betaFeedback

WorksEngineering

Sensing with OFDM Signals: How Do the Constellations Affect the Sensing Accuracy?

Abstract

This paper investigates how communication constellations affect joint range-velocity estimation with data-bearing orthogonal frequency-division multiplexing (OFDM) signals. To this end, starting from the continuous-time sensing process, we derive a discrete observation model and specify its validity conditions. For multiple targets with unknown delays, Doppler shifts, and complex gains, the complete Fisher information matrix (FIM) admits a structural factorization separating deterministic channel derivatives from random symbol-energy weights. This structure yields a Jensen-type lower bound on the average conditional Cramer-Rao bound (CRB), which is attained exactly by constant-modulus constellations. Single-target closed-form expressions further reveal the dependence of the bounds on the transmitted energy distribution and physical OFDM parameters. We also formulate two-dimensional fast Fourier transform (FFT), concentrated maximum-likelihood (ML), and sliding-subarray multiple signal classification (MUSIC) processing under the derived observation model. Numerical results support the theoretical optimality of constant-modulus constellations but also show that the CRB gap between phase-shift keying (PSK) and quadrature amplitude modulation (QAM) becomes negligible when the numbers of subcarriers and OFDM symbols are considerably large. The performance of ML estimation closely approaches the derived bounds, which confirms the effectiveness of the theoretical analysis. FFT processing exhibits grid-induced error floors. In the considered scenarios, MUSIC provides lower estimation errors than FFT and exhibits a more pronounced constellation dependence.

Introduction

Integrated sensing and communications (ISAC) is envisioned as an important capability of sixth-generation wireless networks, enabling communication and sensing to share spectrum, hardware, and transmitted waveforms [1], [2]. One of the fundamental challenges in ISAC is to design dual-functional waveforms that can simultaneously perform accurate target sensing and achieve reliable communications. Existing waveform design strategies can be classified into three categories: sensing-centric [3], communication-centric [4]–[6], and joint design approaches [7]–[9]. Among these, communication-centric ISAC is one of the most promising approaches. On the one hand, it can reuse existing network infrastructure and cellular protocols with only minor modifications, thereby facilitating practical deployment. On the other hand, it exploits data-bearing communication signals for sensing, reducing the need for dedicated sensing transmissions and additional spectrum resources. In particular, orthogonal frequency-division multiplexing (OFDM) is attractive in this context because its multicarrier structure supports data transmission and provides a convenient representation for sensing [10].

Representative studies have optimized transmit waveforms to balance radar beampattern requirements and communication interference [8], or minimized sensing Cramér–Rao bounds (CRBs) subject to communication quality-of-service constraints [9]. The CRB-rate tradeoff for point and extended targets in MIMO ISAC was characterized through transmit covariance optimization in [11]. An OFDM-based waveform design using null-space sensing precoding was proposed in [4] to enhance sensing performance without degrading communication performance. Nevertheless, the above studies assumed that the transmitted signals are deterministic while neglecting the randomness of communication symbols. The authors in [12] further considered the CRB-rate tradeoff for bistatic ISAC with deterministic sensing signals and Gaussian information, assuming that the information-symbol realizations are unknown but their covariance is known to the sensing receiver. From an information-estimation-theoretic perspective, the CRB-rate region reveals a deterministic-random tradeoff associated with the randomness required to convey information [13]. Such randomness matters even when the sensing receiver knows the transmitted data realization. Indeed, precoding designs that explicitly account for random communication data can outperform designs based on a deterministic approximation of the signal sample covariance [14]. These results motivate examining how practical communication components influence sensing accuracy over a finite observation frame. More recent studies have characterized this influence through the correlation and ambiguity properties of random communication waveforms. For CP-equipped orthogonal waveforms, it has been proven that OFDM achieves the lowest average ranging sidelobes under quadrature amplitude modulation (QAM) and phase-shift keying (PSK) constellations [15]. These findings were further extended to pulse-shaped signals in [16], where a metaphorical “iceberg hidden in the sea” structure was introduced to describe the statistical behavior of the ACF. Specifically, the “iceberg” corresponds to the squared mean of the ACF, determined by the squared ACF of the adopted pulse-shaping filter, whereas the “sea level,” arising from the randomness of data payloads, represents the variance of the ACF. For OFDM, probabilistic constellation shaping provides a means of controlling the sensing-communication tradeoff through the fourth moment of the constellation amplitudes [17]. Collectively, these studies establish that modulation, constellation statistics, and pulse shaping affect the sensing properties of random communication waveforms in different ways. The two-dimensional ambiguity-function analysis in [18] provides a particularly useful model for studying delay and Doppler jointly. It distinguishes the discrete periodic ambiguity function from the fast-slow-time ambiguity function, the latter relying on a small intra-block Doppler approximation. Under this approximation, delay acts as a circular shift along fast time, whereas Doppler produces phase rotations along slow time. The analysis establishes constellation-dependent sidelobe properties under general orthogonal modulation bases. However, sidelobe optimality does not by itself establish optimality with respect to a joint estimation CRB, which depends on the derivatives of the observation mean and the coupling among all unknown target parameters [19].

The closest estimation-theoretic result is the ranging CRB analysis in [20]. For joint estimation of multiple target delays and complex gains, that work factors the Fisher information matrix (FIM) into a deterministic target-parameter Jacobian and random frequency-domain energy weights. This structure yields a Jensen-type lower bound on the average conditional CRB, which CP-OFDM with PSK symbols attains exactly. For QAM and broader sub-Gaussian constellations, the same work develops a large-sample analysis comparing CP-OFDM with frequency-spread waveforms. Its principal design variable is the modulation basis, and its estimation model concerns target delays and complex gains. These results motivate a related fundamental question with a different focus: with the OFDM waveform fixed, which communication constellation is theoretically optimal for joint range–velocity estimation, and how do different constellations affect the range and velocity estimation accuracy of practical algorithms?

Answering this question requires accounting for both communication-symbol randomness and the coupling among unknown target delays, Doppler shifts, and complex gains. Although the transmitted symbols are known to the sensing receiver, their random energy fluctuations can make the FIM vary across finite frames. Since averaging and matrix inversion do not commute, the average FIM alone does not determine the average CRB. Moreover, grid quantization and noise amplification during symbol removal must also be considered when relating constellation optimality to practical range–velocity estimation performance. In this paper, we derive the joint range–velocity CRBs for data-bearing OFDM signals and investigate their dependence on the communication constellation. We further examine FFT, ML, and sliding-subarray MUSIC estimation under the same observation model. Numerical results show that ML estimation closely approaches the corresponding CRBs, confirming the validity of the theoretical analysis, while the comparisons with FFT and MUSIC reveal how constellation choice affects practical estimation accuracy. The main contributions are as follows.

  1. We derive an equivalent discrete sensing model directly from the continuous-time OFDM transmit and receive signals. The derivation explicitly identifies the required CP coverage and the approximation conditions for neglecting range migration and intra-symbol Doppler. We also connect the fast-time-slow-time, frequency-slow-time, and delay-Doppler representations, providing complementary views of the same sensing model.

  1. We derive the FIM for joint estimation of the target delays, Doppler shifts, and complex gains, and reveal a structural factorization that separates the deterministic Jacobian of the sensing response with respect to these parameters from the random symbol-energy weights. This factorization yields a Jensen-type lower bound on the average conditional CRB, which is exactly achieved by constant-modulus constellations such as PSK.

  1. We formulate 2D FFT, concentrated ML, and sliding-subarray MUSIC estimators under the derived observation model. For a single target with constant-modulus symbols, we show that ML estimation is equivalent to an off-grid peak search of the 2D FFT profile.

  1. Numerical results show that PSK yields lower average CRBs than QAM when the numbers of subcarriers and OFDM symbols are small, whereas their gap becomes negligible as these numbers increase. Under both constellations, ML estimation closely approaches the corresponding CRBs, confirming the validity of the theoretical analysis. FFT exhibits pronounced error floors due to its integer-grid constraint, while MUSIC achieves substantially lower estimation errors with a relatively small gap to the corresponding CRBs.

The remainder of this paper is organized as follows. Sections II and III develop the continuous-time and discrete-time OFDM sensing models, respectively. Section IV derives the FIM and CRBs, establishes constant-modulus optimality, and presents single-target scenario. Sections V and VI describe FFT-based and ML estimation, and joint range–velocity estimation using 2D MUSIC, respectively. Numerical results and conclusions are presented in Sections VII and VIII, respectively.

Notation: Bold uppercase and lowercase letters denote matrices and vectors, respectively. The superscripts (⋅)T(\cdot)^{T}, (⋅)H(\cdot)^{H}, and (⋅)∗(\cdot)^{*} denote transpose, Hermitian transpose, and element-wise conjugation. The operators ⊗\otimes, ⊙\odot, and vec⁡(⋅)\operatorname{vec}(\cdot) denote the Kronecker product, Hadamard product, and column-wise vectorization. The operators diag⁡(⋅)\operatorname{diag}(\cdot), Tr⁡(⋅)\operatorname{Tr}(\cdot), and Re⁡{⋅}\operatorname{Re}\{\cdot\} denote diagonalization, trace, and real part. The imaginary unit is j=−1j=\sqrt{-1}, and sinc⁡(x)=sin⁡(πx)/(πx)\operatorname{sinc}(x)=\sin(\pi x)/(\pi x). Matrix inequalities are in the positive-semidefinite sense. Subcarrier, symbol, and time-sample indices start at zero; target indices start at one.

Continuous-Time OFDM Sensing Model

Transmit Signal

Consider a monostatic ISAC system that uses an OFDM frame to communicate and to sense QQ point targets. The frame contains MM OFDM symbols, each occupying NN contiguous subcarriers with subcarrier spacing Δf\Delta f. Let S[k,m]S[k,m] denote the communication symbol transmitted on the kkth subcarrier of the mmth OFDM symbol, where k=0,…,N−1k=0,\ldots,N-1 and m=0,…,M−1m=0,\ldots,M-1. The transmitted symbols S[k,m]S[k,m] are independently drawn from a given constellation, such as PSK or QAM. We assume a normalized average symbol energy E{∣S[k,m]∣2}=Es\mathbb{E}\{|S[k,m]|^2\}=E_s for all k,mk,m. The symbol duration is Tu=1/ΔfT_u=1/\Delta f, and the bandwidth is B=NΔfB=N\Delta f. Let uu denote time relative to the beginning of the useful part (i.e., before CP appended) of an OFDM symbol. The continuous-time baseband waveform of symbol mm is

sm(u)=1N∑k=0N−1S[k,m]ej2πkΔfurect⁡(uTu),(1)s_m(u)=\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}S[k,m]e^{j2\pi k\Delta f u}\operatorname{rect}\left(\frac{u}{T_u}\right), \tag*{(1)}

where rect⁡(u)=1\operatorname{rect}(u)=1 for 0≤u<10\le u<1 and zero otherwise. To eliminate inter-symbol interference (ISI) caused by propagation delays, a cyclic prefix (CP) of duration TcpT_{\mathrm{cp}} is added to each OFDM symbol. The symbol duration with CP is therefore Tsym=Tu+TcpT_{\mathrm{sym}}=T_u+T_{\mathrm{cp}}. The subcarrier sum in (1) is periodic with period TuT_u. Extending this sum to −Tcp≤u<0-T_{\mathrm{cp}}\le u<0 produces the CP-extended waveform

sm,cp(u)=1N∑k=0N−1S[k,m]ej2πkΔfurect⁡(u+TcpTsym).(2)s_{m,\mathrm{cp}}(u)=\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}S[k,m]e^{j2\pi k\Delta f u}\operatorname{rect}\left(\frac{u+T_{\mathrm{cp}}}{T_{\mathrm{sym}}}\right). \tag*{(2)}

Taking the beginning of the first useful symbol as the time origin, the frame waveform is

sBB(t)=∑m=0M−1sm,cp(t−mTsym),(3)s_{\mathrm{BB}}(t)=\sum_{m=0}^{M-1}s_{m,\mathrm{cp}}(t-mT_{\mathrm{sym}}), \tag*{(3)}

where fcf_c is the carrier frequency. Symbol mm occupies mTsym−Tcp≤t<mTsym+TumT_{\mathrm{sym}}-T_{\mathrm{cp}}\le t<mT_{\mathrm{sym}}+T_u, and its useful window starts at mTsymmT_{\mathrm{sym}}. Finally, the complex passband representation can be written as

sRF(t)=sBB(t)ej2πfct.s_{\mathrm{RF}}(t)=s_{\mathrm{BB}}(t)e^{j2\pi f_ct}.

Continuous-Time Target Echoes

Let us exploit OFDM communication signals to detect QQ point targets. The qqth target has a reference range RqR_q, radial velocity vqv_q, and complex reflection coefficient βq\beta_q, with positive vqv_q indicating motion toward the sensor. The corresponding reference round-trip delay and Doppler frequency are τq=2Rq/c\tau_q=2R_q/c and fD,q=2vq/λf_{D,q}=2v_q/\lambda, respectively, where cc denotes the propagation speed and λ=c/fc\lambda=c/f_c is the carrier wavelength. Under a low-speed, constant-velocity model, the instantaneous round-trip delay satisfies

τq(t)≈τq−2vqct.\tau_q(t)\approx\tau_q-\frac{2v_q}{c}t.

Then, the received complex passband signal is

rRF(t)=∑q=1QβqsBB(t−τq(t))ej2πfc[t−τq(t)]+nRF(t),(4)r_{\mathrm{RF}}(t)=\sum_{q=1}^{Q}\beta_qs_{\mathrm{BB}}(t-\tau_q(t))e^{j2\pi f_c[t-\tau_q(t)]}+n_{\mathrm{RF}}(t), \tag*{(4)}

where nRF(t)n_{\mathrm{RF}}(t) represents receiver noise. After coherent downconversion, the baseband reflected echo is formulated as

yBB(t)=∑q=1QβqsBB(t−τq(t))e−j2πfcτq(t)+w(t),(5)y_{\mathrm{BB}}(t)=\sum_{q=1}^{Q}\beta_qs_{\mathrm{BB}}(t-\tau_q(t))e^{-j2\pi f_c\tau_q(t)}+w(t), \tag*{(5)}

where w(t)w(t) is the corresponding complex baseband noise. During one frame, the target delay varies as the target moves. When this variation is small relative to the inverse signal bandwidth, the resulting range migration can be neglected [21, 22]. Specifically, let TframeT_{\mathrm{frame}} denote the observation interval of one frame. We assume that

Bmax⁡t∈Tframe∣τq(t)−τq∣≪1,q=1,…,Q,(6)B\max_{t\in\mathcal{T}_{\mathrm{frame}}}|\tau_q(t)-\tau_q|\ll1,\quad q=1,\ldots,Q, \tag*{(6)}

and provided that the actual and reference delays remain within the CP-supported receive window, the time-varying delay in the signal envelope can be approximated by its reference value [23]:

sBB(t−τq(t))≈sBB(t−τq).(7)s_{\mathrm{BB}}(t-\tau_q(t))\approx s_{\mathrm{BB}}(t-\tau_q). \tag*{(7)}

The variation of the carrier phase is retained to characterize the Doppler shift. Substituting (4) into the carrier-phase term yields

e−j2πfcτq(t)≈e−j2πfcτqej2πfD,qt.(8)e^{-j2\pi f_c\tau_q(t)}\approx e^{-j2\pi f_c\tau_q}e^{j2\pi f_{D,q}t}. \tag*{(8)}

Define αq≜βqe−j2πfcτq\alpha_q\triangleq\beta_qe^{-j2\pi f_c\tau_q} as the effective complex gain. Substituting (7) and (8) into (5), we obtain

yBB(t)≈∑q=1QαqsBB(t−τq)ej2πfD,qt+w(t).(9)y_{\mathrm{BB}}(t)\approx\sum_{q=1}^{Q}\alpha_qs_{\mathrm{BB}}(t-\tau_q)e^{j2\pi f_{D,q}t}+w(t). \tag*{(9)}

Assuming symbol-level synchronization with the transmitted OFDM frame, we require the CP duration to be no shorter than the maximum target delay, i.e.,

Tcp≥max⁡1≤q≤Qτq.(10)T_{\mathrm{cp}}\ge\max_{1\le q\le Q}\tau_q. \tag*{(10)}

Under this condition, the delayed signal observed within the useful receive window of each OFDM symbol originates entirely from the CP or the useful part of the corresponding transmitted symbol, thereby avoiding ISI.

Subcarrier Demodulation

Let rm(u)r_m(u) denote the received baseband signal within the useful window of the mmth OFDM symbol, defined as

rm(u)≜yBB(mTsym+u),0≤u<Tu.(11)r_m(u)\triangleq y_{\mathrm{BB}}(mT_{\mathrm{sym}}+u),\quad0\le u<T_u. \tag*{(11)}

Substituting the transmitted waveform in (4) into the received signal model in (9), we obtain

rm(u)≈1N∑q=1Q∑k=0N−1αqS[k,m]×ej2πfD,quej2πmTsymfD,q×e−j2πkΔfτqej2πkΔfu+wm(u),(12)\begin{aligned} r_m(u)&\approx\frac{1}{\sqrt{N}}\sum_{q=1}^{Q}\sum_{k=0}^{N-1}\alpha_qS[k,m]\\ &\quad\times e^{j2\pi f_{D,q}u}e^{j2\pi mT_{\mathrm{sym}}f_{D,q}}\\ &\quad\times e^{-j2\pi k\Delta f\tau_q}e^{j2\pi k\Delta f u}+w_m(u), \tag*{(12)} \end{aligned}

where wm(u)≜w(mTsym+u)w_m(u)\triangleq w(mT_{\mathrm{sym}}+u) is the noise within the mmth useful receive window. Factor ① describes the Doppler-induced phase variation along the fast-time coordinate uu within an OFDM symbol. Factor ② describes the Doppler-induced phase progression along the slow-time coordinate mTsymmT_{\mathrm{sym}} across OFDM symbols. These two factors arise from the same physical Doppler shift, since ej2πfD,q(mTsym+u)=ej2πmTsymfD,qej2πfD,que^{j2\pi f_{D,q}(mT_{\mathrm{sym}}+u)}=e^{j2\pi mT_{\mathrm{sym}}f_{D,q}}e^{j2\pi f_{D,q}u}. Factor ③ is the delay-induced phase shift on the kkth subcarrier, whose variation across subcarriers encodes the target delay. Factor ④ is the kkth subcarrier basis function.

Since the fast-time interval of one useful OFDM symbol is limited, the Doppler-induced phase variation within this interval is typically small. We therefore adopt the small fast-time Doppler approximation [18] to obtain a subcarrier-wise sensing model, requiring

2π∣fD,q∣Tu≪1,q=1,…,Q.(13)2\pi\lvert f_{D,q}\rvert T_u \ll1,\qquad q=1,\ldots,Q. \tag*{(13)}

Hence, we have ej2πufD,q≈1e^{j2\pi u f_{D,q}}\approx1 for 0≤u<Tu0\leq u<T_u. This approximation neglects the fast-time (intra-symbol) Doppler phase variation, while the slow-time (inter-symbol) phase term ej2πmTsymfD,qe^{j2\pi mT_{\mathrm{sym}}f_{D,q}} is retained. Its counterpart in the normalized discrete-time model will be given in the next section.

Define H[k,m]H[k,m] as the sensing-channel coefficient associated with the kkth subcarrier and the mmth OFDM symbol:

H[k,m]=∑q=1Qαqe−j2πkΔfτqej2πmTsymfD,q.(14)H[k,m]=\sum_{q=1}^{Q}\alpha_q e^{-j2\pi k\Delta f\tau_q}e^{j2\pi mT_{\mathrm{sym}}f_{D,q}}. \tag*{(14)}

Substituting (14) and the approximation in (13) into (12), the received signal can be rewritten as

rm(u)≈1N∑k=0N−1H[k,m]S[k,m]ej2πkΔfu+wm(u).(15)r_m(u)\approx\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}H[k,m]S[k,m]e^{j2\pi k\Delta f u}+w_m(u). \tag*{(15)}

The received frequency-domain symbol on the ppth subcarrier of the mmth OFDM symbol is obtained by projecting rm(u)r_m(u) onto the corresponding subcarrier over the useful symbol interval:

Ys[p,m]=NTu∫0Turm(u)e−j2πpΔfu du.(16)Y_s[p,m]=\frac{\sqrt{N}}{T_u}\int_{0}^{T_u}r_m(u)e^{-j2\pi p\Delta f u}\,\mathrm{d}u. \tag*{(16)}

Substituting (15) into (16) and using the orthogonality of the subcarriers, we obtain

Ys[p,m]≈H[p,m]S[p,m]+Ws[p,m],(17)Y_s[p,m]\approx H[p,m]S[p,m]+W_s[p,m], \tag*{(17)}

where Ws[p,m]=NTu∫0Tuwm(u)e−j2πpΔfu duW_s[p,m]=\frac{\sqrt{N}}{T_u}\int_{0}^{T_u}w_m(u)e^{-j2\pi p\Delta f u}\,\mathrm{d}u is the noise component after subcarrier projection. Equation (17) gives the frequency-domain sensing model associated with continuous-time subcarrier demodulation. The received symbol Ys[p,m]Y_s[p,m] consists of the transmitted data symbol S[p,m]S[p,m] weighted by the sensing-channel coefficient H[p,m]H[p,m], together with the projected noise Ws[p,m]W_s[p,m]. As shown in (14), the target delays and Doppler shifts are embedded in H[p,m]H[p,m] through the phase variations across subcarriers and OFDM symbols, respectively. Both Ys[p,m]Y_s[p,m] and H[p,m]H[p,m] are represented on the subcarrier-slow-time grid. A delay-Doppler profile can subsequently be obtained by removing the known data symbols and applying an inverse Fourier transform along the subcarrier axis and a Fourier transform along the slow-time axis.

Remark 1: If intra-symbol Doppler is retained, substituting (12) into the continuous-time subcarrier projection in (16) gives the coupling coefficient from the kkth transmitted subcarrier to the ppth demodulated output:

Cp,k(c)(fD,q)=1Tu∫0Tuej2π[(k−p)Δf+fD,q]u du.(18)C^{(c)}_{p,k}(f_{D,q})=\frac{1}{T_u}\int_{0}^{T_u}e^{j2\pi[(k-p)\Delta f+f_{D,q}]u}\,\mathrm{d}u. \tag*{(18)}

Here, the superscript (c)(c) denotes continuous-time integral demodulation. The coefficient Cp,k(c)(fD,q)C^{(c)}_{p,k}(f_{D,q}) represents the projection of the kkth transmitted subcarrier, after the Doppler shift induced by target qq, onto the ppth receive subcarrier. The case p=kp=k corresponds to the response of the desired subcarrier, whereas p≠kp\ne k describes interference from another subcarrier into the ppth demodulated output. Thus, (18) describes subcarrier coupling, whose off-diagonal terms constitute inter-carrier interference (ICI). With a sufficiently long CP and zero Doppler, the propagation delay introduces only the phase factor e−j2πkΔfτqe^{-j2\pi k\Delta f\tau_q} in (12), without changing the subcarrier frequency. Subcarrier orthogonality is therefore preserved, and Cp,k(c)(0)=δpkC^{(c)}_{p,k}(0)=\delta_{pk}, where δpk\delta_{pk} is the Kronecker delta. A nonzero Doppler shift changes the transmitted subcarrier frequency from kΔfk\Delta f to kΔf+fD,qk\Delta f+f_{D,q}, which generally no longer aligns with the receive subcarrier grid. Consequently, ICI may remain even when the CP is sufficiently long. Under the small intra-symbol Doppler condition in (13), the phase variation caused by this frequency shift over the useful symbol interval can be neglected, yielding Cp,k(c)(fD,q)≈δpkC^{(c)}_{p,k}(f_{D,q})\approx\delta_{pk}. This approximate orthogonality should be distinguished from the exact orthogonality obtained when fD,q=0f_{D,q}=0.

The continuous-time signal flow and its corresponding discrete-time implementation are illustrated in Figure 1. In the following section, we establish the discrete frequency-domain sensing model by sampling the received waveform and applying an NN-point FFT after CP removal.

Figure 1. Continuous-time signal operations and their sampled counterparts.

III Discrete-Time OFDM Sensing Model

This section develops the discrete-time sensing model by sampling the transmitted waveform and received echoes and applying an FFT at the receiver. We first retain intra-symbol Doppler to characterize the resulting subcarrier coupling. To obtain a more compact formulation, we introduce normalized delay and Doppler coordinates and then apply the small intra-symbol Doppler approximation to establish the observation model for parameter estimation and CRB analysis. An equivalent factorization relates this model to the transmit, propagation, and receive operations and clarifies the associated signal domains.

III-A Transmitted Samples and OFDM Modulation

Let the baseband sampling interval be

Δt=TuN=1NΔf.(19)\Delta t=\frac{T_u}{N}=\frac{1}{N\Delta f}. \tag*{(19)}

With NCPN_{\mathrm{CP}} samples in the CP, its duration is TCP=NCPΔtT_{\mathrm{CP}}=N_{\mathrm{CP}}\Delta t, and the total OFDM symbol duration is Tsym=(N+NCP)ΔtT_{\mathrm{sym}}=(N+N_{\mathrm{CP}})\Delta t. Sampling the transmitted waveform at u=nΔtu=n\Delta t, the useful samples and their CP extension are given by

x[n,m]=1N∑k=0N−1S[k,m]ej2πkn/N,0≤n<N,(20)x[n,m]=\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}S[k,m]e^{j2\pi kn/N},\quad0\leq n<N, \tag*{(20)}
xCP[n,m]=x[n mod N,m],−NCP≤n<N.x_{\mathrm{CP}}[n,m]=x[n\bmod N,m],\quad-N_{\mathrm{CP}}\leq n<N.

The first expression admits a unitary NN-point IFFT, while the second prepends the last NCPN_{\mathrm{CP}} useful samples to each symbol.

Here, kk, nn, and mm index the subcarriers, fast-time samples, and slow-time symbols, respectively. Let FN∈CN×NF_N \in\mathbb{C}^{N \times N} denote the unitary NN-point DFT matrix, with [FN]k,n=1Ne−j2πkn/N[F_N]_{k,n} = \frac{1}{\sqrt{N}}e^{-j2\pi kn/N}. Collecting the communication symbols S[k,m]S[k,m] and useful transmitted samples x[n,m]x[n,m] into S,X∈CN×MS,X \in\mathbb{C}^{N \times M}, respectively, the OFDM synthesis admits the matrix representation

X=FNHS.(21)X = F_N^H S. \tag*{(21)}

The data matrix SS is defined on the subcarrier-slow-time grid. Left multiplication by FNHF_N^H converts its first coordinate into fast-time samples while preserving the slow-time coordinate. Thus, XX is a sample matrix on fast-time-slow-time grid whose mmth column contains the time-domain samples of symbol mm.

Received Samples and Subcarrier Coupling

We retain the fixed-envelope-delay approximation and CP coverage condition established in the continuous-time model, while initially preserving intra-symbol Doppler. Let global sample index ℓ=m(N+Ncp)+n\ell= m(N + N_{\mathrm{cp}}) + n indicates the local sample nn of symbol mm. Sampling the continuous-time baseband echo yBB(t)y_{\mathrm{BB}}(t) at t=ℓΔtt = \ell\Delta t yields

yframe[ℓ]≈∑q=1Qαqxτq[ℓ]ej2πfD,qℓΔt+w[ℓ],(22)y_{\mathrm{frame}}[\ell] \approx\sum_{q=1}^{Q}\alpha_q x_{\tau_q}[\ell]e^{j2\pi f_{D,q}\ell\Delta t} + w[\ell], \tag*{(22)}

where yframe[ℓ]=yBB(ℓΔt)y_{\mathrm{frame}}[\ell] = y_{\mathrm{BB}}(\ell\Delta t), w[ℓ]w[\ell] is the sampled receiver noise, and xτq[ℓ]=sBB(ℓΔt−τq)x_{\tau_q}[\ell] = s_{\mathrm{BB}}(\ell\Delta t - \tau_q) contains samples of the delayed transmitted waveform. This definition accommodates noninteger τq/Δt\tau_q/\Delta t without rounding the propagation delay to an integer sample shift.

After symbol synchronization and CP removal, sampling rm(u)r_m(u) at u=nΔtu = n\Delta t yields

yr[n,m]≈1N∑q=1Q∑k=0N−1αqS[k,m]e−j2πkΔfτqej2πkn/N×ej2πmTsymfD,qej2πnΔtfD,q+wr[n,m],(23)y_r[n,m] \approx\frac{1}{\sqrt{N}}\sum_{q=1}^{Q}\sum_{k=0}^{N-1}\alpha_q S[k,m]e^{-j2\pi k\Delta f\tau_q}e^{j2\pi kn/N} \times e^{j2\pi mT_{\mathrm{sym}}f_{D,q}}e^{j2\pi n\Delta t f_{D,q}} + w_r[n,m], \tag*{(23)}

where wr[n,m]=w[m(N+Ncp)+n]w_r[n,m] = w[m(N + N_{\mathrm{cp}}) + n]. The two Doppler factors describe the phase variation along fast time nΔtn\Delta t and slow time mTsymmT_{\mathrm{sym}}, respectively. They arise from the same physical Doppler phase expressed in the two time coordinates.

To obtain a matrix representation, define the delay-phase matrix and the fast-time and slow-time Doppler matrices as follows:

Λ(τ)=diag⁡(1,e−j2πΔfτ,…,e−j2π(N−1)Δfτ),\Lambda(\tau) = \operatorname{diag}\left(1,e^{-j2\pi\Delta f\tau},\ldots,e^{-j2\pi(N-1)\Delta f\tau}\right),
Df(fD)=diag⁡(1,ej2πΔtfD,…,ej2π(N−1)ΔtfD),D_f(f_D) = \operatorname{diag}\left(1,e^{j2\pi\Delta t f_D},\ldots,e^{j2\pi(N-1)\Delta t f_D}\right),
Ds(fD)=diag⁡(1,ej2πTsymfD,…,ej2π(M−1)TsymfD).(24)D_s(f_D) = \operatorname{diag}\left(1,e^{j2\pi T_{\mathrm{sym}}f_D},\ldots,e^{j2\pi(M-1)T_{\mathrm{sym}}f_D}\right). \tag*{(24)}

Let Yr,Wr∈CN×MY_r,W_r \in\mathbb{C}^{N \times M} collect the useful received samples and noise samples, respectively. The sampled echo in (23) can then be written in matrix form as

Yr≈∑q=1QαqDf(fD,q)FNHΛ(τq)SDs(fD,q)+Wr.(25)Y_r \approx\sum_{q=1}^{Q}\alpha_q D_f(f_{D,q})F_N^H\Lambda(\tau_q)SD_s(f_{D,q}) + W_r. \tag*{(25)}

We then perform OFDM subcarrier demodulation by applying an NN-point FFT along fast time, yielding the observed data matrix

Ys=FNYr≈∑q=1QαqC(fD,q)Λ(τq)SDs(fD,q)+Ws,(26)Y_s = F_NY_r \approx\sum_{q=1}^{Q}\alpha_q C(f_{D,q})\Lambda(\tau_q)SD_s(f_{D,q}) + W_s, \tag*{(26)}

where Ws=FNWrW_s = F_NW_r is the noise with each element follows CN(0,σw2)\mathcal{CN}(0,\sigma_w^2), and C(fD)=FNDf(fD)FNHC(f_D) = F_ND_f(f_D)F_N^H is the Doppler-induced subcarrier coupling matrix, whose (p,k)(p,k)th entry is

[C(fD)]p,k=1N∑n=0N−1ej2πn[(k−p)/N+ΔtfD],(27)[C(f_D)]_{p,k} = \frac{1}{N}\sum_{n=0}^{N-1}e^{j2\pi n[(k-p)/N+\Delta t f_D]}, \tag*{(27)}

which describes the contribution of transmitted subcarrier kk to FFT output pp and is the discrete counterpart of the continuous-time projection coefficient in (18). The off-diagonal entries characterize the ICI retained in the current model. A sufficiently long CP prevents delay-induced ISI, whereas Doppler-induced subcarrier coupling may remain.

For notational convenience and to simplify the subsequent derivations, we express delay and Doppler in grid-normalized coordinates, which can take integer or fractional values. Define the delay and Doppler-frequency cell sizes as

Δτ=Δt=1NΔf,ΔfD=1MTsym.(28)\Delta\tau= \Delta t = \frac{1}{N\Delta f},\qquad\Delta f_D = \frac{1}{MT_{\mathrm{sym}}}. \tag*{(28)}

The corresponding range and velocity resolutions are

ΔR=cΔτ2=c2NΔf,Δv=λΔfD2=λ2MTsym.(29)\Delta R = \frac{c\Delta\tau}{2} = \frac{c}{2N\Delta f},\qquad\Delta v = \frac{\lambda\Delta f_D}{2} = \frac{\lambda}{2MT_{\mathrm{sym}}}. \tag*{(29)}

Then, the normalized delay and Doppler frequency of target qq are given by

τˉq=τqΔτ=NΔfτq,νˉq=fD,qΔfD=MTsymfD,q.(30)\bar{\tau}_q = \frac{\tau_q}{\Delta\tau} = N\Delta f\tau_q,\qquad\bar{\nu}_q = \frac{f_{D,q}}{\Delta f_D} = MT_{\mathrm{sym}}f_{D,q}. \tag*{(30)}

τˉq\bar{\tau}_q measures delay in units of the fast-time sampling interval, while νˉq\bar{\nu}_q measures Doppler frequency in units of the slow-time DFT spacing. Hence, the target range and velocity are expressed as Rq=ΔRτˉqR_q = \Delta R\bar{\tau}_q, vq=Δvνˉqv_q = \Delta v\bar{\nu}_q, respectively. Both parameters are real-valued and may be noninteger, corresponding to fractional delays and Doppler frequencies. The vectors in (24) can accordingly be expressed in normalized coordinates as

BˉN(τˉ)≜Λ(Δττˉ)=diag⁡({e−j2πkτˉ/N}k=0N−1),Dˉs(νˉ)≜Ds(ΔfDνˉ)=diag⁡({ej2πmνˉ/M}m=0M−1),Dˉf(νˉ)≜Df(ΔfDνˉ)=diag⁡({ej2πnνˉ/[M(N+Ncp)]}n=0N−1).(31)\begin{aligned} \bar{B}_N(\bar{\tau}) &\triangleq\Lambda(\Delta\tau\bar{\tau}) = \operatorname{diag}\left(\left\{e^{-j2\pi k\bar{\tau}/N}\right\}_{k=0}^{N-1}\right),\\ \bar{D}_s(\bar{\nu}) &\triangleq D_s(\Delta f_D\bar{\nu}) = \operatorname{diag}\left(\left\{e^{j2\pi m\bar{\nu}/M}\right\}_{m=0}^{M-1}\right),\\ \bar{D}_f(\bar{\nu}) &\triangleq D_f(\Delta f_D\bar{\nu}) = \operatorname{diag}\left(\left\{e^{j2\pi n\bar{\nu}/[M(N+N_{\mathrm{cp}})]}\right\}_{n=0}^{N-1}\right). \tag*{(31)} \end{aligned}

Substituting these expressions into (25), the received time-domain matrix becomes

Yr≈∑q=1QαqDˉf(νˉq)FNHBˉN(τˉq)SDˉs(νˉq)+Wr.(32)Y_r \approx\sum_{q=1}^{Q}\alpha_q\bar{D}_f(\bar{\nu}_q)F_N^H\bar{B}_N(\bar{\tau}_q)S\bar{D}_s(\bar{\nu}_q)+W_r. \tag*{(32)}

The corresponding subcarrier observation after the receive FFT is

Ys≈∑q=1Qαq[FNDˉf(νˉq)FNH]BˉN(τˉq)SDˉs(νˉq)+Ws.(33)Y_s \approx\sum_{q=1}^{Q}\alpha_q\left[F_N\bar{D}_f(\bar{\nu}_q)F_N^H\right]\bar{B}_N(\bar{\tau}_q)S\bar{D}_s(\bar{\nu}_q)+W_s. \tag*{(33)}

The two expressions describe the fast-time–slow-time observations before the receive FFT and the subcarrier–slow-time observations after it, respectively. Both retain the delay-dependent phase and the intra-symbol and inter-symbol Doppler phases.

Small Fast-Time Doppler Approximation

The small intra-symbol Doppler condition 2π∣fD,q∣Tu≪12\pi\lvert f_{D,q}\rvert T_u \ll1, established in (13) for the continuous-time model, can be expressed in normalized coordinates as

2π∣fD,q∣Tu=2πN∣νˉq∣M(N+Ncp)≪1,q=1,…,Q.(34)2\pi\lvert f_{D,q}\rvert T_u = \frac{2\pi N\lvert\bar{\nu}_q\rvert}{M(N+N_{\mathrm{cp}})} \ll1,\qquad q=1,\ldots,Q. \tag*{(34)}

Then, we have,

Dˉf(νˉq)≈IN,FNDˉf(νˉq)FNH≈IN.(35)\bar{D}_f(\bar{\nu}_q)\approx I_N,\qquad F_N\bar{D}_f(\bar{\nu}_q)F_N^H\approx I_N. \tag*{(35)}

This approximation neglects intra-symbol Doppler phase variation. Substituting (35) into (32) and (33) yields the simplified received signal

Yr≈∑q=1QαqFNHBˉN(τˉq)SDˉs(νˉq)+Wr,(36)Y_r \approx\sum_{q=1}^{Q}\alpha_qF_N^H\bar{B}_N(\bar{\tau}_q)S\bar{D}_s(\bar{\nu}_q)+W_r, \tag*{(36)}

and the corresponding demodulated signals

Ys≈∑q=1QαqBˉN(τˉq)SDˉs(νˉq)+Ws.(37)Y_s \approx\sum_{q=1}^{Q}\alpha_q\bar{B}_N(\bar{\tau}_q)S\bar{D}_s(\bar{\nu}_q)+W_s. \tag*{(37)}

The demodulated matrix YsY_s is represented on the subcarrier–slow-time grid. The diagonal structure in (37) allows the sensing response to be separated from the transmitted data through an element-wise product. To obtain this representation, define the delay steering vector b(τˉ)∈CNb(\bar{\tau})\in\mathbb{C}^{N} and Doppler steering vector d(νˉ)∈CMd(\bar{\nu})\in\mathbb{C}^{M} as

b(τˉ)=[1,e−j2πτˉ/N,…,e−j2π(N−1)τˉ/N]T,d(νˉ)=[1,e−j2πνˉ/M,…,e−j2π(M−1)νˉ/M]T.(38)\begin{aligned} b(\bar{\tau}) &= \left[1,e^{-j2\pi\bar{\tau}/N},\ldots,e^{-j2\pi(N-1)\bar{\tau}/N}\right]^T,\\ d(\bar{\nu}) &= \left[1,e^{-j2\pi\bar{\nu}/M},\ldots,e^{-j2\pi(M-1)\bar{\nu}/M}\right]^T. \tag*{(38)} \end{aligned}

These vectors satisfy BˉN(τˉ)=diag⁡(b(τˉ))\bar{B}_N(\bar{\tau})=\operatorname{diag}(b(\bar{\tau})) and Dˉs(νˉ)=diag⁡(d∗(νˉ))\bar{D}_s(\bar{\nu})=\operatorname{diag}(d^*(\bar{\nu})). Collecting the target contributions gives the sensing-channel matrix

H=∑q=1Qαqb(τˉq)dH(νˉq),(39)H=\sum_{q=1}^{Q}\alpha_qb(\bar{\tau}_q)d^H(\bar{\nu}_q), \tag*{(39)}

with H[k,m]=∑q=1Qαqe−j2πkτˉq/Nej2πmνˉq/MH[k,m]=\sum_{q=1}^{Q}\alpha_qe^{-j2\pi k\bar{\tau}_q/N}e^{j2\pi m\bar{\nu}_q/M}. With the small fast-time Doppler approximation, we therefore obtain the observation model

Ys=H⊙S+Ws.(40)Y_s=H\odot S+W_s. \tag*{(40)}

This model relies on the fixed-envelope-delay approximation and two distinct conditions. The CP coverage condition, equivalently 0≤τˉq≤Ncp0\leq\bar{\tau}_q\leq N_{\mathrm{cp}}, ensures that the delayed echo within each useful receive window follows the periodic extension of the corresponding transmitted symbol, thereby avoiding delay-induced ISI. The small fast-time Doppler condition in (34) allows the resulting DFT to be neglected. Together, these conditions yield the subcarrier-wise multiplicative structure in (37) and a more compact formulation in (40).

Some Insights of the Sensing Model

The signal propagating through the physical channel is always a time-domain waveform. As the signal is transformed between domains at different stages of OFDM processing, the same effects of propagation delay and Doppler shift take different mathematical forms, typically appearing as phase rotations or circular shifts. The original communication symbols, however, are complex constellation points without an intrinsic representation domain. In the general two-dimensional modulation relation X=USVX=USV, the matrices UU and VV determine how the two data coordinates map to fast-time and slow-time samples. For OFDM, U=FNHU=F_N^H and V=IMV=I_M, yielding

X=FNHS.X=F_N^HS.

The matrix XX contains the transmitted fast-time–slow-time samples. Its subcarrier–slow-time representation is obtained by applying a Fourier transform along fast time,

XFT=FNX=S.X_{\mathrm{FT}}=F_NX=S.

Thus, the interpretation of SS as a subcarrier–slow-time matrix follows from the OFDM modulation mapping.

The delay-induced phase progression across subcarriers is equivalent to a circular shift in fast time, as follows from the Fourier diagonalization of circulant matrices.

JN(τˉ)=FNHBN(τˉ)FN,BN(τˉ)=FNJN(τˉ)FNH(41)J_N(\bar{\tau}) = F_N^H B_N(\bar{\tau})F_N, \qquad B_N(\bar{\tau}) = F_N J_N(\bar{\tau})F_N^H \tag*{(41)}

For an integer τˉ\bar{\tau}, JN(τˉ)J_N(\bar{\tau}) implements a circular shift along fast time. For a noninteger τˉ\bar{\tau}, it implements the corresponding fractional circular delay of the periodic OFDM waveform. Substituting this relationship and X=FNHSX = F_N^H S into the received time-domain model in (36) gives

Yr≈∑q=1QαqJN(τˉq)XD‾s(νˉq)+Wr.(42)Y_r \approx\sum_{q=1}^{Q} \alpha_q J_N(\bar{\tau}_q)X\overline{D}_s(\bar{\nu}_q)+W_r. \tag*{(42)}

Here, delay shifts the waveform along fast time, while Doppler induces a phase progression along slow time. Applying the receive FFT gives the equivalent subcarrier–slow-time model

Ys=FNYr≈∑q=1QαqBN(τˉq)XFMTD‾s(νˉq)+Ws,(43)Y_s = F_NY_r \approx\sum_{q=1}^{Q}\alpha_q B_N(\bar{\tau}_q)XF_M^T\overline{D}_s(\bar{\nu}_q)+W_s, \tag*{(43)}

which reduces to (37) for OFDM. The circular delay is thereby converted into a phase progression across subcarriers, while the slow-time Doppler phases are preserved. This relationship reflects the Fourier duality between the subcarrier and fast-time domains.

Slow time and Doppler form another Fourier-dual pair. Transforming the slow-time coordinate gives the DD representations of the transmitted and received waveforms,

XDD=XFM=FNHSFM,X_{\mathrm{DD}} = XF_M = F_N^H SF_M,
YDD=YrFM=FNHYsFM,(44)Y_{\mathrm{DD}} = Y_rF_M = F_N^H Y_sF_M, \tag*{(44)}

where FMF_M is the unitary MM-point DFT matrix. The slow-time Doppler matrix satisfies

JM(−νˉ)=FMHD‾s(νˉ)FM,D‾s(νˉ)=FMJM(−νˉ)FMH.J_M(-\bar{\nu}) = F_M^H\overline{D}_s(\bar{\nu})F_M, \qquad \overline{D}_s(\bar{\nu}) = F_MJ_M(-\bar{\nu})F_M^H.

Consequently, the received DD waveform can be expressed as

YDD≈∑q=1QαqJN(τˉq)XDDJM(−νˉq)+WDD,(45)Y_{\mathrm{DD}} \approx\sum_{q=1}^{Q}\alpha_qJ_N(\bar{\tau}_q)X_{\mathrm{DD}}J_M(-\bar{\nu}_q)+W_{\mathrm{DD}}, \tag*{(45)}

where WDD=WrFMW_{\mathrm{DD}}=W_rF_M. The slow-time phase progression is converted into a shift along the Doppler coordinate, so that both delay and Doppler are represented by circular shifts. Right multiplication by JM(−ℓ)J_M(-\ell) means shifting the columns by ℓ\ell positions in the positive direction. Integer normalized parameters produce grid-aligned shifts, while noninteger parameters produce fractional circular shifts. The relationships among the transmitted and received waveform representations are illustrated in Figure 2.

Figure 2. Domain transformations and representations of delay-Doppler effects at different processing stages.

These representations describe the same received echo under the adopted sensing model. For OFDM, the subcarrier–slow-time representation directly yields the element-wise observation model Ys=H⊙S+WsY_s=H\odot S+W_s in (40), which separates the sensing response from the transmitted symbols and provides the basis for the subsequent parameter estimation and CRB analysis.

The Structure of FIM and CRBs

This section analyzes the joint delay–Doppler estimation CRB in the presence of unknown complex target gains. We first derive a structural factorization of the complete FIM, which provides the basis for the subsequent average conditional CRB analysis.

Structural Factorization of the FIM

We begin with the subcarrier observation model, with the transmitted symbols and the number of targets assumed known. Vectorizing the received matrix in (40) yields

y=B(θ)α+w,(46)\boldsymbol{y}=B(\boldsymbol{\theta})\boldsymbol{\alpha}+\boldsymbol{w}, \tag*{(46)}

where y=vec⁡(Ys)\boldsymbol{y}=\operatorname{vec}(Y_s), α=[α1,…,αQ]T\boldsymbol{\alpha}=[\alpha_1,\ldots,\alpha_Q]^T, and w=vec⁡(Ws)∼CN(0,σW2INM)\boldsymbol{w}=\operatorname{vec}(W_s)\sim\mathcal{CN}(0,\sigma_W^2I_{NM}). The vector θ=[τˉT,νˉT]T∈R2Q\boldsymbol{\theta}=[\bar{\boldsymbol{\tau}}^T,\bar{\boldsymbol{\nu}}^T]^T\in\mathbb{R}^{2Q} collects the normalized target delays and Doppler shifts, where τˉ=[τˉ1,…,τˉQ]T\bar{\boldsymbol{\tau}}=[\bar{\tau}_1,\ldots,\bar{\tau}_Q]^T and νˉ=[νˉ1,…,νˉQ]T\bar{\boldsymbol{\nu}}=[\bar{\nu}_1,\ldots,\bar{\nu}_Q]^T. The data-weighted steering matrix is

B(θ)=diag⁡(vec⁡(S))A(θ)=[vec⁡(S)⊙a1,…,vec⁡(S)⊙aQ]∈CNM×Q,(47)B(\boldsymbol{\theta})=\operatorname{diag}(\operatorname{vec}(S))A(\boldsymbol{\theta}) =[\operatorname{vec}(S)\odot\boldsymbol{a}_1,\ldots,\operatorname{vec}(S)\odot\boldsymbol{a}_Q]\in\mathbb{C}^{NM\times Q}, \tag*{(47)}

with

A(θ)=[a1,…,aQ]∈CNM×Q,(48)A(\boldsymbol{\theta})=[\boldsymbol{a}_1,\ldots,\boldsymbol{a}_Q]\in\mathbb{C}^{NM\times Q}, \tag*{(48)}

and aq=d∗(νˉq)⊗b(τˉq)∈CNM\boldsymbol{a}_q=\boldsymbol{d}^{*}(\bar{\nu}_q)\otimes\boldsymbol{b}(\bar{\tau}_q)\in\mathbb{C}^{NM} denoting the joint delay–Doppler steering vector of target qq. Using vec⁡(bdH)=d∗⊗b\operatorname{vec}(\boldsymbol{b}\boldsymbol{d}^{H})=\boldsymbol{d}^{*}\otimes\boldsymbol{b}, the vectorized sensing channel satisfies

h≜vec⁡(H)=A(θ)α.(49)\boldsymbol{h}\triangleq\operatorname{vec}(H)=A(\boldsymbol{\theta})\boldsymbol{\alpha}. \tag*{(49)}

The normalized target delays, Doppler shifts, and complex gains are jointly estimated through the real parameter vector

ϑ=[τˉT,νˉT,αRT,αIT]T∈R4Q,(50)\boldsymbol{\vartheta}=[\bar{\boldsymbol{\tau}}^T,\bar{\boldsymbol{\nu}}^T,\boldsymbol{\alpha}_R^T,\boldsymbol{\alpha}_I^T]^T\in\mathbb{R}^{4Q}, \tag*{(50)}

where

αR=[Re⁡{α1},…,Re⁡{αQ}]T,(51)\boldsymbol{\alpha}_R=[\operatorname{Re}\{\alpha_1\},\ldots,\operatorname{Re}\{\alpha_Q\}]^T, \tag*{(51)}
αI=[Im⁡{α1},…,Im⁡{αQ}]T.\boldsymbol{\alpha}_I=[\operatorname{Im}\{\alpha_1\},\ldots,\operatorname{Im}\{\alpha_Q\}]^T.

We further denote ξ=[αRT,αIT]T\boldsymbol{\xi}=[\boldsymbol{\alpha}_R^T,\boldsymbol{\alpha}_I^T]^T, so that ϑ=[θT,ξT]T\boldsymbol{\vartheta}=[\boldsymbol{\theta}^T,\boldsymbol{\xi}^T]^T. The effective gains are treated as independent unknown complex parameters. Accordingly, the carrier-delay phase absorbed into these gains is not treated as separately available delay information. Conditioned on the transmitted symbols, the observation is proper complex Gaussian with mean

μ(ϑ;S)=B(θ)α.(52)\boldsymbol{\mu}(\boldsymbol{\vartheta};S)=B(\boldsymbol{\theta})\boldsymbol{\alpha}. \tag*{(52)}

For notational convenience, we use AA and A(θ)A(\boldsymbol{\theta}), as well as BB and B(θ)B(\boldsymbol{\theta}), interchangeably hereafter.

Since the noise covariance is independent of ϑ\boldsymbol{\vartheta}, the (i,j)(i,j)th element of the FIM J(S)∈R4Q×4QJ(S)\in\mathbb{R}^{4Q\times4Q} is

[J(S)]i,j=2σW2Re⁡{(∂μ∂ϑi)H∂μ∂ϑj},i,j=1,…,4Q.(53)[J(S)]_{i,j}=\frac{2}{\sigma_W^2}\operatorname{Re}\left\{\left(\frac{\partial\boldsymbol{\mu}}{\partial\vartheta_i}\right)^H\frac{\partial\boldsymbol{\mu}}{\partial\vartheta_j}\right\},\qquad i,j=1,\ldots,4Q. \tag*{(53)}

To obtain a compact FIM, let K,L∈RNM×NMK,L\in\mathbb{R}^{NM\times NM} collect the subcarrier and OFDM-symbol indices, respectively,

K=IM⊗diag⁡(0,…,N−1),K=I_M\otimes\operatorname{diag}(0,\ldots,N-1),
L=diag⁡(0,…,M−1)⊗IN.(54)L=\operatorname{diag}(0,\ldots,M-1)\otimes I_N. \tag*{(54)}

The delay and Doppler Jacobians of μ(ϑ;S)\boldsymbol{\mu}(\boldsymbol{\vartheta};S) are

Dτ≜∂μ∂ϑτT=−j2πNKBdiag⁡(α),Dν≜∂μ∂ϑνT=j2πMLBdiag⁡(α).(55)\mathbf{D}_{\tau} \triangleq\frac{\partial\boldsymbol{\mu}}{\partial\boldsymbol{\vartheta}_{\tau}^{T}}=-\frac{j2\pi}{N}\mathbf{K}\mathbf{B}\operatorname{diag}(\boldsymbol{\alpha}), \qquad \mathbf{D}_{\nu} \triangleq\frac{\partial\boldsymbol{\mu}}{\partial\boldsymbol{\vartheta}_{\nu}^{T}}=\frac{j2\pi}{M}\mathbf{L}\mathbf{B}\operatorname{diag}(\boldsymbol{\alpha}). \tag*{(55)}

Denote D=[Dτ,Dν]∈CNM×2Q\mathbf{D}=[\mathbf{D}_{\tau},\mathbf{D}_{\nu}]\in\mathbb{C}^{NM\times2Q}. Including the derivatives with respect to the real and imaginary gain components gives the complete mean Jacobian

G≜∂μ∂ϑT=[Dτ,Dν,B,jB]∈CNM×4Q.(56)\mathbf{G}\triangleq\frac{\partial\boldsymbol{\mu}}{\partial\boldsymbol{\vartheta}^{T}}=[\mathbf{D}_{\tau},\mathbf{D}_{\nu},\mathbf{B},j\mathbf{B}]\in\mathbb{C}^{NM\times4Q}. \tag*{(56)}

The complete FIM consequently admits the compact expression

J(S)=2σW2Re⁡{GHG}.(57)\mathbf{J}(S)=\frac{2}{\sigma_{W}^{2}}\operatorname{Re}\{\mathbf{G}^{H}\mathbf{G}\}. \tag*{(57)}

To reveal how the target parameters and transmitted symbols contribute to this FIM, we factor the symbol weights out of the mean Jacobian. The diagonal matrices K\mathbf{K} and L\mathbf{L} commute with diag⁡(vec⁡(S))\operatorname{diag}(\operatorname{vec}(S)), giving

G=diag⁡(vec⁡(S))G0,(58)\mathbf{G}=\operatorname{diag}(\operatorname{vec}(S))\mathbf{G}_{0}, \tag*{(58)}

where G0∈CNM×4Q\mathbf{G}_{0}\in\mathbb{C}^{NM\times4Q} is the Jacobian of the vectorized sensing channel with respect to the complete target-parameter vector, given by

G0≜∂h∂ϑT=[−j2πNKAdiag⁡(α),j2πMLAdiag⁡(α),A,jA].(59)\mathbf{G}_{0}\triangleq\frac{\partial\mathbf{h}}{\partial\boldsymbol{\vartheta}^{T}} = \left[ -\frac{j2\pi}{N}\mathbf{K}\mathbf{A}\operatorname{diag}(\boldsymbol{\alpha}), \frac{j2\pi}{M}\mathbf{L}\mathbf{A}\operatorname{diag}(\boldsymbol{\alpha}), \mathbf{A},j\mathbf{A} \right]. \tag*{(59)}

For fixed target parameters, G0\mathbf{G}_{0} is deterministic and describes the sensitivity of the sensing channel to changes in the target delays, Doppler shifts, and complex gains. Substituting this factorization into (57) gives the equivalent expression

J(S)=2σW2Re⁡{G0Hdiag⁡(∣vec⁡(S)∣2)G0}.(60)\mathbf{J}(S)=\frac{2}{\sigma_{W}^{2}}\operatorname{Re}\left\{\mathbf{G}_{0}^{H}\operatorname{diag}\left(\left|\operatorname{vec}(S)\right|^{2}\right)\mathbf{G}_{0}\right\}. \tag*{(60)}

This factorization separates the deterministic derivatives of the sensing channel from the random symbol-energy weights. The complete FIM depends linearly on the energy of each transmitted symbol, while the symbol phases cancel. The energy distribution across subcarriers and OFDM symbols is retained, so equal total frame energy alone does not imply equal conditional CRBs.

Average CRB and Constant-Modulus Optimality

Let ϑ^\widehat{\boldsymbol{\vartheta}} be an estimator of the complete target-parameter vector that is locally unbiased conditioned on the transmitted symbols. Under the usual regularity conditions and assuming J(S)≻0\mathbf{J}(S)\succ0, its conditional error covariance satisfies

Cov⁡(ϑ^∣S)⪰J−1(S).(61)\operatorname{Cov}\left(\widehat{\boldsymbol{\vartheta}}\mid S\right)\succeq\mathbf{J}^{-1}(S). \tag*{(61)}

The delay and Doppler bounds are extracted from this complete inverse FIM using

Pτ=[IQ,0,0,0]∈RQ×4Q,Pν=[0,IQ,0,0]∈RQ×4Q.(62)\mathbf{P}_{\tau}=[\mathbf{I}_{Q},\mathbf{0},\mathbf{0},\mathbf{0}]\in\mathbb{R}^{Q\times4Q}, \qquad \mathbf{P}_{\nu}=[\mathbf{0},\mathbf{I}_{Q},\mathbf{0},\mathbf{0}]\in\mathbb{R}^{Q\times4Q}. \tag*{(62)}

For a∈{τ,ν}a\in\{\tau,\nu\}, let ta=Paϑ\mathbf{t}_{a}=\mathbf{P}_{a}\boldsymbol{\vartheta} and t^a=Paϑ^\widehat{\mathbf{t}}_{a}=\mathbf{P}_{a}\widehat{\boldsymbol{\vartheta}}. The conditional CRB for their total squared estimation error is

CRB⁡a(S)≜Tr⁡(PaJ−1(S)PaT),(63)\operatorname{CRB}_{a}(S)\triangleq\operatorname{Tr}\left(\mathbf{P}_{a}\mathbf{J}^{-1}(S)\mathbf{P}_{a}^{T}\right), \tag*{(63)}

which accounts for the uncertainty in all jointly estimated parameters, including the unknown complex gains. For a given transmitted frame,

EWs[∥t^a−ta∥22∣S]≥CRB⁡a(S).(64)\mathbb{E}_{W_s}\left[\left\|\widehat{\mathbf{t}}_{a}-\mathbf{t}_{a}\right\|_{2}^{2}\mid S\right]\geq\operatorname{CRB}_{a}(S). \tag*{(64)}

To evaluate the delay–Doppler block of the complete inverse FIM directly, we use its Schur-complement representation. Recall that ϑ=[θT,ξT]T\boldsymbol{\vartheta}=[\boldsymbol{\theta}^{T},\boldsymbol{\xi}^{T}]^{T}, so the complete FIM can be partitioned as

J(S)=[JθθJθξJθξTJξξ].(65)\mathbf{J}(S)= \begin{bmatrix} \mathbf{J}_{\theta\theta} & \mathbf{J}_{\theta\xi}\\ \mathbf{J}_{\theta\xi}^{T} & \mathbf{J}_{\xi\xi} \end{bmatrix}. \tag*{(65)}

When the complete FIM is positive definite, the block-inversion identity gives

[J−1(S)]θθ=Jθ−1(S),(66)\left[\mathbf{J}^{-1}(S)\right]_{\theta\theta}=\mathbf{J}_{\theta}^{-1}(S), \tag*{(66)}

where the equivalent FIM for the delays and Doppler is

Je(S)=Jθθ−JθξJξξ−1JθξT(67)J_{\mathrm{e}}(\mathbf{S})=J_{\boldsymbol{\theta}\boldsymbol{\theta}}-J_{\boldsymbol{\theta}\boldsymbol{\xi}}J_{\boldsymbol{\xi}\boldsymbol{\xi}}^{-1}J_{\boldsymbol{\theta}\boldsymbol{\xi}}^{T} \tag*{(67)}

These two expressions therefore provide the same joint delay–Doppler CRB, with the uncertainty in the unknown complex gains fully accounted for. The following lemma gives a projection representation of this Schur complement.

Lemma 1. For a full-column-rank BB, the equivalent FIM can be expressed as

Je(S)=2σW2Re⁡{DHΠB⊥D}(68)J_{\mathrm{e}}(\mathbf{S})=\frac{2}{\sigma_{W}^{2}}\operatorname{Re}\left\{D^{H}\Pi_{B}^{\perp}D\right\} \tag*{(68)}

where

ΠB⊥=INM−B(BHB)−1BH(69)\Pi_{B}^{\perp}=I_{NM}-B\left(B^{H}B\right)^{-1}B^{H} \tag*{(69)}

is the orthogonal projector onto the complement of the column space of BB.

Proof. See Appendix A.

This projection retains the components of the delay- and Doppler-induced changes in the received mean that cannot be explained by changes in the unknown gains.

To characterize sensing accuracy under random communication symbols, we average the conditional CRB over the transmitted data:

CRB‾a≜ES[CRBa(S)]=Tr⁡(PaES[J−1(S)]PaT)(70)\overline{\mathrm{CRB}}_{a}\triangleq\mathbb{E}_{S}\left[\mathrm{CRB}_{a}(\mathbf{S})\right]=\operatorname{Tr}\left(P_{a}\mathbb{E}_{S}\left[J^{-1}(\mathbf{S})\right]P_{a}^{T}\right) \tag*{(70)}

The target parameters remain fixed in this expectation, while the transmitted symbols follow the prescribed constellation distribution. Averaging the conditional MSE inequality gives

ES,Ws[∥t^a−ta∥22]≥CRB‾a.(71)\mathbb{E}_{S,W_{s}}\left[\left\|\hat{\mathbf{t}}_{a}-\mathbf{t}_{a}\right\|_{2}^{2}\right]\geq\overline{\mathrm{CRB}}_{a}. \tag*{(71)}

This average conditional CRB is a Miller–Chang-type bound [13]. Its evaluation requires averaging the inverse FIM across transmitted frames, which generally differs from inverting the average FIM.

The structural factorization derived above allows the average FIM to be evaluated directly:

J‾≜ES[J(S)]=2EsσW2Re⁡{G0HG0}(72)\overline{J}\triangleq\mathbb{E}_{S}\left[J(\mathbf{S})\right]=\frac{2E_{s}}{\sigma_{W}^{2}}\operatorname{Re}\left\{G_{0}^{H}G_{0}\right\} \tag*{(72)}

where Es=E{∣S[k,m]∣2}>0E_{s}=\mathbb{E}\{|S[k,m]|^{2}\}>0 is the average symbol energy. The average FIM therefore depends on the constellation only through its average symbol energy. This observation leads to a universal lower bound on the average conditional CRB.

Proposition 1. Suppose that J(S)≻0J(\mathbf{S})\succ0 almost surely and ES{∥J−1(S)∥}<∞\mathbb{E}_{S}\{\|J^{-1}(\mathbf{S})\|\}<\infty. Then

ES[J−1(S)]⪰J‾−1.(73)\mathbb{E}_{S}\left[J^{-1}(\mathbf{S})\right]\succeq\overline{J}^{-1}. \tag*{(73)}

Consequently, for a∈{τ,ν}a\in\{\tau,\nu\},

CRB‾a≥Tr⁡(PaJ‾−1PaT)≜CRBa,J‾.(74)\overline{\mathrm{CRB}}_{a}\geq\operatorname{Tr}\left(P_{a}\overline{J}^{-1}P_{a}^{T}\right)\triangleq\mathrm{CRB}_{a,\overline{J}}. \tag*{(74)}

The matrix bound and both scalar bounds are exactly achieved by any constant-modulus constellation satisfying ∣S[k,m]∣2=Es|S[k,m]|^{2}=E_{s} for every k,mk,m, including PSK and 4-QAM.

Proof. The matrix inverse is operator convex on the positive-definite cone. Jensen’s inequality therefore gives

ES[J−1(S)]⪰(ES[J(S)])−1=J‾−1.(75)\mathbb{E}_{S}\left[J^{-1}(\mathbf{S})\right]\succeq\left(\mathbb{E}_{S}\left[J(\mathbf{S})\right]\right)^{-1}=\overline{J}^{-1}. \tag*{(75)}

Premultiplying by PaP_{a}, postmultiplying by PaTP_{a}^{T}, and taking the trace establishes the scalar bounds. For constant-modulus constellations,

diag⁡(∣vec⁡(S)∣2)=EsINM(76)\operatorname{diag}\left(|\operatorname{vec}(\mathbf{S})|^{2}\right)=E_{s}I_{NM} \tag*{(76)}

for every transmitted frame. The factorized FIM consequently satisfies

J(S)=J‾,ES[J−1(S)]=J‾−1,(77)J(\mathbf{S})=\overline{J},\qquad\mathbb{E}_{S}\left[J^{-1}(\mathbf{S})\right]=\overline{J}^{-1}, \tag*{(77)}

which establishes equality in the matrix and scalar bounds.

Constant-modulus constellations eliminate the data-induced fluctuations of the complete FIM and attain the lower bound on its average inverse. They therefore simultaneously minimize the average conditional delay and Doppler CRBs, including the individual variance bounds for every target.

To express these bounds in physical range and velocity units, define

TRv=diag⁡(ΔRIQ,ΔvIQ).T_{Rv}=\operatorname{diag}\left(\Delta R I_{Q},\Delta v I_{Q}\right).

The corresponding range–velocity CRB matrix is

CRv(S)=TRv[J−1(S)]θθTRvT=TRvJe−1(S)TRvT,C_{Rv}(\mathbf{S})=T_{Rv}[J^{-1}(\mathbf{S})]_{\boldsymbol{\theta}\boldsymbol{\theta}}T_{Rv}^{T}=T_{Rv}J_{\mathrm{e}}^{-1}(\mathbf{S})T_{Rv}^{T},

and the individual target bounds are

CRB(Rq∣S)=ΔR2[J−1(S)]q,q,\mathrm{CRB}(R_{q}\mid\mathbf{S})=\Delta R^{2}[J^{-1}(\mathbf{S})]_{q,q},
CRB(vq∣S)=Δv2[J−1(S)]Q+q,Q+q.\mathrm{CRB}(v_{q}\mid\mathbf{S})=\Delta v^{2}[J^{-1}(\mathbf{S})]_{Q+q,Q+q}.

Taking a principal submatrix and applying a deterministic congruence transformation preserve the positive-semidefinite order. The average conditional range–velocity CRB matrix therefore inherits the lower bound and constant-modulus optimality in Proposition 1, as do the individual range and velocity bounds.

To make the dependence of these bounds on the symbol-energy distribution and physical OFDM parameters explicit, we next derive their single-target closed forms.

Single-Target Closed-Form CRBs

For a single target with an unknown nonzero complex gain α\alpha, the conditional range and velocity CRBs are

CRB(R∣S)=σW2N2ΔR28π2∣α∣2SmmD,\mathrm{CRB}(R\mid\mathbf{S})=\frac{\sigma_{W}^{2}N^{2}\Delta R^{2}}{8\pi^{2}|\alpha|^{2}}\frac{S_{mm}}{\mathcal{D}},
CRB(v∣S)=σW2M2Δv28π2∣α∣2SkkD,(78)\mathrm{CRB}(v\mid\mathbf{S})=\frac{\sigma_{W}^{2}M^{2}\Delta v^{2}}{8\pi^{2}|\alpha|^{2}}\frac{S_{kk}}{\mathcal{D}}, \tag*{(78)}

where D=SkkSmm−Skm2>0\mathcal{D}=S_{kk}S_{mm}-S_{km}^{2}>0, and

Skk=∑k=0N−1∑m=0M−1pk,m(k−kˉ)2,S_{kk}=\sum_{k=0}^{N-1}\sum_{m=0}^{M-1}p_{k,m}(k-\bar{k})^{2},
Smm=∑k=0N−1∑m=0M−1pk,m(m−mˉ)2,S_{mm}=\sum_{k=0}^{N-1}\sum_{m=0}^{M-1}p_{k,m}(m-\bar{m})^{2},
Skm=∑k=0N−1∑m=0M−1pk,m(k−kˉ)(m−mˉ)(79)S_{km}=\sum_{k=0}^{N-1}\sum_{m=0}^{M-1}p_{k,m}(k-\bar{k})(m-\bar{m}) \tag*{(79)}

are the energy-weighted central moments, with pk,m=∣S[k,m]∣2p_{k,m} = \lvert S[k,m]\rvert^2 and

kˉ=∑k,mpk,mk∑k,mpk,m,mˉ=∑k,mpk,mm∑k,mpk,m.(80)\bar{k} = \frac{\sum_{k,m} p_{k,m}k}{\sum_{k,m} p_{k,m}}, \qquad\bar{m} = \frac{\sum_{k,m} p_{k,m}m}{\sum_{k,m} p_{k,m}}. \tag*{(80)}

The quantities SkkS_{kk} and SmmS_{mm} characterize the energy-weighted spreads along the subcarrier and slow-time dimensions, respectively, while SkmS_{km} describes their coupling. These expressions retain the influence of the symbol-energy distribution on estimation accuracy. For fixed positive SkkS_{kk} and SmmS_{mm}, a nonzero SkmS_{km} increases both bounds.

Proposition 1 establishes that constant-modulus constellations minimize the average conditional CRBs. For a single target, these minimum bounds admit the following closed forms.

Corollary 1. Under the conditions of Proposition 1, consider a single target with N,M≥2N,M \ge2 and α≠0\alpha\ne0. Constant-modulus constellations satisfying ∣S[k,m]∣2=Es\lvert S[k,m]\rvert^2 = E_s attain the minimum average conditional range and velocity CRBs. For every transmitted frame, the corresponding conditional bounds are

CRB⁡(R∣S)=3c28π2ρNM(N2−1)Δf2,\operatorname{CRB}(R \mid S) = \frac{3c^2}{8\pi^2\rho NM(N^2-1)\Delta f^2},
CRB⁡(v∣S)=3λ28π2ρNM(M2−1)Tsym2,(81)\operatorname{CRB}(v \mid S) = \frac{3\lambda^2}{8\pi^2\rho NM(M^2-1)T_{\mathrm{sym}}^2}, \tag*{(81)}

where ρ=∣α∣2Es/σW2\rho= \lvert\alpha\rvert^2 E_s/\sigma_W^2 denotes the received SNR per subcarrier and OFDM symbol.

Proof. See Appendix B.

Under constant-modulus transmission, Skm=0S_{km}=0, so delay and Doppler decouple in the equivalent FIM. The resulting CRBs are independent of the specific symbol realizations and therefore coincide with their averages over the communication data. We observe from (81) that the square roots of the range and velocity CRBs are proportional to the squared radar range resolution ΔR=c2NΔf\Delta R = \frac{c}{2N\Delta f} and speed resolution Δv=λ2MTsym\Delta v = \frac{\lambda}{2MT_{\mathrm{sym}}}, respectively. This correspondence links the physical resolution scales of radar sensing to the fundamental statistical limits of range and velocity estimation [24, 25].

V 2D FFT and ML Estimation for Delay and Doppler

Dividing the received subcarrier samples by the known nonzero communication symbols gives

Z[k,m]=YS[k,m]S[k,m]=∑q=1Qαqe−j2πkτˉq/Nej2πmνˉq/M+η[k,m],(82)Z[k,m] = \frac{Y_S[k,m]}{S[k,m]} = \sum_{q=1}^{Q} \alpha_q e^{-j2\pi k\bar{\tau}_q/N}e^{j2\pi m\bar{\nu}_q/M} + \eta[k,m], \tag*{(82)}

where η[k,m]=WS[k,m]/S[k,m]\eta[k,m] = W_S[k,m]/S[k,m]. Conditioned on the transmitted symbols, the noise samples remain independent and satisfy

η[k,m]∣S∼CN(0,σW2pk,m),pk,m=∣S[k,m]∣2.(83)\eta[k,m] \mid S \sim\mathcal{CN}\left(0,\frac{\sigma_W^2}{p_{k,m}}\right), \qquad p_{k,m} = \lvert S[k,m]\rvert^2. \tag*{(83)}

The delay and Doppler phase progressions are resolved by an NN-point IFFT across subcarriers and an MM-point FFT across OFDM symbols, yielding

V=FNHZFM.(84)V = F_N^H Z F_M. \tag*{(84)}

The corresponding entries are

V[r,s]=1NM∑k=0N−1∑m=0M−1Z[k,m]ej2πkr/Ne−j2πms/M,(85)V[r,s] = \frac{1}{\sqrt{NM}}\sum_{k=0}^{N-1}\sum_{m=0}^{M-1} Z[k,m]e^{j2\pi kr/N}e^{-j2\pi ms/M}, \tag*{(85)}

where r=0,…,N−1r = 0,\ldots,N-1, and the Doppler index is expressed in the signed range s=−⌊M/2⌋,…,⌈M/2⌉−1s = -\lfloor M/2\rfloor,\ldots,\lceil M/2\rceil-1, with column indices interpreted modulo MM. To characterize the target responses, substituting (82) into (85) yields

V[r,s]=1NM∑q=1QαqDN(r−τˉqN)DM(νˉq−sM)+η~[r,s],(86)V[r,s] = \frac{1}{\sqrt{NM}}\sum_{q=1}^{Q}\alpha_q D_N\left(\frac{r-\bar{\tau}_q}{N}\right)D_M\left(\frac{\bar{\nu}_q-s}{M}\right) + \tilde{\eta}[r,s], \tag*{(86)}

where η~[r,s]\tilde{\eta}[r,s] denotes the transformed noise and

DL(u)=∑a=0L−1ej2πau=ejπ(L−1)usin⁡(πLu)sin⁡(πu),(87)D_L(u) = \sum_{a=0}^{L-1}e^{j2\pi au} = e^{j\pi(L-1)u}\frac{\sin(\pi Lu)}{\sin(\pi u)}, \tag*{(87)}

with singular points interpreted by continuity. A target with integer normalized delay and Doppler contributes to a single transform bin with amplitude NMαq\sqrt{NM}\alpha_q. Fractional parameters produce responses distributed across neighboring bins. The 2D delay-Doppler profile is formed as

PFFT[r,s]=∣V[r,s]∣2.(88)P_{\mathrm{FFT}}[r,s] = \lvert V[r,s]\rvert^2. \tag*{(88)}

For QQ resolvable targets, the dominant distinct local maxima within the admissible search region provide the estimates (r^q,s^q)(\hat{r}_q,\hat{s}_q). Their physical range and velocity estimates are

R^q=ΔRr^q,v^q=Δvs^q.(89)\hat{R}_q = \Delta R\hat{r}_q,\qquad\hat{v}_q = \Delta v\hat{s}_q. \tag*{(89)}

FFT-based estimates are restricted to discrete delay-Doppler grid points, which introduces quantization errors for off-grid targets. To evaluate continuous-parameter estimation accuracy against the derived CRBs, we further introduce ML estimation as follows.

Using the vectorized observation model (46), the joint ML estimate for a known number of targets is obtained by minimizing

(θ^,α^)=arg⁡min⁡θ,α ∥y−B(θ)α∥22.(90)(\hat{\boldsymbol{\theta}},\hat{\boldsymbol{\alpha}}) = \underset{\boldsymbol{\theta},\boldsymbol{\alpha}}{\arg\min}\ \lVert\boldsymbol{y} - B(\boldsymbol{\theta})\boldsymbol{\alpha}\rVert_2^2. \tag*{(90)}

Here, the dependence of BB on the candidate delay-Doppler parameters is made explicit. For each candidate θ\boldsymbol{\theta} at which B(θ)B(\boldsymbol{\theta}) has full column rank, the complex gains admit the conditional ML estimate

α^(θ)=[BH(θ)B(θ)]−1BH(θ)y.(91)\hat{\boldsymbol{\alpha}}(\boldsymbol{\theta}) = \left[B^H(\boldsymbol{\theta})B(\boldsymbol{\theta})\right]^{-1}B^H(\boldsymbol{\theta})\boldsymbol{y}. \tag*{(91)}

Substituting (91) into (90) eliminates the gain variables and yields the concentrated ML criterion

θ^=arg⁡min⁡θ yHΠB(θ)⊥y=arg⁡max⁡θ yHB(θ)[BH(θ)B(θ)]−1BH(θ)y.(92)\begin{aligned} \hat{\boldsymbol{\theta}} = \underset{\boldsymbol{\theta}}{\arg\min}\ \boldsymbol{y}^H\boldsymbol{\Pi}_{B(\boldsymbol{\theta})}^{\perp}\boldsymbol{y} \\ &= \underset{\boldsymbol{\theta}}{\arg\max}\ \boldsymbol{y}^H B(\boldsymbol{\theta})\left[B^H(\boldsymbol{\theta})B(\boldsymbol{\theta})\right]^{-1}B^H(\boldsymbol{\theta})\boldsymbol{y}. \tag*{(92)} \end{aligned}

where ΠB(θ)⊥\boldsymbol{\Pi}_{\mathbf{B}(\theta)}^{\perp} is the orthogonal projector defined in the preceding CRB analysis, evaluated at the candidate parameters. The optimization involves 2Q2Q real delay-Doppler parameters. For single target scenario, (92) reduces to the two-dimensional search

(τ^,ν^)=arg⁡max⁡(τˉ,νˉ)∈Ω∣∑k,mS∗[k,m]Ys[k,m]ej2πkτˉ/Ne−j2πmνˉ/M∣2∑k,m∣S[k,m]∣2(93)(\widehat{\tau},\widehat{\nu})=\underset{(\bar{\tau},\bar{\nu})\in\Omega}{\arg\max}\frac{\left|\sum_{k,m}S^{*}[k,m]Y_{\mathrm{s}}[k,m]e^{j2\pi k\bar{\tau}/N}e^{-j2\pi m\bar{\nu}/M}\right|^{2}}{\sum_{k,m}|S[k,m]|^{2}} \tag*{(93)}

where Ω\Omega denotes the admissible continuous delay–Doppler search region. The identity S∗[k,m]Ys[k,m]=pk,mZ[k,m]S^{*}[k,m]Y_{\mathrm{s}}[k,m]=p_{k,m}Z[k,m] shows that (93) weights the division output by the symbol energies, compensating for the unequal conditional noise variances in (83).

Under constant-modulus transmission, pk,m=Esp_{k,m}=E_{\mathrm{s}}, and the objective in (93), evaluated at (τˉ,νˉ)=(r,s)(\bar{\tau},\bar{\nu})=(r,s), becomes

∣∑k,mEsZ[k,m]ej2πkr/Ne−j2πms/M∣2NMEs=Es∣V[r,s]∣2.(94)\frac{\left|\sum_{k,m}E_{\mathrm{s}}Z[k,m]e^{j2\pi kr/N}e^{-j2\pi ms/M}\right|^{2}}{NME_{\mathrm{s}}}=E_{\mathrm{s}}|V[r,s]|^{2}. \tag*{(94)}

Thus, for a single target, constant-modulus transmission makes ordinary FFT peak selection equivalent to ML estimation constrained to the FFT grid. Continuous ML estimation removes that grid restriction. For nonconstant-modulus transmission, the symbol-energy weights must also be retained to obtain the ML criterion.

The FFT peak locations provide initial estimates for continuous optimization of (92), or of (93) in the single-target case. This procedure is referred to as FFT-initialized ML refinement; local optimization does not guarantee the global ML solution. The refined normalized parameters are converted to range and velocity using the same physical scaling as in (89).

VI 2D MUSIC for Joint Range-Velocity Estimation

We apply two-dimensional MUSIC to the division output ZZ, exploiting the orthogonality between the target steering vectors and the noise subspace [26]. To construct a covariance matrix from a single frame, we extract overlapping subarrays of size Ns×MsN_{\mathrm{s}}\times M_{\mathrm{s}}, where 2≤Ns≤N2\le N_{\mathrm{s}}\le N and 2≤Ms≤M2\le M_{\mathrm{s}}\le M. Each subarray contains Ps=NsMsP_{\mathrm{s}}=N_{\mathrm{s}}M_{\mathrm{s}} samples, and the number of sliding positions is Ls=(N−Ns+1)(M−Ms+1)L_{\mathrm{s}}=(N-N_{\mathrm{s}}+1)(M-M_{\mathrm{s}}+1).

Let ZjZ_j denote the subarray with origin (ij,hj)(i_j,h_j), whose entries are [Zj]u,v=Z[ij+u,hj+v][Z_j]_{u,v}=Z[i_j+u,h_j+v], for u=0,…,Ns−1u=0,\ldots,N_{\mathrm{s}}-1 and v=0,…,Ms−1v=0,\ldots,M_{\mathrm{s}}-1. Column-wise vectorization gives

zj=vec⁡(Zj)=Ascj+ηj,j=1,…,Ls,(95)z_j=\operatorname{vec}(Z_j)=A_{\mathrm{s}}c_j+\eta_j,\qquad j=1,\ldots,L_{\mathrm{s}}, \tag*{(95)}

where ηj∈CPs\eta_j\in\mathbb{C}^{P_{\mathrm{s}}} contains the corresponding noise samples, and

As=[as(τˉ1,νˉ1),…,as(τˉQ,νˉQ)]∈CPs×Q,A_{\mathrm{s}}=[a_{\mathrm{s}}(\bar{\tau}_1,\bar{\nu}_1),\ldots,a_{\mathrm{s}}(\bar{\tau}_Q,\bar{\nu}_Q)]\in\mathbb{C}^{P_{\mathrm{s}}\times Q},
as(τˉ,νˉ)=ds(νˉ)⊗bs(τˉ),a_{\mathrm{s}}(\bar{\tau},\bar{\nu})=d_{\mathrm{s}}(\bar{\nu})\otimes b_{\mathrm{s}}(\bar{\tau}),
cj=[α1e−j2πijτˉ1/Nej2πhjνˉ1/M⋮αQe−j2πijτˉQ/Nej2πhjνˉQ/M]∈CQ.(96)c_j= \begin{bmatrix} \alpha_1e^{-j2\pi i_j\bar{\tau}_1/N}e^{j2\pi h_j\bar{\nu}_1/M}\\ \vdots\\ \alpha_Qe^{-j2\pi i_j\bar{\tau}_Q/N}e^{j2\pi h_j\bar{\nu}_Q/M} \end{bmatrix}\in\mathbb{C}^{Q}. \tag*{(96)}

Here, bs(τˉ)b_{\mathrm{s}}(\bar{\tau}) and ds(νˉ)d_{\mathrm{s}}(\bar{\nu}) contain the first NsN_{\mathrm{s}} and MsM_{\mathrm{s}} entries of the delay and Doppler steering vectors defined in (38), respectively. Stacking all subarray snapshots as columns yields

Z=AsCs+N,(97)\mathcal{Z}=A_{\mathrm{s}}C_{\mathrm{s}}+\mathcal{N}, \tag*{(97)}

where

Z=[z1,…,zLs]∈CPs×Ls,\mathcal{Z}=[z_1,\ldots,z_{L_{\mathrm{s}}}]\in\mathbb{C}^{P_{\mathrm{s}}\times L_{\mathrm{s}}},
Cs=[c1,…,cLs]∈CQ×Ls,C_{\mathrm{s}}=[c_1,\ldots,c_{L_{\mathrm{s}}}]\in\mathbb{C}^{Q\times L_{\mathrm{s}}},
N=[η1,…,ηLs]∈CPs×Ls.(98)\mathcal{N}=[\eta_1,\ldots,\eta_{L_{\mathrm{s}}}]\in\mathbb{C}^{P_{\mathrm{s}}\times L_{\mathrm{s}}}. \tag*{(98)}

To account for the unequal noise variances introduced by division, we normalize the stacked observations and form the smoothed covariance matrix

R^w=1LsΛ−1ZZHΛ−1,(99)\widehat{R}_{w}=\frac{1}{L_{\mathrm{s}}}\Lambda^{-1}\mathcal{Z}\mathcal{Z}^{H}\Lambda^{-1}, \tag*{(99)}

where Λ\Lambda is the positive diagonal square root of

Λ2=1Ls∑j=1Lsdiag⁡(vec⁡(∣Sj∣−2)).(100)\Lambda^{2}=\frac{1}{L_{\mathrm{s}}}\sum_{j=1}^{L_{\mathrm{s}}}\operatorname{diag}\left(\operatorname{vec}(|S_j|^{-2})\right). \tag*{(100)}

Here, SjS_j is the transmitted-symbol submatrix at the same position as ZjZ_j, and the inverse squared magnitudes are taken element-wise. By the conditional noise variances in (83),

E[NNHLs∣S]=σw2Λ2.(101)\mathbb{E}\left[\left.\frac{\mathcal{N}\mathcal{N}^{H}}{L_{\mathrm{s}}}\right|S\right]=\sigma_{w}^{2}\Lambda^{2}. \tag*{(101)}

Consequently, the expected normalized covariance is

E[R^w∣S]=AwCsCsHLsAwH+σw2IPs,(102)\mathbb{E}\left[\left.\widehat{R}_{w}\right|S\right]=A_{w}\frac{C_{\mathrm{s}}C_{\mathrm{s}}^{H}}{L_{\mathrm{s}}}A_{w}^{H}+\sigma_{w}^{2}I_{P_{\mathrm{s}}}, \tag*{(102)}

where Aw=Λ−1AsA_{w}=\Lambda^{-1}A_{\mathrm{s}}. The normalization makes the averaged conditional noise contribution spatially white. For constant-modulus symbols, Λ=Es−1/2IPs\Lambda=E_{\mathrm{s}}^{-1/2}I_{P_{\mathrm{s}}}, so it reduces to a common scalar scaling.

Assume that Q<PsQ<P_{\mathrm{s}} and both AsA_{\mathrm{s}} and CsC_{\mathrm{s}} have rank QQ. The noise subspace of (102) is then orthogonal to the columns of AwA_{w}. In practice, we estimate this subspace from R^w\widehat{R}_{w}. Ordering its eigenvalues as γ^1≤⋯≤γ^Ps\widehat{\gamma}_1\le\cdots\le\widehat{\gamma}_{P_{\mathrm{s}}}, with corresponding orthonormal eigenvectors e^i\widehat{e}_i, gives the noise-subspace basis

E^n=[e^1,…,e^Ps−Q]∈CPs×(Ps−Q).(103)\widehat{E}_{n}=[\widehat{e}_1,\ldots,\widehat{e}_{P_{\mathrm{s}}-Q}]\in\mathbb{C}^{P_{\mathrm{s}}\times(P_{\mathrm{s}}-Q)}. \tag*{(103)}

The resulting two-dimensional MUSIC pseudo-spectrum is

PMUSIC(τˉ,νˉ)=∥aw(τˉ,νˉ)∥22∥E^nHaw(τˉ,νˉ)∥22,(104)P_{\mathrm{MUSIC}}(\bar{\tau},\bar{\nu})=\frac{\|a_{w}(\bar{\tau},\bar{\nu})\|_{2}^{2}}{\|\widehat{E}_{n}^{H}a_{w}(\bar{\tau},\bar{\nu})\|_{2}^{2}}, \tag*{(104)}

where aw(τˉ,νˉ)=Λ−1as(τˉ,νˉ)a_{w}(\bar{\tau},\bar{\nu})=\Lambda^{-1}a_{\mathrm{s}}(\bar{\tau},\bar{\nu}) is the consistently normalized steering vector. The QQ dominant distinct local maxima over the admissible delay–Doppler region provide paired estimates (τ^q,ν^q)(\widehat{\tau}_q,\widehat{\nu}_q). The corresponding physical estimates are

R^q=ΔRτ^q,v^q=Δvν^q.(105)\widehat{R}_q=\Delta R\widehat{\tau}_q,\qquad\widehat{v}_q=\Delta v\widehat{\nu}_q. \tag*{(105)}

The joint search associates each estimated delay directly with its Doppler shift, avoiding a separate pairing procedure.

Numerical Results

In this section, we evaluates the constellation-dependent CRBs and compares them with the estimation accuracy of FFT, ML, and MUSIC approaches. The carrier frequency and subcarrier spacing are set to fc=30f_c = 30 GHz and Δf=960\Delta f = 960 kHz, respectively, with a CP length of N/4N/4. Unless otherwise specified, we consider independently generated 16-PSK and 16-QAM symbols normalized to the same average energy Es=1E_s = 1, and define the SNR as Es/σW2E_s/\sigma_W^2.

The average conditional CRBs are evaluated over the transmitted communication data. All Monte Carlo simulations are performed with 10410^4 trials. The frame dimensions and number of targets are specified for each experiment.

Two-Dimensional FFT and MUSIC Spectra

Figure 3 presents the range-velocity spectra obtained by 2D FFT and MUSIC under 16-QAM transmission, with N=256N = 256, M=16M = 16, and Q=2Q = 2. The two targets are located at 20 m and 30 m, with radial velocities of 200 m/s and −200-200 m/s, respectively. Both methods produce distinct peaks near the true target positions, demonstrating joint range-velocity estimation under the adopted observation model. Compared with FFT, MUSIC produces more concentrated spectral peaks. The corresponding top views further illustrate the agreement between the estimated and true target locations.

Figure 3. Range–velocity spectra of 2D FFT and MUSIC under 16-QAM transmission with N=256N = 256, M=16M = 16, and Q=2Q = 2. (a) FFT spectrum. (b) FFT spectrum, top view. (c) MUSIC spectrum. (d) MUSIC spectrum, top view.

Average CRBs and Constant-Modulus Optimality

Figure 4 compares the average range and velocity CRBs of 16-PSK and 16-QAM with the Jensen lower bound. Two configurations are considered: N=16N = 16, M=4M = 4, Q=12Q = 12, and N=128N = 128, M=64M = 64, Q=60Q = 60. In both configurations, the PSK curves coincide with the Jensen lower bound, confirming equality in (74) under constant-modulus transmission, as established in Proposition 1. However, the CRB advantage of PSK over QAM is not always pronounced. As shown in Figure 4(c) and Figure 4(d), the PSK–QAM CRB gap becomes negligible in the configuration with larger numbers of subcarriers and OFDM symbols, despite the increased number of targets. This observation suggests that large OFDM frames can substantially reduce the influence of the communication constellation on the range and velocity CRBs.

Average range and velocity CRBs of 16-PSK and 16-QAM

Figure 4. Average range and velocity CRBs of 16-PSK and 16-QAM compared with the Jensen lower bound. (a)(b) N=16N = 16, M=4M = 4, and Q=12Q = 12. (c)(d) N=128N = 128, M=64M = 64, and Q=60Q = 60.

Single-Target Estimation Performance

Figure 5 compares the range and velocity RMSEs of FFT and continuous ML estimation with the corresponding CRBs for N=16N = 16, M=4M = 4, and Q=1Q = 1. The PSK bounds are obtained directly from (81), whereas the QAM bounds are calculated by averaging the conditional expressions in (78) over the transmitted data. The PSK and QAM CRBs nearly coincide, indicating little constellation dependence in this single-target setting.

Single-target range and velocity RMSE versus CRB

Figure 5. Single-target (Q=1Q = 1) FFT and ML estimation performance compared with the square roots of the average CRBs under 16-PSK and 16-QAM transmission, with N=16N = 16 and M=4M = 4. (a) Range RMSE. (b) Velocity RMSE.

As the SNR increases, the ML results closely approach the corresponding bounds, providing numerical validation of the single-target CRB expressions. FFT instead exhibits an error floor because its estimates in (89) are restricted to discrete delay–Doppler grid points. For the off-grid target considered here, increasing the SNR cannot eliminate this quantization error. Note that at some low-SNR points, the FFT RMSE falls below the root CRB. This behavior does not contradict (61), which assumes locally unbiased estimation, since FFT grid selection can produce biased estimates.

Two-Target Estimation Performance

Figure 6 extends the comparison to two targets with N=256N = 256 and M=16M = 16, including FFT, MUSIC, and joint ML estimation based on (92). The PSK and QAM CRBs remain nearly indistinguishable. For both constellations, joint ML estimation closely approaches the theoretical bounds at sufficiently high SNR, supporting the multi-target CRB analysis in the presence of unknown complex gains.

Two-target range and velocity RMSE versus CRB

Figure 6. Two-target (Q=2Q = 2) FFT, MUSIC, and joint ML estimation performance compared with the square roots of the average CRBs under 16-PSK and 16-QAM transmission, with N=256N = 256 and M=16M = 16. (a) Range RMSE. (b) Velocity RMSE.

MUSIC achieves lower RMSE than FFT, whose performance remains limited by the discrete estimation grid. Unlike the nearly overlapping CRBs, the MUSIC results exhibit a more visible constellation dependence, with PSK yielding lower errors than QAM. This difference is consistent with the nonuniform noise amplification introduced by symbolwise division under QAM transmission, as characterized by (83). Thus, similar estimation CRBs do not necessarily imply similar performance for a particular sensing algorithm for different constellations.

Conclusion

This paper investigated how communication constellations affect joint range–velocity estimation with OFDM signals. Starting from the continuous-time transmit and receive signals, we established an equivalent discrete sensing model and specified its validity conditions. Under joint estimation of target delays, Doppler shifts, and unknown complex gains, we proved that constant-modulus constellations minimize the average conditional CRBs by exactly attaining the Jensen lower bound. The numerical results showed that this theoretical optimality does not always produce a pronounced performance advantage over QAM. The PSK–QAM CRB gap becomes negligible for the larger numbers of subcarriers and OFDM symbols considered. ML estimation closely approaches the corresponding CRBs under both constellations, confirming the validity of the theoretical analysis. FFT exhibits error floors imposed by its integer-grid constraint, whereas MUSIC achieves substantially lower estimation errors and retains a more visible constellation dependence than the CRBs. These findings show that the benefit of constellation selection depends on both the OFDM parameters and the estimation algorithm.

Proof of Lemma 1

With the complete Jacobian G=[D,B,jB]\boldsymbol{G} = [\boldsymbol{D},\boldsymbol{B},j\boldsymbol{B}], the FIM blocks associated with the delay–Doppler and gain parameters are

Jθθ=2σW2Re⁡{DHD},\boldsymbol{J}_{\theta\theta} = \frac{2}{\sigma_W^2}\operatorname{Re}\{\boldsymbol{D}^{H}\boldsymbol{D}\},
Jθξ=2σW2[Re⁡{DHB},−Im⁡{DHB}],\boldsymbol{J}_{\theta\xi} = \frac{2}{\sigma_W^2}\left[\operatorname{Re}\{\boldsymbol{D}^{H}\boldsymbol{B}\},-\operatorname{Im}\{\boldsymbol{D}^{H}\boldsymbol{B}\}\right],
Jξξ=2σW2[Re⁡{BHB}−Im⁡{BHB}Im⁡{BHB}Re⁡{BHB}].(106)\boldsymbol{J}_{\xi\xi} = \frac{2}{\sigma_W^2} \begin{bmatrix} \operatorname{Re}\{\boldsymbol{B}^{H}\boldsymbol{B}\} & -\operatorname{Im}\{\boldsymbol{B}^{H}\boldsymbol{B}\}\\ \operatorname{Im}\{\boldsymbol{B}^{H}\boldsymbol{B}\} & \operatorname{Re}\{\boldsymbol{B}^{H}\boldsymbol{B}\} \end{bmatrix}. \tag*{(106)}

Since B\boldsymbol{B} has full column rank, BHB\boldsymbol{B}^{H}\boldsymbol{B} is invertible. For any invertible complex matrix R\boldsymbol{R}, its real block representation satisfies

\begin{bmatrix} \operatorname{Re}\{\boldsymbol{R}\} & -\operatorname{Im}\{\boldsymbol{R}\}\\ \operatorname{Im}\{\boldsymbol{R}\} & \operatorname{Re}\{\boldsymbol{R}\} \end{bmatrix}^{-1} = \begin{bmatrix} \operatorname{Re}\{\boldsymbol{R}^{-1}\} & -\operatorname{Im}\{\boldsymbol{R}^{-1}\}\\ \operatorname{Im}\{\boldsymbol{R}^{-1}\} & \operatorname{Re}\{\boldsymbol{R}^{-1}\} \end{bmatrix}. \tag*{(107)}\end{bmatrix}

Applying this identity with R=BHB\boldsymbol{R} = \boldsymbol{B}^{H}\boldsymbol{B} gives

Jξξ−1=σW22[Re⁡{(BHB)−1}−Im⁡{(BHB)−1}Im⁡{(BHB)−1}Re⁡{(BHB)−1}].(108)\boldsymbol{J}_{\xi\xi}^{-1} = \frac{\sigma_W^2}{2} \begin{bmatrix} \operatorname{Re}\{(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\} & -\operatorname{Im}\{(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\}\\ \operatorname{Im}\{(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\} & \operatorname{Re}\{(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\} \end{bmatrix}. \tag*{(108)}

Thus, we further have

JθξJξξ−1JθξT=2σW2Re⁡{DHB(BHB)−1BHD}.(109)\boldsymbol{J}_{\theta\xi}\boldsymbol{J}_{\xi\xi}^{-1}\boldsymbol{J}_{\theta\xi}^{T} = \frac{2}{\sigma_W^2}\operatorname{Re}\{\boldsymbol{D}^{H}\boldsymbol{B}(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\boldsymbol{B}^{H}\boldsymbol{D}\}. \tag*{(109)}

Substituting this expression and Jθθ\boldsymbol{J}_{\theta\theta} into the Schur complement in (67) gives

Je(S)=2σW2Re⁡{DH[INM−B(BHB)−1BH]D}=2σW2Re⁡{DHΠB⊥D}.(110)\begin{aligned} \boldsymbol{J}_{e}(\boldsymbol{S}) &= \frac{2}{\sigma_W^2}\operatorname{Re}\left\{\boldsymbol{D}^{H}\left[\boldsymbol{I}_{NM}-\boldsymbol{B}(\boldsymbol{B}^{H}\boldsymbol{B})^{-1}\boldsymbol{B}^{H}\right]\boldsymbol{D}\right\}\\ &= \frac{2}{\sigma_W^2}\operatorname{Re}\{\boldsymbol{D}^{H}\boldsymbol{\Pi}_{\boldsymbol{B}}^{\perp}\boldsymbol{D}\}. \tag*{(110)} \end{aligned}

which obtains (68) and completes the proof.

Proof of Corollary 1

To compute the single-target equivalent FIM in (68), we expand the projector in (69) and evaluate the two products DHD\boldsymbol{D}^{H}\boldsymbol{D} and BHD\boldsymbol{B}^{H}\boldsymbol{D}. For a single target, B\boldsymbol{B} has one column, and the delay–Doppler Jacobian becomes

D=j2πα[−1NKB,1MLB].(111)\boldsymbol{D} = j2\pi\alpha\left[-\frac{1}{N}\boldsymbol{K}\boldsymbol{B},\frac{1}{M}\boldsymbol{L}\boldsymbol{B}\right]. \tag*{(111)}

Using the energy-weighted centers in (80), the gain–parameter coupling product is

BHD=j2πα[BHKBN−BHLBM]=j2πα(BHB)[kˉN−mˉM](112)B^{H}D = j2\pi\alpha\left[\frac{B^{H}KB}{N}-\frac{B^{H}LB}{M}\right] = j2\pi\alpha\left(B^{H}B\right)\left[\frac{\bar{k}}{N}-\frac{\bar{m}}{M}\right] \tag*{(112)}

where BHB=∑k,mpk,mB^{H}B=\sum_{k,m}p_{k,m}.

Since KK and LL are diagonal with entries kk and mm, respectively, and the squared entry magnitudes of BB equal pk,mp_{k,m}, we have

BHK2B=∑k,mpk,mk2,BHKLB=∑k,mpk,mkm,B^{H}K^{2}B=\sum_{k,m}p_{k,m}k^{2},\qquad B^{H}KLB=\sum_{k,m}p_{k,m}km,
BHLKB=∑k,mpk,mkm,BHL2B=∑k,mpk,mm2.B^{H}LKB=\sum_{k,m}p_{k,m}km,\qquad B^{H}L^{2}B=\sum_{k,m}p_{k,m}m^{2}.

The product DHDD^{H}D therefore admits the expression

DHD=4π2∣α∣2[BHK2BN2−BHKLBNM−BHLKBNMBHL2BM2]=4π2∣α∣2∑k,mpk,m[k2/N2−km/(NM)−km/(NM)m2/M2].(113)\begin{aligned} D^{H}D=4\pi^{2}|\alpha|^{2}\left[ \begin{array}{cc} \frac{B^{H}K^{2}B}{N^{2}} & -\frac{B^{H}KLB}{NM}\\ -\frac{B^{H}LKB}{NM} & \frac{B^{H}L^{2}B}{M^{2}} \end{array} \right] \\ &=4\pi^{2}|\alpha|^{2}\sum_{k,m}p_{k,m} \left[ \begin{array}{cc} k^{2}/N^{2} & -km/(NM)\\ -km/(NM) & m^{2}/M^{2} \end{array} \right]. \tag*{(113)} \end{aligned}

Substituting (69), (112), and (113) into (68) yields

Je(S)=2σW2Re⁡{DHD−(BHD)H(BHD)BHB}=8π2∣α∣2σW2[Skk/N2−Skm/(NM)−Skm/(NM)Smm/M2].(114)\begin{aligned} J_{\mathrm{e}}(S)=\frac{2}{\sigma_{W}^{2}}\operatorname{Re}\left\{D^{H}D-\frac{(B^{H}D)^{H}(B^{H}D)}{B^{H}B}\right\} \\ &=\frac{8\pi^{2}|\alpha|^{2}}{\sigma_{W}^{2}} \left[ \begin{array}{cc} S_{kk}/N^{2} & -S_{km}/(NM)\\ -S_{km}/(NM) & S_{mm}/M^{2} \end{array} \right]. \tag*{(114)} \end{aligned}

The subtracted coupling term converts the weighted second moments into the central moments defined in (79). Thus, the effect of the unknown gain is accounted for through the energy-weighted centering of the subcarrier and slow-time indices.

For D=SkkSmm−Skm2>0\mathcal{D}=S_{kk}S_{mm}-S_{km}^{2}>0, inversion gives

Je−1(S)=σW28π2∣α∣2D[N2SmmNMSkmNMSkmM2Skk].(115)J_{\mathrm{e}}^{-1}(S)=\frac{\sigma_{W}^{2}}{8\pi^{2}|\alpha|^{2}\mathcal{D}} \left[ \begin{array}{cc} N^{2}S_{mm} & NMS_{km}\\ NMS_{km} & M^{2}S_{kk} \end{array} \right]. \tag*{(115)}

The diagonal entries are the normalized delay and Doppler CRBs with the unknown gain accounted for. Scaling by ΔR2\Delta R^{2} and Δv2\Delta v^{2}, respectively, gives the general-frame bounds in (78).

For constant-modulus transmission, pk,m=Esp_{k,m}=E_{\mathrm{s}}, so the weighted centers are kˉ=(N−1)/2\bar{k}=(N-1)/2 and mˉ=(M−1)/2\bar{m}=(M-1)/2. The central moments reduce to

Skk=EsMN(N2−1)12,S_{kk}=\frac{E_{\mathrm{s}}MN(N^{2}-1)}{12},
Smm=EsNM(M2−1)12,S_{mm}=\frac{E_{\mathrm{s}}NM(M^{2}-1)}{12},
Skm=0.(116)S_{km}=0. \tag*{(116)}

Using ΔR=c/(2NΔf)\Delta R=c/(2N\Delta f), Δv=λ/(2MTsym)\Delta v=\lambda/(2MT_{\mathrm{sym}}), and ρ=∣α∣2Es/σW2\rho=|\alpha|^{2}E_{\mathrm{s}}/\sigma_{W}^{2} then gives the closed forms in (81).

References

Biographies

Junjie Chen

Portrait of Junjie Chen

Figure 7 (PDF p. 15).

Junjie Chen received the B.Eng. degree from the Beijing Institute of Technology (BIT), Beijing, China, in 2017, and the M.E. degree from the University of Science and Technology of China (USTC), Hefei, China, in 2020. From 2020 to 2022, he was with the China Electronics Technology Group Corporation (CETC), where he conducted research on radar and communication systems. He received the Ph.D. degree from Sun Yat-sen University (SYSU), Guangzhou, China, in 2026. He is currently a lecturer with the School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics, Nanchang, China. His research interests include signal processing, wireless communications, and machine learning.

References

  1. [1]F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and S. Buzzi, “Integrated sensing and communication: Toward dual-functional wireless networks for 6g and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728–1767, 2022.DOI
  2. [2]F. Liu, Y.-F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, S. Buzzi, Y. C. Eldar, and S. Jin, “Sensing with communication signals: From information theory to signal processing,” IEEE J. Sel. Areas Commun., vol. 44, pp. 1–30, 2026.arxiv.org/abs/2502.10819
  3. [3]A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Dual-functional radar-communications: Information embedding using sidelobe control and waveform diversity,” IEEE Trans. Signal Process., vol. 64, no. 8, pp. 2168–2181, 2015.
  4. [4]Y. He, G. Yu, Z. Tang, J. Wang, and H. Luo, “A dual-functional sensing-communication waveform design based on OFDM,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 16962–16975, 2024.DOI
  5. [6]———, “Ofdm-isac beyond cp limit: Performance analysis and mitigation algorithms,” IEEE Trans. Signal Process., 2026.arxiv.org/abs/2511.17878
  6. [7]A. R. Chiriyath, B. Paul, and D. W. Bliss, “Radar-communications convergence: Coexistence, cooperation, and co-design,” IEEE Transactions on Cognitive Communications and Networking, vol. 3, no. 1, pp. 1–12, 2017.DOI
  7. [8]F. Liu, L. Zhou, C. Masouros, A. Li, W. Luo, and A. Petropulu, “Toward dual-functional radar-communication systems: Optimal waveform design,” IEEE Trans. Signal Process., vol. 66, no. 16, pp. 4264–4279, 2018.DOI
  8. [9]F. Liu, Y.-F. Liu, A. Li, C. Masouros, and Y. C. Eldar, “Crǎmer-rao bound optimization for joint radar-communication beamforming,” IEEE Trans. Signal Process., vol. 70, pp. 240–253, 2021.DOI
  9. [10]K. M. Braun, “Ofdm radar algorithms in mobile communication networks,” Ph.D. dissertation, Karlsruhe, Karlsruher Institut für Technologie (KIT), Diss., 2014, 2014.DOI
  10. [11]H. Hua, T. X. Han, and J. Xu, “MIMO integrated sensing and communication: CRB-rate tradeoff,” IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 2839–2854, 2023.arxiv.org/abs/2209.12721
  11. [12]X. Song, X. Yu, J. Xu, and D. W. K. Ng, “CRB-rate tradeoff for bistatic ISAC with gaussian information and deterministic sensing signals,” IEEE Trans. Wireless Commun., vol. 25, pp. 11 768–11 782, 2026.
  12. [13]Y. Xiong, F. Liu, Y. Cui, W. Yuan, T. X. Han, and G. Caire, “On the fundamental tradeoff of integrated sensing and communications under gaussian channels,” IEEE Trans. Inf. Theory, vol. 69, no. 9, pp. 5723–5751, 2023.arxiv.org/abs/2204.06938
  13. [14]S. Lu, F. Liu, F. Dong, Y. Xiong, J. Xu, Y.-F. Liu, and S. Jin, “Random ISAC signals deserve dedicated precoding,” IEEE Trans. Signal Process., vol. 72, pp. 3453–3469, 2024.arxiv.org/abs/2311.01822
  14. [15]F. Liu, Y. Zhang, Y. Xiong, S. Li, W. Yuan, F. Gao, S. Jin, and G. Caire, “CP-OFDM achieves the lowest average ranging sidelobe under QAM/PSK constellations,” IEEE Trans. Inf. Theory, vol. 71, no. 9, pp. 6950–6967, 2025.arxiv.org/abs/2407.06691
  15. [16]F. Liu, Y. Xiong, S. Lu, S. Li, W. Yuan, C. Masouros, S. Jin, and G. Caire, “Uncovering the iceberg in the sea: Fundamentals of pulse shaping and modulation design for random ISAC signals,” IEEE Trans. Signal Process., vol. 73, pp. 2511–2526, 2025.arxiv.org/abs/2501.01721
  16. [17]Z. Du, F. Liu, Y. Xiong, T. X. Han, Y. C. Eldar, and S. Jin, “Reshaping the ISAC tradeoff under OFDM signaling: A probabilistic constellation shaping approach,” IEEE Trans. Signal Process., vol. 72, pp. 4782–4797, 2024.arxiv.org/abs/2312.15941
  17. [18]Y. Zhang, F. Liu, Y. Xiong, W. Yuan, S. Li, L. Zheng, T. X. Han, C. Masouros, and S. Jin, “On discrete ambiguity functions of random communication waveforms,” arXiv preprint arXiv:2512.08352, 2025.arxiv.org/abs/2512.08352
  18. [19]S. M. Kay, “Statistical signal processing: estimation theory,” Upper Saddle River, NJ, USA: Prentice-Hall, 1998.
  19. [20]F. Liu, Y. Xiong, Y.-F. Liu, J. Yang, C. Masouros, and S. Jin, “CP-OFDM achieves lower ranging CRB than frequency-spread waveforms in the large-sample regime,” arXiv preprint arXiv:2605.14541, 2026.
  20. [21]D. Bok, D. O’Hagan, and P. Knott, “Effects of movement for high time-bandwidths in batched pulse compression range-doppler radar,” Sensors, vol. 21, no. 7, p. 2492, 2021.
  21. [22]Z. Du, J. Xu, Y. Xiong, J. Wang, M. F. Keskin, H. Wymeersch, F. Liu, and S. Jin, “Probabilistic constellation shaping for OFDM ISAC signals under temporal-frequency filtering,” IEEE Transactions on Wireless Communications, vol. 25, pp. 12914–12929, 2026.arxiv.org/abs/2510.12204
  22. [23]G. Hakobyan and B. Yang, “A novel narrowband interference suppression method for OFDM radar,” in Proc. 24th European Signal Processing Conference (EUSIPCO), 2016, pp. 2230–2234.DOI
  23. [24]A. Sakhniini, A. Albaba, A. Bourdoux, and S. Pollin, “A general technique for radar resolution analysis based on a quadratic approximation,” IEEE Transactions on Radar Systems, vol. 2, pp. 251–262, 2024.
  24. [25]A. Dogandzic and A. Nehorai, “Cramer-rao bounds for estimating range, velocity, and direction with an active array,” IEEE transactions on Signal Processing, vol. 49, no. 6, pp. 1122–1137, 2001.DOI
  25. [26]R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Transactions on Antennas and Propagation, vol. 34, no. 3, pp. 276–280, 1986.DOI

Paper details

vlnc:W_xsvu67gf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents