On The Fundamental Tradeoff of Integrated Sensing and Communications Under Gaussian Channels

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO.
9, SEPTEMBER 2023 5723
On the Fundamental Tradeoff of Integrated Sensing

and Communications Under Gaussian Channels
Yifeng Xiong , Member, IEEE, Fan Liu , Member, IEEE, Yuanhao Cui , Member, IEEE,
Weijie Yuan , Member, IEEE, Tony Xiao Han, Senior Member, IEEE, and Giuseppe Caire , Fellow, IEEE
Abstract— Integrated Sensing and Communication (ISAC) is these two points can be achieved by the conventional Gaussian
recognized as a promising technology for the next-generation signalling and a novel strategy relying on the uniform distribution
wireless networks, which provides significant performance gains over the set of semi-unitary matrices, i.e., the Stiefel manifold,
over individual sensing and communications (S&C) systems via respectively. Based on the above-mentioned analysis, we provide
the shared use of wireless resources. The characterization of an outer bound and various inner bounds for the achievable
the S&C performance tradeoff is at the core of the theoretical CRB-rate regions. Our main results reveal a two-fold tradeoff in
foundation of ISAC. In this paper, we consider a point-to-point ISAC systems, consisting of the subspace tradeoff (ST) and the
(P2P) ISAC model under vector Gaussian channels, and propose deterministic-random tradeoff (DRT) that depend on the resource
to use the Cramér-Rao bound (CRB)-rate region as a basic allocation and data modulation schemes employed for S&C,
tool for depicting the fundamental S&C tradeoff. In particular, respectively. Within this framework, we examine the state-of-the-
we consider the scenario where a unified ISAC waveform is art ISAC signalling strategies and study a number of illustrative
emitted from a dual-functional ISAC transmitter (Tx), which examples, which are validated through numerical simulations.
simultaneously communicates information to a communication
receiver (Rx) and senses targets with the help of a sensing Rx. Index Terms— Integrated sensing and communication, Gaus-
In order to perform both S&C tasks, the ISAC waveform is sian channels, CRB-rate region, deterministic-random tradeoff,
required to be random to convey communication information, subspace tradeoff.
with realizations being perfectly known at both the ISAC Tx and
the sensing Rx as a reference sensing signal as in typical radar I. I NTRODUCTION
systems. In this context, we treat the ISAC waveform as a random
but known nuisance parameter in the sensing signal model, and A. Background and Related Works
define a Miller-Chang type CRB for the analysis of the sensing
performance. As the main contribution of this paper, we charac-
terize the S&C performance at the two corner points of the CRB-
I N AN ISAC system, wireless communications and sensing
functionalities are performed by using a single hardware
platform and a common radio waveform over the same
rate region, namely, PSC indicating the maximum achievable
communication rate constrained by the minimum CRB, and frequency band, which considerably improves the energy-,
PCS indicating the minimum achievable CRB constrained by spectral-, and hardware-efficiencies [1], [2], [3], [4], [5]. Due
the maximum communication rate. In particular, we derive the to the substantial performance gain attainable for both S&C
high-SNR communication capacity at PSC , and provide lower capabilities, ISAC is well-recognized as a key enabler for a
and upper bounds for the sensing CRB at PCS . We show that
variety of emerging applications including vehicular networks,
Manuscript received 27 October 2022; revised 22 March 2023; accepted smart home, and smart cities. Despite the fact that S&C have
31 May 2023. Date of publication 9 June 2023; date of current ver- long been considered as two isolated fields, e.g., wireless
sion 18 August 2023. This work was supported in part by the National
Natural Science Foundation of China under Grant 62101234, in part by networks and radar systems, they are indeed intertwined with
the Guangdong Province “Pearl River” Young Talent Support Program each other as an “odd couple” in an information-theoretic
under Grant 2021QN02X128, and in part by the Shenzhen Science and sense [6].
Technology Program under Grant RCBS20210609103227018 and Grant
20220815100308002. An earlier version of this paper was presented in For decades, S&C researchers are mostly working on a
part at the 2022 IEEE Global Communications Conference (GLOBECOM) generic linear Gaussian model, expressed as
[DOI: 10.1109/GLOBECOM48099.2022.10001144]. (Corresponding author:
Fan Liu.) Y = HX + Z, (1)
Yifeng Xiong is with the School of Information and Electronic Engineering,
N ×T N ×M M ×T N ×T
Beijing University of Posts and Telecommunications, Beijing 100876, China, where Y ∈ C ,H∈C ,X∈C , and Z ∈ C
and also with the Department of Electronic and Electrical Engineering, denote the received signal, target response/communication
Southern University of Science and Technology, Shenzhen 518055, China
(e-mail: yifengxiong@bupt.edu.cn). channel, transmitted signal, and Gaussian noise, respectively,
Fan Liu, Yuanhao Cui, and Weijie Yuan are with the Department which can be in scalar, vector, or matrix forms. From the
of Electronic and Electrical Engineering, Southern University of Science communication perspective, the basic problem is given channel
and Technology, Shenzhen 518055, China (e-mail: liuf6@sustech.edu.cn;
cuiyuanhao@bupt.edu.cn; yuanwj@sustech.edu.cn). H to design an optimal transmission strategy X that maxi-
Tony Xiao Han is with Huawei Technologies Company Ltd., Shenzhen mizes the channel capacity. From the sensing perspective, the
518063, China (e-mail: tony.hanxiao@huawei.com). basic problem is to estimate H, or, more relevant to radar
Giuseppe Caire is with the Chair of Communications and Information
Theory, Technical University of Berlin, 10623 Berlin, Germany (e-mail: sensing, to estimate the target parameters (e.g., amplitude,
caire@tu-berlin.de). delay, angle, and Doppler) contained in H as accurate as
Communicated by A. Sabharwal, Associate Editor for Communications. possible, based on the knowledge of X. Indeed, S&C are
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TIT.2023.3284449. connected with each other given the duality between signals
Digital Object Identifier 10.1109/TIT.2023.3284449 and linear systems, which are mathematically interchangeable.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
5724 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 9, SEPTEMBER 2023
Nonetheless, the distinct physical roles of X and H in S&C and

the resulting performance measures lead to a number of unique
performance tradeoffs and design criteria [7], [8], [9], [10] to
the ISAC signal processing, particularly for simultaneously
realizing S&C functionalities via a unified waveform. Despite
that capacity characterization and parameter estimation limits
have been independently investigated for S&C for years, the
depiction of the inherent information-estimation tradeoff and
the corresponding performance bounds in an ISAC system Fig. 1. The ISAC scenarios considered in this paper, where the
remains long-standing open. dual-functional waveform X is known to the sensing receiver.
The pioneering effort to reveal the fundamental connection
between information and estimation theory can be traced back is random. To cope with this issue, we propose to treat X as a
to the early results on the I-MMSE equation√[11], which states random known nuisance parameter in the sensing model, and
that for the linear Gaussian model Y = snrHX + Z, the resort to a Miller-Chang type CRB to measure the average
derivative of the mutual information I (Y; X) with respect to sensing performance [18].
the signal-to-noise ratio (SNR) is equal to the minimum mean Under this framework, our main contributions are summa-
squared error (MMSE) for estimating HX by observing Y. rized as follows:
While the I-MMSE equation implies maximizing the MMSE • We define a Miller-Chang type CRB for measuring the
of HX also maximizes the scaling ratio of the mutual infor- sensing performance limit in an ISAC system, and show
mation I (Y; X), which bridges the information theory and that its optimum is achieved when the sample covariance
estimation theory, the equation itself mainly concentrates on matrix RX = T −1 XXH has a deterministic trace, and the
the estimation performance of unknown data symbol X rather support of the distribution p (RX ) (and hence p(X)) is
than on the target parameters in H, since H is assumed to be restricted to the optimal solution set of a deterministic
deterministic and known [11]. Consequently, the connection CRB minimization problem. In particular, if the solution
between I (Y; X) and the sensing performance limits for H is unique, the sensing-optimal sample covariance matrix
was not fully investigated. More relevant to this work, one may RX itself should be deterministic;
model the ISAC channel as a state-dependent channel [12], • We define the CRB-rate region as the set of all achievable
which characterizes the information-conveyed signal reflected pairs of the communication rate and the sensing CRB, and
by state-varying targets [13]. The goal for such an ISAC propose a pentagon inner bound of the CRB-rate region
system is to transmit information through the channel while that can be achieved through simple time-sharing strategy.
estimating its state under the minimum distortion. On top Within this framework, we study the ISAC performance
of that, a capacity-distortion (C-D) tradeoff is acquired to at the two corner points of the CRB-rate region, namely,
depict the S&C performance bounds [14], [15], [16], [17]. PCS indicating the minimum achievable CRB constrained
While these studies could capture the S&C tradeoff from by the maximum communication rate, and PSC indicating
the information-theoretic perspective, the connections among the maximum achievable communication rate constrained
communication capacity, target channel states, and to-be- by the minimum CRB.
estimated parameters in sensing tasks are still not clearly • We derive the high-SNR communication capacity for
indicated. the sensing-optimal point PSC , and prove that it can be
asymptotically achieved by a strategy based on uniform
sampling over the set of semi-unitary matrices, i.e., the
B. Contribution of This Paper Stiefel manifold [19]. As a further step, we provide
To gain a better understanding of the interweaved S&C lower and upper bounds for the sensing CRB at the
functionalities, in this paper, we shed light on the fundamental communication-optimal point PCS ;
limits of ISAC by analyzing the tradeoff between the com- • We reveal that the S&C tradeoff in ISAC systems is
munication capacity and sensing CRB, two key performance a two-fold tradeoff, namely, the subspace tradeoff (ST)
metrics at the cores of information theory and estimation balancing the resource allocation between the subspaces
theory. We consider a general point-to-point (P2P) system spanned by S&C channels, and the deterministic-random
setting, where a random signal X is emitted from an ISAC tradeoff (DRT) depicting the exploitable S&C degrees of
Tx and received both at a communication Rx and a sensing freedom (DoFs) in ISAC signals. Based on these two
Rx, thus to realize simultaneous S&C functionalities. Since X tradeoffs, we further propose an outer bound and a variety
encodes useful information intended for the communication of inner bounds for the CRB-rate region.
Rx, it is unknown to the communication Rx, yet is perfectly • We unveil the connection between the above funda-
known to both the ISAC Tx and sensing Rx due to the fact mental tradeoffs and existing ISAC waveform designs,
that they are colocated or collaborative in typical monostatic or and provide illustrative examples to demonstrate the
bistatic radar settings. While the P2P communication capacity behaviour of ST and DRT in typical ISAC application
is known to be the maximum mutual information over all scenarios.
possible distributions of X, the optimal achievable sensing To the best of our knowledge, this is the first work
performance is less understood, especially when the signal X that addresses the fundamental tradeoff in ISAC from both
XIONG et al.: ON THE FUNDAMENTAL TRADEOFF OF INTEGRATED SENSING AND COMMUNICATIONS 5725
information- and estimation-theoretic perspectives. Our hope Yc ∈ CNc ×T and Ys ∈ CNs ×T denote the received communi-
is that this paper can serve as a stepping stone towards the fully cation and sensing signals, respectively. X ∈ CM ×T denotes
characterization of the CRB-rate performance region, as well the transmitted dual-functional waveform emitted from an
as towards the design of bound-achieving ISAC transmission ISAC TX for performing both S&C tasks. In this paper,
strategies. we assume that X is known to both the ISAC Tx and sensing
The rest of this paper is organized as follows. The system Rx, but unknown to the communication Rx. This is an abstrac-
model and the performance metrics for both sensing and com- tion of the practical scenarios portrayed in Fig. 1. A more
munication are introduced in Section II. Then, in Section III, general scenario where the sensing task can be performed by
we present our main results characterizing the corner points the communication receiver itself will be addressed in our
PSC and PCS of the CRB-rate region, while the corresponding future works. The random matrix Zc ∈ CNc ×T denotes the
proofs are deferred to the Appendices. Based on these results, received communication noise, modelled as independently and
we discuss the ST and DRT in Section IV-A, provide some identically distributed (i.i.d.) circularly symmetric complex
strategies achieving inner bounds of the CRB-rate region in Gaussian random variables with zero mean, namely vec(Zc ) ∼
Section IV-B, and unveil the connections between the S&C CN (0, σc2 I Nc ×T ). Similarly, we also model the sensing noise
tradeoff and existing ISAC system designs in Section IV-C. Zc as vec(Zs ) ∼ CN (0, σs2 I Ns ×T ). The communication sub-
This is followed by a case study in Section V, where we system aims for transmitting as much information as possible
discuss the ST and DRT in practical ISAC application sce- (reliably) to the receiver, while the sensing subsystem aims
narios. We illustrate and validate the analytical results using for estimating the sensing parameters η ∈ RK contained in
numerical examples in Section VI, and finally conclude the Hs satisfying
paper in Section VI. Hs = g(η)
to the highest possible accuracy, where g(·) is an injective
Notations
mapping from RK to CNs ×M .1 We consider a block fading
Throughout this paper, a, a, A, and A represent random vari- model for both the target response matrix and the commu-
ables (scalars), random vectors, random matrices and random nication channel. Specifically, we assume that parameters η
sets, respectively; The corresponding deterministic quantities, varies every T symbols in an i.i.d. manner, following a known
are denoted by a, a, A, and A, respectively. The m-by-n distribution pη (η) which has a finite variance, and that the
matrix of zeros (resp. ones) is denoted by 0m×n (resp. 1m×n ). communication channel Hc also varies every T symbols in
The m-dimensional vector of zeros (resp. ones) is denoted by an i.i.d. manner.2 We will refer to T as the coherent sensing
0m (resp. 1m ). The m-by-m identity matrix is denoted by period in the rest of the paper. We limit the average power of
I m . The Kronecker product between matrices A and B is each transmitted symbol to be PT , hence we have
denoted by A⊗B. ∥x∥p denotes the lp norm, which represents n o
the l2 norm by default when the subscript is omitted. The tr R e X = E{tr {RX }} = PT M, (3)
notation Ex {·} denotes the expectation with respect to x. [·]†
denotes the Moore-Penrose pseudo inverse of its argument, e X := E{RX } denotes the covariance matrix of X, with
where R
[·]∗ denotes the complex conjugate of its argument, while [·]H RX := T −1 XXH representing the sample covariance matrix.
denotes the Hermitian transpose of its argument. A:,i denotes
the i-th column of A, while Ai,: denotes the i-th row. The B. S&C Performance Metrics
notation diag(·) denotes the matrix obtained by placing its
arguments on the main diagonal of a sqaure matrix, mdiag(·) The performance of the communication subsystem is con-
denotes the vector consisting of the main diagonal entries of ventionally characterized by the ergodic achievable rate (under
its argument, while blkdiag(·) denotes the matrix obtained certain design constraints), which can be expressed as
similarly, but with matrix arguments. A ≽ B implies that RF = max T −1 I(Yc ; X|Hc ), s.t. pX (X) ∈ F, (4)
A − B is positive semidefinite. tr{·} stands for the trace of pX (X)
a square matrix. vec{A} denote the column-stacked vector where I(Yc ; X|Hc ) denotes the mutual information between
of the matrix A, while matM ×N {v} denotes the M × N Yc and X conditioned on Hc , and F denotes the feasibility
matrix satisfying vec{matM ×N (v)} = v. The subscripts in region of pX (X) determined by the design constraints. If not
the aforementioned notations may be omitted when they are stated otherwise, in the rest of this paper, the feasibility region
clear from the context. F is the set of all pX (X)’s that satisfy the power constraint (3),
II. S YSTEM M ODEL AND P ERFORMANCE M ETRICS 1 This would impose upper bounds for the maximum number K
max of
identifiable parameters. A trivial upper bound would be Kmax ⩽ 2M Ns .
A. System Model Tighter bounds may be obtained by incorporating specific knowledge of the
Let us consider the general ISAC system model of observation model. We refer interested readers to [20].
2 These assumptions correspond to the memoryless channel model widely
Yc = Hc X + Zc , (2a) used in information-theoretical ISAC studies (cf. [17], [21]). This model
mainly aims for portraying practical scenarios in which the sensing target
Ys = Hs X + Zs , (2b) is relatively close to the sensing link, and hence the sensing parameters vary
synchronously with the communication channel [22]. We note that our analysis
where Hc ∈ CNc ×M and Hs ∈ CNs ×M denote the commu- also applies to the case where the communication channel varies every kT
nication channel and the target response matrix, respectively, symbols where k ∈ Z+ .
and the subscript F in RF is omitted whenever there is no

confusion.
The performance of the sensing subsystem is typically
characterized by the estimation mean-squared error (MSE) of
the parameters η, given by
MSEη (η̂) := E{∥η − η̂∥2 }. (5)
However, this metric depends on the specific choice of the
estimator η̂, which hinders the essential relationship between
the sensing performance and the design of pX (X). To this
end, we consider the Bayesian Cramér-Rao bound (BCRB) of
η, which constitutes a lower bound for the MSE of weakly
Fig. 2. Graphical illustration of various possible CRB-rate regions.
unbiased estimators [23], taking the following form:
n o
MSEη (η̂) ⩾ E tr J−1 , (6)
η|X Fig. 2. Specifically, the point (ϵmin , 0) represents the minimum
where the expectation is taken with respect to X, and Jη|X achievable sensing MSE regardless of the communication
denotes the Bayesian Fisher information matrix (BFIM) of η performance, while the point (+∞, Rmax ) represents the max-
given by imum achievable rate regardless of the sensing performance.
The boundary A may be seen in scenarios where the integra-
∂ ln pYs |X,η (Y s |X, η) ∂ ln pYs |X,η (Y s |X, η)

Jη|X := E X tion of sensing and communication does not provide additional
∂η ∂η T performance gain (but we have some a priori knowledge

∂ ln pη (η) ∂ ln pη (η) about the sensing parameters, hence the maximum CRB is
+E .
∂η ∂η T not infinity), while the boundary B may be seen in idealistic
scenarios where both sensing and communication performance
It is well-known that certain practical estimators (e.g.
can achieve their optimum without eroding the other. In most
the maximum a posteriori (MAP) estimator) are capable
practical scenarios, the boundary of CRB-rate regions may
of achieving the BCRB in the asymptotic limit of high
have a similar shape as that of the boundary C.
SNR [23].3 In light of this, in this paper, we choose the BCRB
of η, namely
n o III. M AIN R ESULTS
ϵ := E tr J−1 η|X , (7) In this section, we summarize the main analytical results of
this paper. Before delving into details, we first give a general
as the metric of sensing accuracy.
description of an inner bound of the CRB-rate region.
At this point, it is worthwhile to clarify our motivation of
Proposition 1 (Pentagon Inner Bound): Any point (ϵ, R)
choosing the specific form of BCRB in (7), as well as to
satisfying the inequalities
justify this choice. Indeed, there exist multiple versions of the
BCRB when the parameters (or a part of the parameters) are ϵ ⩾ ϵmin , (8a)
random [24]. What we are using is in fact the Miller-Chang R ⩽ Rmax , (8b)
type bound [18], which treats X as a nuisance parameter, ϵCS − ϵmin
and takes n theoexpectation with respect to the conditional ϵ ⩾ ϵmin + (R − RSC ) (8c)
Rmax − RSC
CRB tr J−1 η|X . This specific form is particularly relevant
is achievable, where we define
to the scenario where X is known to the sensing receiver.
To elaborate, let us consider a specific realization of X, ϵmin := min ϵ, (9a)
pX (X)∈F
denoted as X i . Since the sensing receiver always have the full
knowledge o X, the BCRB for this realization is given by
about Rmax := max T −1 I(Yc ; X|Hc ), (9b)
n pX (X)∈F
tr J−1η|X=X i . Taking the expectation over all realizations of ϵCS := min ϵ, s.t. T −1 I(Yc ; X|Hc ) = Rmax , (9c)
X, we obtain (7). pX (X)∈F
To reveal the fundamental S&C performance tradeoff, RSC := max T −1 I(Yc ; X|Hc ), s.t. ϵ = ϵmin , (9d)
we define the CRB-rate region, which is the set of all feasible pX (X)∈F
ordered pairs (ϵ, R), where R and ϵ are the communication and
rate and the sensing CRB, respectively. Typically, we are
interested in the boundary of the CRB-rate region, which PSC := (ϵmin , RSC ), PCS := (ϵCS , Rmax ). (10)
may be viewed as the Pareto front constituted by all optimal Proof: The points in (10) are obviously achievable (but
performance tradeoffs. To provide further intuitions about the note that in certain scenarios we may have RSC = 0 or
boundary, we portray some conceptual CRB-rate regions in ϵCS = ∞). The line segment connecting PSC and PCS
3 When the SNR is lower, BCRB is less tight due to the fact that it mainly
(characterized by (8c)) can be achieved by using the cele-
exploits the local information of the posterior distribution around the mode. brated time-sharing strategy [25, Chap. 4], namely applying
Therefore, in this paper, we will focus on the analysis of high-SNR scenarios. the strategy corresponding to PSC with probability p1 , while
When the trace of RX is fixed, we have the following

alternative representation of Φ(·).
Corollary 1: When the trace of RX is fixed, i.e., tr{RX } =
γ, we may express Φ(·) as follows
r3
H
X
Φ(A) = Φγ (A) := F̄ i AḠi , (13)
i=1
where r3 ⩽ KM , and the matrices F̄ i and Ḡi are given

by (134).
Proof: Please refer to Appendix B.
Using Proposition 2 and Corollary 1, we may now charac-
terize the minimum achievable CRB as follows.
Fig. 3. The pentagon inner bound of CRB-rate regions. Proposition 3 (Sensing-Optimal RX ): The minimum
achievable CRB ϵmin is achieved when the o of RX is
n trace
deterministic, namely when tr {RX } = tr R e X . Moreover,
applying the strategy corresponding to PCS with probability
1 − p1 . This completes the proof. the support of p(RX ) should be restricted to the optimal
solution set of the following deterministic convex optimization
problem
A. The Structure of Sensing-Optimal Signals −1
As it can be observed from Fig. 3, the points PSC and PCS min tr ΦPT M (Re X) , (14a)
R
eX
correspond to the sensing-optimal and the communication-
e H.
n o
optimal strategies, respectively, which serve as important s.t. tr R e X = PT M, R e X ≽ 0, R eX = R
X (14b)
control points of the CRB-rate region. Naturally, a detailed
Proof: (sketch) We first show that the function
characterization of these points is desired. In particular,
f (γ) := min tr (Φγ (R))−1 , s.t. tr {R} = γ, R = RH

to obtain PSC –achieving strategies, we should first find the
R≽0
conditions that a sensing-optimal signal X must satisfy. Espe-
cially, the ISAC scenario imposes a unique challenge that is convex. The optimal objective function value of the prob-
the signal X has to be random for carrying information, lem (14) can then be expressed as f (PT M ). Thus for positive
while existing analysis of pure sensing (e.g. radar sensing) semidefinite Hermitian RX , we obtain
scenarios typically assume deterministic signal or neglect the min E tr [Φ(RX )]−1 , s.t. tr {RX } = PT M

randomness. To this end, we derive the general form of the p(RX )

BFIM as follows.
min tr [Φγ (RX )]−1

⩾ min E , s.t. E{γ} = PT M
Proposition 2 (BFIM Structure): The BFIM of η condi- p(γ) tr{RX }=γ
tioned on X takes the following form = min E {f (γ)} , s.t. E{γ} = PT M
p(γ)
T
Jη|X = 2 Φ(RX ), (11) ⩾ f (PT M ),
σs
where the last line follows from Jensen’s inequality. For a
where Φ(·) is an affine map characterized by detailed proof, please refer to Appendix C.
r1 r2 Proposition 3 tells us that the sensing-optimal RX has a
eH + eH + J
X X
Φ(A) = e i AT F
F G
e j AG e P, (12) deterministic trace. Moreover, when the problem (14) has a
i j
i=1 j=1 unique optimal solution, it holds that the sensing-optimal RX
e P = σ 2 T −1 J P , and the term J P is contributed by is itself deterministic, namely RX = R e X . Next, we present a
where J s sufficient and necessary condition for the uniqueness of the
the prior distribution pη (η), given by4
optimal solution of (14).
Proposition 4 (Uniqueness of the Sensing-Optimal RX ):

∂ ln pη (η) ∂ ln pη (η)
JP = E . Consider a maximum-rank optimal solution of (14), denoted
∂η ∂η T
as Ropt . It is also the unique optimal solution of (14), if and
2 2
The matrices Fe i and G e i are given by (122) and (124). only if the matrix Ξ(U ∗opt ⊗U opt ) ∈ CK ×r has full column
Furthermore, we have r1 , r2 ⩽ KM . 2 2
rank, where Ξ ∈ CK ×M is given by
Proof: Please refer to Appendix A. r1 r2
e∗ ⊗ F e∗ ⊗ G
X X
4 In Ξ= (F i
e i )K + G j
e j, (15)
this paper, we consider the scenario where Hs and Hc are not
statistically correlated, although they may be physically correlated in terms i=1 j=1
of subspace overlap (as will be detailed in Section IV-A). When they do r is the rank of Ropt , K is an elementary matrix satisfying
have a statistical correlation, the knowledge of Hc may be modelled as an
observation of η via p(Hc |Hs ) and Hs = g(η). Let us denote the contribution Kvec(A) = vec(AT ), U opt is obtained by the eigendecom-
of p(Hc |Hs ) to the BFIM as Jη|Hc . Since Jη|Hc does not depend on the position of Ropt
transmitted signal X, it can be absorbed into the term J P , and hence the
proposed framework also applies to this case after proper modifications. Ropt = U opt Λopt U H
opt , (16)
and Λopt ∈ Cr×r is a diagonal matrix containing all non-zero B. Point PSC –Achieving Strategy
eigenvalues of Ropt . In general, the optimization problem (14) has to be solved
Proof: (sketch) Observe that the objective function (14a) on a case-by-case basis. Considering the limited scope of this
is convex with respective to R e X . Therefore, it suffices to show
paper, we defer the discussion of the detailed structure of the
that it is strictly convex in a set constituted by R e X satisfying
optimal solution to our future works, and simply denote the
the aforementioned conditions. This would in turn require optimal sample covariance matrices as RSCX . Note that
Φ(·) be an injective map. For a detailed proof, please refer
to Appendix D. I(Yc ; X|Hc ) = I(Yc ; X, RSC
X |Hc ), (22)
Proposition 4 is both sufficient and necessary for the unique-
since RSC
X is a deterministic function of X, and thus using the
ness of the optimal solution, but may not be convenient for
chain rule of mutual information, we have
verifying the solution-uniqueness of specific problems, since
it requires the information about the maximum-rank optimal I(Yc ; X|Hc ) = I(Yc ; RSC SC
X |Hc ) + I(Y c ; X|Hc , RX ). (23)
solutions. In light of this, next we provide some sufficient
The rate RSC can now be expressed as
conditions.
Proposition 5 (Sufficient Conditions for Uniqueness): The 1
RSC = max I(Yc ;RSC
X |Hc )+I(Y c ;X|Hc ,R SC
X ) , (24a)
problem (14) has a unique optimal solution, if any of the pX (X) T
following conditions holds: s.t. RSC
X ∈ RSC , (24b)
1) (Generic) Ξ(U ∗R ⊗U R ) has full column rank, where U R
where RSC is the set of optimal solutions of problem (14),
contains the eigenvectors corresponding to the maximum
given by
eigenvalue of ΦaPT M (Z A ), namely n o
RSC = R|tr (ΦPT M (R))−1 = min tr (ΦPT M (R))−1

UH a H a
R ΦPT M (Z A )U R = I · max v ΦPT M (Z A )v, (17) R∈D
∥v∥=1
where D is constituted by all feasible R’s satisfying (14b).
with Z A being the optimal solution of the following
Especially, we observe that when the sensing-optimal sample
problem
covariance matrix is unique, i.e.
n 1 o 2
max tr Z A2 , RSC e SC
ZA X = RX with probability 1, (25)
s.t. PT M ΦaPT M (Z A ) ≼ I, ZA = ZH
A, Z A ≽ 0, (18) I(Yc |Hc ; RSC
the term X )in (24a) is zero, and hence we have
where ΦaPT M (·)
is the adjoint operator of ΦPT M (·), 1 e SC
I Yc ; X Hc , RSC

RSC = max X = RX , (26a)
characterized by pX (X) T
r3 e SC = argmin tr (ΦP M (R))−1 . (26b)

X H s.t. R X
ΦaPT M (A) = Ḡi AF̄ i . (19) R∈D
T
i=1
When (25) holds, we may provide a generic characterization
2) (K ⩾ M ) Ξ has full column rank; of the point PSC in the high-SNR regime, as detailed in the
3) (K = 1) The eigenspace corresponding to the maximum following proposition.
eigenvalue of the following matrix Theorem 1 (Sensing-Limited High-SNR Ergodic Capacity):
r1 r2 When the sensing-optimal sample covariance matrix is unique,
e HF e HG
X X
B 1 := (F e i )T + G ej (20) in the high-SNR regime, namely when PT /σc2 → ∞, the rate
i j
i=1 j=1 RSC can be expressed as
has dimensionality 1. n MSC e SC HH |+c0 +O(σ 2 ).
o
RSC = E 1− log |σc−2 Hc R X c c
Proof: (sketch) For the K ⩾ M case, it is straightforward 2T
that when Ξ has full column rank, the map Φ(·) is injective. (27)
For the general case, the problem (18) is the dual of the
where MSC denotes the rank of Hc R e SC HH , and the term
original problem (14). Since strong duality holds, these two X c
problems have identical optimal solutions, but the uniqueness MSC h MSC T √ i
condition derived from the dual problem is more convenient as c0 = T− log −log Γ(T )+log(2 π) (28)
T 2 e
it does not involve the maximum-rank solution. The condition converges to zero as T → ∞.
for the K = 1 case follows from a simplification of the general Proof: Please refer to Appendix G.
condition. For a detailed proof, please refer to Appendix E. Theorem 1 also enables us to propose a scheme that
When the sensing-optimal covariance matrix is unique, asymptotically achieves the point PSC in the high-SNR regime,
we have the following result. as given in the following corollary.
Corollary 2: If the sensing-optimal R e X is unique, its rank
Corollary 3 (Point PSC –Achieving Strategy): In the high-
is upper-bounded by SNR regime, when (25) holds, the rate RSC may be asymp-
e X ) ⩽ min{K, M }.
rank(R (21) totically achieved by a waveform XSC generated as follows:
√
Proof: Please refer to Appendix F. XSC = T H†c U e + X⊥ ,
e ΣQ (29)
where H†c denotes the Moore-Penrose pseudo-inverse of Hc , e CS ). Then, the communication-limited

distribution CN (0, R X
U
e contains the first MSC columns of U given by the eigende- minimum CRB is given by
composition of Hc R e SC HH
σs2
−1
X c CS
ϵCS = E tr Φ(RX ) . (35)
e SC HH = UΛUH , T
Hc R X c (30)
√ Next, we provide an upper bound for ϵCS at the high-SNR
e = diag {σ1 , . . . , σM }, σi = λi , λi is the i-th largest
Σ limit.
SC
eigenvalue of the matrix Hc R e SC HH , and X⊥ ∈ CM ×T is any Theorem 2 (Sensing DoF Loss): When the columns in X
X c
matrix satisfying are independent of one another, and identically follow a cir-
cularly symmetric complex Gaussian distribution CN (0, Re X ),
X⊥ (H†c U e H = 0,
e ΣQ) (31a) −1
if [Φ(RX ) − J P ] exists almost surely and J P is invertible,
e e
T −1
X⊥ XH + e SC H† Hc
H†c Hc R = e SC .
R (31b) we may obtain the following bound:
⊥ X c X
n o
e X )−1 ⩽ E tr Φ(RX )−1

The matrix Q ∈ CMSC ×T contains the modulated data, uni- tr Φ(R
formly sampled from the following rescaled Stiefel manifold n
e X )−1
o
T · tr Φ(R
V = Q ∈ CMSC ×T |QQH = I MSC . ⩽ . (36)

(32)
T − min{K, rank(R e X )}
Moreover, the uniform sampling over V can be implemented Proof: Please refer to Appendix J.
as follows: To fully unveil the implications of Theorem 2, let us
1) Construct A ∈ CMSC ×T , in which each column is formally define the concept of sensing DoF. Upon denoting
independently drawn from the standard circular complex the SNR as γ, the communication DoF is known to be defined
Gaussian distribution CN (0, I MSC ); as the asymptotic ratio [28]
2) Obtain Q by performing the LQ decomposition of A,
R(γ)
namely A = LQ, where L is a lower-triangular matrix νc = lim , (37)
γ→∞ log(1 + γ)
with real diagonal entries, and Q is a semi-unitary matrix
satisfying QQH = I MSC . where R(γ) denotes the maximum achievable rate given the
Proof: Please refer to Appendix I. SNR γ. Similar to the communication DoF, we define the
A PSC –achieving strategy in the form of (29) seems to rely sensing DoF as the following high-SNR limit
on the knowledge of Hc . Next, we show that this knowledge
n o
tr [Φ(R e X )]−1
is not necessary, by providing a specific form of XSC which νs = lim −1 . (38)
does not incorporate Hc . σs →0 T E{tr {[Φ(RX )]−1 }}
Corollary 4: One of the possible XSC satisfying (29) is Using this definition, we may now interpret the quantity
given by min{K, rank(R e X )} in (36) as an upper bound for the sensing
√ 1 DoF loss induced by employing the communication-optimal
XSC,1 = T U s Λs2 Q, (33)
transmission scheme. Intuitively, when RX = R e X , the sensing
where U s and Λs are obtained from the eigendecomposition DoF is T , which may be viewed as the number of independent
e SC as
of R observations. In general, for a random RX corresponding to
X
e SC = U s Λs U H , a sensing DoF of νs , the sensing CRB equals to that of the
R X s deterministic covariance matrix scenario (i.e. RX = R e X ) with
and Q is uniformly sampled from the set V given in (32). T = νs . This suggests that νs may be interpreted as the
Proof: Please refer to Appendix I. effective number of independent observations. The a priori
knowledge about the sensing parameters clearly improves the
sensing performance, but it does not contribute to the sensing
C. Point PCS –Achieving Strategy DoF in general. This can be observed from (12), where the
In contrast to that of achieving point PSC , the strategy term Je P converges to zeros as σs → 0.
of achieving point PCS is well-known in the literature [26]. The first equality (the best sensing DoF) in (36) is achieved
Specifically, we may simply employ the capacity-achieving when RX is deterministic. For the second equality (the worst
strategy in the communication-only scenario, which gives the case) to be achieved, some sufficient conditions are summa-
maximum achievable rate rized as follows.
Corollary 5 (Sufficient Conditions for Maximum Sensing
Rmax = max E{log I + σc−2 Hc R
e X HH }, s.t. (3),
c DoF Loss): The second equality in (36) is achieved if either
R
eX
of the following conditions holds:
e CS HH },
= E{log I + σc−2 Hc R (34)
X c 1) K ⩽ rank(R e X ), r1 + r2 = 1;
where the maximization with respect to R e X can be carried out 2) The covariance matrix R e X is invertible, and the matrix
using water filling for each realization of Hc [27, Sec. 9.4], F st = [F
e 1, . . . , F
er , G
e 1, . . . , G
er ]
1 2
and each column xi in the PCS –achieving X (denoted by
XCS following the circularly symmetric complex Gaussian is unitary.
Proof: Please refer to Appendix K. following the notations in (39), any sensing-optimal signal
By presenting Corollary 5, we hope that it may provide should take the following form
further intuitions regarding the sensing DoF, since in the worst √ 1
case we have precise knowledge about the sensing DoF loss. X = T U s Λs2 Q, (41)
In Section V-B we will see that, the task of target response where Q ∈ Cr×T is a semi-unitary matrix satisfying
matrix estimation is a typical example of scenarios where the QQH = I. Among these sensing-optimal signals, those
condition 2) in Corollary 5 is applicable. achieving the highest communication rate are given by Corol-
lary 3. By contrast, at point PCS , the communication-optimal
IV. D ISCUSSIONS signal is given by
In this section, we discuss the intuitions and implications 1
of the results presented in Section III. X = U c Λc2 D, (42)

where D ∈ Crank(Hc )×T contains i.i.d. standard circularly
A. S&C Tradeoff as a Two-fold Tradeoff: ST and DRT symmetric complex Gaussian entries.
A central implication of the results in Section III is that the Comparing (41) with (42), it can be observed that the
S&C tradeoff is two-fold, detailed as follows. sensing-optimal signal X sacrifices some communication
1) Subspace Tradeoff (ST): According to the discussions in DoFs, in the sense that the rows in Q are forced to have
Section III-B and Section III-C, at point PSC , any sensing- unit norms and to be orthogonal to one another. On the other
optimal R
e X has the following structure hand, as indicated by Theorem 2, the communication-optimal
X also sacrifices some sensing DoFs. We refer to this part of
e X = U s Λs U H ,
R (39) the S&C tradeoff as the deterministic-random tradeoff (DRT).
s
A noteworthy issue is that, when the coherent sensing period
where U s ∈ CM ×r contains the eigenvectors of the maximum-
T is sufficiently long compared to MSC , the DRT becomes
rank (= r) sensing-optimal RX corresponding to the non-
less prominent, since even a communication-optimal Gaussian
zero eigenvalues, while Λs is a positive semidefinite Hermitian
waveform X would have asymptotically orthogonal rows as
matrix that ensures R e X to be an optimal solution of the prob-
T /MSC → ∞ [30]. A related fact is that most existing
lem (14). In particular, when the sensing optimal Re X is unique,
contributions on the joint design of ISAC systems treat the
Λs is a diagonal matrix containing the non-zero eigenvalues
sample covariance matrix RX as a deterministic matrix [7],
of R
e X . By contrast, at point PCS , the communication-optimal
[31], [32], [33]. Such designs may now be interpreted as
RX takes the following form
e
neglecting the DRT, and hence can be viewed as infinite-T
e X = U c Λc U H ,
R (40) approximations to the original ISAC design problem.
c
3) ST and DRT in Terms of DoF: Next, we provide some
where U c ∈ CM ×rank(Hc ) contains the left-singular vectors quantified characterizations of the ST and DRT. Let us first
of the communication channel matrix Hc corresponding to take the point PSC as an example. According to (27), the
the non-zero singular values, while Λc is a diagonal matrix sensing-limited rate RSC in the high-SNR regime is given by
determined by the water-filling power allocation strategy.
From the discussion above, we may observe that the matrix
n MSC e SC HH | + c0 + O(σ 2 ),
o
RSC = E 1 − log |σc−2 Hc R X c c
U s characterizes the sensing subspace, while U c character- 2T
(43)
izes the communication subspace. Given a specific statistical
covariance matrix R e X , when its column space is more closely which implies that the
number of communication DoFs at
aligned with the sensing subspace, we may achieve a more the point PSC is MSC 1 − M2T
SC
. Hence we may define the
favorable sensing performance at the cost of degraded com- communication DoF efficiency at point PSC as
munication performance, and vice versa. This depicts a part of
the entire picture of the S&C tradeoff, which will be referred
MSC
ζSC := αSC 1 − , (44)
to as the subspace tradeoff (ST). 2T
2) Deterministic-Random Tradeoff (DRT): Given a specific where αSC = α(R e SC ) is the communication subspace overlap
X
Re X , the S&C tradeoff can be further adjusted by controlling
coefficient at point PSC , defined as
the “degree of randomness” of the signal X. To obtain a
higher communication rate, one should modulate as much e X HH )
rank(Hc R c
α(R
e X) = . (45)
information as possible onto the carrier signal, hence the rank(Hc HHc )
transmitted waveform should be “as random as possible”.
However, sensing systems prefer deterministic signals for e CS HH ), we see that
With the definition MCS = rank(Hc R X c
achieving a stable sensing performance. This has been an
αSC = MSC /MCS . (46)
intuitive insight consistent with both the engineers’ experience
and S&C signal processing theory developed during the past Observe that in general, the communication subspace overlap
few decades [29]. Indeed, Proposition 3 tells us that to achieve coefficient α(R e X ) takes value in [αSC , 1], which may be
the optimal sensing performance, the transmitted waveform viewed as an indicator of the ST. The DRT is quantified by
has to be deterministic to a degree, in the sense that its sample the term (1 − MSC /2T ) in (44), which is less than 1 as long
covariance matrix RX should be deterministic. Consequently, as T is finite.
Regarding the point PCS , we see from (36) that where α ∈ [0, 1] controls the preference between the sensing
n o and the communication performance. This is a convex opti-
T tr [Φ(R e X )]−1 mization problem, which may be efficiently solved by off-the-
E{tr [Φ(RX )]−1 } ⩽

, (47) shelf numerical solvers [34], [35]. Note that when α = 0, the
T − min{K, rank(R e X )}
problem (48) degenerates to the sensing-only problem (14).
which implies that the sensing DoF is lower bounded by T − An alternative formulation is given by
min{K, rank(R e X )}. Note that at PCS , we have rank(Re X) =
H max log I + σc−2 Hc R e X HH ,
c (49a)
rank(Hc RX Hc ). Thus when K ⩽ rank(RX ), we may
e e
RX
e
observe an interesting symmetry between PSC and PCS :
e H,
n o
the maximum sensing DoF loss at PCS is MCS , while the s.t. tr R e X = PT M, R e X ≽ 0, R eX = R
X (49b)
communication DoF loss at PSC is determined by MSC . When nh i−1 o
K > rank(R e X ), the maximum sensing DoF loss becomes K. tr ΦPT M (R e X) ⩽ ϵα , (49c)
Intuitively, we may interpret the ratio min{K, rank(R e X )}/T nh i−1 o
as the sensing load, which depicts the number of effective where ϵα is the value of tr ΦPT M (R e X (α)) , with
sensing parameters per unit time. As the sensing load tends RX (α) being the optimal RX of the problem (48) given a
e e
to zero (e.g. T → ∞), we see that the maximum sensing specific value of α.
DoF loss also converges to zero. This corroborates the intu- Once a specific statistical covariance matrix R e X (α) is
ition obtained from the discussion of PSC , namely, the DRT obtained from (48) or (49) for a given α, we can first obtain
becomes negligible in the large T regime, where pure Gaussian a natural outer bound of the CRB-rate region as follows:
signals are sufficient for achieving the Pareto-optimal S&C
Rout (α) = E log I + σc−2 Hc R e X (α)HH ,

performance (i.e. the boundary of the CRB-rate region). c (50a)
2 nh −1
σ i o
ϵout (α) = s tr ΦPT M (R e X (α)) . (50b)
T
B. Achievable Inner Bounds Connecting PSC and PCS
Next, we may decide the specific signalling strategy. One of
Besides PSC and PCS , other points on the boundary of the possible strategies is to transmit the following signal
the CRB-rate region are also of significant practical interest.
However, obtaining the genuine boundary would in general X = XG + XG,⊥ , (51)
require solving the following complicated functional optimiza- where XG has i.i.d. columns following the circularly sym-
tion problem: e X H† Hc ),
metric complex Gaussian distribution CN (0, H†c Hc R c
while XG,⊥ ∈ CM ×T is any matrix satisfying
min E tr [Φ(RX )]−1 − λ · I(Yc ; X|Hc )

† †
p(X)
T −1 XG,⊥ XH
G,⊥ + Hc Hc RX Hc Hc = RX .
e e (52)
s.t. tr {RX } = PT M, RX ≽ 0, RX = RH
X,
This leads to an achievable inner bound referred to as the
which is typically intractable in the vector Gaussian channel “Gaussian inner bound”, as follows:
scenario considered in this paper, due to the lack of explicit Rin,G (α) = Rout (α), (53a)
form of the mutual information I(Y c ; X|Hc ) and the com-
σ 2 n h −1 io
plexity of the CRB term. In light of this, in this subsection, ϵin,G (α) = s E tr Φ(RX (α)) . (53b)
we discuss two strategies that may be applied either separately T
or jointly, forming inner bounds of the actual boundary. Another possible strategy is given by
1) Time Sharing: In Section III, the time-sharing strategy √ 1
X = T H†c U eΛe 2 Q + XU,⊥ , (54)

has been applied to conceive the pentagon inner bound.
It can be generalized to any pair of given signalling schemes where Q ∈ CMU ×T contains the modulated data, which is
corresponding to a pair of points in the CRB-rate region, uniformly sampled from
by employing one of the schemes with probability p and MU ×T
VR
e X = {Q ∈ C |QQH = I MU },
the other with probability 1 − p. In general, once a specific
achievable inner bound is given, its convex envelope is also MU = rank(Hc R e X HH ), Λ
e ∈ CMU ×MU is a diagonal matrix
c
achievable with the aid of time sharing. e X HH , and
consisting of all non-zero eigenvalues of Hc R c
2) Statistical Covariance Shaping: It is also possible to H
adjust the S&C tradeoff by altering the statistical covariance e X HH = U
Hc R eΛeU
e .
c
matrix Re X . Specifically, given a realization of Hc , a flexible
The matrix XU,⊥ ∈ CM ×T is any matrix satisfying
tradeoff may be struck by solving the following optimization 1
problem XU,⊥ (H†c U
eΛe 2 Q)H = 0, (55a)
† †
T −1 XU,⊥ XH
i−1 o
U,⊥ + Hc Hc RX Hc Hc = RX . (55b)
nh e e
min (1 − α)tr ΦPT M (R e X)
R
This corresponds to another inner bound referred to as the
eX
− α log I + σc−2 Hc R
e X HH ,
c (48a) “semi-unitary inner bound”, characterized by
e H,
n o
s.t. tr R
e X = PT M, R e X ≽ 0, R eX = R (48b)
n MU
X Rin,U (α) = E 1 − log |σc−2 Hc R
e X (α)HH |
c
2T
o
+ c0,U + O(σc2 ), (56a) the sensing objective is the entire channel Hs . By contrast,
the maximum communication throughput of this scheme is
ϵin,U (α) = ϵout (α), (56b)
log2 M ! bits per transmission (i.e. each length-T block), which
where is typically lower than the sensing-optimal capacity that is
MU h MU T √ i asymptotically achieved by the uniform distribution over the
c0,U = T− log −log Γ(T )+log(2 π) . entire ensemble of M × T semi-unitary matrices, as detailed
T 2 e
(57) in Corollary 3. In light of this, we may conclude that the
Furthermore, we may obtain a refined inner bound on the IM-based scheme follows a design philosophy that aims for
basis of the Gaussian inner bound and the semi-unitary inner achieving the PSC point. However, due to its suboptimal
bound, by applying the time-sharing strategy. Specifically, this communication throughput, the scheme would actually achieve
inner bound, referred to as the “semi-unitary–Gaussian inner a point right below PSC in the CRB-rate region.
bound”, is constructed by computing the convex envelope of 2) Communication-Centric Designs: When an ISAC sys-
the union of the Gaussian inner bound and the semi-unitary tem is built upon commercialized communication networks,
inner bound. Each point on this refined inner bound is communication-centric designs are practically more attractive,
achievable by applying a suitable time-sharing strategy, which whose primary objective is to guarantee the communica-
interpolates between a signalling scheme adapted from the tion performance. Such designs typically rely on existing
Gaussian inner bound and another signalling scheme adapted communication-oriented protocol stacks, for example, Wi-Fi
from the semi-unitary inner bound. and 5G new radio (NR).
Regarding the tightness of the semi-unitary–Gaussian inner A representative approach to communication-centric designs
bound, we have the following result. is to exploit the widely-used orthogonal frequency-division
Proposition 6: When the optimal objective function value multiplexing (OFDM)-based waveform [39]. In the absence
of (49) is not identical for all α ∈ [0, 1], the semi-unitary– of the cyclic prefix (CP), both the communication channel
Gaussian inner bound is a tighter bound compared to the and the target response matrix can be represented by Toeplitz
pentagon inner bound obtained by connecting PSC and PCS . matrices in the following form:
Proof: Please refer to Appendix L. 
h1 0 ... 0

Proposition 6 suggests that, under mild technical assump-  .. .. 
tions, it would be beneficial to combine the time-sharing strat-  h2 h1 . . 
..
 
egy with statistical covariance shaping, rather than applying  . . 
 . h2 . 0 
them separately. 
.. ..

H =  hT , (60)
 
 CIR . . h1  
C. The Connection Between Existing ISAC Schemes and .
hTCIR . .
 
 0 h2 
PSC – & PCS –Achieving Strategies 
 .. .. .. ..


In this subsection, we will discuss how the design philoso-
 . . . . 
phy of current ISAC systems is related to different CRB-rate 0 ... 0 hTCIR
boundary-approaching strategies. where TCIR is the duration of the channel impulse response
1) Sensing-Centric Designs: Sensing-centric schemes are (CIR). When the CP is attached to the data symbols, and it is
typically implemented relying on existing sensing infrastruc- longer than the duration of the CIR, the channels can also be
tures, such as radars [36], [37], [38]. They are designed expressed in the form of circulant matrices (with M = Nc =
aiming for incorporating communication functionalities into Ns ), and hence could be diagonalized by the discrete Fourier
the system, without compromising the sensing performance. transform (DFT) matrices as follows
A representative sensing-centric design is based on index
modulation (IM) [37], [38], in which the communication Hc = U H
F,M Dc U F,M , (61a)
information is encoded into the index of the waveform chosen Hs = UH
F,M Ds U F,M , (61b)
from a preset codebook. More precisely, the transmitted signal
in the IM-based scheme designed for colocated multiple-input where Dc and Ds are diagonal matrices representing the
multiple-output (MIMO) radars takes the following form frequency-domain communication channel and the target
p response, respectively, while U F,M denotes the M -point DFT
X = PT PU , (58)
matrix. Before the transmission, the raw data symbols are
where U ∈ CM ×T represents the radar waveform code- pre-processed by the inverse DFT, and thus can be expressed
book, which is a deterministic semi-unitary matrix satisfying as
U U H = I M , while P is an M × M permutation matrix
that conveys the communication information. Observe that the X = UH
F,M XF , (62)
sample covariance matrix RX is given by
where each row of XF contains the data symbols modulated on
1 a single subcarrier. Each data symbol is typically chosen from
RX = XXH = PT I M , (59)
T a fixed constellation, for example, phase shift keying (PSK)
which is deterministic. As will be discussed in Section V-B, and quadrature amplitude modulation (QAM), according to the
this choice yields the optimal sensing performance when requirement of communication throughput.
Let us illustrate the DRT of the OFDM-based scheme by where Q ∈ CM ×T is a semi-unitary matrix and A ∈ CM ×M
considering a specific sensing task, namely that of estimating is a diagonal matrix being invariant under permutation of its
the CIR, which is a widely-used pre-processing procedure for diagonal entries, which is asymptotically proportional to an
delay estimation [39]. In this example, the sensing parameter identity matrix in the high-SNR regime [41]. The specific
is η = U HF,M mdiag(Ds ). Note that designs of the semi-unitary matrix have then been referred
T to as USTM schemes, which are related to coding on the
η = UH
F,M P K U vec(Hs ), (63) Grassmann manifold GT,M defined as the quotient space
2
where K U = U F,M ⊗ U ∗F,M , and P ∈ RM ×M is a matrix
2
between two Stiefel manifolds ST,M and SM,M under the
satisfying P vec(A) = mdiag(A). Thus the linear map Φ(RX ) equivalence relationship
is given by P = U Q ⇐⇒ P is equivalent to Q, P , Q ∈ ST,M ,
where U ∈ SM,M is a unitary matrix, and the Stiefel manifold
Φ(RX ) = U F,M P K H (I ⊗ RT H H
X )K U P U F,M
U ST,M is defined as
T ∗
= U F,M P I ⊗ U T H H
F,M RX U F,M P U F,M
n o
ST,M = Q ∈ CM ×T |QQH = I .
T ∗
= U F,M mat mdiag U T R U
F,M X F,M UH
F,M ,
Comparing (67) with (41), we see that the USTM and the
(64)
PSC –achieving signals are similar in form. Nevertheless, it is
hence the sensing-optimal sample covariance matrix should noteworthy that they are substantially different in the following
satisfy aspects:

T ∗ • Precoding: In the non-coherent communication scenario,
mat mdiag U T F,M RX U F,M = PT I M , (65)
the communication channel Hc is isotropic. Conse-
which implies that quently, the precoding matrix A in (67) is also isotropic
in terms of its diagonal entries. By contrast, in the PSC –
∥[XF ]i,: ∥2 = T PT , ∀i = 1, . . . , M. (66) √ 1
achieving strategy, the precoding matrix T U s Λs2 is
In other words, the transmit power should be equally allocated determined by the sensing-optimality constraint, which
to each subcarrier for achieving the optimal CIR estimation aligns the signal with the sensing subspace;
performance. In a typical communication-oriented OFDM • Reason of optimality: An intuitive interpretation for
system, the statistical covariance matrix would in general the optimality of USTM is provided in [42], which
be a diagonal matrix corresponding to the specific power states that in the high-SNR regime, Hc X should be
allocation strategy. When the power is equally allocated to entrywise i.i.d. Gaussian distributed for achieving the
each subcarrier, we have R e X = PT I M , since the data non-coherent channel capacity, which implies that X
symbols are independent of one another. However, the sample should take the form of (67). By contrast, in this
covariance matrix is not necessarily the same as R e X due to paper, the reason of using semi-unitary signals fol-
the randomness of the data. The issue could be resolved by lows from the sensing-optimality constraint, while the
applying PSK constellations, at the cost of a lower commu- communication-optimality is achieved by applying uni-
nication throughput, since the amplitude of the symbols does form sampling over Stiefel manifolds instead of coding
not carry information in PSK. By varying the power allocation over Grassmann manifolds.
strategy as well as the constellation (possibly accompanied by • Communication DoF: In the non-coherent communica-
error-correcting codes for rate adaptation), we may strike a tion scenario, it has been shown [42] that an alternative
flexible S&C tradeoff from point PSC to point PCS . optimal strategy is to estimate the channel using pilot
symbols before transmitting data symbols. This would
D. Semi-Unitary Signalling: The PSC –Achieving Strategy vs. impose a pilot cost of M Ns symbols since Hc has M Ns
Non-Coherent Communication unknown entries, and hence results in a communication
The PSC –achieving strategy discussed in Section III-B DoF loss of M 2 /T . This may also be inferred from the
bears some resemblance to the unitary space-time modulation fact that the dimensionality of the Grassmann manifold is
(USTM) [40] designed for communicating over non-coherent M (T −M ). By contrast, in this paper, the communication
channels. To elaborate, let us consider the scenario where the DoF is determined by the dimensionality of the Stiefel
communication channel Hc follows an entrywise i.i.d. zero- manifold, which is MSC (T − 12 MSC ). Consequently, the
M2SC
mean, unit-variance complex Gaussian distribution, namely communication DoF loss of 2T , which is half of the
vec(H c ) ∼ CN (0, I M Nc ), and is not known to the either the communication DoF loss in the non-coherent communi-
transmitter or the receiver. In this context, it has been shown cation scenario when MSC = M .
that the capacity is achieved by signals satisfying5
X = AQ, (67) V. C ASE S TUDY
A. Target Angle Estimation
5 For simplicity of discussion, in the rest of this subsection we assume
T ⩾ 2M . When this is not the case, the analysis becomes tedious and less Let us first consider the scenario, where the sensing task is
relevant to this paper. Interested readers are referred to [41] and [42]. to estimate the angles of targets using a MIMO radar equipped
with co-located antennas. In this scenario, the target response where M = E{M∗ (θ)}, and
matrix Hs admits the following parametrization [43]:
M(θ) = ∥ȧ(θ)∥2 v(θ)vH (θ) + ȧH (θ)a(θ)v̇(θ)vH (θ)
NT
Hs =
X
αn a(θn )vT (θn ), (68) + aH (θ)ȧ(θ)v(θ)v̇H (θ) + ∥a(θ)∥2 v̇(θ)v̇H (θ). (76)
n=1
We may simplify the expression of M(θ) by choosing the
where NT represents the number of targets, θn denotes the phase reference point of the transmitting and receiving arrays
angle of the n-th target relative to the radar, αn denotes the such that v̇H (θ)v(θ) = 0 and ȧH (θ)a(θ) = 0, and hence
complex amplitude of the echo received from the n-th target,
while v(·) and a(·) denote the mappings from the angle to M(θ) = ∥ȧ(θ)∥2 v(θ)vH (θ) + ∥a(θ)∥2 v̇(θ)v̇H (θ). (77)
the steering vectors of the transmitting and receiving antennas, 1) Sensing DoF: According to Theorem 2, the sensing DoF
respectively. We assume the transmitting and receiving antenna in the single-target angle estimation scenario at point PCS is
arrays are conjugate symmetric. For the simplicity of discus- lower-bounded by
sion, we consider the case of NT = 1, which corresponds to
the parameter vector νs ⩾ T − K = T − 1. (78)
η = [θ, Re{α}, Im{α}]T , In particular, when the communication channel Hc is rank-

1 (e.g. line-of-sight (LoS) MIMO channel), the lower bound
where θ = θ1 , α = α1 . According to [43], we have in (78) is achieved. To elaborate, note that when Hc is rank-
e CS is also
 
Re{f11 } Re{f12 } −Im{f12 } 1, the communication-optimal covariance matrix R X
Jη|X = 2  Re{f12 } Re{f22 } −Im{f22 }  + J P , rank-1.
In this case, since we have RX = RCS X , the term
−Im{f12 } −Im{f22 } Re{f22 } tr M RX in (75) is given by
(69) T
PT M r H M r X
|ni |2 ,

tr M RX = (79)
where T i=1
T |α|2 2
n
H T
o
f11 = ∥ȧ(θ)∥ tr v(θ)v (θ)R X where r is the eigenvector corresponding to the maxi-
σs2 e CS , and ni ’s are mutually independent
n n oo mum eigenvalue of R X
+2Re ȧH (θ)a(θ)tr v̇(θ)vH (θ)RT X zero-mean circularly symmetric complex Gaussian random

2
n
H T
o variables with unit variance. Hence tr M RX is proportional
+∥a(θ)∥ tr v̇(θ)v̇ (θ)RX , (70)
to a chi-squared distributed random variable having DoF 2T ,
T α∗ n o which implies that
f12 = 2 ȧH (θ)a(θ)tr v(θ)vH (θ)RT X
σs n o −1 T
CS
E tr M RX = o. (80)
n o
+∥a(θ)∥2 tr v(θ)v̇H (θ)RT , (71) e CS
n
X (T − 1)tr M R X
T n o
f22 = 2 ∥a(θ)∥2 tr v(θ)vH (θ)RT X , (72) Thus the sensing DoF can be calculated as
σs
n −1
σs2 JθP
o
where ȧ(θ) = ∂a(θ)/∂θ, and v̇(θ) = ∂v(θ)/∂θ. Let us TE CS
tr M RX + T E{|α|2 }
assume that the prior distributions of θ, Re{α} and Im{α} νs = lim −1
are independent of one another, and hence σs →0
e CS + σs Jθ 2
2 P
n o
tr M R X T E{|α| }
J P = diag JθP , JRe{α}
P P
, JIm{α} . (73) = T − 1, (81)
Furthermore, we assume that the complex amplitude α is which indeed achieves the lower bound in (78).
circularly symmetric, which implies that E{α} = E{α∗ } = 0, 2) Communication DoF: Observe from (75) that, at point
P P
and that JRe{α} = JIm{α} = JαP . Since we are only interested PSC , the columns of RSCX should be spanned by the subspace
in the angle θ, we consider its equivalent BFIM (by treating corresponding to the largest eigenvalue of M . Therefore,
α as a nuisance parameter) given by [44] when the largest eigenvalue of M has multiplicity 1, the
sensing-optimal sample covariance matrix RSC X is unique and
Je (θ) = 2E{f11 } + JθP
rank-1. In this case, the high-SNR sensing-limited capacity is
∗
− 4E{f12 }(2E{f22 } + JαP )−1 E{f12 }, (74) asymptotically given by

which equals to the first diagonal entry (corresponding to θ) 2T − 1 −2 SC H
in J−1 RSC = E log |σc Hc RX Hc | + c0 + O(σc2 ),
e
η|X . Using the assumption that α is circularly symmetric, 2T
we obtain that E{f12 } = 0, and thus (82)
Jeθ|X = 2E{f11 } + JθP
when the column space of RSC e SC
X (= RX ) is not orthogonal to
2T E{|α|2 } the column space of Hc , otherwise the capacity is zero. We see
= tr M RX + JθP , (75)
σs2 from (82) that the communication DoF is (2T − 1)/2T , and
hence the communication DoF loss induced by the DRT is where we have used the base-2 logarithm to ensure that the
1/2T . rate is in the unit of bit per channel use (bpcu). By contrast,
To further understand the S&C tradeoff in the task of when we apply the power allocation strategy to the Gaussian
target angle estimation, let us consider a simplistic scenario, signal, we obtain the following inner bound
where the communication receiver is equipped with a single
Rin,G (λ) = log2 (1 + ∥hc ∥−2 Nc |r H (λ)hc |2 SNRc ), (90a)
antenna. We further assume that the communication channel n −1 o
is deterministic and denoted as hc , hence the (per antenna) ϵin,G (λ) = E χ22T SNRs r H (λ)M r(λ) + JθP
maximum sensing and communication receiving SNR may be
computed as = (2SNRs r H (λ)M r(λ))−1 ζ T −1 eζ Γ(1 − T, ζ)
−1
SNRs = M PT E{|α|2 }σs−2 , (83a) = 2(T −1)SNRs r H (λ)M r(λ) (1+rζ ), (90b)
SNRc = M PT ∥hc ∥2 σc−2 , (83b)
where χ22T is a chi-squared
R ∞ distributed random variable with
respectively. Furthermore, we assume that the maximum DoF 2T , Γ(a, x) = x ta−1 e−t dt denotes the incomplete
eigenvalue of M has multiplicity 1. Since Nc = 1, we now Gamma function [45], and ζ = JθP (2SNRs r H (λ)M r(λ))−1 .
see that both the sensing-optimal and the communication- The correction term rζ is on the order of O(ζ), given by
optimal sample covariance matrices are rank-1. Therefore, the T −2
communication subspace is spanned by hc , while the sensing
X (−1)n ζ n eζ ζ T −1 Γ(0, ζ)
rζ = Qn + (−1)T −1 · ,
subspace is spanned by the sensing-optimal steering vector us n=1 i=1 (T − i − 1) Γ(T − 1)
characterized by
| {z }
O(ζ T −1 log ζ)
M us = λ1 {M }us , (84) which can be derived from [45, Sec. 8.352]. We can also
where λ1 (·) denotes the maximum eigenvalue of its argument. obtain another inner bound by applying the power allocation
In other words, us is the eigenvector corresponding to the max- strategy to the semi-unitary signal (a constant-norm vector
imum eigenvalue of M . In this scenario, the overlap between in this scenario, since the sensing-optimal sample covariance
the communication subspace and the sensing subspace can be matrix is rank-1)
depicted by the following quantity 2T − 1
Rin,U (λ) = log2 (Nc ∥hc ∥−2 |r H (λ)hc |2 SNRc )
2T
hHc M hc
ρ= . (85) + c0 + O(σc2 ), (91a)
∥hc ∥2 λ1 {M } −1
Under the aforementioned assumptions, besides the pen- ϵin,U (λ) = 2T SNRs r H (λ)M r(λ)} + JθP , (91b)
tagon inner bound obtained by employing time-sharing strate- where the correction term c0 is given by
gies between PSC and PCS , we can also obtain some refined
1 h 1 T √ i
inner bounds of the CRB-rate region using the statistical c0 = T− log −log Γ(T )+log(2 π) ,
covariance shaping method discussed in Section IV-B, as well T 2 e
2
as an outer bound. Specifically, the statistical covariance and the O(σc ) residual term can be computed numerically.
shaping problem now takes the following form Moreover, we can obtain a tighter inner bound, by employ-
n o ing time-sharing strategies between the Gaussian signal and
max tr M R e X + λhH R e X hc
R
eX
c the semi-unitary signal. This inner bound can be obtained by
numerically computing the convex envelope of the union of
e H,
n o
s.t. tr Re X = PT M, R e X ≽ 0, ReX = R
X (86) the regions determined by (90) and (91).
where λ ∈ [0, +∞) controls the preference between the
sensing and the communication performance. The solution to B. Target Response Matrix Estimation
this problem can be expressed as follows: Next, let us consider the task of target response matrix
e X (λ) = PT M r(λ)r H (λ), estimation commonly seen in the literature of statistical MIMO
R (87)
radar [46], where the sensing objective is to estimate the entire
where r(λ) is the eigenvector corresponding to the largest matrix Hs . More precisely, we have
eigenvalue of the matrix h iT
η = Re{vec(HT s )}T
, Im{vec(H T T
s )} . (92)
M (λ) = M + λhc hH
c . (88)
We further assume that the a priori distribution of each entry
By employing this Pareto-optimal power allocation strategy
in Hs is identically CN (0, σp2 ), and hence
and ignoring the DRT, according to (75), we obtain an outer
bound characterized as follows: 2
e P = σs I 2N M .
J P = σp−2 I 2Ns M , J 2 s
λ ∈ [0, +∞), (89a) σp T
−2 H 2
Rout (λ) = log2 (1 + ∥hc ∥ Nc |r (λ)hc | SNRc ), (89b) Note that
−1 1
ϵout (λ) = 2T SNRs r H (λ)M r(λ) + JθP , (89c) η = √ U Had hs , (93)
2
where hs := [vec(HT T T H T
s ) , vec(Hs ) ] , and U Had is a unitary
matrix given by

1 1 1
U Had = √ ⊗ I Ns M .
2 ı −ı
In light of this, the affine map Φ(·) can be expressed explicitly
as

T
Φ(RX ) = 2U H Had I Ns ⊗ blkdiag RX , RX U Had
2σs2
+ I 2Ns M . (94)
σp2 T
Thus the CRB of η is given by
2
σ
ϵ = tr E s [Φ(RX )]−1
T Fig. 4. The minimum achievable CRB ϵCS and its Marchenko-Pastur approx-
( " −1 #) imation (100) versus the a priori precision σp−2 , where RCS
X ∼ CW M (I, T ),
2
σ s Ns σs2
= tr E RX + 2 I , (95) σs2 = 1, and β = M/T is fixed at 1/4.
T σp T
which follows from the fact that the unitary matrix U Had
preserves trace, and that the transpose operation preserves where z = σs2 (σp2 T )−1 . As it can be observed from Fig. 4,
trace. Next, let us consider the sensing and communication this limit serves as an excellent approximation even when M is
performance at points PSC and PCS , respectively. rather small (e.g. M = 2). In the more general scenario RCS
X ∼
1) Sensing DoF: At point PSC , the optimal R
e X is obtained CW M (Σ, T ), if the eigenvalues of Σ are upper and lower
SC bounded by some positive constants as M → ∞, we have
at RX = PT I M . Moreover, we have
e
−1 #) −1 ) M
( " (
σs2 2 σs2 Ns X
r
σ s z z z
tr E RX + 2 I = tr R
eX + I ϵCS → β − −1+ ( +β −1)2 + 4 ,
σp T σp2 T 2T zβ i=1 σi σi σi
M (101)
= , (96)
PT + σs2 (σp2 T )−1 where σi is the i-th largest eigenvalue of Σ.
since the sensing-optimal sample covariance matrix RX in this As for the sensing DoF, observe that
scenario is deterministic. Hence the minimum CRB can be ( −1 )
2
expressed as −1 2 CS σ s
νs,CS = lim ϵCS σs Ns tr R e +
X I (102a)
σs →0 σp2 T
σs2 Ns M
ϵmin = . (97) n CS o n h io
T PT + σs2 σp−2 = T tr (R e )−1 /tr E (RCS )−1
X X (102b)
The sensing DoF at point PSC is clearly νs,max = T . Observe = T − M, (102c)
that the minimum CRB may be alternatively expressed as
CS
Ns M 1 when Re
X has full rank (otherwise the sensing DoF is zero),
ϵmin = · −2 , (98)
T PT σs + (σp2 T )−1 where (102c) follows from the fact that
which implies that the a priori knowledge contributes to an h
−1
i T e CS )−1 .
E (RCSX ) = (R (103)
additional effective SNR of (σp2 T )−1 . T −M X
By contrast, at point PCS , the minimum achievable CRB is
Hence we may conclude that the sensing DoF loss in the task
given by
of target response matrix estimation is
( " −1 #)
σs2 Ns CS σs2
ϵCS = tr E RX + 2 I , (99) νs,max − νs,CS = M, (104)
T σp T
which achieves the upper bound indicated by Theorem 2.
where RCS X is a complex Wishart distributed matrix having
2) Communication DoF: At the point PCS , the maximum
degree of freedom T and scale matrix R e CS . When RCS
X X achievable rate is given by the channel capacity in the absence
follows the complex Wishart distribution CW M (I, T ), it is of sensing constraints, as follows:
known that the eigenvalue distribution of CW M (I, T ) con-
verges to the Marchenko-Pastur distribution as M → ∞, e CS HH |},
Rmax = E{log |I + σc−2 Hc R (105)
X c
if β = M/T is a constant [47]. In the same asymptotic limit,
CS
according to the Stieltjes transform of the Marchenko-Pastur where R X is determined by the water-filling strategy. The
e
distribution [47], the CRB ϵCS satisfies communication DoF at the point PCS is MSC .
σ 2 Ns M p As for the point PSC , according to Proposition III-B, when
ϵCS → s (β − z − 1 + (z + β − 1)2 + 4z), (100) the SNR is high, the corresponding achievable rate may be
2T zβ
asymptotically expressed as TABLE I

C ONFIGURATIONS FOR THE N UMERICAL E XAMPLES
n MCS o
OF THE TARGET A NGLE E STIMATION P ROBLEM
RSC = E 1 − log |PT σc−2 Hc HH 2
c | + c0 + O(σc ),
2T
(106)
where we have used the fact that
e SC HH ) = rank(Hc HH ) = MCS
MSC = rank(Hc R X c c
holds in the high-SNR regime. In this case, the communication

DoF is
MSC
νc,SC = MSC 1 − . (107) e CS is rank-
2T Gaussian signals. To elaborate, in this case, since R X
The corresponding communication DoF loss at point PSC is deficient, the sensing DoF is zero. In particular, at point PCS ,
the ISAC signal cannot provide any information about the
M2SC target response matrix in the null space of R e CS , and hence
νc,max − νc,SC = . (108) X
2T we have to rely entirely on the a priori knowledge when
Observe that the communication subspace overlap coefficient estimating this part of the target response matrix. As we move
αSC achieves its maximum 1, hence the sensing-induced from PCS to PSC on the outer bound, R e X (α) becomes full-
communication performance loss is mainly due to the rank. However, the power allocated to the null space of R e CS
X
row-orthogonality of the sensing-optimal coding strategy. does not contributed to the communication rate. Therefore,
Similar to the target angle estimation problem, we can we may transmit semi-unitary signals instead of Gaussian
also conceive an outer bound based on statistical covariance signals in the null space, without eroding the communication
shaping, namely solving the following Pareto optimization rate at all.
problem [33]
n σ 2 −1 o VI. N UMERICAL R ESULTS
min tr RX + 2s I +α log I +σc−2 Hc RX HH c
RX σp T A. Target Angle Estimation
s.t. tr {RX } ⩽ PT M, RX ≽ 0, RX = RH
X, (109) We first demonstrate the S&C tradeoff in the task of target
angle estimation discussed in Section V-A. We consider the
for each realization of Hc , where α ∈ [0, ∞) is a parameter scenario where the sensing Rx and the ISAC Tx are co-
controlling the power allocation strategy. Upon denoting the located, both equipped with uniform linear arrays. There is
optimal solution to (109) as R e X (α) and the corresponding
a single target with a bearing angle of θ, for which the
sample covariance matrix as Rα , we could obtain the Gaussian a priori distribution is a von Mises distribution with mean 30◦
inner bound and standard deviation of 5◦ . The communication channel is
assumed to be a rank-1 LoS channel. The configurations that
n o
Rin,G (α) = E log2 I + σc−2 Hc R e X (α)HH ,
c (110a)
are identical across all numerical examples are summarized
σ 2 Ns n h −1 io in Table I. For all numerical examples in this subsection,
ϵin,G (α) = s tr E RX (α) + σs2 (σp2 T )−1 I ,
T we observe that the maximum eigenvalue of M has multi-
(110b) plicity 1, and thus the discussions in Section V-A are fully
applicable.
the semi-unitary inner bound Let us first fix the bearing angle of the communication Rx at
n Mα θc = 42◦ , which corresponds to a correlation coefficient of ρ ≈
Rin,U (α) = E 1 − log2 I + σc−2 Hc R e X (α)HH
c
2To 0.61, and consider the case of T = 3. The outer bound (89),
+ c0 + O(σc2 ), (111a) the Gaussian inner bound (90) and the semi-unitary inner
bound (91), as well as the refined inner bound obtained by
2

σ Ns −1
ϵin,U (α) = s tr Re X (α) + σ 2 (σ 2 T )−1 I , employing the time-sharing strategy between the two inner
s p
T bounds, are portrayed in Fig. 5. We have numerically com-
(111b) puted the rate of the semi-unitary signalling strategy, accom-
panied by an approximation obtained by neglecting the O(σc2 )
and the outer bound term in (91). Observe that the approximation error is rather
Rout (α) = RinG (α), ϵout (α) = ϵinU (α), (112) small, since the SNR is high. A noteworthy phenomenon is
that the semi-unitary–Gaussian inner bound is tighter than the
following a similar line of reasoning as (89), (90) and (91), PSC –PCS inner bound. An intuitive interpretation is that the
where Mα = rank(Hc R e X (α)HH ), and the expectation semi-unitary–Gaussian inner bound benefits from amalgamat-
c
in (110b) may be approximated using (101). ing the time-sharing strategy with power allocation, whereas
Remark 1 (Null Space Completion by Semi-Unitary Sig- the PSC –PCS inner bound only employs the time-sharing
nals): According to (51), when MCS < M , the Gaussian strategy. The vertical dashed line and the horizontal dotted line
inner bound–achieving signalling scheme does not send pure depict the ultimate loss of sensing DoF and communication
Fig. 5. Inner and outer bounds of the CRB-rate region for the task of Fig. 7. Comparison between the inner bounds of the CRB-rate region for
single-target angle estimation. The correlation coefficient is ρ ≈ 0.61. the task of single-target angle estimation, with various values of ρ.
Fig. 6. The outer bound and the unitary-Gaussian time-sharing inner bound Fig. 8. The outer bound and the unitary-Gaussian time-sharing inner bound
of the CRB-rate region for the task of single-target angle estimation, with of the CRB-rate region for the task of single-target angle estimation, with
various values of ρ. various values of T .
DoF caused by employing the Gaussian and the semi-unitary

signalling strategies, respectively. corresponds to the optimal adjustment of ST. This is portrayed
By varying the bearing angle of the communication Rx in Fig. 7, where the curved shape of the naïve time-sharing
(hence the value of the correlation coefficient ρ), we may inner bound is due to the logarithmic scale of the abscissa.
observe the dependence of the ISAC integration gain on the We may observe from Fig. 7 that the gap between these two
overlap between the sensing subspace and the communication inner bounds grows as the correlation coefficient ρ decreases.
subspace, as portrayed in Fig. 6. In this example we also This suggests that the performance gain of statistical covari-
fix the coherent sensing period at T = 3. We see that as ρ ance shaping over naïve time-sharing is larger under weaker
increases, both the semi-unitary–Gaussian inner bound and the S&C subspace correlation.
outer bound tend to become closer to the rectangular boundary, Next we demonstrate the dependence of the CRB-rate region
which corroborates the intuition that the ISAC integration gain on the length T of the coherent sensing period. In this example
should be higher when sensing subspace has a larger overlap we fix the bearing angle of the communication Rx at θc = 50◦ ,
with the communication subspace. A noteworthy fact is that which corresponds to a correlation coefficient of ρ ≈ 0.22.
the inner bound is not rectangular when ρ ≈ 1. In this case, the As it can be observed from Fig. 8, as T increases, the
sensing subspace is identical to the communication subspace, gap between the semi-unitary–Gaussian inner bound and the
and hence the S&C tradeoff is completely determined by the outer bound vanishes. This suggests that the DRT becomes
DRT. The two corner points of the inner bound correspond to less prominent when the coherent sensing period is long.
the sensing-optimal and the communication-optimal signalling Ultimately, it would become completely irrelevant to the S&C
strategies, respectively. tradeoff as T → ∞.
In fact, the gap between the semi-unitary–Gaussian inner Finally, let us investigate the relationship between the
bound and the naïve PSC –PCS time-sharing inner bound CRB-rate region and the number of antennas. In this example
may be viewed as the room of improvement upon the naïve we choose θc = 42◦ . The maximum sensing receiving SNR
time-sharing relying on statistical covariance shaping, which is (10 + 10 log10 M ) dB per antenna, while the maximum
Fig. 9. The inner and outer bounds of the CRB-rate region for the task of Fig. 10. Inner and outer bounds of the CRB-rate region for the task of target
single-target angle estimation, with various values of M and Ns . The bearing response matrix estimation, where the communication channel is subject to
angle of the communication Rx is θc = 42◦ . spatially uncorrelated Rayleigh fading.
TABLE II
C ONFIGURATIONS FOR THE N UMERICAL E XAMPLES OF
THE TARGET R ESPONSE M ATRIX E STIMATION P ROBLEM
communication receiving SNR is (23 + 10 log10 M ) dB per

antenna. These parameter values are chosen to be consistent
with the example shown in Fig. 5. The inner and outer bounds
obtained under various values of M and Ns are portrayed in
Fig. 9. Observe that as M and Ns increase, the gap between the Fig. 11. Inner and outer bounds of the CRB-rate region for the task of target
semi-unitary–Gaussian inner bound and the naïve time-sharing response matrix estimation, where the communication channel is a rank-1 LoS
channel.
inner bound expands, and the rate at point PSC decreases.
These phenomena originate from the fact that the sensing and
communication subspaces become asymptotically orthogonal
illustration, we normalize the CRB as follows
to each other as M and Ns tend to infinity, which may be
intuitively interpreted as that the resolution of the system ϵ
ϵnormalized = .
improves as the number of antennas increases. This may M Ns
also be quantitatively depicted by the value of the correlation
coefficient ρ, which decreases from ρ ≈ 0.82 to ρ ≈ 0.61 as In this example, we set the coherent sensing period as T =
M = Ns increases from 7 to 10. Consequently, the room of 4M = 16. The outer bound and the inner bounds of the
improvement due to ST adjustment also expands as M and Ns CRB-rate region are portrayed in Fig. 10, accompanied by
increase, as suggested by the previous discussion about Fig. 7. the approximated Gaussian inner bound computed according
to (101). Observe that the outer bound is rather close to
a rectangle in this scenario, which is reminiscent of the
B. Target Response Matrix Estimation ρ = 1 scenario in the target angle estimation task. We may
Next, let us consider the task of target response matrix esti- understand this phenomenon intuitively, by noticing that both
mation discussed in Section V-B. We assume that the a priori the communication channel and the target response matrix
distributions of each entry in the target response matrix Hs have full rank, and that the communication-optimal water-
are independent of one another, and are identically given by filling strategy becomes similar to the sensing-optimal uniform
CN (0, 1), which implies that σp2 = 1. The configurations of power allocation strategy in the high SNR regime.
other parameters are summarized in Table II. To further validate this intuition, let us consider another
We first consider the scenario where the communication scenario, where the communication channel is a rank-1 LoS
channel is subject to spatially uncorrelated Rayleigh fading, channel taking the form of
namely that each entry in Hc is independent of one another,
and follows the distribution of CN (0, M −1 ). For the sake of Hc = αav T ,
where a ∈ CNs and v ∈ CM are the steering vectors of whose BFIM can be expressed as
the communication receiver and the transmitter, respectively,
J θ = σs−2 I 2Ns T . (114)
in which every entry has amplitude 1. Note that in this
example, the Gaussian inner bound is achieved by the null Hence the BFIM of hs := [vec(HT T
s ) , vec(HT H T
s ) ] can be
space completing technique discussed in Remark 1. Indeed, written as
here we have MCS = 1 < M . Consequently, the null space ∗ T
∂θ ∂θ
has a dimensionality of M − 1, in which semi-unitary signals Jhs |X = Ehs Jθ
are transmitted. We assume that the amplitude α is a random ∂hs ∂hs
I Ns ⊗ RT

variable following CN (0, 1), and set the coherent sensing T X 0
= 2 , (115)
period as T = M = 4. From the CRB-rate region plotted in σs 0 I N s ⊗ RX
Fig. 11 we may see that the boundaries are obviously non- ∂h∗
according to the convention in [48]. Let us denote F := ∂η
s
∈
rectangular. This corroborate the aforementioned intuition,
since the communication subspace in this scenario is only one- CK×2Ns M , and partition F in the following form
dimensional. Consequently, the overlap between the sensing F = [F1 , . . . , F2Ns ],
and the communication subspaces is significantly smaller than
the one in the previous scenario of the spatially uncorrelated where each Fi has a dimensionality of K × M .
Rayleigh channel. Now, taking the contribution of the prior distribution back
into account, we have
I Ns ⊗ RT

VII. C ONCLUSION T 0
Jη|X = 2 Eη F X FH +J P
σs 0 I N s ⊗ RX
In this paper, we have proposed a general framework for )
X Ns
the analysis of the S&C performance tradeoff. In particular, T T H H
= 2 Eη Fi RX Fi +FNs +i RX FNs +i +J P . (116)
we have introduced the CRB-rate region for characterizing the σs i=1
S&C tradeoff, and have shown that it has a pentagon inner
bound obtained by connecting the communication-optimal Let us define a pair of linear super-operators Φ1 (·) and Φ2 (·),
point PCS and the sensing-optimal point PSC . Especially for characterized by
PSC , we have shown that it is achieved when the transmitted Ns
X
waveform has a sample covariance matrix with deterministic Φ1 (A) = Fi AFH
i , (117a)
trace. In particular, when the CRB minimization problem has a i=1
unique solution, the sample covariance matrix is deterministic, XNs
and hence the ISAC waveform should be modulated by sym- Φ2 (A) = FNs +i AFH
Ns +i . (117b)
bols generated based on the uniform distribution of the Stiefel i=1
manifold. For PCS , we have shown that the maximum sensing Without loss of generality, in the following discussion, we will
DoF loss is determined by the number of sensing parameters focus on Φ1 (·). Observe that the following matrix
K, as well as the rank of the statistical covariance matrix M X
X M
Re X . We have also conceived signalling strategies based on Ψ1 = E ij ⊗ Φ1 (E ij ) (118)
statistical covariance shaping, which are shown to be capable i=1 j=1
of achieving more favorable S&C tradeoffs than the pentagon
is a faithful representation of Φ1 (known as the Choi represen-
inner bound when applied in conjunction with time sharing.
tation, according to the Choi-Jamiołkowski isomorphism [49],
Based on these analytical results, we have highlighted that
[50]), where E ij is an M × M matrix with its (i, j)-th
the S&C performance trade-off is two-fold, in the sense that
entry equals to 1, while all other entries are zero. According
it is determined by the subspace alignment of the transmitted
to (117a), we have
waveform in general, and also by the deterministic-random
Ns
tradeoff in the finite coherent sensing period regime, which X
may shed light on the design of practical Pareto-optimal ISAC Ψ1 = vec(Fi )vec(Fi )H , (119)
i=1
signalling strategies.
which implies that
A PPENDIX A Ns
X
E vec(Fi )vec(Fi )H .

e 1 := E{Ψ1 } =
Ψ
P ROOF OF P ROPOSITION 2
i=1
Proof: The sensing model (2b) may be rewritten as
Consider the eigendecomposition of Ψ
e 1 as follows
vec(YT T T T
s ) = (I Ns ⊗ X )vec(Hs ) + vec(Zs ). (113)
e 1 = U (1) Λ(1) (U (1) )H
Ψ
For the moment, let us ignore the contribution of the prior r1 q q H
(1) (1) (1) (1)
X
distribution pη (η) to the BFIM. To facilitate further derivation, = λi ui λi ui (120)
we define an extended parameter vector i=1
where
(I Ns ⊗ XT )vec(HT

s )
θ := , (1) (1) (1) (1)
(I Ns ⊗ XH )vec(HHs ) U (1) = [u1 , . . . , uKM ], Λ(1) = diag(λ1 , . . . , λKM ).
(1) (1) (1)

We assume that λ1 ⩾ λ2 ⩾ . . . ⩾ λKM , and the number As for the term ΦP (γ, A), we note that its Choi representation
of non-zero eigenvalues is denoted as r1 . Since there exist a is given by
linear one-to-one correspondence between every realization of
e 1 = E{Ψ1 }
Φ1 and its Choi representation Ψ1 , we see that Ψ e P (γ) = 1 I ⊗ J P .
Ψ (130)
γ
is also the Choi representation of E{Φ1 (A)}, and thus
Now, we may write the Choi representation of Φγ (A) as
r1
X
e H, follows
E{Φ1 (A)} = F
e i AF
i (121)
i=1 Ψ
eγ = Ψ
e 1T + Ψ
e2 + Ψ
e P (γ), (131)
where r1 is the rank of Ψ
e 1 , and where Ψ
e 2 denotes the Choi representation of Φ2 (A), given
q by
e i = λ(1) mat(u(1) ).
F (122) r2
i i X
Ψ
e2 = vec(G e j )H .
e j )vec(G (132)
Similar arguments also apply to Φ2 (·), and hence we have
j=1
r2
e H,
X
E{Φ2 (A)} = G
e i AG (123) Note that Ψ e γ is not Hermitian, since Ψ e 1T and Ψ e P are
i
i=1 not. Therefore, the eigendecomposition in (120) is no longer
applicable, and we have to resort to the singular value decom-
where
q position of Ψeγ
(2) (2)
G
ei = λi mat(ui ), (124) e γ = U γ Σγ V H,
Ψ γ (133)
(2) (2) where U γ = [u1 , . . . , uKM ], V γ = [v 1 , . . . , v KM ],
under the corresponding definitions of λi and ui . Conse-
quently, we obtain and Σ γ = diag(σ1 , . . . , σKM ). We may then obtain the
r1 r2
representation in (13), where
T X eH +
e i RT F
X
e H + J P , (125)

√ √
Jη|X = F X i G
e j RX G
j F̄ i = σi mat(ui ), Ḡi = σi mat(v i ). (134)
σs2 i=1 j=1
e ∈ CKM ×KM .
Finally, we have r3 ⩽ KM , since Ψ
Finally, it is obvious that r1 ⩽ KM and r2 ⩽ KM , since
Ψ
e 1 and Ψ e 2 are of the size KM × KM . Hence the proof is
A PPENDIX C
completed. P ROOF OF P ROPOSITION 3
Proof: The problem of finding the distribution for the
A PPENDIX B sensing-optimal RX may be formulated as
P ROOF OF C OROLLARY 1
min ERX {tr [Φ(RX )]−1 },

(135a)
Proof: First let us consider the following decomposition p(RX )
Φγ (A) = Φ1 (ΦT (A)) + Φ2 (A) + ΦP (γ, A), (126) s.t. tr {E(RX )} = PT M, RX ≽ 0, RX = RH

X, (135b)
Observe that
where
min E{tr [Φ(RX )]−1 }

M
M X
X p(RX )
ΦT (A) = AT = E ij AE H
ji , (127a) n o
i=1 j=1 = min E E tr [Φγ (RX )]−1 |tr {RX } = γ
p(RX )
r1
e H,
X n o
min tr [Φγ (RX )]−1

Φ1 (A) = F
e i AF
i (127b) ⩾ min E (136)
p(γ) tr{RX }=γ
i=1
r2
holds under the constraints (135b), when the following deter-
e H,
X
Φ2 (A) = G
e i AG
i (127c) ministic optimization problem
i=1 −1
tr{A} min tr Φγ (R) (137a)
ΦP (γ, A) = J P. (127d) R
γ
s.t. tr {R} = γ, R ≽ 0, R = RH (137b)
From (127a) and (127b), we have
r1 X
M X
M has valid optimal solutions for every positive γ. Furthermore,
eH we have
X
Φ1 (ΦT (A)) = e i E jk AE H F
F kj i , (128)
i=1 j=1 k=1 Ef (γ) ⩾ f (E{γ}), (138)
which has the following Choi representation where
r1 X
M X
M
f (γ) := min tr (Φγ (R))−1 , s.t. tr {R} = γ, R = RH .
X
Ψ
e 1T = vec(F e i E kj )H .
e i E jk )vec(F (129) R≽0
i=1 j=1 k=1 (139)
The result (138) follows from the following lemma. is orthogonal to the null space of Ξ. Moreover, we see
Lemma 1: The function f (γ) is convex. that dim(V) = r2 − 1, since the complex vector vec(D)
Proof: Consider the optimal R1 corresponding to f (γ1 ) is linearly independent of its complex conjugate vec(D)∗ ,
and the optimal R2 corresponding to f (γ2 ), which satisfy and the zero-trace constraint tr {D} = 0 only reduces the
dimensionality by 1. Furthermore, vec(I) is never in the
αf (γ1 ) + (1 − α)f (γ2 ) ⩾ tr [Φ(αR1 + (1 − α)R2 )]−1

null space of Ξ (otherwise we would have Φ(I) = 0
(140) implying that Jη|X = 0 for all RX ), hence the column rank
for α ∈ [0, 1], since tr [Φγ (A)]−1 is a convex function of of Ξ(U ∗opt ⊗ U opt ) is r2 , which is exactly the full column

A. Furthermore, we have rank.
tr [Φγ (αR1 + (1 − α)R2 )]−1 ⩾ f (αγ1 + (1 − α)γ2 ),

A PPENDIX E
(141) P ROOF OF P ROPOSITION 5
which follows from the definition of f (γ) in (139), and the Proof:
fact that 1) (Generic): According to Proposition 4, to prove the
sufficiency of condition 1), it suffices to show that
tr {αR1 + (1 − α)R2 } = αγ1 + (1 − α)γ2 .
Span{U opt } ⊆ Span{U R }. (145)
Hence the proof is completed.
The equality in (138) is achieved
n when
o γ is determinis- Let us first rewrite problem (14) as follows
tic, namely when tr {RX } = tr RX . Due to the power
e
min tr A−1 ,

n o
A,R
eX
constraint of tr R e X = PT M , we see that the optimal
distribution p(RX ) may now be simplified as p(RX |tr {RX } = s.t. A = ΦPT M (R
e X ),
e H.
n o
PT M ). Furthermore, p(RX ) should be defined over the set of tr R
e X = PT M, R e X ≽ 0, R
eX = R
X (146)
optimal solutions of the problem (137) with γ = PT M . Thus
the proof is completed. The Karush-Kuhn-Tucker (KKT) conditions of problem (146)
can be expressed as
A PPENDIX D
ΦaPT M (A−2 ) = λI − Z R , (147a)
P ROOF OF P ROPOSITION 4 n o
Proof: We will first show the sufficiency, and then show ΦPT M (Re X ) = A, tr R
e X = PT M, R
e X ≽ 0, (147b)
the necessity.
n o
tr Z R Re X = 0, (147c)
1) Sufficiency: Since Ropt is a maximum-rank solution,
we can express any optimal solution as follows Z R ≽ 0. (147d)
R= U opt DU H
opt , (142) From (147a), (147c) and (147d), we see that
where D ∈ Cr×r is a Hermitian matrix. Otherwise, if we have ΦaPT M (A−2 )R
e X = λR
e X, (148)
any other solution R′ that cannot be represented as (142), then
(Ropt + R′ )/2 is also an optimal solution, but rank{(Ropt + and that λ is the maximum eigenvalue of ΦaPT M (A−2 ).
R′ )/2} > rank{Ropt }, which contradicts to the assumption Applying (148) to the maximal-rank solution Ropt , we have
that Ropt is a maximum-rank optimal −2
solution. UH a
opt ΦPT M (A )U opt = λI. (149)
Next, note that the function tr X −1 is strictly convex
with respect to X. Therefore, the optimal solution is unique, From (149), we see that it now suffices to show that
when the linear map from D to ΦPT M (U opt DU H opt ) is UH a −2
R ΦPT M (A )U R = λI rA , (150)
injective. Observe that
where rA is the multiplicity of the maximum eigenvalue of
vec(ΦPT M (U opt DU H
opt )) ΦaPT M (A−2 ). To this end, we write the Lagrange dual function
= Ξ(U ∗opt ⊗ U opt )vec(D) + vec(J
e P ), (143) of problem (146) as follows
hence the map is indeed injective if Ξ(U ∗opt ⊗ U opt ) has full g(λ̃, Z A , Z R )
column rank.
n n o
= inf tr A−1 + tr {Z A A} − tr Z R R eX
2) Necessity: If Ropt is the unique optimal solution, there A,ReX
does not exist any R taking the form of (142) corresponding n o n o o
to a zero-trace D that satisfies − tr Z A ΦPT M (R e X ) + λ̃ tr R e X − PT M
n
Ξvec(R) = 0. (144) = inf tr A−1 + tr {Z A A} − λ̃PT M
A,ReX
Otherwise, one may construct another optimal solution R′ = n
e X − Z R − Φa
o o
+ tr R (Z ) + λ̃I
Ropt + αR, where α ∈ R is chosen to ensure that R′ is ( n 1o
PT M A
positive semidefinite. This implies that the subspace 2tr Z A − λ̃PT M, ΦaPT M (Z A ) = λ̃I − Z R ;
2
=
V = {(U ∗opt ⊗U opt )vec(D)|D ∈ Cr×r , D = D H , tr {D} = 0} −∞, otherwise.
Hence the dual problem of (146) can be written as which implies that the eigenspace corresponding to the max-
n 1o imum eigenvalue of ΦaPT M (z) has dimensionality 1. In this
max 2tr Z A2 − λ̃PT M case, (155) always holds, otherwise the optimal solution
λ̃,Z A
of (14) would be unbounded, since ΦPT M (Re X ) is not invert-
s.t. ΦaPT M (Z A ) − λ̃I ≼ 0, Z A ≽ 0, Z A = Z H
A , (151) ible. Thus the proof is completed.
which is equivalent to (18) after a maximization over λ̃, while
λ is the optimal value of λ̃. Obviously, strong duality holds A PPENDIX F
for the primal-dual pair (146) and (18), which implies that P ROOF OF C OROLLARY 2
Proof: It is straightforward that rank(R
e X ) ⩽ M when
ΦaPT M (Z A ) = ΦaPT M (A−2 ), (152) K ⩾ M . Hence it suffices to consider the case of K < M
if A is an optimal solution of (146) and Z A is an optimal only. Consider the eigendecomposition of the optimal R
e X as
solution of (18). Since λ is the maximum eigenvalue of follows
ΦaPT M (Z A ), we have e X = U opt Λopt U H .
R (160)
opt
UH a
R ΦPT M (Z A )U R = λI rA , (153) According to Appendix D, when the optimal solution of (14) is
unique, the column rank of Ξ(U ∗opt ⊗ U opt ) is (rank(R
e X ))2 ,
which implies (150).
2 2 which implies that
2) (K ⩾ M ): Note that Ξ ∈ CK ×M , therefore, it is
indeed possible that Ξ has full column rank when K ⩾ M . r2 ⩽ K 2 , (161)
When this is true, we also have that 2 e X ))2
since Ξ(U ∗opt ⊗ U opt ) ∈ CK ×(rank(R
.
rank{Ξ(U ∗R ⊗ U R )} = r , 2
(154)
A PPENDIX G
which implies that rank{Ξ(U ∗R ⊗ U R )} has full column
P ROOF OF T HEOREM 1
rank, since r2 ⩽ M 2 ⩽ K 2 . Thus we can conclude that
problem (14) has a unique optimal solution according to Proof: Under the aforementioned assumptions, rate RSC
Proposition 4. may be expressed as
3) (K = 1): When K = 1, the optimal Z A of the dual 1 e SC .
problem (18) is a scalar, denoted as z. Furthermore, Ξ is RSC = max I(Yc ; X|Hc ), s.t. XXH = T R X (162)
pX (X) T
now a row-vector, denoted as ξ T . Hence the generic sufficient
Consider the singular value decomposition of Hc X:
condition for solution-uniqueness becomes √
Hc X = T UΣVH HX . (163)
(U ∗R ⊗ U R )H ξ ̸= 0, (155)
Note that only VH
HX is not yet determined, since we have
where U R contains the eigenvectors corresponding to the max-
imum eigenvalue of ΦaPT M (z), which can now be expressed e SC HH = T −1 Hc XXH HH
Hc R X c c
as (according to (19))
= UΣΣH UH ,
r3
H
X
ΦaPT M (z) = z Ḡi F̄ i . (156) which does not depend on VH HX . Let us denote the first MSC
i=1 columns of U corresponding to the non-zero singular values as
Furthermore, since J P is now a scalar (denoted as JP ), the U.
e Without loss of generality, we may apply a linear combiner
Choi representation of ΦP (γ, A) is now given by e H at the receiver side, and obtain
U
√
e P (γ) = JP I, Ue H Yc = T Σ eVeH + U e H Zc , (164)
Ψ (157) HX
γ H
where Ue Zc ∈ CMSC ×T has i.i.d. entries which are circularly
which has no impact on the eigenspace structure of Ψ
e γ . As for
symmetric complex Gaussian distributed, Σ e ∈ CMSC ×MSC
the transpose map ΦT (·), we now have denotes the MSC × MSC submatrix containing non-zero sin-
r1 gular values, and V
e HX contains the first MSC columns of
e ∗ AF
e T,
X
Φ1 (ΦT (A)) = F (158) H
i i VHX . Note that U
e Yc is a sufficient statistic of Yc for the
i=1
estimation of X, hence the mutual information I(Yc |Hc ; X)
according to (128), since F
e i ’s are now row-vectors. Hence we may be expressed as
obtain H
r1 r2
I(Yc ; X|Hc ) = I(U
e Yc ; X|Hc )
e HF e HG
X
e H Yc |Hc ) − h(U
e H Yc |Hc , X).
X
ΦaPT M (z) = k (F i
e i )T + G j
ej = h(U (165)
i=1 j=1
√
= kB 1 , (159) Upon denoting Yec = U e H Yc , X
e = TΣ eVe H , and Z
ec =
HX
H
U
e Zc , from (164) we have
where k is a positive constant with respect to B 1 . Finally,
let us consider the case when U R has a single column, Y
ec = X
e +Z
ec, (166)
and hence density function is given by

I(Yc ; X|Hc ) = I(Y
e c ; X|Hc ) (167a) 1 2
f (x) = 2 xMSC /2−1 e−x/2 .
e c |Hc ) − h(Y
= h(Y e c |Hc , X). (167b) 2MSC /2 Γ(M2SC /2)
We may now express the maximum differential entropy
From (167b) we obtain e c |Hc ) as follows
h(Y
max I(Yc ; X|Hc ) Z
pX (X)
max h(Y e c |Hc ) = log ∂Vol(Sϵ ) f (ϵ)dϵ + h(ε)

e c |Hc ) − h(Y
e c |Hc , X)
pX (X) ∂ϵ
= max h(Y 1 1
pX (X)
= log Vol(S)−log Γ M2SC +1 + M2SC log π+log M2SC
e c |Hc ) − T MSC log(πeσ 2 ).
= max h(Y (168) 2 2
c
pX (X) 1 2 σc2 √
+ (MSC − 1) E{log x} + log +h x
e c |Hc ) based on the follow-
Next, we provide estimates of h(Y 2 2
1 2 2
ing lemma. + log(σc /2) + O(σc )
Lemma 2: The logarithmic volume of the rescaled complex 2
1
e 2 + 1 M2 log π + log VT,M
Stiefel manifold = T − MSC log T Σ
√ 2 2 SCZ SC
∈ CMSC ×T , QQH = I MSC

1 1
S= T ΣQ|Q
− log Γ M2SC +1 + (M2SC −1) f (x) log(xσc2 /2)dx
e

e ∈ CMSC ×T |Q eH = TΣ e2 2 2
= Q eQ (169) 1 2
+ [MSC −(MSC −1)ψ(M2SC /2)] log e+log(σc2 /2)
2
is given by 2
1 1
e 2 + log VT,M , (170) + log Γ(M2SC /2) + log M2SC − log 2 + O(σc2 ),

log Vol(S) = T − MSC log T Σ (175)
2 SC 2
where VT,MSC denotes the volume of the complex Stiefel where ∞
1 X 1 1
manifold (without rescaling), given by ψ(x) = −γ − + −
x n=1 n n + x
T
Y 2π k
VT,MSC = . (171) is the digamma function [45], with γ = 0.572 . . . being the
(k − 1)!
k=T −MSC +1 Euler-Mascheroni constant. Note that
Z
The result (171) is known in [42].
f (x) log xdx = ψ(M2SC /2) log e + log 2, (176)
Proof: Please refer to Appendix H.
Now, let us consider the following set and thus we obtain
n o
e c |Hc ) = T − 1 MSC log T Σ e 2 + log VT,M

Sϵ = A + ϵB|A ∈ S, ∥B∥F ⩽ 1 , (172) max h(Y SC
pX (X) 2
which is known as the ϵ-tube about the manifold S [51]. 1
According to Theorem 9.23 in [51], for a small ϵ, the volume + M2SC log(πeσc2 ) + O(σc2 ), (177)
2
of the ϵ-tube can be approximated as follows after some simplification, which implies that
1 2
(πϵ2 ) 2 MSC max I(Yc ; X|Hc ) (178)
Vol(Sϵ ) = Vol(S)(1 + O(ϵ2 )). (173)
Γ( 21 M2SC + 1) pX (X)
n 1
e 2 + log VT,M
o
Hence we have = E T − MSC log σc−2 Σ SC
2
∂Vol(Sϵ ) 1
1 T
= log Vol(S) − log Γ M2SC + 1
n o
log + E MSC T − MSC log + O(σc2 ). (179)
∂ϵ 2 2 πe
1 2 1 2
+ MSC log π + (MSC − 1) log ϵ2 Next, we will show that the term
2 2
+ log M2SC + O(ϵ2 ). (174)
1 T
c1 := log VT,MSC + MSC T − MSC log (180)
2 πe
According to (166), the maximum differential entropy can be
is on the order of o(T ) as T → ∞. To see this, we first note
expressed as
Z that
∂Vol(Sϵ ) M
max h(Yc |Hc ) = log f (ϵ)dϵ +h(ε), XSC
pX (X) ∂ϵ log VT,MSC = log VT −i+1,1 , (181)

| {z }
h(Yc |Hc ,ε) i=1
which implies that

where ε is a random variable such that x = 2ε2 σc−2 is a
chi-squared distribution having DoF M2SC ,6 whose probability
1
log VT,MSC = MSC log 2 + MSC T − MSC log π
2
6 Therandom variable ε represents ∥Ze ⊥ ∥F , where Z
e ⊥ is the component M SC T −1
c c 1 X X
of Zc being orthogonal to the tangent space of S at X. The dimensionality
e + MSC log π − log j, (182)
⊥
e is M2 .
of Z 2 i=1 j=1
c SC
Z q
and hence = J Q,Q
e
e Q |(dQ)∧ ,
|G (191)
MSC T −1 V
1 X X
c1 = MSC T − MSC log T /e − log j (183) where |J Q,Q
e | is the Jacobian determinant of the transforma-
2 i=1 j=1
√ tion from Q on V to Q e on S. Especially, when the Jacobian
+ MSC log(2 π) determinant does not depend on the specific points Q and
T −1 e in the sense that J e = |J | for all Q ∈ V and the
Q,
h MSC X √ i Q,Q
= MSC T − log T /e − log j + log(2 π) . corresponding Qe ∈ S, we have
2 j=1
(184) Vol(S) = |J |Vol(V). (192)
Now we have Next, let us consider the structure of dQ and dQ.

e For the
T −1 √ differential dQ, we see that
c1 h MSC 1 X log(2 π) i
= MSC 1 − log T /e − log j +
T 2T T j=1 T d(QQH ) = 0
h 1 i = (dQ)QH + QdQH . (193)
→ MSC log T /e − log Γ(T )
T Without loss of generality, we consider the tangent space at
→ 0, (185) Q0 = [I, 0MSC ×(T −MSC ) ]. Thus from (193) we have
as T → ∞, where the last line follows from the fact that
dQ0 = [∆∥ , ∆⊥ ], (194)
1
lim ln T − ln Γ(T ) = lim 1 + ln T − ψ(T ) where ∆∥ ∈ CMSC ×MSC is a skew-Hermitian matrix, and
T →∞ T T →∞
= 1, (186) ∆⊥ ∈∈ CMSC ×(T −MSC ) is an arbitrary matrix. As for the
tangent space at an arbitrary Q ∈ V, we have Q = Q0 U ,
Thus we arrive at where U is a T × T unitary matrix, and hence
n MSC e 2 − MSC log Γ(T )
RSC = E 1 − log σc−2 Σ dQ = (dQ0 )U = [∆∥ , ∆⊥ ]U . (195)
2T T√
h MSC T log(2 π) io
+ MSC 1 − log + + O(σc2 ). Similarly, the matrix Qe 0 ∈ S corresponding to Q0 takes the
2T e T form of
(187) √
Qe 0 = [ T Σ,
e 0M ×(T −M ) ]. (196)
Hence the proof is completed. SC SC
Its differential
A PPENDIX H dQ
e = [∆
e ∥, ∆
e ⊥]
0
P ROOF OF L EMMA 2
Proof: Let us first consider the Riemannian metric tensor satisfies
√
on S, which can be written in a matrix form in the tangent ∆
e ∥ = T Σ∆
e ∥, (197a)
space TQ
e S at each point Q ∈ S, denoted as GQ
e e . The volume
√
∆
e ⊥ = T Σ∆
e ⊥. (197b)
of V can then be computed as [19, Sec. 1.2]
Z q e ∈ S, we also have
e ∧ For a generic Q
Vol(S) = |GQe |(dQ) , (188)
S
dQ
e = (dQ
e 0 )U = [∆
e ∥, ∆
e ⊥ ]U . (198)
where dQ e can be viewed as a tangent vector in the tangent
q
e ∧ is a differential form Note that the inner product between matrices is preserved by
space TQ e S, the term |GQe |(dQ)
right-multiplying unitary matrices, and thus the Riemannian
known as the volume form [19, Sec. 1.2] on the manifold metric of V at Q is equal to that at Q0 , namely we have GQ =
S induced by its Riemannian metric GQ e ∧
e , and (dQ) is the GQ0 . Moreover, the exterior product is also preserved by
exterior product over all components in dQ, which serves as
e right-multiplying
q unitary matrices. Consequently, the volume
a volume form on the tangent space (as a Euclidean space). form |G e Q |(dQ)∧ on V is preserved by right-multiplying
According to [19, Sec. 1.2], (169) may be viewed as an q
unitary matrices. Similarly, the volume form |G e |(dQ) e ∧
alternative parametrization of the original Stiefel manifold Q
V = Q ∈ CMSC ×T |QQH = I MSC .

(189) on S at Qe = Q e 0 U should be equal to that at Qe 0 as well.
Thus the result (192) is now applicable, and we have
Upon denoting the Riemannian metric of V at Q as G
e Q,
we have Vol(S) = J Q Vol(V). (199)
0 ,Q0
e e
Z q
Vol(V) = e Q |(dQ)∧ ,
|G (190) In order to obtain the Jacobian determinant J Q ,
0 ,Q0
e
V we consider the real vector representations of dQ0 and dQ
e 0.
and hence Naturally, we may choose the vector representation of dQ0 as
Z q
Vol(S) = |GQ e ∧
e |(dQ) dq 0 = [v ∥ , Re{vec(∆⊥ )}, Im{vec(∆⊥ )}]T , (200)
S
2
where v ∥ ∈ RMSC is characterized by This is exactly the signalling scheme given in (29).
As for the uniform sampling procedure detailed in Corol-
Bv ∥ = vec(∆∥ ), (201) lary 3, it is a well-known method for obtaining a random sam-
B is a matrix constituted by orthonormal columns forming a ple from uniform distribution with respect to the Haar measure
basis of the space of all MSC ×MSC skew-Hermitian matrices. over the complex Stiefel manifold V defined in (32) [52]. Thus
Specifically, we may express B as follows the proof is completed.

√1 vec(E jk − E kj ), i = jMSC + k;


 √2 B. Proof of Corollary 4
M (M −1)
B :,i = √−1 vec(E jk + E kj ), i = SC 2SC + jMSC + k;
2 Proof: First let us introduce the following lemma.
√1 vec(E kk ),

i = MSC (MSC − 1) + k,


2 Lemma 3 (Unitary Equivalence of Hermitian Square
(202) Roots): Given a positive semidefinite Hermitian matrix H ∈
Cm×m , any pair of its square roots A ∈ Cm×n and B ∈
where j < k. Now, we may write the vector representation of
Cm×n (where m ≤ n) satisfying AAH = BB H = H, are
dQ
e as
0 equivalent up to a unitary transformation, in the sense that
de
q 0 = [e
v ∥ , Re{e v ⊥ }]T ,
v ⊥ }, Im{e (203) there exists a unitary matrix U ∈ Cn×n satisfying A = BU .
Proof: Consider the singular value decomposition of A
where and B as follows
√
e∥ = B H (I MSC ⊗ T Σ)Bv
v e ∥ (204a) A = U AΣAV H H
A, B = U BΣBV B , (210)
√
e⊥ = (I T −MSC ⊗ T Σ)vec(∆
v e ⊥ ). (204b) where U A ∈ C m×rH
, UB ∈ C m×rH
, VA ∈ C n×rH
, VB ∈
Thus we obtain the Jacobian determinant as follows Cn×rH are semi-unitary matrices, rH is the rank of H, while
√ √ Σ A ∈ CrH ×rH and Σ B ∈ CrH ×rH are positive definite
J Q ,Q
e = |B H (I MSC ⊗ T Σ)B|
e e2
· |I T −MSC ⊗ T Σ| diagonal matrices. Since AAH = BB H , we have
0 0
√ √
= |I M ⊗ T Σ|SC
e · |I T −M ⊗ T Σ| e2
SC
U A Σ 2A U H 2 H
A = U BΣBU B , (211)
1
2 T − 2 MSC and hence
= TΣ
e , (205)
U AΣAU H H
A = U BΣBU B . (212)
where the second line follows from the fact that B ∈
2 2
CMSC ×MSC is constituted by orthonormal columns, and that We can now rewrite A and B as
it is a square matrix, hence B is unitary. A = P W A, B = P W B, (213)
Finally, using (199), we arrive at
1
where P = U A Σ A U H H
A = U B Σ B U B , while W A =
2 T − 2 MSC H m×n H
Vol(S) = T Σ
e Vol(V), (206) U AV A ∈ C and W B = U B V B ∈ Cm×n are semi-
unitary matrices. We may then construct the following aug-
which implies (170). mented matrices
f A = [W T ∆T ]T , W
W f B = [W T ∆T ]T , (214)
A A A B
A PPENDIX I
H
P ROOF OF C OROLLARY 3 AND C OROLLARY 4 such that W
f A and WgB are unitary. Let U = Wf W
B
f A , we see
A. Proof of Corollary 3 that W A = W B U , and U is a unitary matrix. This also
f f
implies that W A = W B U , and hence we obtain A = BU
Proof: Observe that ΣQ e in (29) has an identical dis-
eVe H in Appendix G that according to (213). This completes the proof.
tribution as that of the optimal Σ HX Now, observe that
maximizes the mutual information I(U e H Yc ; X) in the high-
SNR regime. Hence the PSC –achieving X satisfies (H†c Hc XSC,1 )(H†c Hc XSC,1 )H
√ √
√ = ( T H† U c
e ΣQ)(
e T H† U
c
e H
e ΣQ)
Hc X = T U e ΣQ.
e (207) SC
† †
e (H Hc )H .
= T Hc Hc R X c
One of the solutions to (207) is
According to Lemma 3, we have
X = Xinf + X⊥ , (208a) √ 1 √
√ H†c Hc XSC,1 U = T H†c Hc U s Λs2 QU = T H†c Ue ΣQ,
e (215)
Xinf = T H†c U
e ΣQ,
e (208b)
√ 1
for some unitary matrix U . We see that T H†c Hc U s Λs2 QU
where Xinf denotes the part of the signal that actually carries √ 1
the information, while X⊥ serves as a non-informative padding has the same distribution as T H†c Hc U s Λs2 Q does, since Q
in a subspace of Col(R e SC ) that is orthogonal to Col(Hc X), is sampled from the Haar measure on the complex Stiefel
X
manifold V, which is invariant under multiplication of unitary
which can be any matrix that satisfies
matrices. Finally, we may choose X⊥ as follows
X⊥ XH
inf = 0, (209a)
X⊥ = (I − H†c Hc )XSC,1 ,
T −1 (X⊥ XH H e SC .
⊥ + Xinf Xinf ) = R X (209b) which ensures that XSC,1 is a valid candidate of XSC .
T e X )]−1 .
A PPENDIX J = [Φ(
e R (227)
P ROOF OF T HEOREM 2 T − rank(R
e X)
Proof: First, let us denote Next, we will show that when K ⩽ rank(R
e X ), we have
Φ(A)
e = Φ(A) − J
e P. (216) e X )]−1 } ≼ T e X )]−1 .
E{[Φ(R [Φ(
e R (228)
T −K
According to (12), Φ(·)
e is a positive linear map characterized
Observe that
by
r1 r2
r1 r2
X X
X
eH
e i AT F
X
e H. Φunital (A) = F i AT F H
i + Gj AGH
j , (229)
Φ(A)
e = F i + G
e j AG
j (217) i=1 j=1
i=1 j=1
1 1
where F i = L−1 F e i U Λ 2 , G = L−1 G
i
e i U Λ 2 . Using the
Consider the eigendecomposition of R
eX
singular value decompositions
e X = U ΛU H ,
R (218) F i = U iΣiV H H
i , Gi = U i+r1 Σ i+r1 V i+r1 , (230)
rank(R
eX )×rank(R
eX )
where Λ ∈ R is a diagonal matrix containing where for all i, U i ∈ CK×K , Σ i ∈ CK×K , and V i ∈
all non-zero eigenvalues of R e X , and U ∈ CM ×rank(R e X)
Crank(RX )×K , we obtain an alternative representation of
e
is a semi-unitary matrix constituted by the corresponding Φunital (A) as follows

eigenvectors. Now, observe that the sample covariance matrix
RX may be expressed as Φunital (A) = F V (A)F H , (231)
1 1
RX = U Λ 2 WΛ 2 U H , (219) where
where W satisfies T W ∼ CW rank(R F = [U 1 Σ 1 , . . . , U r1 +r2 Σ r1 +r2 ],

e X ) (I, T ), namely T W
is complex Wishart distributed. Next, we construct the linear V (A) = blkdiag(V H T H
1 A V 1 , . . . , V r1 +r2 AV r1 +r2 ).
map (232)
1 1
Φunital (W) = L−1 Φ(U
e Λ WΛ U H )(LH )−1 ,
2 2 (220) Since A = I implies V (A) = I, Φunital (A) may also
be viewed as a unital positive linear map of V (A), based
which L is obtained by the Cholesky decomposition
on (220). Thus we have
LLH = Φ(
e Re X ). (221) −1
[Φunital (W)] = [F V (W)F H ]−1
−1
The inverse L exists since Φ(R e X ) is invertible. Note that ≼ F [V (W)]−1 F H . (233)
Φunital is a unital map in the sense that
Note that
Φunital (I) = L−1 Φ(
e Re X )(LH )−1 = I. (222) T
E{[V (W)]−1 } = I r +r ⊗ I K , (234)
Moreover, Φunital maps one positive semidefinite matrix to T −K 1 2
another, and hence it is a positive unital linear map. According since each diagonal block in V (W) satisfies
to Choi [53], such a map satisfies T
TV H
i W V i ∼ CW K (I, T ), (235a)
−1 −1
[Φunital (W)] ≼ Φunital (W ), (223) TV H
∼ CW K (I, T ).
i WV i (235b)
which implies that Hence
e X )]−1 L ≼ Φunital (W−1 ),
LH [Φ(R (224) e X )]−1 } ≼ L−1 F E{[V (W)]−1 }F H (LH )−1
E{[Φ(R
and hence T
= L−1 Φunital (I)(LH )−1
e X )]−1 ≼ [Φ( e Λ 12 W−1 Λ 12 U H )[Φ(
e X )]−1 Φ(U e X )]−1 . T −K
[Φ(R e R e R
T
= [Φ(
e Re X )]−1 , (236)
(225) T −K
Since T W ∼ CW rank(R
e X ) (I, T ), we obtain
which implies (228). Combining (227) and (228), we obtain
T [Φ(
e Re X )]−1
T e X )]−1 } ≼
E{W−1 } = T E{(T W)−1 } = I. (226) E{[Φ(R . (237)
T − rank(R
e X) T − min{K, rank(R
e X )}
Thus we have Next, let us consider the spectral decomposition of J

eP
e X )]−1 }
E{[Φ(R e P = U P ΛP U H .
J (238)
P
≼ [Φ(
e R e Λ 12 E(W−1 )Λ 12 U H )[Φ(
e X )]−1 Φ(U e Re X )]−1 Using (238), we obtain
T e X )]−1 Φ( e X )]−1 −1 io
= [Φ(
e R e Re X )[Φ(
e R n h
T − rank(R
e X) E tr Φ(Re X) + J eP
n h −1 io
= E tr U H P Φ(RX )U P + ΛP
e A PPENDIX K
−1 P ROOF OF C OROLLARY 5
−1 e − 12
n h io
= E tr ΛP 2 Φ 0 (R X ) + I Λ P Proof: From (231) we see that, when K ⩽ rank(R
e X ),
−1 i the equality in (36) is achieved if
− 12 −1
n h o
= tr ΛP E Φ e 0 (RX ) + I ΛP 2 , (239)
E{[F V (W)F H ]−1 } = E{F [V (W)]−1 F H }. (244)
−1 − 21
where Φ e 0 (RX ) = ΛP 2 U H
P Φ(RX )U P ΛP .
e Note that now it
suffices to show that Note that F F H = I since Φunital is unital with respect to
−1 V (W). Thus when r1 + r2 = 1, we have that F ∈ CK×K ,
h −1 i T Φe 0 (R
e X) + I which implies that F is unitary, and hence (244) holds.
E Φ e 0 (RX ) + I ≼ . e P can be neglected in the limit of σs → 0, let
Next, since J
T − min{K, rank(R
e X )}
us assume that J
e P = 0 and obtain
Let us define
I ⊗ AT

0
Φ(A) = F st (F st )H . (245)
Φ
e 1 (RX ) = Φ
e 0 (RX ) + αI, (240) 0 I ⊗A
which linear map, with the notation α = Therefore, if F st is unitary, we may conclude that
is a positive

tr RX /E tr RX . Using again the previous arguments,
E{[Φ(RX )]−1 }
we obtain
I ⊗ (E{R−1 T

e X )]−1 = F st X }) 0
(F st )H
e 1 (RX )]−1 } ≼ T [Φ
e 1 (R
0 I ⊗ E{R−1 }
E{[Φ X
T − min{K, rank(R e X )} T −1
e X ) + I]−1 = [Φ(RX )] ,
e (246)
T [Φ
e 0 (R T −M
= . (241)
T − min{K, rank(R
e X )} where the last line follows from the fact that
Next we will show that T RX ∼ CW M (R
e X , T ).
h −1 i
E Φ e 0 (RX ) + I e 1 (RX )]−1 },
≼ E{[Φ This implies the equality in (36), since rank(R
e X ) = M when
R
e X is invertible.
which can be rewritten as
h −1 i h −1 i
E Φ e 0 (RX ) + E(α)I ≼E Φ e 0 (RX ) + αI . A PPENDIX L
P ROOF OF P ROPOSITION 6
(242)
Proof: To prove this proposition, it suffices to show that
To this end, we consider the quantity there exists at least one point (ϵ, R) on the Gaussian inner
h −1 −1 i bound or the semi-unitary inner bound that lies above the line
E Φ e 0 (RX ) + E(α)I − Φe 0 (RX ) + αI
segment connecting PSC and PCS . Once we found such a point
P = (ϵ, R), we see that the time-sharing strategies between
h i
e 0 (RX )+αI −1

= E (α−E(α)) Φ e 0 (RX )+I Φ
n P , PSC and PCS readily constitute a tighter inner bound than
o
e 0 (RX )+αI −1 the pentagon inner bound.

= cov αI, Φ e 0 (RX )+I Φ
n n oo Next we show that there do exist such points, under mild
e 0 (RX )+αI −1 |α ,

= cov αI, E Φe 0 (RX )+I Φ assumptions. To this end, let us consider the slope of the semi-
unitary–Gaussian inner bound at a specific value of α. For
(243)
our purpose, it suffices to show that at PSC (resp. PCS ), the
where cov(·, ·) is defined as cov(A, B) = E(AB)−E(A)E(B), slope of the semi-unitary–Gaussian inner bound is larger (resp.
and the quantity smaller) than the slope of the line segment connecting PSC and
n o PCS . First note that, as long as the optimal objective function
e 0 (RX )+αI −1 |α

E Φ
e 0 (RX )+I Φ value is not identical for all α ∈ [0, 1], the slope of the outer
bound (50) at a specific value of α is given by
can be seen to be a matrix-valued monotonically decreasing
function of α. Next, note that cov {A(α), B(α)} ≽ 0 when ∂Rout (α) T
= 2 · λα , (247)
A(α) and B(α) are mutually commuting matrix-valued mono- ∂ϵout (α) σs
tonically increasing functions of α [54]. Correspondingly, where λα is the optimal Lagrange dual variable corresponding
when A(α) is increasing and B(α) is decreasing, we would to the constraint (49c). According to the correspondence
have cov {A(α), B(α)} ≼ 0. This implies that between the formulations (48) and (49) (c.f. [55, Sec. 5.3.3]),
h −1 −1 i λα can be explicitly obtained as
E Φ e 0 (RX ) + E(α)I − Φ e 0 (RX ) + αI ≼ 0,
1−α
λα = , (248)
and hence the proof is completed. α
which tends to infinity as α → 0. Now, the slope of the [16] H. Joudeh and F. M. J. Willems, “Joint communication and binary state
semi-unitary inner bound can be expressed as detection,” IEEE J. Sel. Areas Inf. Theory, vol. 3, no. 1, pp. 113–124,
Mar. 2022.
∂Rin,U (α) MU (1 − α)T [17] M. Ahmadipour, M. Kobayashi, M. Wigger, and G. Caire,
= 1− + O(σc2 ), (249) “An information-theoretic approach to joint sensing and
∂ϵin,U (α) 2T ασs2 communication,” IEEE Trans. Inf. Theory, early access, May 30, 2022,
doi: 10.1109/TIT.2022.3176139.
which also tends to infinity as α → 0. Since the semi-
[18] R. Miller and C. Chang, “A modified Cramér–Rao bound and its
unitary–Gaussian inner bound equals to the semi-unitary inner applications (Corresp.),” IEEE Trans. Inf. Theory, vol. IT-24, no. 3,
bound around PSC , we may conclude that the slope of the pp. 398–400, May 1978.
semi-unitary–Gaussian inner bound at PSC is positive infinity. [19] M. P. Do Carmo and J. F. Francis, Riemannian Geometry, vol. 6. Boston,
MA, USA: Springer, 1992.
Following a similar line of reasoning, we can show that the
[20] J. Li, P. Stoica, L. Xu, and W. Roberts, “On parameter identifiability of
slope at PCS is zero. Therefore, no matter what the slope of the MIMO radar,” IEEE Signal Process. Lett., vol. 14, no. 12, pp. 968–971,
line segment connecting PSC and PCS is, it should be upper Dec. 2007.
and lower bounded by that of the semi-unitary–Gaussian inner [21] Y. Liu, M. Li, A. Liu, J. Lu, and T. X. Han, “Information-theoretic
limits of integrated sensing and communication with correlated sensing
bound at PSC and PCS , respectively, completing the proof. and channel states for vehicular networks,” IEEE Trans. Veh. Technol.,
vol. 71, no. 9, pp. 10161–10166, Sep. 2022.
ACKNOWLEDGMENT [22] Z. Han, Z. Lu, X. Wen, L. Guo, and J. Zhao, “Towards 3D centimeter-
level passive gesture tracking with two WiFi links,” IEEE Trans. Mobile
The authors would like to acknowledge Prof. Ya-Feng Liu Comput., vol. 22, no. 5, pp. 3031–3045, May 2023.
and Dr. Chunyi Wang for their valuable suggestions. [23] H. L. Van Trees, Detection, Estimation, and Modulation Theory, Part
I: Detection, Estimation, and Linear Modulation Theory. Hoboken, NJ,
USA: Wiley, 2004.
R EFERENCES [24] B. Z. Bobrovsky, E. Mayer-Wolf, and M. Zakai, “Some classes of
global Cramér–Rao bounds,” Ann. Statist., vol. 15, no. 4, pp. 1421–1438,
[1] F. Liu et al., “Integrated sensing and communications: Toward dual- Dec. 1987.
functional wireless networks for 6G and beyond,” IEEE J. Sel. Areas
[25] A. El Gamal and Y.-H. Kim, Network Information Theory, 1st ed.
Commun., vol. 40, no. 6, pp. 1728–1767, Jun. 2022.
Cambridge, U.K.: Cambridge Univ. Press, 2011.
[2] Y. Cui, F. Liu, X. Jing, and J. Mu, “Integrating sensing and communi-
cations for ubiquitous IoT: Applications, trends, and challenges,” IEEE [26] A. Goldsmith, Wireless Communications, 1st ed. Cambridge, U.K.:
Netw., vol. 35, no. 5, pp. 158–167, Sep. 2021. Cambridge Univ. Press, 2005.
[3] A. Liu et al., “A survey on fundamental limits of integrated sensing [27] T. M. Cover, Elements of Information Theory, 2nd ed. Hoboken, NJ,
and communication,” IEEE Commun. Surveys Tuts., vol. 24, no. 2, USA: Wiley, 2006.
pp. 994–1034, 2nd Quart., 2022. [28] S. A. Jafar and S. Shamai (Shitz), “Degrees of freedom region of the
[4] Y. Xiong, N. Wu, Y. Shen, and M. Z. Win, “Cooperative localiza- MIMO x channel,” IEEE Trans. Inf. Theory, vol. 54, no. 1, pp. 151–170,
tion in massive networks,” IEEE Trans. Inf. Theory, vol. 68, no. 2, Jan. 2008.
pp. 1237–1258, Feb. 2022. [29] H. Li, “Conflict and trade-off of waveform uncertainty in joint commu-
[5] Y. Xiong, N. Wu, Y. Shen, and M. Z. Win, “Cooperative network nication and sensing systems,” in Proc. IEEE Int. Conf. Commun. (ICC),
synchronization: Asymptotic analysis,” IEEE Trans. Signal Process., May 2022, pp. 5561–5566.
vol. 66, no. 3, pp. 757–772, Feb. 2018. [30] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
[6] D. W. Bliss, “Cooperative radar and communications signaling: The esti- bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
mation and information theory odd couple,” in Proc. IEEE Radar Conf., no. 11, pp. 3590–3600, Nov. 2010.
Covington, KY, USA, May 2014, pp. 50–55. [31] H. Hua, X. Song, Y. Fang, T. X. Han, and J. Xu, “MIMO integrated
[7] F. Liu, Y.-F. Liu, A. Li, C. Masouros, and Y. C. Eldar, “Cramér–Rao sensing and communication with extended target: CRB-rate tradeoff,”
bound optimization for joint radar-communication beamforming,” IEEE in Proc. IEEE Global Commun. Conf. (GLOBECOM), Rio de Janeiro,
Trans. Signal Process., vol. 70, pp. 240–253, 2022. Brazil, Dec. 2022, pp. 4075–4080.
[8] J. A. Zhang et al., “An overview of signal processing techniques for joint [32] Z. Ren, X. Song, Y. Fang, L. Qiu, and J. Xu, “Fundamental CRB-rate
communication and radar sensing,” IEEE J. Sel. Topics Signal Process., tradeoff in multi-antenna multicast channel with ISAC,” in Proc. IEEE
vol. 15, no. 6, pp. 1295–1315, Nov. 2021. Globecom Workshops (GC Wkshps), Rio de Janeiro, Brazil, Dec. 2022,
[9] F. Liu, W. Yuan, C. Masouros, and J. Yuan, “Radar-assisted predic- pp. 1261–1266.
tive beamforming for vehicular links: Communication served by sens- [33] H. Hua, T. X. Han, and J. Xu, “MIMO integrated sensing and commu-
ing,” IEEE Trans. Wireless Commun., vol. 19, no. 11, pp. 7704–7719, nication: CRB-rate tradeoff,” 2022, arXiv:2209.12721.
Nov. 2020. [34] M. Grant and S. Boyd. (Mar. 2014). CVX: MATLAB Software for
[10] W. Yuan, F. Liu, C. Masouros, J. Yuan, D. W. K. Ng, and Disciplined Convex Programming, Version 2.1. [Online]. Available:
N. González-Prelcic, “Bayesian predictive beamforming for vehicular http://cvxr.com/cvx
networks: A low-overhead joint radar-communication approach,” IEEE
Trans. Wireless Commun., vol. 20, no. 3, pp. 1442–1456, Mar. 2021. [35] M. C. Grant and S. P. Boyd, “Graph implementations for nonsmooth
convex programs,” in Recent Advances in Learning and Control (Lecture
[11] D. Guo, S. Shamai (Shitz), and S. Verdu, “Mutual information and
Notes in Control and Information Sciences), V. Blondel, S. Boyd, and
minimum mean-square error in Gaussian channels,” IEEE Trans. Inf.
H. Kimura, Eds. Springer-Verlag, 2008, pp. 95–110. [Online]. Available:
Theory, vol. 51, no. 4, pp. 1261–1282, Apr. 2005.
http://stanford.edu/~boyd/graph_dcp.html
[12] A. Sutivong, M. Chiang, T. M. Cover, and Y.-H. Kim, “Channel capacity
and state estimation for state-dependent Gaussian channels,” IEEE Trans. [36] G. N. Saddik, R. S. Singh, and E. R. Brown, “Ultra-wideband multi-
Inf. Theory, vol. 51, no. 4, pp. 1486–1495, Apr. 2005. functional communications/radar system,” IEEE Trans. Microw. Theory
[13] M. Kobayashi, G. Caire, and G. Kramer, “Joint state sensing and Techn., vol. 55, no. 7, pp. 1431–1437, Jul. 2007.
communication: Optimal tradeoff for a memoryless case,” in Proc. IEEE [37] E. BouDaher, A. Hassanien, E. Aboutanios, and M. G. Amin, “Towards
Int. Symp. Inf. Theory (ISIT), Vail, CO, USA, Jun. 2018, pp. 111–115. a dual-function MIMO radar-communication system,” in Proc. IEEE
[14] M. Kobayashi, H. Hamad, G. Kramer, and G. Caire, “Joint state sensing Radar Conf. (RadarConf), Philadelphia, PA, USA, May 2016, pp. 1–6.
and communication over memoryless multiple access channels,” in [38] T. Huang, N. Shlezinger, X. Xu, Y. Liu, and Y. C. Eldar, “MAJoRCom:
Proc. IEEE Int. Symp. Inf. Theory (ISIT), Paris, France, Jul. 2019, A dual-function radar communication system using index modulation,”
pp. 270–274. IEEE Trans. Signal Process., vol. 68, pp. 3423–3438, 2020.
[15] M. Ahmadipour, M. Wigger, and M. Kobayashi, “Joint sensing and [39] C. Sturm and W. Wiesbeck, “Waveform design and signal processing
communication over memoryless broadcast channels,” in Proc. IEEE aspects for fusion of wireless communications and radar sensing,” Proc.
Inf. Theory Workshop (ITW), Apr. 2021, pp. 1–5. IEEE, vol. 99, no. 7, pp. 1236–1259, Jul. 2011.
[40] B. M. Hochwald and T. L. Marzetta, “Unitary space-time modulation for member of the IMT-2030 (6G) ISAC Task Group. He was a recipient of
multiple-antenna communications in Rayleigh flat fading,” IEEE Trans. the 2023 IEEE Communications Society Stephan O. Rice Prize, the 2023 IEEE
Inf. Theory, vol. 46, no. 2, pp. 543–564, Mar. 2000. ICC Best Paper Award, the 2022 First Prize of Science and Technology
[41] T. L. Marzetta and B. M. Hochwald, “Capacity of a mobile multiple- Progress of China Institue of Communications, the 2021 IEEE Signal Pro-
antenna communication link in Rayleigh flat fading,” IEEE Trans. Inf. cessing Society Young Author Best Paper Award, the 2019 Best Ph.D. Thesis
Theory, vol. 45, no. 1, pp. 139–157, Jan. 1999. Award of Chinese Institute of Electronics, and the 2018 EU Marie Curie
[42] L. Zheng and D. N. C. Tse, “Communication on the Grassmann mani- Individual Fellowship. He is the Founding Academic Chair of the IEEE
fold: A geometric approach to the noncoherent multiple-antenna chan- ComSoc ISAC Emerging Technology Initiative (ISAC-ETI). He has served as
nel,” IEEE Trans. Inf. Theory, vol. 48, no. 2, pp. 359–383, Feb. 2002. an Organizer and the Co-Chair for numerous workshops, special sessions, and
[43] J. Li, L. Xu, P. Stoica, K. W. Forsythe, and D. W. Bliss, “Range compres- tutorials in flagship IEEE/ACM conferences, including ICC, GLOBECOM,
sion and waveform optimization for MIMO radar: A Cramér–Rao bound ICASSP, and MobiCom. He is the TPC Co-Chair of the second and third
based study,” IEEE Trans. Signal Process., vol. 56, no. 1, pp. 218–232, IEEE Joint Communication and Sensing (JC&S) Symposium, the Track Chair
Jan. 2008. of the IEEE GLOBECOM 2023, and the Track Co-Chair of the IEEE WCNC
[44] Y. Shen, H. Wymeersch, and M. Z. Win, “Fundamental limits of 2024. He was listed in the World’s Top 2% Scientists by Stanford University
wideband localization—Part II: Cooperative networks,” IEEE Trans. Inf. for citation impact in 2021 and 2022. He is an Associate Editor of the IEEE
Theory, vol. 56, no. 10, pp. 4981–5000, Oct. 2010. O PEN J OURNAL OF S IGNAL P ROCESSING and the IEEE C OMMUNICATIONS
[45] D. Zwillinger and A. Jeffrey, Table of Integrals, Series, and Products, L ETTERS and a Guest Editor of the IEEE J OURNAL ON S ELECTED A REAS
7th ed. Amsterdam, The Netherlands: Elsevier, 2007. IN C OMMUNICATIONS and the IEEE W IRELESS C OMMUNICATIONS .
[46] B. Tang and J. Li, “Spectrally constrained MIMO radar waveform design
based on mutual information,” IEEE Trans. Signal Process., vol. 67,
no. 3, pp. 821–834, Feb. 2019.
[47] D. N. C. Tse and S. V. Hanly, “Linear multiuser receivers: Effective
interference, effective bandwidth and user capacity,” IEEE Trans. Inf.
Theory, vol. 45, no. 2, pp. 641–657, Mar. 1999.
Yuanhao Cui (Member, IEEE) received the B.Eng. degree (Hons.) from
[48] A. van den Bos, “A Cramér–Rao lower bound for complex parameters,”
Henan University and the Ph.D. degree from the Beijing University of
IEEE Trans. Signal Process., vol. 42, no. 10, p. 2859, Oct. 1994.
Posts and Telecommunications (BUPT). He is currently a Research Assistant
[49] M.-D. Choi, “Completely positive linear maps on complex matrices,” Professor with the Department of Electrical and Electronic Engineering,
Linear Algebra Appl., vol. 10, no. 3, pp. 285–290, Jun. 1975. Southern University of Science and Technology, Shenzhen, China. He was
[50] A. Jamiołkowski, “Linear transformations which preserve trace and the CTO and the co-founder of two startup companies, where he has invested
positive semidefiniteness of operators,” Rep. Math. Phys., vol. 3, no. 4, more than $3 million dollars. He holds more than 20 granted patents. His
pp. 275–278, Dec. 1972. research interests include precoding and protocol designs for ISAC. He is
[51] A. Gray, Tubes, vol. 221. Springer, 2003. a member of the IMT-2030 (6G) ISAC Task Group. He received the Best
[52] G. W. Stewart, “The efficient generation of random orthogonal matrices Paper Award from IWCMC 2021. He is the Founding Secretary of the IEEE
with an application to condition estimators,” SIAM J. Numer. Anal., ComSoc ISAC Emerging Technology Initiative (ISAC-ETI) and the CCF
vol. 17, no. 3, pp. 403–409, Jun. 1980. Science Communication Working Committee. He was an Organizer and the
[53] M.-D. Choi, “A Schwarz inequality for positive linear maps on C ∗ - Co-Chair of a number of workshops and special sessions in flagship IEEE
algebras,” Illinois J. Math., vol. 18, no. 4, pp. 565–574, Dec. 1974. conferences, including Mobicom, ICC, ICASSP, WCNC, and VTC. He is
[54] K. D. Schmidt, “On the covariance of monotone functions of a ran- the Lead Guest Editor of the Special Issue on Integrated Sensing and Com-
dom variable,” Professoren des Inst. für Math. Stochastik, Technische munication for 6G of the IEEE O PEN J OURNAL OF THE C OMMUNICATIONS
Universität Dresden, Dresden, Germany, Tech. Rep., 2003. S OCIETY and the Guest Editor of the Special Issue on Integrated Sensing and
[55] S. Boyd and L. Vandenberghe, Convex Optimization, 1st ed. Cambridge, Communications for Future Green Networks of the IEEE T RANSACTIONS ON
U.K.: Cambridge Univ. Press, 2004. G REEN C OMMUNICATIONS AND N ETWORKING.
Yifeng Xiong (Member, IEEE) received the M.S. degree in information Weijie Yuan (Member, IEEE) received the B.E. degree from the Beijing
and communication engineering from the Beijing Institute of Technology, Institute of Technology, China, in 2013, and the Ph.D. degree from the
Beijing, China, in 2018, and the Ph.D. degree in electronic and electrical University of Technology Sydney, Australia, in 2019. In 2016, he was a
engineering from the University of Southampton, Southampton, U.K., in 2022. Visiting Ph.D. Student with the Institute of Telecommunications, Vienna
He is currently an Associate Professor with the School of Information and University of Technology, Austria. He was a Research Assistant with the
Communication Engineering, Beijing University of Posts and Telecommuni- University of Sydney, a Visiting Associate Fellow with the University of
cations. He is also a Visiting Researcher with the Department of Electrical Wollongong, and a Visiting Fellow with the University of Southampton,
and Electronic Engineering, Southern University of Science and Technology, from 2017 to 2019. From 2019 to 2021, he was a Research Associate with the
Shenzhen, China. His research interests include integrated sensing and com- University of New South Wales. He is currently an Assistant Professor with
munications, quantum computation, and quantum information theory. He was the Department of Electrical and Electronic Engineering, Southern University
a recipient of the Best Master Thesis Award of Chinese Institute of Electronics. of Science and Technology, Shenzhen, China. He was a recipient of the
He serves as a Track Co-Chair for the IEEE GLOBECOM 2023. Best Ph.D. Thesis Award from the Chinese Institute of Electronics, the Best
Editor from IEEE C OMMUNICATIONS L ETTERS, the Exemplary Reviewer
from IEEE T RANSACTIONS ON C OMMUNICATIONS/IEEE W IRELESS C OM -
MUNICATIONS L ETTERS, and the Best Paper Award from IEEE ICC 2023.
He was an Organizer/the Chair of several workshops and special sessions
on orthogonal time frequency space (OTFS) and integrated sensing and
Fan Liu (Member, IEEE) received the B.Eng. and Ph.D. degrees from the communication (ISAC) in flagship IEEE and ACM conferences, including
Beijing Institute of Technology (BIT), Beijing, China, in 2013 and 2018, IEEE ICC, IEEE/CIC ICCC, IEEE SPAWC, IEEE VTC, IEEE WCNC, IEEE
respectively. He is currently an Assistant Professor with the Department of ICASSP, and ACM MobiCom. He is the Founding Chair of the IEEE ComSoc
Electronic and Electrical Engineering, Southern University of Science and Special Interest Group on Orthogonal Time Frequency Space (OTFS-SIG).
Technology (SUSTech). Previously, he has held academic positions with Uni- He currently serves as an Associate Editor for the IEEE C OMMUNICATIONS
versity College London (UCL), as a Visiting Researcher, from 2016 to 2018, L ETTERS, and an Associate Editor and an Award Committee Member for the
and then as a Marie Curie Research Fellow, from 2018 to 2020. His research EURASIP Journal on Advances in Signal Processing. He has led the guest
interests lie in the general area of signal processing and wireless communica- editorial teams for three special issues in IEEE Communications Magazine,
tions, and in particular in the area of integrated sensing and communications IEEE T RANSACTIONS ON G REEN C OMMUNICATIONS AND N ETWORKING,
(ISAC). He has ten papers selected as IEEE ComSoc Best Readings. He is a and China Communications.
Tony Xiao Han (Senior Member, IEEE) received the B.E. degree in electrical Giuseppe Caire (Fellow, IEEE) was born in Torino, in 1965. He received the
engineering from Sichuan University and the Ph.D. degree in communications B.Sc. degree in electrical engineering from Politecnico di Torino in 1990, the
engineering from Zhejiang University, Hangzhou, China. He is currently a M.Sc. degree in electrical engineering from Princeton University in 1992, and
Research Expert and a Project Manager with Huawei Technologies Company the Ph.D. degree from Politecnico di Torino in 1994. He was a Post-Doctoral
Ltd. He was a Post-Doctoral Research Fellow with the National University Research Fellow with the European Space Agency (ESTEC), Noordwijk,
of Singapore, Singapore. His research interests include wireless communi- The Netherlands, from 1994 to 1995; an Assistant Professor of telecommuni-
cations, signal processing, integrated sensing and communication (ISAC), cations with Politecnico di Torino; an Associate Professor with the University
and standardization of wireless communication. He was the Chair of IEEE of Parma, Italy; a Professor with the Department of Mobile Communications,
802.11 WLAN Sensing Topic Interest Group (TIG) and the 802.11 WLAN Eurecom Institute, Sophia-Antipolis, France; and a Professor of electrical
Sensing Study Group (SG). He is currently serving as the Chair of IEEE engineering with the Viterbi School of Engineering, University of Southern
802.11bf WLAN Sensing Task Group (TG). He is also the Founding Industry California, Los Angeles. He is currently an Alexander von Humboldt Professor
Chair of IEEE ComSoc ISAC Emerging Technology Initiative (ISAC-ETI) with the Faculty of Electrical Engineering and Computer Science, Technical
and the Vice Chair of IEEE WTC Special Interest Group (SIG) on ISAC. University of Berlin, Germany. His main research interests include commu-
He has served as the Co-Chair for many workshops, such as IEEE GLOBE- nications theory, information theory, and channel and source coding, with a
COM 2020 Workshop on ISAC. He is the Guest Editor of the IEEE J OURNAL particular focus on wireless communications. He received the Jack Neubauer
ON S ELECTED A REAS IN C OMMUNICATIONS Special Issue on Integrated Best System Paper Award from the IEEE Vehicular Technology Society in
Sensing and Communications (ISAC). 2003, the IEEE Communications Society and Information Theory Society
Joint Paper Award in 2004 and 2011, the Okawa Research Award in 2006,
the Alexander von Humboldt Professorship in 2014, the Vodafone Innovation
Prize in 2015, an ERC Advanced Grant in 2018, the Leonard G. Abraham
Prize for Best IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS
Paper in 2019, the IEEE Communications Society Edwin Howard Armstrong
Achievement Award in 2020, and the 2021 Leibniz Prize of the German
National Science Foundation (DFG). He has served on the Board of Governors
for the IEEE Information Theory Society, from 2004 to 2007, and an Officer,
from 2008 to 2013. He was the President of the IEEE Information Theory
Society in 2011.

On The Fundamental Tradeoff of Integrated Sensing and Communications Under Gaussian Channels

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

On The Fundamental Tradeoff of Integrated Sensing and Communications Under Gaussian Channels

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO.

9, SEPTEMBER 2023 5723

On the Fundamental Tradeoff of Integrated Sensing

Nonetheless, the distinct physical roles of X and H in S&C and

and the subscript F in RF is omitted whenever there is no

When the trace of RX is fixed, we have the following

where r3 ⩽ KM , and the matrices F̄ i and Ḡi are given

where H†c denotes the Moore-Penrose pseudo-inverse of Hc , e CS ). Then, the communication-limited

of the results presented in Section III. X = U c Λc2 D, (42)

X = T H†c U eΛe 2 Q + XU,⊥ , (54)

η = [θ, Re{α}, Im{α}]T , In particular, when the communication channel Hc is rank-

asymptotically expressed as TABLE I

holds in the high-SNR regime. In this case, the communication

DoF caused by employing the Gaussian and the semi-unitary

communication receiving SNR is (23 + 10 log10 M ) dB per

(1) (1) (1)

Φγ (A) = Φ1 (ΦT (A)) + Φ2 (A) + ΦP (γ, A), (126) s.t. tr {E(RX )} = PT M, RX ≽ 0, RX = RH

A. Furthermore, we have rank.

tr [Φγ (αR1 + (1 − α)R2 )]−1 ⩾ f (αγ1 + (1 − α)γ2 ),

and hence density function is given by

∈ CMSC ×T , QQH = I MSC

pX (X) ∂ϵ log VT,MSC = log VT −i+1,1 , (181)

which implies that

Now we have Next, let us consider the structure of dQ and dQ.

V = Q ∈ CMSC ×T |QQH = I MSC .

is a semi-unitary matrix constituted by the corresponding Φunital (A) as follows

where W satisfies T W ∼ CW rank(R F = [U 1 Σ 1 , . . . , U r1 +r2 Σ r1 +r2 ],

Thus we have Next, let us consider the spectral decomposition of J

You might also like