly the tight frames since in this case the
frame operator becomes a multiple of the identity�
Noting that !��i � ��iS�� and stacking !���� !�
��� � � �� !�
�M in a matrix� the frame
operator associated with !� is
!F � FS��� ��
Since !F � !F � S��� ��� shows that B�� and A�� are frame bounds for !��Another important concept is that of a pseudo�inverse F y� It is the analysis
frame operator associated with the dual frame�
F y � !F �� ���
Note that for any matrix F with rows ��i
S � F �F �MXi��
�i��i � ���
This identity will prove to be useful in many proofs�
���� The Role of Eigenvalues
The product S � F �F will appear everywhere and its eigenstructure will play animportant role� Denote by �k �s the eigenvalues of S � F �F � We now summarizethe important eigenvalue properties�
General Frame� For any frame in HN � the sum of the eigenvalues of S � F �F �equals the sum of the lengths of the frame vectors�
NXk��
�k �
MXi��
k�ik�� � ��
Uniform Frame� For a uniform frame� that is� when k�ik � c� i � � � � � �M �
NXk��
�k �
MXi��
k�ik� � M � c�� � �
CASAZZA AND KOVACEVI�C
Tight Frame� Since tightness means A � B� for a TF� we have from � �
MXi��
jhf� �iij� � Akfk�� � ��
for all f � HN � Moreover� according to ���� a frame is a TF if and only if
F �F � A � IN � � ��
Thus� for a TF� all the eigenvalues of the frame operator S � F �F are equal to A�Then� using � ��� the sum of the eigenvalues of S � F �F is as follows�
N � A �NXk��
�k �MXi��
k�ik�� � ��
Normalized Tight Frame� If a frame is an NTF� that is� A � B � � then
MXi��
jhf� �iij� � kfk�� � ��
for all f � HN � In operator notation� a frame is an NTF if and only if
S � F �F � IN � � �
For an NTF� all the eigenvalues of S � F �F are equal to �Then� using � ��� the sum of the eigenvalues of S � F �F is as follows�
N �
NXk��
�k �
MXi��
k�ik�� � �
Uniform Tight Frame� From � � and � ��� we see that
N �A �
NXk��
�k �
MXi��
k�ik� � M � c�� � ��
Then� from � �� and � ���
MXi��
jhf� �iij� � M
Nc�kfk�� � ��
for all f � HN � The redundancy ratio is then
A �M
N� c�� ����
Since S � F �F � �M�N�I � the following is obvious�
MXi��
j�ik j� � M
N� �� �
UNIFORM TIGHT FRAMES WITH ERASURES
Uniform Normalized Tight Frame� If a frame is a UNTF� that is� we also ask
for A � B � � and if the frame vectors have norm c� then
N �
NXk��
�k �
MXi��
k�ik� � c�M�
Thus� a UNTF with norm c � is an orthonormal basis�
���� Examples of Uniform Normalized Tight Frames
We start with a simple example of a frame" � vectors in � dimension� The
particular frame we examine is termed Mercedes�Benz �MB� frame �for obvious
reasons� just draw the vectors�� The UNTF version of it is given by the analysis
frame operator
F �
r�
�
�� �
�p��� � ��p��� � ��
�A � ����
This is obviously a UNTF since �����F �F � I � We could make this frame just a
UTF with norm �as in ����� by having the analysis frame operator be simply F �
We now review two general classes of uniform tight frames which are commonly
used� The �general� harmonic frames and the tight Gabor frames� More general
classes of uniform tight frames are the full�diversity constellations that form a group
given in � ��
���� Harmonic Frames
Harmonic tight frames �HTF� are obtained by keeping the �rst N coordinates of
an M � M discrete Fourier transform basis as in Naimark�s Theorem �� � They
have been proven to be useful in applications � ���
An HTF is given by�
�k � �wk� � w
k� � � � � wk
N �� ����
for k � �� � � � �M � � where wi are distinct Mth roots of unity� This is a TF �F �F � MI � A more general de�nition of the harmonic frame �general harmonic
frame� is as follows�
Definition ���� Fix M � N � jcj � and fbigNi�� with jbij � �pM� Let fcigNi��
be distinct Mth roots of c� and for � � k �M � � let
�k � �ck�b�� ck�b�� � � � � c
kNbN ��
Then f�kgM��k�� is a uniform normalized tight frame for HN called a general har�
monic frame �GHF��
Note the di�erence between the HTFs and GHFs" the former ones are only tight�
while the latter ones are uniform� normalized and tight� We will now examine the
general harmonic frames in detail� Our goal is to determine when two such frames
are unitarily equivalent� We start by showing that the harmonic frames �not GHFs�
are unique up to a permutation of the orthonormal basis �that is� if and only if their
columns are permutations of each other��
� CASAZZA AND KOVACEVI�C
Proposition ���� Let f�kgM��k�� and f�kgM��
k�� be harmonic frames on HN �
Then f�kgM��k�� is equivalent to f�kgM��
k�� if and only if there is a permutation � of
f � �� � � � � Ng so that �kj � �k��j�� for all � � k �M� and all � j � N � Hence�
two harmonic frames are equivalent if and only if they are unitarily equivalent�
Proof �Proposition �� �� Let fejgNj�� be the natural orthonormal basis for HNand
�k � �wk� � w
k� � � � � wk
N � �
NXj��
wkj ej �
and
�k � �vk� � vk� � � � � � vkN � �
NXj��
vkj ej �
where wj � vj are sets of distinct Mth roots of unity� If
fwj j � j � Ng �� fvj j � j � Ng�
then� without loss of generality� we may assume that v� �� fwj j � j � Ng�Therefore�
M��Xk��
vk��k �
NXj��
�M��Xk��
�v�wj�k
�ej �
NXj��
� � ej � ��
On the other hand�
M��Xk��
vk��k� � �
M��Xk��
jvk� j� � M�
It follows that f�kgM��k�� is not equivalent to f�kgM��
k�� � The other direction is imme�
diate�
We now show that every general harmonic frame is unitarily equivalent to a
simple variation of a harmonic frame�
Proposition ���� Every general harmonic frame is unitarily equivalent to a
frame of the form fck�kgM��k�� � where jcj � and f�kgM��
k�� is a harmonic frame�
Proof �Proposition ����� Let M � N � jcj � and jbkj � �pM � for all � k �
N � Let fcjgNj�� be distinct Mth roots of c and let the GHF be
�k � �ck�b�� ck�b�� � � � � c
kNbN ��
for all � � k � M � � If c � ei�� let d � ei��M � Then there exist distinct Mth
roots of unity w�� w� � � � � wN with cj � ei��Mwj � dwj � For all sets of complex
numbers fakgM��k�� we have�
kM��Xk��
ak�kk� �NXj��
jM��Xk��
akckj bj j� �
M
NXj��
jM��Xk��
akckj j� �
UNIFORM TIGHT FRAMES WITH ERASURES �
M
NXj��
jM��Xk��
akeik��Mwk
j j� �
MkM��Xk��
akdk�kk��
where �k � �wk� � w
k� � � � � � w
kN �� Thus� the GHF f�kgM��
k�� is unitarily equivalent to�pMf�kgM��
k�� � with f�kgM��k�� an HTF�
Finally� we show that the frames given in Proposition ��� are not unitarily equiv�
alent to each other or to harmonic frames except in the trivial case�
Proposition ���� Let f�kgM��k�� and f�kgM��
k�� be harmonic frames and let jcj � � Then fck�kgM��
k�� is equivalent to f�kgM��k�� if and only if c is an M th root of
unity and there is a permutation � of f � �� � � � � Ng so that �kj � �k��j�� for all
� � k � M � and all � j � N � In particular� a general harmonic frame is
equivalent to a harmonic tight frame if and only if it equals a harmonic tight frame�
Proof �Proposition ����� Let �k �PN
j�� wkj ej � for all � � k �M� � If c � w��j �
for some j� we are done� Otherwise� sincePM��
k�� �k � �� if fck�kg is equivalent tof�kg then also
PM��k�� ck�k � �� Hence� for all � j � N we have
M��Xk��
ckwkj �
M��Xk��
�cwj�k �
� �cwj�M � cwj
� ��
Hence� �cwj�M � cMwM
j � cM � � The proposition now follows from Proposi�
tion �� �
We now have immediately�
Corollary ���� Let f�kgM��k�� and f�kgM��
k�� be harmonic frames and let jcj �jdj � � The frames fci�kgM��
k�� and fdi�kgM��k�� are equivalent if and only if c � d
and there is a permutation � of f � �� � � � � Ng so that �kj � �k��j�� for all � � k �M � �The next proposition gives a classi�cation of GHFs�
Proposition ���� A family f�igM��i�� in HN is a GHF if and only if for all
� � i� j �M � we have
h�i� �ji � h�i��� �j��i�
where �M � ���
Proof� If f�igM��i�� in HN is a generalized harmonic frame then there is a unitary
operator U on HN so that U i�� � �i� for all � � i � M � � Hence� for all� � i� j �M � we have
h�i��� �j��i � hU i����� Uj����i � hU i��� U
j��i � h�i� �ji�
� CASAZZA AND KOVACEVI�C
Conversely� given our equality� for any sequence of scalars faigM��i�� �
kM��Xi��
ai�i��k� � hM��Xi��
ai�i���M��Xi��
ai�i��i
�
M��Xi�j��
aiajh�i��� �j��i �M��Xi�j��
aiajh�i� �ji
� hM��Xi��
ai�i�
M��Xi��
ai�ii � kM��Xi��
ai�ik��
It follows that U�i � �i�� is a unitary operator and Ui�� � �i� for all � � i �M�
�
As we will see in the next section� UTFs with M � N � are all unitarily
equivalent� Thus� as a direct consequence of Theorem ���� any UTF withM � N�
is unitarily equivalent to the HTF with M � N � � This is a very useful result
since we have HTFs for any N and M " thus� for M � N � � we always have an
expression for all UTFs� For example� this means that the MB frame we introduced
earlier is equivalent to the HTF with M � � and N � ��
Another interesting property of an HTF is that it is the only UNTF such that
its elements are generated by a group of unitary operators with one generator� as
we will see in Section �� That is� � � f�igMi�� � fU i��gMi��� where U is a unitary
operator�
Moreover� HTFs have a very convenient property when it comes to erasures� We
can erase any e � �M �N� elements from the original frame and what is left is still
a frame �Theorem ��� from ������ We provide a complete classi�cation of frames in
terms of their robustness to erasures in Section ��
���� Gabor or Weyl�Heisenberg Frames
For the other general class of frames� we introduce two special operators on L��R��
Fix � � a� b and for f � L��R� de�ne translation by a as
Taf�t� � f�t� a��
and modulation by b as
Ebf�t� � e��imbtf�t��
Now� �x g � L��R�� If fEmbTnaggm�n�Z is a frame for L��R�� we call it a Gabor
frame �or a Weyl�Heisenberg frame�� It is clear that in this class the frames are
uniform� Also� since the frame operator S for a Gabor frame fEmbTnaggm�n�Zmust commute with translation and modulation� each Gabor frame is equivalent
to the �uniform� normalized tight Gabor frame fEmbTnaS����ggm�n�Z� For an
introduction to Gabor frames we refer the reader to �� and � ���
UNIFORM TIGHT FRAMES WITH ERASURES
�� UNIFORM TIGHT FRAMES
���� What is Known about UTFs�
As we have mentioned in the introduction� UNTFs have become popular in appli�
cations� In particular� in ����� the authors attack the problem of robust transmission
over the Internet by using frames� Being redundant sets of vectors� they provide
robustness to losses� which are modeled as erasures of certain frame coe�cients� It
is further shown in ���� that� assuming a particular quantization model� a uniform
frame with quantized coe�cients� minimizes mean�squared error if and only if it is
tight� Moreover� the same is true when we consider one erasure and look at both
the average� and worst�case MSE�
As a result� our aim is to construct useful families of frames for such applications�
To that end� it is important to classify uniform normalized tight frames in order to
facilitate the search for useful families�
Our notion of usefulness includes any families which would be computationally
e�cient" for example� those that can be obtained from one or more generating
vectors or those with a simple structure� for example� such that any of the frame
operators could be expressed as a product of sparse matrices�
Finally� we are interested in the robustness of our frames to coe�cient �frame
vector� erasures� These and other issues will be explored in the rest of the paper�
������ Construction of UTFs
There is a general well known method for getting �nite UNTFs� We state this
result here in its standard form and will give a new proof of the result in Section ��
In Section � we will also extend this result to uniform frames�
Theorem ���� There is a unique way to get UNTFs with M elements in HN �
Take any orthonormal set fkgNk�� in HM which has the property
NXk��
jkij� � c� for all i�
Thinking of the k as row vectors� switch to the M column vectors and divide bypc� This family is a uniform tight frame for HM with M elements� and all UNTFs
for HN with M elements are obtained in this way�
Corollary ���� If f�igMi�� is a uniform normalized tight frame for HN then
k�ik� � M
N�
There is a detailed discussion concerning the uniform normalized tight frames for
R� in ����� In � �� there is a deep classi�cation of groups of unitary operators which
generate uniform normalized tight frames� The simplest case of this is the harmonic
frames� In Section � we do not assume that we have a �group� of unitaries� but
instead conclude that our family of unitaries must be a group�
The picture becomes much more complicated if the uniform normalized tight
frame is generated by a group of unitaries with more than one generator �see � ��
� CASAZZA AND KOVACEVI�C
or worse� if the uniform tight frame comes from a subset of the elements of such a
group�
����� Classi�cation of UNTFs
Although we would like to classify all uniform tight frames� especially those which
can be obtained by reasonable algorithms� this is an impossible task in general� The
following theorem shows that every �nite family of norm� vectors in a Hilbert space
can be extended to become a uniform tight frame�
Theorem ���� If f�igMi�� is a family of norm� vectors in a Hilbert space H �
then there is a uniform tight frame for H which contains the family f�igMi���
Proof �Theorem ����� For each � i �M choose an orthonormal basis feijgj�Jfor H which contains the vector �i� Now the family feijgMi���j�J is made up ofnorm� vectors and for any f � H we have
MXi��
Xj�J
jhf� eijij� �MXi��
kfk� � Mkfk��
In the above construction we get as a tight frame bound the number of elements
M in the family f�igMi��� In general� this is the best possible� For example� just let�i � �j be norm� vectors for all � i� j �M � Then
MXj��
jh�i� �jij� � M�
Hence� any tight frame containing the family f�ig has the tight frame bound atleast M �
There is another general class of uniform tight frames �see ������ The result below
tells us that all UNTFs with M � N � are unitarily equivalent� It is a direct
consequence of Theorem �� from ���� where it is stated for UTFs with norm �
Theorem ��� �Goyal� Kova�cevi�c and Kelner ������ A family f�igN��i�� is a uni�
form normalized tight frame for HN if and only if f�igN��i�� is unitarily equivalent
to the frame fPeigN��i�� where feigN��
i�� is an orthonormal basis for HN�� and P
is the orthogonal projection of HN�� onto the orthogonal complement of the one�
dimensional subspace of HN�� spanned byPN��
i�� ei�
Since there are HTFs with N � �elements in HN � it follows that every UNTF
for HN with N � �elements is unitarily equivalent to a HTF�
���� Normalized and Uniform Tight Frames and Subspaces of the
Hilbert Space
Here� we begin to classify NTFs and UNTFs by providing a correspondence with
subspaces of the original Hilbert space� We give two results� one for NTFs and
another for UNTFs with the correspondence being one�to�one� The material in this
section grew out of conversations between the �rst author and V� Paulsen�
UNIFORM TIGHT FRAMES WITH ERASURES �
As we saw in Theorem �� � there is a unique way to get normalized tight frames
in HN with M elements� Namely� we take an orthonormal basis feigMi�� for HMand take the orthogonal projection PHN of HM onto HN � Then fPHN eigMi�� is anormalized tight frame for HN with M elements� In particular� there is a natural
one�to�one correspondence between the normalized tight frames for HN with M
elements and the orthonormal bases for HM � Then� the uniform normalized tight
frames for HN are the ones for which kPeik � kPejk� for all � i� j � M � We
use this to exhibit a natural correspondence between these families and certain
subspaces of HM � Here we treat two frames as the same if they are unitarily
equivalent�
Theorem ���� Let P be a rank�N orthogonal projection on HM and let feigMi��be an orthonormal basis for HM � There is a natural one�to�one correspondence
between the normalized tight frames for PHM with M elements and the family of
all N�dimensional subspaces of HM �
Proof �Theorem ����� If f�igMi�� is any normalized tight frame for PHM � then byNaimark�s Theorem� there is an orthonormal basis fe�
igMi�� for HM so that Pe�
i � �i�
De�ne a unitary operator U on HM by Uei � e�
i� Now� �i � PUei which is uni�
tarily equivalent to ��
i � U�PUei� So we will associate our normalized tight framef�ig with the subspace U�PUHM � Now we need to check that this correspon�dence is one�to�one� Let f�ig be another normalized tight frame for PHM which
is associated with the same subspace of HM � namely U�PUHM � Then there is an
orthonormal basis fe��
i g and a unitary operator V ei � e��
i with V �PV � U�PUand PV ei � e
��
i � Hence� V�PV ei � V ��i � U�PUei � U��i� This implies that
the normalized tight frames f�ig and f�ig are unitarily equivalent �and hence thesame�� Finally� we need to see that this correspondence covers all subspaces� If
W is any subspace of HM of dimension N � we can de�ne a unitary operator U
on HM so that UW � PHM � Then U�PU � PW while fPUeigMi�� is a nor�malized tight frame for PHM �which under our association corresponds to W ��
One of the consequences of the above result is that if we have one normalized
tight frame for HN � then all the others can be obtained from it by this process� We
now turn our attention to uniform frames�
Theorem ���� Fix an orthonormal basis feigMi�� for HM so that if P is the
orthogonal projection of HM onto HN then fPeigMi�� is a uniform normalized tight
frame for HN � Then there is a natural one�to�one correspondence between the
uniform normalized tight frames for HN with M elements and the subspaces W of
HM for which kPW eik� �M�N � for all � i �M �
Proof �Theorem ����� We already have our classi�cation of the normalized tight
frames in terms of all subspaces� Now we need to see which of these subspaces
correspond to the uniform normalized tight frames� Suppose U is any unitary
operator on H so that fPUeig is a uniform normalized tight frame for PHM � Thenthis frame is associated to the subspace W � U�PHM and PW � U�PU � Hence�PUei � UPW ei� for all i� But� fPUeig is a uniform normalized tight frame for
� CASAZZA AND KOVACEVI�C
PHM and so kPUeik � N�M � for all i� Hence�
kPUeik � N
M� kUPW eik � kPW eik�
So our association between normalized tight frames and subspaces given in The�
orem ��� identi�es the subspacesW of HM for which kPW eik � N�M � for all i�
���� Uniform Dual Frames
Another method for classifying UTFs uses a back door� get UTFs as alternate
dual frames for a given frame� To do this� we need the de�nition of alternate dual
frames�
Definition ���� Let f�igi�I be a frame for a Hilbert space H � A family f�igi�Iis called an alternate dual frame for f�igi�I if
f �Xi�Ihf� �ii�i� for all f � H �
In De�nition �� � if we let F� �respectively� F�� be the analysis frame operators
for f�ig �respectively f�ig�� then we see that F �� F� � I �
There are many alternate dual frames for a given frame� In fact� a frame has a
unique alternate dual frame �the canonical dual frame� if and only if it is a Riesz
basis � �� Moreover� no two distinct alternate dual frames for a given frame are
equivalent � �� For a normalized tight frame� its canonical dual frame is the frame
itself since !F � FS�� � F � I � F � Moreover�
Proposition ���� If f�igi�I is a normalized tight frame for HN � then the only
normalized tight alternate dual frame for f�igi�I is f�igi�I itself�
Proof �Proposition �� �� Let F� be the analysis frame operator for f�igi�I �Let f�igi�I be any normalized tight alternate dual frame for f�igi�I with analysisframe operator F�� It follows that F�� F� are isometries� Now�
F �� �F� � F�� � F �� F� � F �� F� � I � I � ��
Hence� for every f� g � H we have
hF �� g� �F� � F��fi � hg� F �� �F� � F��fi � hg� �i � ��
Since f�igi�I is an NTF� for every f � H we have
kfk� �MXi��
jhf� �iij� � kF�fk�� ����
We can further expand this as
kF�fk� � k�F� F��fk�� kF�fk� � k�F� � F��fk� � hF �� �F� � F��f� fi� �z �
�
� hF �� �F� � F��f� fi�� �z ��
�
UNIFORM TIGHT FRAMES WITH ERASURES �
However� since F� corresponds to an NTF as well�
kF�fk� � k�F� � F��fk� � kfk� � k�F� � F��fk�� ����
Equating ���� and ����� we get
k�F� � F��fk� � ��
Hence� F� � F� and so �i � �i� for all i � I �
If we ask for the dual frame just to be tight� but not normalized� then�
Proposition ���� If f�igMi�� is a normalized tight frame for HN and M � �N �
then the only tight dual frame for f�igMi�� is f�igMi�� itself� If M � �N then there
are in�nitely many �nonequivalent� tight alternate dual frames for f�igMi���
Proof �Proposition ����� With the notation of the proof of Proposition �� � the
only change is that� since the dual frame is only tight and not normalized tight�
there is a frame bound A � so that
kF�k� � Akfk�� for all f � H �
Hence� as in the proof of Proposition �� we have
kfk� � kF�fk� � kF�fk� � k�F� � F��fk� � Akfk� � k�F� � F�k��
Hence� k�F� � F��fk� � � �A�kfk� for all f � H � Hence� either F� � F� � � and
A � which gives us the NTF as in the previous proposition� �and so �i � �i for
all � i �M� or � �p �A��F� � F�� is an isometry on HN � In the latter case it
follows that dim �F�HN �� � N � and so M � �N � Thus� if M � �N � the only tight
dual frame is the frame itself�
On the other hand� ifM � �N given any F � HN �M� which is a constant times
an isometry� and with F�HN � FHN � F��F de�nes a tight frame which is an alter�
nate dual frame for f�ig�However� since it is really the uniform case we are interested in�
Proposition ���� If f�igMi�� is a uniform normalized tight frame for HN and
M � �N � then there are in�nitely many uniform tight alternate dual frames for
f�ig�Ni���
Proof �Proposition ����� Again we use the notation of the proof of Proposi�
tion �� � Choose any F� � HN �M� which is a constant multiple of an isometry
and with F�HN � �F�HN ��� Then F� � F� de�nes a tight alternate dual frame
for f�igMi��� say �i � �F �� � F �� �ei� for all � i � M � We just need to check
that this frame is uniform� Let P � �M� F�HN be the orthogonal projection�
so that Pei � F��i� for all � i � M � It follows that kPeik� � N�M and
k�I � P �eik� � �N�M � for all � i �M � Now�
�i � F �� ei � F �� ei � �i � F �� �I � P �ei�
CASAZZA AND KOVACEVI�C
Also� there is an a � so that for every � i �M we have
k�ik� � k�ik� � kF �� �I � P �eik� � N
M� ak�I � P �eik�
�N
M� a
� � N
M
� a� � � a�
N
M�
Hence� f�igMi�� is a uniform tight frame�It can be shown that the results of this section actually classify when NTFs or
UNTFs have tight �respectively uniform tight� alternate dual frames� that is� all
tight alternate dual frames for f�ig are obtained by the methods of this section� Wedo not know in general which frames �not tight� have tight �respectively uniform
tight� alternate dual frames�
���� Frames Equivalent to UTFs
Another approach to the classi�cation of UTFs looks into families of frames
obtained from a UTF by equivalence relations� For example� we know that every
frame is equivalent to a normalized tight frame� That is� given any frame f�igi�Iwith the frame operator S� the frame fS�����igi�I is a normalized tight framewhich is equivalent to f�igi�I � Therefore� it is natural to try to �nd ways to turnframes into uniform normalized tight frames� As it turns out� this is not possible
in most cases�
Theorem ���� If a frame f�igi�I with the frame operator S is equivalent to
a UTF� then fS�����igi�I is a UNTF� In particular� a tight frame which is not
uniform cannot be equivalent to any UNTF�
Proof �Theorem ���� It is known that fS�����kg is a normalized tight framewhich is equivalent to f�kg� So if f�kg is equivalent to a uniform tight frame� sayf�kg� and k�kk � c� for all k� then fS�����kg is equivalent to f�kg� That is� thereis an invertible operator T on H so that TS�����k � �k�
Now we show that T�pA is a unitary operator� Let A be the tight frame constant
of fTS�����kg� Then for all f � H � the tightness of fTS�����kg implies
Akfk� �Xk
jhf� TS�����kij� �Xk
jhT �f� S�����kij�� ���
However� since fS�����kg is an NTF� then this leads us toXk
jhT �f� S�����kij� � kT �fk�� ���
Equating ��� and ���� we get that
Akfk� � kT �fk��
that is� T�pA is unitary� Hence� the frame
S�����k � T���k�
UNIFORM TIGHT FRAMES WITH ERASURES
is a UNTF since
kT���kk � pAk�kk � p
Ac� for all k � � � � � �M�
�� UNIFORM TIGHT FRAMES WITH GROUP STRUCTURE
In this section� we examine frames which are generated by a group of unitary
operators applied to a �xed vector in the Hilbert space� These classes are espe�
cially important in applications� The traditional use of one of these classes is in the
Gabor frames used in signal#image processing� These are groups with two genera�
tors� There is also a class of wavelet frames which has two generators ���� Recently�
several new applications have arisen� In ����� the authors proposed using the redun�
dancy of frames to mitigate the losses in packet�based communication systems such
as the Internet� In � �� frames are used for multiple antenna coding#decoding �see
the discussion Section ��� for further explanation�� In ���� ��� connections between
quantum mechanics and tight frames are established� Although each of these ap�
plications requires a di�erent class of UNTFs� they all have one important common
constraint� Namely� calculations for the frame must be easily implementable on
the computer� Since UNTFs generated by a group of unitary operators satisfy this
constraint� this class is one of the most important to understand�
In this section we will show that the UNTFs generated by the family fUk��gM��k��
�where U is unitary and �� � H � are precisely the GHFs� We then extend our
discussion to UNTFs generated by fUkV j��g and higher numbers of generators�In the discussion Section ��� we will relate the results of this section to the literature�
���� UTFs with a Single Generator
Here� we give a complete classi�cation of UNTFs of the form fUk��gM��k�� where
U is a unitary operator on HN � We will� in fact� �nd that in this case fUkgM��k��
must be a group� Moreover� we will see the the GHFs will be that special class�
Since the proofs of the following results are long� the reader can �nd them in the
Appendix� Our �rst result classi�es GHFs as those frames generated by powers of
a special class of unitary operators applied to a �xed vector in H �
Proposition ���� f�kgM��k�� is a general harmonic frame for HN if and only if
there is a vector �� � HN with k��k� � NM � an orthonormal basis feigNi�� for HN
and a unitary operator U on HN with Uei � ciei� with fcigNi�� distinct M th roots
of some jcj � so that �k � Uk��� for all � � k �M � �
Proof� See Appendix A� �
We now show that the only UNTFs with group structure and a single generator
are GHFs� This result is an excellent result and a disappointment at the same time�
On the one hand� the fact that GHFs are the only ones with such a group structure
proves once more why they are so universally used� It also spares us the trouble
of looking for other such UNTFs� On the other hand� we cannot �nd any other
useful families via this route as we were hoping for� Another important property of
� CASAZZA AND KOVACEVI�C
GHFs �Theorem ��� from ����� is that GHFs are robust to the maximum number
of possible erasures� that is� a GHF with M elements in HN is robust to e erasures
for any e � �M �N��
Our next theorem completes the classi�cation of GHFs as being those UNTFs of
the form fUk��gM��k�� where U is a unitary operator on H � As a consequence� we
discover that� in this case� fUkgM��k�� is a group�
Theorem ���� Let U be a unitary operator on HN � �� � HN and assume
fUk��gM��k�� is a UNTF for HN � Then UM � cI for some jcj � and fUk��gM��
k��
is a general harmonic frame� That is� the GHFs are the only UNTFs generated by
a group of unitary operators with a single generator�
Proof� See Appendix A���
���� UTFs with Two or More Generators
Having completely characterized UTFs which come from a group of unitary oper�
ators with one generator� we now turn our attention to those with more than one�
The fundamental examples of frames with two generators are the Gabor frames
fEmbTmaggm�n�Z �or� in our case� the �nite discrete Gabor frames�� Each of theseframes is equivalent to the UNTF Gabor frame fEmbTnaS
��ggm�n�Z where S isthe frame operator for fEmbTmaggm�n�Z�We will classify the UNTFs of the form fU iV j��gL���M��
i���j�� where fV j��gM��j�� is
a UNTF for its span� In the process of proving these results� we introduce a general
method for using special families of �nite�rank projections to produce frames for
a space� If we associate a frame f�igMi�� with the family of rank� orthogonalprojections Pi taking H onto span �i� then we can view the results of this section
as a generalization of the very notation of a frame to the case where the Pi have
arbitrary rank�
Our �rst result states that given a GHF generated by a unitary operator V � a
unitary operator U and an integer M � we can always �nd a corresponding UNTF
generated by U� V �
Proposition ���� Let V be a unitary operator on HN � let m � N and �� � H
so that fV i��gm��i�� is a UNTF for HN � Then for every unitary operator U on HN
and every M � N� there is a vector �� � HN such that fUkV j��gM���m��k���j�� is a
uniform� normalized tight frame for HN �
Proof �Proposition ����� Let �� � ���pM � Now� for every f � HN we have
M��Xk��
m��Xj��
jhf� UkV j��ij� �M��Xk��
m��Xj��
jhU�kf� V j��ij� �
M
M��Xk��
m��Xj��
jhU�kf� V j��ij��
Since fV j��gm��j�� is a UNTF� the previous sum equals
M
M��Xk��
kU�kfk� �
M
M��Xk��
kfk� � kfk��
UNIFORM TIGHT FRAMES WITH ERASURES �
It is easily seen that an invertible operator on H maps a frame for H to another
frame for H � We next observe that there is a natural relationship between the frame
operators of these equivalent frames�
Proposition ���� If f�kg is a frame for H with the frame operator S� and T is
an invertible operator on H � then TST � is the frame operator for the frame fT�kg�In particular� if f�kg is a UNTF then every frame unitarily equivalent to f�kg is
also a UNTF�
Proof �Proposition ����� Let L be the frame operator for fT�kg� For any f � H
we have�
Lf �Xk
hf� T�kiT�k � T
�Xk
hT �f� �ki�k�� T �S�T ��f ��
As a consequence of Proposition ���� we can apply a unitary operator to any
UNTF and get a new UNTF� If this unitary operator is part of the generating
set for our UNTF� then our new UNTF has a special form as given in the next
proposition�
Proposition ���� If U is unitary and fUkV j��g is a UNTF for HN � then for
every M � Z and for every f � HN we have�
f �Xk�j
hf� Uk�MV j��iUk�MV j���
that is� fUk�MV j��gk�j is a UNTF as well�
Proof �Proposition ����� Since UM is unitary and fUkV j��gk�j is a UNTF forHN � it follows that fUMUkV j��gk�j is also a UNTF for HN �To explain where we are going from here� let us revisit the notion of a UNTF
and look at it from a slightly di�erent point of view� Suppose f�igi�I is a UNTFfor a ��nite or in�nite�dimensional� Hilbert space H with k�ik � c� for all i � I �
For each i � I � let Pi be the orthogonal projection of H onto the one�dimensional
subspace of H spanned by �i� Then� for all f � HXi�I
kPifk� �Xi�I
khf� �ii�ik�
�Xi�I
jhf� �iij�k�ik� � c�Xi�I
jhf� �iij� � c�kfk��
Conversely� let fPigi�I be rank� orthogonal projections on H satisfyingXi�I
kPifk� � akfk�� for all f � H �
�� CASAZZA AND KOVACEVI�C
Then f�igi�I is a UNTF for H where �i � PiH and k�ik� � �a� That is� if f � H
then Xi�I
jhf� �iij� �Xi�I
jhPif� �iij�Xi�I
kPifk�k�ik�
�
a
Xi�I
kPifk� �
aakfk� � kfk��
It follows that the condition on f�ig which guarantees that we have a UNTF isreally a condition on the rank� orthogonal projections of H onto the span of �i�
It is this which we now generalize to �nd our next general class of UNTFs� Since
this material is also new for general frames� we start here�
Proposition ���� Let fPigi�I be a family of projections �of any rank� on a
��nite or in�nite�dimensional� Hilbert space H and assume there are constants � �
A�B satisfying�
Akfk� �Xi�I
kPifk� � Bkfk�� for all f � H �
Let f�ijgMi
j�� be a frame for PiH with frame bounds Ai� Bi� for all i � I� Then
f�ijgMi
j���i�I has frame bounds�
A infi�IAi� B supi�IBi�
Proof� For any f � H we have
MiXj���i�I
jhf� �ijij� �Xi�I
MiXj��
jhf� �ijij�
�Xi�I
MiXi��
jhPif� �ijij� �Xi�I
BikPifk�
� supi�IBi
Xi�I
kPifk� � B �supi�IBi� kfk��
The lower frame bound is computed similarly�
The cases of interest to us are the next two corollaries� the �rst for NTFs and
the second for UNTFs�
Corollary ���� Let fPigi�I be a family of projections on a Hilbert space H
satisfying� Xi�I
kPifk� � a�kfk�� for all f � H �
Then a � � and if f�ijgMi
j�� is an NTF for PiH � for all i � I� then f �a�ijgMi
j���i�Iis an NTF for H �
UNIFORM TIGHT FRAMES WITH ERASURES �
In the next corollary we assume that the projections Pi all have the same rank
and the frames for PiH all have the same number of elements� This assumption
is necessary to guarantee that our frame is uniform� That is� as we observed in
Section �� � if dim PiH � N and f�i�jgj�J is a UNTF for PiH then
k�ijk� � N
jJ j �
Corollary ���� Let fPigi�I be a sequence of projections �all with the same
rank� on a Hilbert space H satisfying�Xi�I
kPifk� � a�kfk�� for all f � H �
Then a � � and if f�ijgj�J is a UNTF for PiH � for all i � I� then f �a�ijgj�J�i�Iis a UNTF for H �
We are now ready to consider UNTFs with two generators�
Theorem ���� Let U be a unitary operator on H � and let fV j�gM��j�� be a UTF
�with tight frame bound �a� for its closed linear span in H � Let P� be the orthogonal
projection of H onto this span� Assume that fU iV j��gL���M��i���j�� is a UNTF for H �
If Pi is the orthogonal projection of H onto the span of U iP�H � for all � i � L� �then
L��Xi��
kPifk� � akfk�� for all f � H �
Proof� Note that for all � i � K � we have Pi � U iP�U�i� Now� for all
f � H we compute�
L��Xi��
kPifk� �L��Xi��
kU iP�U�ifk�
�L��Xi��
kP�U�ifk� �L��Xi��
aM��Xj��
jhU�if� V j��ij�
� a
K��Xi��
M��Xj��
jhf� U iV j��ij� � akfk��
The next theorem is the converse of the above�
Theorem ���� Let fPigL��i�� be a family of orthogonal projections on H satisfying
L��Xi��
kPifk� � a�kfk�� for all f � H �
Assume that U is a unitary operator on H so that U iP�H � PiH � Let �� � P�H
and let V be a unitary operator on P�H such that fV j��gM��j�� is a UNTF for P�H �
Then f �aU iV j��gL���M��i���j�� is a UNTF for H �
�� CASAZZA AND KOVACEVI�C
Proof� For all f � H we have�
L��Xi��
M��Xj��
jhf� aU iV j��ij� �
a�
L��Xi��
M��Xj��
jhU�if� V j��ij� �
a�
L��Xi��
kP�U�ifk� �
a�
L��Xi��
kU iP�U�ifk� �
a�
L��Xi��
kPifk� � kfk��
���� Discussion
Some remarks are in order�
� The assumptions in Theorems ��� and ��� that fV j��gM��j�� is a UNTF for its
span is su�cient for the conclusion of the theorems but not necessary� For exam�
ple� it is known to Gabor frame specialists that for ab � � there are Gabor frames
fEmbTnaggm�n�Z for which neither of fEmbTnaggm�Znor fEmbTnaggn�Z is a framefor its closed linear span� The equivalent UNTF Gabor frame fEmbTnaS
����ggm�n�Zfails the hypotheses of Theorems ��� and ��� but satis�es the conclusion�
�� The results of Section ��� generalize to three or more generators� For example�
for three generators� Theorem ��� becomes�
Theorem ���� Let U� V be unitary operators with fU iV j��gL���M��i���j�� a UNTF for
its span� Let P� be the projection of the Hilbert space H onto this span� Let W be
a unitary operator on H so that
K��Xk��
kW kP�W�kfk� � a�kfk�� for all f � H �
Then f �aW kU iV j��gK���L���M��k���i���j�� is a UNTF for H �
�� It would be very interesting to classify when a �nite group of unitaries G
generates a UNTF for H of the form fU��gU�G� Since �nite Abelian groups areisomorphic to a direct product of cyclic groups� the results of this section give
su�cient conditions for fU��gU�G to generate a UNTF for H � Frames generatedby Abelian groups of unitaries were studied in ��� where they are called geometrically
uniform frames �GU frames�� It is shown there that the canonical dual frame of
a GU frame is also a GU frame and that the equivalent NTF frame is also GU�
Since GU frames have strong symmetry properties� they are particularly useful
in applications� In ���� it is further shown that the frame bounds resulting from
removing a single vector of a GU frame are the same regardless of the particular
vector removed� This result for HTFs was observed in �����
�� Frames generated by possibly noncommutative groups of unitaries are impor�
tant in multiple antenna coding and decoding � �� Here� one needs classes of
UNIFORM TIGHT FRAMES WITH ERASURES ��
unitary space�time signals �that is� frames generated by classes of unitary opera�
tors� called constellations� It is also desirable in this setting to have full transmitter
diversity� meaning that
det �I � U� �� �� for all I �� U � G�
Groups with this property are called �xed�point free groups� In a tour�de�force�
the authors in � � classify all full�diversity constellations that form a group� for
all rates and numbers of transmitting antennas� Along the way they correct some
errors in the classi�cation theory of �xed point free groups�
�� CLASSIFICATION OF UNIFORM TIGHT FRAMES WITH
ERASURES
As mentioned in the introduction� one of the main applications that motivated us
to examine uniform tight frames is that of robust data transmission� This means
that at some point� the system experiences losses� These losses are modeled as
erasures of transmitted frame coe�cients� Since at the receiver side� this looks like
the original frame was the one without vectors corresponding to erased coe�cients�
we examine the structure of our frames after losses�
The �rst question is whether after erasures what we have is still a frame� For
the MB frame� if we erase any one element� the remaining two are enough for
reconstruction� However� if we erase any two� we have lost one subspace and recon�
struction is not possible� We can provide more robustness by adding more vectors�
For example� take a uniform tight frame in R� made up of two orthonormal bases
� � ��� ��� �� �� � ��� ���� �� This frame is robust to any one erasure" what is left
contains at least one of the two bases� However� if two elements are erased� the
situation is not that clear anymore� We could erase coe�cients corresponding to
one of the bases and still have a basis able to reconstruct� If� on the other hand�
we lose coe�cients corresponding to two collinear vectors� we are stuck" we have
lost one entire subspace and cannot reconstruct� We could� however� take another
UNTF with M � � vectors in R� � the HTF with the analysis frame operator
F �
�BB�
�p���
p���
�
�p��� p���
�CCA � ����
This frame is also made up of two orthonormal bases� However� this frame is robust
to any two erasures� It is thus more robust than the MB frame" we pay for this
added bene�t with more redundancy �more vectors�� It is also more robust than
the previous frame with � vectors�
These simple examples demonstrate the types of questions we will be asking in
this section� One might think that starting with a tight frame might provide some
resilience to losses �as opposed to starting with a general frame�� As we have seen
in our example� this is not the case� Thus� we are searching for frames robust to
a certain number of erasures� That is� frames which remain frames after erasures�
There are frames on HN with M elements which are robust to M � N erasures�
�� CASAZZA AND KOVACEVI�C
namely the HTFs �Theorem ��� from ������ Such frames are the best� This is clearly
optimal since any more erasures would not leave enough elements to span HN � In
this section� we will characterize frames based on their robustness to erasures�
���� UTFs and UNTFs Robust to One Erasure
In this section we will consider UNTFs which are robust to one erasure� This
contains the basic idea for the general case�
We start with the notion of a frame robust to k erasures� or a k�robust frame�
Definition ���� A frame f�igMi�� is said to be robust to k erasures if f�igi�Icis still a frame� for I any index set of k erasures� I � f � �� � � �Mg and jI j � k�
The following property tells us we do not destroy the robustness of the frame
by projection� This observation lead us to the idea that we could classify frames
by starting from a large space and �step down�using projections� For example� we
could start with a frame robust to one erasure and step down to frames �hopefully�
robust to two erasures� then once more to frames robust to three erasures and so
on� Although this will not be the case� the idea of stepping down lead us to the
results in this section�
Note that it is clear from the de�nition of a frame that if we apply an orthogonal
projection P on H to a frame we will get a frame for PH with the same frame
bounds�
Proposition ���� Let f�igMi�� be a frame in HN robust to k erasures and let P
be an orthogonal projection on HN � Then fP�igMi�� is a frame for PHN robust to
k erasures�
Proof� Let I be the index set of erasures� I � f � �� � � �Mg with jI j � k� Now�
f�igi�Ic spans HN and so fP�igi�Ic is a frame for PHN with the same frame
bounds�
The main ingredient for classifying frames robust to one erasure is contained in
the next proposition�
Proposition ���� Let f�igMi�� be a set of vectors in HN � The following are
equivalent�
�f�igMi�� is a frame robust to one erasure�
��There are nonzero scalars ai �� �� for � i �M so that
MXi��
ai�i � ��
Proof� � �� ���� Choose I � f � �� � � �Mg maximal for which there are nonzeroai�s� i � I and
g �Xi�I
ai�i � ��
UNIFORM TIGHT FRAMES WITH ERASURES ��
We claim that jI j � M � We proceed by contradiction� If I �� f � �� � � �Mg� choosem � Ic� Since f�igMi�� is robust to one erasure� there are scalars fbig� not all zero�so that if �m is erased� it can be recovered from the rest as
�m �Xi��m
bi�i�
or
h � �m �Xi�I
bi�i �X
m��i�Icbi�i � ��
We have two cases�
Case I� Assume that bi � � for all i � I �
Then� h � �m �Pm��i�Ic bi�i � �� In this case� g � h � � and has nonzero
coe�cients on every �i� i � I � plus a nonzero coe�cient on �m contradicting the
maximality of I �
Case II� At least one bi �� � for some i � I �
Since ai �� � for all i � I � we can choose an � � so that
� �� ai�bi� for all i � I�
Now� g � �h � � and has nonzero coordinates on �i� for all i � I � as well as � for a
coordinate on �m� again contradicting the maximality of I �
���� � �� Assume ai �� �� for all � i �M andXi�I
ai�i � ��
Then for each m � I we have�
�m � �X
m��i�I
aiam
�i�
that is� any vector lost can be recovered using the rest and so f�igMi�� is robust to theerasure of �m� for an arbitrary m � I �
Note that to construct a frame guaranteed not to be robust to one erasure� it is
enough to put one vector� say �M � orthogonal to the span of the rest� Erasing that
vector destroys the frame property� Thus� we cannot �nd a nonzero coe�cient aMsuch that
PMi�� ai�i � ��
For example� we know that the MB frame f�ig�i�� is robust to one erasure �The�orem �� from ������ Hence� we can �nd a� � a� � a� � such that
P�i�� ai�i � ��
The next results will characterize frames robust to one erasure� Since we men�
tioned the idea of stepping down� we will start from an orthonormal basis feigMi��for HM and project it using a projection operator P �or I �P �� We thus look into
a few facts connecting P and �I � P ��
� If P is an orthogonal projection� then �I � P � is an orthogonal projection as
well�
� CASAZZA AND KOVACEVI�C
�� The subspaces P and �I � P � project onto are orthogonal� For any f � H �
h�I � P �f� Pfi � hP ��I � P �f� fi � h�P � P �f� fi � ��
�� f�I � P �eigMi�� is uniform if and only if fPeigMi�� is uniform� To see this wecompute� for all i� j
h�I � P �ei� �I � P �eii � h�I � P �ej � �I � P �ejih�I � P ���I � P �ei� eii � h�I � P ���I � P �ej � eji
h�I � P �ei� eii � h�I � P �ej � eji � hPei� eii � � hPej � ejihP �Pei� eii � hP �Pej � ejihPei� P eii � hPej � P eji�
Here we used the fact that P is an orthogonal projection and thus P � P � � P �P ��� Both fPeigMi�� and f�I � P �eigMi�� are NTFs�
By Theorem �� � all NTFs are of the form fPeigMi��� where feigMi�� is an or�thonormal basis for H and P is an orthogonal projection on H � The UNTFs form
a subclass of these frames� In the rest of this paper we will identify this subclass�
So we will be working with frames of the form above� We will discover that the
subspace PH determines when the frame is a UNTF� From our earlier discussion� if
P is an orthogonal projection on H � we may work either with fPeig or f�I �P �eigto classify UNTFs� We will freely switch between these classes because each of
them has certain advantages for speci�c results� The class fPeig works well forclassi�cations when the dimension of PH is �large� relative to the dimension of H �
On the other hand� f�I�P �eig is easier to work with when the dimension of PH is�small� relative to the dimension of H � Also� as we will see� important information
about fPeig is often contained in �I � P �H �and vice�versa�� So we will need to
bring both of these into our discussion�
Now we give the general classi�cation for frames robust to one erasure�
Corollary ���� Let feigMi�� be an orthonormal basis for HM and let P be an
orthogonal projection on HM � The following are equivalent�
�f�I � P �eigMi�� is a frame robust to one erasure�
��There are nonzero scalars ai �� �� for � i �M so that
MXi��
ai�I � P �ei � ��
��There is an f � span��i�MPei so that
hf� eii �� �� for all � i �M�
UNIFORM TIGHT FRAMES WITH ERASURES �
Proof� The equivalence of � � and ��� comes from Proposition ���� Let us check
the equivalence of ��� and ����
���� ���� Given ���� choose f to be
f �
MXi��
aiei�
By our assumption�PM
i�� aiei �PM
i�� aiPei and thus
f �
MXi��
aiPei � PHM �
Moreover� hf� eii � ai �� �� for all � i �M �
���� ���� Choose f � PHM as in ��� and call ai � hf� eii �� �� for all � i �M �
Now
MXi��
ai�I � P �ei �
MXi��
hf� eii�I � P �ei �
MXi��
hf� eiiei �MXi��
hPf� eiiei � ��
since f � PHM � and thus Pf � f �
Corollary �� classi�es frames which are robust to one erasure� We now extend this
to a classi�cation of UNTFs robust to one erasure�
Corollary ���� Let feigMi�� be an orthonormal basis for HM and let P be an
orthogonal projection on HM � The following are equivalent�
�f�I � P �eigMi�� is a frame robust to one erasure�
��There is a g �PM
i�� aiei with nonzero scalars ai �� �� for all � i � M such
that Pf � hf� gig� for all f � HM � and kgk � �Moreover� if P is rank� � f�I � P �eigMi�� is a uniform frame if and only if jaij � �pM � for all � i �M �
Proof� � � ��� is immediate from Corollary �� �
For the moreover part� we know that f�I � P �eigMi�� is uniform if and only if
fPeigMi�� is uniform� Also� since P is rank� � for all � i � M and using the
expression expression Pf � hf� gig in Part �� with f � ei we have�
Pei � aig�
Hence� kPeik � kPejk is true if and only if jaij � jaj j� for all � i� j � M �
That is� fPeigMi�� is uniform �and hence f�I � P �eigMi�� is uniform� if and onlyif jaij � jaj j� for all � i� j �M � Finally� kPeik � jaijkgk � jaij � �
pM �
���� UTFs Robust to More than One Erasure
We now try to apply the same ideas from the previous section to classify frames
robust to more than one erasure� The classi�cation of UNTFs with k erasures in
�� CASAZZA AND KOVACEVI�C
this section is useful if we haveM vectors in an N �dimensional space andM �N is
�small�� In the next section we will give another classi�cation of this family which
works best when M �N is �large��
In what follows� I � f � �� � � � �Mg will denote the erasure index set�Proposition ���� Let feigMi�� be an orthonormal basis for HM � Let P be an or�
thogonal projection of HM onto an L�dimensional subspace H L � Fix I � f � �� � � � �Mgwith jI j � k � L and let K � spani�Icei� If f�jgLj�� is any orthonormal basis for
H L and we have the following L�M matrix
A � �h�j � eii� L Mj���i�� �
then
dim �ker �I � P �jK � � L� �row rank of the k columns of A indexed by I ��
Proof� We have
ker �I � P �jK � ff � K j�I � P �f � �g � �ker �I � P �� � K �
Now� apply row reduction �relative to feigMi��� to f�igLi�� using the elements of thecolumns in I � This gives a linearly independent set fgjgsj�� � fgjg L
j�s�� spanning
K with
s � row rank of �h�j � eii� Lj���i�I �gj � K for s� � j � L and
�span��j�sgj� � K � ��
Hence�
ker �I � P � � K � spans���j�Lgj �Therefore�
dim �ker �I � P �� � K � dim �ker �I � P �jK � L� �s� � � � L� s�
We are looking for the NTFs which are robust to k erasures� Proposition ���
actually yields a stronger result� Namely� it gives necessary and su�cient conditions
for fPeig to be robust to one particular choice of k erasures� We state this strongerresult in two di�erent forms in the next two propositions�
Proposition ���� With the notation of Proposition ���� the following are equiv�
alent�
�f�I � P �eigMi�� is robust to the erasure of the elements f�I � P �eigi�I ���We have rank �I � P �jK �M � L�
��We have dim �ker �I � P �jK � � L� k�
UNIFORM TIGHT FRAMES WITH ERASURES ��
Proof� � � ���� f�I � P �eigMi�� is robust to the erasure of the elementsf�I � P �eigi�I if and only if �I � P �jK is full rank� which must be rank �I � P � �
M � L�
��� ���� By Proposition ����
dim �ker �I � P �jK � L� �row rank of the k columns of A indexed by I � � L� k�
On the other hand� �I �P �jK is full rank if and only if �I �P ��I �Pi� is full rank�
where Pi is the orthogonal projection of HM onto spani�Iei� Hence�
dim �ker �I � P ��I � Pi�� � L�
But� ei � ker �I � Pi�� for all i � I � Hence�
dim �ker �I � P ��I � Pi� � dim �ker �I � P � � k � L�
Therefore�
dim �ker �I � P � � L� k�
Proposition ��� gives precise information when a frame is robust to the erasure of
one particular choice of k elements in terms of the coe�cients of f�ig relative tothe unit vector basis of H L � The following corollary is a statement for a �xed choice
of erasures as well�
Corollary ���� With the notation of Proposition ���� the following are equiv�
alent�
�f�I � P �eigMi�� is robust to the erasure of the elements f�I � P �eigi�I ���The row rank of �h�j � eii�Lj���i�I � k�
Applying the above result to every choice of erasures results in Theorem �� and
a classi�cation of when our frame is robust to any k erasures� We are dealing here
with general frames� We will deal with uniform frames right afterwards�
Theorem ���� With the notation of Proposition ���� the following are equivalent�
�f�I � P �eigMi�� is robust to k erasures�
��For every I � f � �� � � �Mg with jI j � k� the row rank of A � �h�j � eii�Lj���i�Iequals k�
Note the crucial role played by erased elements in the theorem� since in the
statement� the indices are from I �
Next we see what it takes for such frames to be uniform� Here� we call the �angle�
between two vectors the inner product when this is really the cosine of the angle�
Note that this result classi�es when f�I � P �eig is a UNTF without any referenceto erasures�
Theorem ���� With the notation of Proposition ���� the following are equivalent�
�� CASAZZA AND KOVACEVI�C
�fPeigMi�� is uniform �and hence f�I � P �eigMi�� is uniform��
��For every � i �M we have�
�A�A�ii �LXj��
jh�j � eiij� � L
M�
That is� the columns of A � f�jgLj�� all have the same square sums�
Moreover� the angle between �I � P �ei and �I � P �ej is given by the inner product
of the ith and jth columns of f�jgLj���
Proof� � � ���� We know that f�I�P �eigMi�� is uniform if and only if fPeigMi��which� in turn� is uniform if and only if the diagonal elements of the matrix for P
relative to feigMi�� have constant modulus� We have�
Pei �
LXn��
hei� �ni�n �
MXm��
�LX
n��
hei� �nih�n� emi�em�
The diagonal element of Pei is�
LXn��
hei� �nih�n� eii �LX
n��
jh�n� eiij��
So fPeigMi�� is a uniform tight frame if and only if for all � i� j �M we have
LXn��
jh�n� eiij� �LX
n��
jh�n� ejij��
For the moreover part� we compute for � i �� j �M �
h�I � P �ei� �I � P �eji � hei� eji � hPei� eji � hei� P eji� hPei� P eji
� hPei� P eji � hei� P eji
hei�LX
n��
hej � �ni�ni � hei�MXm��
�LX
n��
hej � �nih�n� emi�emi �
LXn��
hej � �nih�n� eii�
The right�hand side of the above equality is precisely the inner product of the ith
and jth columns of f�jgLj���The following consequence of the above is surprising at �rst� It says that we can
almost never get �in the real case� frames robust to k erasures by stepping down
from frames robust to one erasure� then to frames robust to two erasures etc�
Corollary ���� In the real case� if M � �� there do not exist rank� orthogonal
projections P�� P� with P�P� � � so that f�I � P��eigMi�� is uniform and f�I �P���I � P��eigMi�� is uniform and robust to two erasures�
UNIFORM TIGHT FRAMES WITH ERASURES �
Proof� Assume such P�� P� exist� Since f�I � P��eigMi�� is uniform� by The�orem ���� there is a vector �� �
PMi�� aiei � HM with P�f � hf� ��i��� for all
f � HM and jaij � jaj j� for all � i� j � M � Now choose �� �PM
i�� biei � HM
so the P�f � hf� ��i��� for all f � HM � Since P�� P� are rank� � P�P� � � implies
P�P� � �� That is� h��� ��i � �� Let P � P��P�� Then �I �P���I �P�� � I �P �
Since f�I � P �gMi�� is uniform� by Theorem ��� we have
a�i � b�i � a�j � b�j � for all a � i� j �M�
Hence� b�i � b�j � for all � i� j �M � Since k��k � k��k � � it follows that
a�i � b�i �
M� for all � i� j �M�
Since f�I � P �eigMi�� is robust to two erasures� by Theorem ���
aibj � ajbi �� �� for all � i� j �M� ����
However� for all i� j we have ai � aj � bi� Hence� there is some i� j witheither ai � aj and bi � bj or ai � �aj and bi � �bj � In either case ���� fails�Corollary ��� is heavily dependent on having a real Hilbert space� In the complex
case� we will show in Corollary �� that the GHFs have the property that we can step
down to the frame by successively applying rank� projections to the orthonormal
basis� Corollary ��� shows that in general we cannot step down one dimension at
a time to construct frames robust to k erasures� It would be interesting to extend
this to a classi�cation of when we can step down from k� erasures to k� erasures
and so on� One reason is that this has implications for entangled states in quantum
detection theory � �� ����
This raises the question of whether we can step down one dimension at a time
as in Corollary ��� if we only ask for each level to be uniform� The answer is no as
the next example shows�
Example ���� Let feig�i�� be an orthonormal basis for H � and let
�� � p�e� �
�e� �
�e��
and
�� � p�e� �
�e� �
� �e��
Then h��� ��i � �� and if P is the orthogonal projection of H � onto the span of
f��� ��g� then by Theorem ���� f�I�P �eig�i�� is a uniform normalized tight frame�However� there does not exist a rank� orthogonal projection P� of H � into range
of P so that f�I � P��eig�i�� is a uniform frame�
�� CASAZZA AND KOVACEVI�C
Proof� By Theorem ���� in order for such a P� to exist� there must exist a vector
in PH � of the formP�
i�� biei with jbij � jbj j� for all � i� j � �� For any a�� a��
a��� � a��� �a�p�e� �
a�p�e� �
a� � a��
e� �a� � a��
e��
If �ja� � a�j��� � �ja� � a�j���� then one of ja�j� ja�j equals � so ja�j �� ja�j�The next question is whether we can step down one dimension at a time as in
Corollary ��� if all we want is for our frame to be robust to j erasures at the jth
level but are willing to give up uniformity at each level� Surprisingly� the answer
here is yes�
Proposition ���� Let P be an orthogonal projection of HM onto an L�dimensional
subspace H L � Let feigMi�� be an orthonormal basis for HM � If f�I � P �eigMi�� is
robust to k erasures �k � L�� then there are mutually orthogonal rank� projections
fPigLi�� taking HM into H L so that f�I �Pj��I � Pj��� � � � �I �P��eigMi�� is robust
to j erasures� for all � j � k� and robust to k erasures for all k � j � L�
Proof� See Appendix A���
���� UTFs Robust to More than One Erasure� An Alternative
Approach
In the previous sections we used the orthogonal projection P on HM to classify
when f�I � P �eigMi�� is robust to k erasures and when it is uniform� However� ifthe dimension of PHM is large �that is� close to the dimension of HM � these results
become di�cult to implement since they require knowledge about the orthonormal
bases for the large�dimensional space PHM � In this case� f�I � P �eigMi�� is a largenumber of vectors in a small�dimensional space �I � P �HM � So it is easier in this
case to work directly with �I � P �H and f�I � P �eig� Or� equivalently� with PH
and fPeig� where dim PH is small�
Proposition ���� Let f�igNi�� be an orthonormal sequence in HM � let feigMi�� bean orthonormal basis for HM and let P be the orthogonal projection of HM onto the
span of f�igNi��� Then the normalized tight frame fPeigMi�� is unitarily equivalent
to fgigMi�� where for � j �M we have
gj �
NXi��
h�i� ejiei�
That is� fPeigMi�� is the frame obtained by turning the columns of f�igNi�� into row
vectors in HN �
Proof� For any � i �M we have�
Pei �
LXn��
hei� �ni�n�
UNIFORM TIGHT FRAMES WITH ERASURES ��
Since f�ngLn�� is an orthonormal sequence� the operator T�n � en� for � n � L
is a unitary operator which takes Pei to fgjgMj�� where for � j �M and we have
gj �
NXi��
h�i� ejiei�
Corollary ���� Given the conditions in Proposition �� we have�
�fgjgMj�� is a uniform frame for HN if and only if
NXi��
jh�I � ejij� � N
M� for all � j �M�
��The following are equivalent�
�a� fgjgMj�� is robust to the erasure of fgjgj�I for some I � f � �� � � �Mg��b� We have that
�h�i� eji�Ni���j�Ichas row rank N �
Hence� fgjgMj�� is robust to any choice of k erasures� k �M �N � if and only if
every set of M � k columns of the matrix
A � �h�i� eji� N Mi���j��
has row rank N �
Finally� the angle between gi and gj is given by the inner product of the ith and
jth columns of A�
It would be interesting and useful to classify the GHFs in the format of this
section� That is� precisely when is f�I �P �eigMi�� a GHF� One important propertyof GHFs is contained in the following result which says they have the step�down
property of Section �� �
Corollary ���� Let P be an orthogonal projection of HM onto an L�dimensional
subspace H L and let feigM��i�� be an orthonormal basis for HM � If f�I � P �eigM��
i��
is a GHF then there is an orthogonal sequence of rank� projections fPigLi�� of HMinto H L so that for all � j � L and for all permutations � of f � �� � � � � Lg we
have that
fj��Ym��
�I � P��m��eigM��i�� �
is a GHF and hence is a UNTF which is robust to j erasures�
Proof� We will do the proof for HTFs since the GHF case requires only nota�
tional changes but obscures the basic ideas of the proof� By Proposition ��� there
�� CASAZZA AND KOVACEVI�C
is a unique way to get HTFs� Namely� let feigM��i�� be the natural orthonormal basis
for HM � Let fwigM��i�� be distinctMth roots of unity and consider the orthonormal
basis f�igM��i�� for HM given by�
�i �
M��Xj��
wji ej �
Without loss of generality� we may as well assume that H L � span f�igLi���Now� turning the row vectors of f�igM��
i�L into column vectors gives a HTF for
HM�L � Moreover� again by Proposition ��� this HTF is unitarily equivalent tof�I �P �eigM��
i�� where I �P is the orthogonal projection onto �H L ��� Now� let Pi
be the orthogonal rank� projection of HM onto span �i for � � i � L� � Fix apermutation � of f � �� � � � � Lg and let Ij � f����� �� �� � � � � ��j � �g� Then� againby Proposition ���
fj��Ym��
�I � P��m��eigM��i�� �
is unitarily equivalent to the HTF obtained by turning the row vectors of f�igi��Ijinto column vectors�
It is possible that the property in Corollary �� characterizes GHFs� but we do
not have a proof for this� We can show that the use of permutations is necessary
in Proposition ��� That is� there are UNTFs which have the step�down property
for erasures and for uniformness for one �xed ordering of the rank� projections Piwhile failing to be equivalent to a GHF� This is the point of the next example�
Example ���� In C � � let
�� �
�Xi��
ei� �� �
�Xi��
wiei� �� � e� � e� � e� � e��
where w� � � w� � � � w� � i and w� � �i� Let fPig�i�� be the rank� orthogonalprojections of C � onto the span of �i� Then it follows easily from our results that�
� f�I � P��eig�i�� is a UNTF which is robust to one erasure��� f�I � P���I � P��eig�i�� is a UNTF which is robust to � erasures��� f�I�P���I�P���I�P��eig�i�� is a UNTF which is robust to � erasures� but is
not a harmonic frame since f�I �P���I �P��eig�i�� is a UNTF which is not robust
to � erasures�
Example ���� If �i � � � wi� in C� for � i � M � where jwij � and the
fwigMi�� are distinct� then f�igMi�� is a UNTF for C � if and only ifPM
i�� wi � ��
In this case� if Pi is an orthogonal rank� projection onto the span �i� then f�I �Pj�eigMi�� is a UNTF which is robust to one erasure while f�I � Pj��I � Pk�eigMi��is a UNTF which is robust to � erasures for all � j �� k �M �
UNIFORM TIGHT FRAMES WITH ERASURES ��
Proof� This family is certainly uniform� The assumptionPM
i�� wi � � guaran�
tees that the vectors
� � � � � � � � and �w�� w�� � � � � wM �
are orthogonal in HM �and conversely�� Hence� f�igMi�� is a UNTF for C � by Propo�sition ���
Example ���� In general� the frames constructed in Example ��� are not equivalent
to harmonic frames even after a permutation�
Proof� The reason is that f�jgM��j�� is a GHF implies that
h�j � �j��i � h�j��� �j��i� for all � � j �M � ��
�see Proposition ����� This would imply in Example ��� that
h�j � �j��i � � wjwj�� � � wj��wj�� � h�j��� �j��i�
Hence� wjwj�� � wj��wj��� Now� let w� � � w� � � �� �p���i and w� �
� �� � p���i� Then jwj j � andP�
j�� wj � �� Hence� by Example ����
f� � wj�g�j�� is a UNTF for C � � However� for any permutation � of f � �� �g� wedo not have w��j�w��j��� � w��j���w��j���� for all � � j � � �with w� � w���
It is not hard to see that in the case of R� � the condition in Corollary �� is su�cient�
APPENDIX� PROOFS
A��� PROOF OF PROPOSITION ���
First note that if fcigNi�� are distinct Mth roots of c with jcj � � then
M��Xk��
cki � ��M��Xk��
jcki j� � M�M��Xk��
i��l
�cicl�k � �� �A� �
and thus
M��Xk��
�cicl�k � M�i�l� �A���
We now check the necessity of our condition for a frame to be a general harmonic
frame� If f�kgMk�� is a general harmonic frame for HN then by the de�nition�
�k � �ck�b�� ck�b�� � � � � ckNbN� �
NXi��
cki biei�
� CASAZZA AND KOVACEVI�C
where feigNi�� is the natural unit vector orthonormal basis of HN � jbij � �pM �
and fcigNi�� are distinct Mth roots of c with jcj � � Now let
�� �
nXi��
biei�
De�ne a unitary operator U on HN by Uei � ciei� Then� for all � � k �M � wehave Ukei � cki ei� and thus� U
k�� � �k for all � � k � M � � So our frame hasthe form given in the proposition�
The su�ciency of the condition is checked similarly� If we assume that f�kg isof the form described in the proposition� then�
k�kk � kUk��k � k��k �r
N
M�
since U is a unitary operator� So f�kg is a uniform frame� To see that f�kg is anormalized tight frame� we let f �
PNi�� aiei � HN and compute�
M��Xk��
jhf� Uk��ij� �M��Xk��
jNXi��
aicki bij� �
�
M��Xk��
NXi��
jaicki bij� �Xm���
amb�
M��Xk��
�cmc��k
� MNXi��
jaij�jbij� � � � MNXi��
jaij�j pMj� � kfk��
where in the next to last equality we used the fact that the fcig is a family ofdistinct Mth roots of c �see �A����� Hence� f�kg is a UNTF� Thus� using � �� andwriting f � ej � we see that
kejk� � �
MXk��
jhej � Uk��ij� �MXk��
jhej �NXi��
cki bieiij� �MXk��
jckj bj j� � M jbj j��
So jbj j � �pM � for all j� � � j � N � � This further means that
�� � �b�� b�� � � � � bN�
with jbj j � �pM and since �k � Uk���
�k � �ck�b�� ck�b�� � � � � c
kNbN ��
A��� PROOF OF THEOREM ���
We prove the theorem in steps�
Step I� We �rst prove that there is a constant jcj � so that UM � cI � This will
show that the operators fU igM��i�� � in fact� form a � Since the frame is an NTF� for
UNIFORM TIGHT FRAMES WITH ERASURES �
every f � HN we have�
f �
M��Xk��
hf� Uk��iUk���
Hence�
Uf �
M��Xk��
hUf� Uk��iUk�� �
�
M��Xk��
hf� Uk����iUk��� since U� � U���
� U
�M��Xk��
hf� Uk����iUk����
�� U
�M��Xk���
hf� Uk��iUk��
�� �A���
Since U is one�to�one�
f �M��Xk���
hf� Uk��iUk�� �M��Xk��
hf� Uk��iUk��� �A���
From �A��� and �A���� it follows that�
hf� UM����iUM���� � hf� U����iU�����
Applying U we have
hf� UM����iUM�� � hf� U����iU��� � hf� U����i��� �A���
Hence� there is a c � C so that UM�� � c��� Replacing f by U��f in �A��� gives�
hU��f� UM����iUM�� � hf� UM��iUM��
� hf� c��ic�� � jcj�hf� ��i��� hU��f� U����i�� � hf� ��i���
So jcj� � � Also� for all � � k �M � �
UMUk�� � UkUM�� � Ukc�� � cUk���
Since fUk��g spans HN � that is� for any f � HN f �P
khf� Uk��iUk��� it follows
that UM � cI � This completes Step I�
Step II� We want to prove�
� U is diagonalizable with respect to an orthonormal basis feigNi�� with diag�onal elements fcigNi�� with ci an Mth root of c�
�� �� �PN
i�� biei and jbij �q
�M � for all � i � N �
These two steps give us frame elements as in De�nition �� �
�� CASAZZA AND KOVACEVI�C
Since U is unitary �and hence normal�� and our space is �nite dimensional� it follows
that U is diagonalizable with respect to an orthonormal basis feigNi�� with diagonalelements fcigNi��� Writing �� as in ���� we have that
Uk�� �NXi��
cki biei�
Since UM � cI � it follows that each ci is an Mth root of c� Also� since the frame is
an NTF� for all � j � N we have
�
MXk��
jhej � Uk��ij� �MXk��
jhej �NXi��
cki bieiij� �NXk��
jckj bj j� � M jbj j��
So jbj j � �pM � This completes Step II�
Step III� We �nally prove that the cj are distinct Mth roots of c� This will
complete the proof�
Fix � m �� � �M � Since the columns of a normalized tight frame are orthogonal
vectors� that is� F �F � I � we have for columns m� ��
� �
M��Xk��
bmckmb�c
k� � bmb�
M��Xk��
�cmc��k�
Hence�
M��Xk��
�cmc��k � ��
Thus� cm and c� are distinct Mth roots of c� This completes the proof of Step III
and the theorem�
A��� PROOF OF PROPOSITION ���
For all � j � L� let
gj �MXi��
hgj � eiiei�
be an orthonormal basis for H L � We will construct the projections fPjgLj�� by�nite induction� We start with P�� Since f�I � P �eigMi�� is robust to erasure� byTheorem �� � for every � i � M there exists a � j � L so that hgj � eii �� ��Choose scalars a�� a� � � � � aL so that for any � j � L� if hgj � eii �� � for some
� i �M � then
jajhgj � eiij � �
j��Xk��
jakhgk� eiij�
It follows easily that
�� �
PLj�� ajgj
kPLj�� ajgjk
�
UNIFORM TIGHT FRAMES WITH ERASURES ��
then h��� eii �� �� for all � i �M � Now� let P� be the orthogonal projection onto
the span f��g� then f�I � P��eigMi�� is robust to erasure by Corollary �� �Now assume we have found orthonormal vectors f�jgnj��� n � k � so that the
rank� projections Pj onto span f�jg satisfy the proposition and span f�jgnj�� �span fgjgnj��� We will construct �n�� and Pn��� Let fgjgLj�� be an orthonormalbasis for the orthogonal complement of span f�jgnj�� in H L � Fix � i� � i� �
� � � � in�� �M � Since
f�I � Pn��I � Pn��� � � � �I � P��eigMi��
is robust to n erasures� if we row reduce f�jgnj�� using the columns fi�� i�� � � � ing�and switching rows if necessary� we obtain vectors fhjgnj�� with span fhjgnj�� �span f�jgnj�� and for every � j � n we have�
hhj � eili � �j��
Now we verify the following claim�
Claim� For all but �nitely many � � x � � if
hn�� �
LXj�n��
xjgj �
then the row rank of
�hhj � ei�i�n��j���� �
is n� �
Proof of Claim� Fix � � x � � We row reduce this matrix by taking� for all
� � �M �
�hLX
j�n��
xjgj � ei�ih� � hn���
We then arrive at a matrix with s in the jth row and the ijth column for � j � n�
zeroes otherwise� zeroes in the �n � �st row for columns ij for all � j � n� and
in the �n� � n� �st position we have�
hhn��� ein��i �nX���
hLX
j�n��
xjgj � ei�ihh�� ei�i�
If this number is nonzero� then our matrix is of rank n� � Now we check what it
takes for this to be nonzero�
hhn��� ein��i �nX���
hLX
j�n��
xjgj � ei�ihh�� ei�i
�
LXj�n��
xjhgj � ein��i �LX
j�n��
xjnXj��
hgj � ei�ihh�� ei�i
�� CASAZZA AND KOVACEVI�C
�
LXj�n��
xj
hgj � ein��i �
nX���
hgj � ei�ihh�� ei�i��
By the hypotheses on the proposition�
�hgj � ei�i� L n��j������ �
has rank n� � It follows that there is a n� � j � L so that
hgj � ein��i �nX���
hgj � ei�ihh�� ei�i �� ��
Hence� the number of x�s with
hhn��� ein��i �nX���
hLX
j�n��
xjgj � ei�ihh�� ei�i � ��
is �nite� This establishes the claim�
Applying the claim to all choices of � i� � i� � � � � in�� � M � we see that for
all but �nitely many � � x � � for all � i� � i� � � � � � in�� � L� the rank of
�hhj � ei�i� M n��j������
is n� � Let
�n�� �hn��khn��k �
and let Pn�� be the orthogonal projection of HM onto span f�n��g� Now�
f�I � Pn����I � Pn� � � � �I � P��eiiMi��
is robust to �n � � erasures by Theorem �� � This completes the proof of the
proposition�
ACKNOWLEDGMENT
The �rst author would like to thank V� Paulsen for fruitful discussions�
REFERENCES
�� N� I� Akhiezer and I� M� Glazman� Theory of Linear Operators in Hilbert Spaces� volume ��Frederick Ungar Publisher� �����
�� J�J� Benedetto and D� Colella� Wavelet analysis of spectogram seizure chips� In Proc� SPIE
Conf� on Wavelet Appl� in Signal and Image Proc�� pages �������� San Diego� CA� July �����
�� J�J� Benedetto and G�E� Pfander� Wavelet periodicity detection algorithms� In Proc� SPIE
Conf� on Wavelet Appl� in Signal and Image Proc�� pages ����� San Diego� CA� July ����
�� H� Bolcskei and Y� Eldar� Geometrically uniform frames� Preprint� �����
�� H� Bolcskei� F� Hlawatsch� and H�G� Feichtinger� Frame�theoretic analysis of oversampled �lterbanks� IEEE Trans� Signal Proc�� ����������������� December ����
�� P�G� Casazza� The art of frame theory� Taiwanese Journ� of Math�� ������������� �����
UNIFORM TIGHT FRAMES WITH ERASURES �
�� P�G� Casazza� Modern tools for Weyl�Heisenberg �Gabor� frame theory� Adv� in Imag� and
Electron� Physics� ���������� �����
� Z� Cvetkovi�c and M� Vetterli� Oversampled �lter banks� Signal Proc�� ���������������� May���� Submitted�
�� I� Daubechies� The wavelet transform� time�frequency localization and signal analysis� IEEETrans� Inform� Th�� ��������������� September �����
��� I� Daubechies� Ten Lectures on Wavelets� SIAM� Philadelphia� PA� �����
��� R� J� Du�n and A� C� Schae�er� A class of nonharmonic Fourier series� Trans� Amer� Math�
Soc�� ����������� �����
��� Y� Eldar and G�D� Forney� Jr� Optimal tight frames and quantum measurement� Preprint������
��� G�J� Foschini and M�J� Gans� On the limits of wireless communications in a fading environmentwhen using multiple antennas� Wireless Personal Comm�� ������������� March ����
��� V� K Goyal� M� Vetterli� and N� T� Thao� Quantized overcomplete expansions in RN � Analysis�synthesis� and algorithms� IEEE Trans� Inform� Th�� ������������ January ����
��� V�K Goyal� Single and Multiple Description Transform Coding with Bases and Frames� ��������
��� D� Han and D�R� Larson� Frames� bases and group representations� In Memoirs AMS� Provi�dence� RI� �����
��� B� Hassibi� B� Hochwald� A� Shokrollahi� and W� Sweldens� Representation theory for high�ratemultiple�antenna code design� IEEE Trans� Inform� Th�� ���������������� �����
�� C� Heil and D� Walnut� Continuous and discrete wavelet transforms� SIAM Rev�� ��������������
��� B� Hochwald� T� Marzetta� T� Richardson� W� Sweldens� and R� Urbanke� Systematic designof unitary space�time constellations� IEEE Trans� Inform� Th�� ����� Submitted�
��� J�R� Holub� Pre�frame operators� Besselian frames and near�Riesz bases in Hilbert spaces�Proc� AMS� �������������� �����
��� I�E� Telatar� Capacity of multiantenna gaussian channels� European Trans� on Telecomm��������������� November �����
��� E� Soljanin� Frames and quantum information theory� Preprint� �����
��� T� Strohmer� Modern Sampling Theory� Mathematics and Applications� chapter Finite andin�nite�dimensional models for oversampled �lter banks� Birkhauser� �����
��� V�K Goyal and J� Kovacevi�c� Optimal multiple description transform coding of Gaussianvectors� In Proc� Data Compr� Conf�� pages ������ Snowbird� UT� March ����
��� V�K Goyal� J� Kovacevi�c� and J�A� Kelner� Quantized frame expansions with erasures� Journalof Appl� and Comput� Harmonic Analysis� �������������� May �����
��� V�K Goyal� J� Kovacevi�c� and M� Vetterli� Quantized frame expansions as source�channel codesfor erasure channels� In Proc� Data Compr� Conf�� pages ������ Snowbird� UT� March �����