WorksEngineering
Wideband Terahertz Pinching-Antenna Systems (PASS): Bandwidth Limits, Ergodic Rate, and Placement Optimization
Abstract
Terahertz (THz) bands offer the vast bandwidth that sixth-generation (6G) networks demand, but their severe path loss makes bringing the radiator close to the user decisive. Pinching-antenna systems (PASS) do exactly this, turning path loss into a programmable parameter by sliding radiating points along a dielectric waveguide, which makes THz their native regime. That regime, however, is inherently wideband and the waveguide is dispersive, whereas every existing PASS analysis is single-frequency and assumes a non-dispersive guide. This paper develops the first wideband, dispersion-aware theory of PASS and uncovers a limitation intrinsic to the pinching mechanism. Because co-phasing by physical position equalizes path length only modulo a wavelength, a pinching array is inherently beam-squint limited, and we show its frequency response is a Dirichlet kernel with a closed-form coherence bandwidth and an irreducible dispersion floor that even true-time-delay (TTD) correction cannot remove. A single unifying law governs the behavior, in which array gain trades against both coherence bandwidth and spatial robustness through one invariant that TTD correction breaks, reaching the floor with far fewer delay elements than a conventional array. We further characterize the wideband ergodic rate under random user locations, show that molecular absorption acts as a deterministic taper that multiplies the squint, and derive a simple rule for the rate-optimal number of pinches. Finally, by exploiting the integer structure of co-phasing, we recast the non-convex placement problem as a mixed-integer program that can be solved exactly in polynomial time. Together, these results move PASS from the narrowband, dispersion-blind setting to the wideband, dispersive regime that THz operation actually requires.
Introduction
The Sixth-generation (6G) wireless networks envision terabit-per-second connectivity and high-resolution sensing, motivating the use of the terahertz (THz) band (0.1–10 THz) and its vast available bandwidth [1]. However, THz propagation faces two fundamental limitations: severe attenuation and blockage sensitivity. Free-space spreading and molecular absorption substantially constrain communication range [1, 2, 3], while quasi-optical propagation and weak diffraction produce deep shadows, with human-body blockage causing tens of decibels of additional loss [1]. These challenges motivate short, reconfigurable line-of-sight (LOS) links that mitigate propagation loss without requiring prohibitively large antenna arrays.
Existing technologies address these limitations only partially. Ultra-massive MIMO and hybrid beamforming increase array gain [1, 4, 5], while holographic surfaces improve aperture utilization [6]. However, these co-located architectures primarily concentrate energy along existing propagation paths and cannot independently bypass LOS blockages. Reconfigurable intelligent surfaces (RISs) can establish alternative paths around obstructions, but incur cascaded propagation losses [7]. Fluid and movable antennas offer spatial adaptation by repositioning radiating elements [8, 9], yet their typically wavelength-scale movement primarily exploits small-scale channel variations rather than substantially reducing propagation distance. Consequently, these approaches do not provide the meter-scale flexibility needed to dynamically shorten the serving link and mitigate large-scale THz attenuation.
In pinching-antenna systems (PASS), signals propagate along a low-loss dielectric waveguide and are radiated toward users through reconfigurable pinching elements placed at selected positions. This enables on-demand LOS links and makes large-scale path loss controllable through antenna placement [10, 11]. By repositioning radiating points over meter-scale distances, PASS directly address the severe distance-dependent attenuation of THz channels. Crucially, this mobility provides placement freedom rather than a correspondingly large coherent aperture, since the radiating cluster remains centimeter-scale. However, most existing PASS analyses assume narrowband operation over a non-dispersive waveguide, an assumption that becomes increasingly restrictive for wideband THz transmission.
Related Work
Table 1 compares this work with representative PASS and wideband-array studies, highlighting a clear research gap between two largely separate lines of investigation that the proposed framework unifies.
Table 1. Positioning of this work against representative wideband-array and pinching-antenna studies. ✓ fully (✓) partially ✗ not addressed.
The first investigates PASS primarily under narrowband propagation models. Early works established closed-form rate expressions and multiuser beamforming through antenna placement [11], uplink placement and resource allocation [12], NOMA-assisted activation [13], array-gain and coupling limits [14], and architectural principles [10]. Subsequent studies extended PASS to MIMO beamforming [15], capacity analysis [16], and robust degree-of-freedom optimization [17]. Moreover, [18] incorporated waveguide attenuation and random user locations, but retained a single-frequency, single-pinching-antenna model. These studies do not explicitly characterize frequency-dependent guided-wave phase propagation or its implications for wideband beamforming.
The second direction addresses wideband beam squint in conventional phased and hybrid arrays. Frequency-independent phase shifters produce frequency-dependent beam directions, degrading coherent array gain over large bandwidths [2, 5]. True-time-delay (TTD) and delay-phase precoding mitigate this effect by providing frequency-dependent phase compensation [5], with recent extensions to flexible-antenna architectures [19]. However, these approaches do not account for the frequency-dependent phase accumulated along a dielectric waveguide, which governs coherent radiation in PASS. Consequently, the interaction between waveguide dispersion, wideband beam squint, and pinching-antenna placement remains largely unexplored. In particular, it is unclear how dispersion affects the beam direction and coherent array gain across frequency, limits the usable bandwidth, and influences optimal pinching positions. This paper addresses these questions through a unified wideband PASS framework that explicitly incorporates guided-wave dispersion and attenuation, thereby bridging the two research directions summarized in Table I.
Contributions
We develop the first wideband, dispersion-aware theory of PASS. Our contributions are as follows.
Wideband dispersive channel model: We derive a PASS channel that jointly captures the free-space roll-off, frequency-dependent waveguide attenuation, and waveguide dispersion, reducing the dispersion to a few physical constants through a group-delay expansion. This is the first PASS model valid across a wideband.
Inherent beam squint and coherence bandwidth: We prove that single-waveguide co-phasing equalizes path length only modulo a wavelength, so a pinching array is inherently squint-limited. The array response is a Dirichlet kernel in frequency, which yields a closed-form coherence bandwidth and an -invariant gain-bandwidth product.
Irreducible dispersion floor: We isolate the group-velocity-dispersion (GVD) term as the single component that TTD correction cannot remove, and derive the resulting bandwidth floor that bounds any corrected pinching array.
Wideband rate, absorption, and a design rule for N: We characterize the ergodic rate under random user location and prove that the user-position statistics and the wideband squint decouple, the squint entering only through a deterministic efficiency scalar. We further show molecular absorption multiplies this scalar as a deterministic taper with a closed-form absorption-limited bandwidth, and we prove that the band-integrated rate is maximized when the coherence bandwidth meets the signal bandwidth.
Delay-phase mitigation: We propose a delay-phase architecture of coarse feed TTD plus pinch positioning that attains the dispersion floor with only delay elements rather than one per pinch.
Optimal and robust placement: Exploiting the integer structure of path-modulo-wavelength co-phasing, we reformulate the non-convex wideband placement problem exactly as a mixed-integer program, prove that minimal-step consecutive orders are optimal under an explicit spacing-feasibility condition so the problem is polynomially solvable, and establish a spatial position-coherence region whose gain-robustness product mirrors the gain-bandwidth law.
System Model and Problem Formulation
We introduce in this section the unified system model used throughout the paper, summarized in Figure 1, in which a pinch cluster far smaller than the waveguide is positioned near the user, the feature that distinguishes PASS from a fixed aperture.
Figure 1. Wideband dispersive PASS: the coherent cluster is centimeter-scale on a meters-long guide, giving placement freedom rather than aperture. Each pinch combines a hop and in-guide run at angle .
Pinching-Antenna Array and Wideband Signal Model
We consider a downlink in which an access point serves a single-antenna user through a pinching-antenna array built on dielectric waveguides [10, 11]. The array comprises $W \ge 1 parallel dielectric waveguides indexed by . Waveguide is aligned with the axis at height and is fed at the point through a TTD element of delay . It carries radiating points, the pinches, located at for , with axial positions ordered as . The total number of pinches is , and a minimum spacing separates any two pinches so that electromagnetic coupling between radiating sites is negligible [14]. Unlike fluid or movable antennas, whose displacement spans only a few wavelengths and therefore reshapes small-scale fading [8], a pinch can be applied anywhere along a waveguide that may extend over meters, so that the radiating site can be brought close to the user and the large-scale pathloss becomes a controllable quantity [10, 18].
The single-antenna user lies in the plane at , within a rectangular service region of size , so that and . The free-space distance from pinch to the user is
In the single-waveguide case we drop the index and write , , and . This case underlies § 3-A–IV, whereas the multi-waveguide case supports the designs of § 5–VI.
We study a wideband system of total bandwidth centered at the carrier , consistent with the ultra-broadband operation envisioned for THz links [2, 20, 21]. Orthogonal frequency-division multiplexing with subcarriers is assumed, the -th subcarrier occupying frequency
and a generic in-band frequency is written with . We denote by the speed of light and by the carrier wavelength.
Wideband Dispersive Channel Model
Three physically distinct mechanisms make the link frequency selective. The narrowband pinching-antenna literature evaluates all three at the single frequency [11, 15, 16], an approximation that is untenable across a THz band. We retain their frequency dependence.
Free-Space Spreading
Free-Space Spreading: The first mechanism is free-space spreading, whose reference gain at unit distance is
decaying as and representing the growth of free-space loss with frequency [18].
In-Waveguide Attenuation
In-Waveguide Attenuation: The second mechanism is in-waveguide attenuation, modeled by the frequency-dependent power-attenuation coefficient
where is the attenuation at and captures the rise of dielectric loss with frequency [22, 23].
Guided Propagation and Dispersion
Guided Propagation and Dispersion: The third mechanism is the guided phase accumulated along the waveguide, described by the propagation constant
which sets the phase per unit length of the guided wave and whose dependence on frequency through the effective index constitutes waveguide dispersion [22].
Collecting free-space spreading, in-waveguide attenuation, the free-space propagation phase, and the guided phase, the channel contributed by pinch at frequency is
where is the amplitude and the total phase gathers the free-space hop and the guided run,
Here
is the electrical path of the pinch, equal to the physical free-space hop plus the electrically scaled in-waveguide run. The combined channel at the user, including the feed delays, is
To separate the first-order index from genuine dispersion we expand about following standard guided-wave practice [22, 23],
where , the prime denotes differentiation at , and is the group index. A non-dispersive guide has . The group index governs delay rather than phase, a distinction that proves central in § 3.
For the squint analysis we adopt the following standard approximation, which isolates the phase behavior that drives beam squint from the slowly varying amplitude [5].
Assumption 1. *For pinches clustered near the pathloss-optimal site and over the coherence bandwidth of interest, the amplitudes in (6) are approximately constant across pinches and across in-band frequencies, that is . Amplitude taper is treated as a second-order refinement in Section VI.
Under Assumption 1, a single waveguide with and yields the normalized array gain
which equals unity when the contributions combine coherently.
Problem Formulation
With the model above, the received signal on subcarrier is , where is the transmitted symbol and is additive noise. Under an equal power split of the total power across the subcarriers and the pinches of a passive feed, consistent with the array-gain accounting of [18], the received signal-to-noise ratio (SNR) on subcarrier at the pathloss-optimal placement, where the cluster sits at so that the common distance is , is
The paper addresses three problems built on this model. The first is the characterization of the usable wideband range, that is the coherence bandwidth over which remains within 3 dB of its peak. The second is the wideband ergodic spectral efficiency
averaged over a random user location. The third is the joint optimization of pinch positions and feed delays that maximizes the band-averaged received power
subject to the region, spacing, and causality constraints. All three problems hinge on a feature absent from prior PASS analyses: the frequency dependence of . A single-frequency model sees only , where and the array gives its full coherent gain , so rate, bandwidth, and placement reduce to their narrowband counterparts. Across a wide band for , and the three become three views of one object: the first characterizes it via the coherence bandwidth and the post-correction dispersion floor (Section III), the second integrates it into the ergodic rate (13) (Section IV), and the third shapes it through placement and feed delays (14) (Sections V–VI). The parts thus form one loop around the same gain profile, not separate studies.
Coherence Bandwidth and Dispersion Floor
Inherent Beam Squint and Coherence Bandwidth
This section establishes the central limitation of a pinching array over a wideband. We show that single-waveguide co-phasing is intrinsically frequency selective, so the array is beam-squint limited, and we obtain its coherence bandwidth in closed form. We take throughout.
A pinch controls phase only through its physical position, so co-phasing equalizes path length modulo a wavelength rather than true delay. The following result makes this precise.
Lemma 1. Coherent combining of pinches and at the design frequency requires the electrical-path difference to be an integer multiple of the carrier wavelength,
At any other frequency the corresponding phase difference is
which is a multiple of only at and its harmonics.
Proof. From (7), . Co-phasing at means , which with gives (15). For a non-dispersive guide does not depend on frequency, so , which lies in only when is an integer, proving (16).
Lemma 1 is the crux of the wideband behavior. Because alignment is set by path length modulo , it holds exactly only at the frequency for which it was designed. A TTD array equalizes physical delay and would hold at all frequencies [5], but a single passive waveguide offers position as its only control and cannot. The resulting squint is quantified exactly, to first order in , by the differential group delay between propagation paths.
Lemma 2. Let the pinches be co-phased at , and define the differential free-space hop and the differential guide length . To first order in , the residual phase of pinch relative to pinch 1 is
where is the differential group delay. The group index , not the phase index , sets the squint rate.
Proof. Using (7) and (10),
where the co-phasing identity from Lemma 1 was used with . Subtracting the coherent part and writing gives . To first order we set in the second term and use together with . The two terms combine into , which is (17).
We now specialize to the canonical placement, in which the pinches are co-phased with consecutive orders so that . This placement uses the densest packing that keeps the pinches co-phased (subject to the spacing-feasibility condition of Section VI-C; order-skipping applies when it is violated), and Section VI proves it is optimal for the wideband gain, so it is adopted here without loss. The next result is the central characterization of this paper.
Theorem 1. Under Assumption 1 and the canonical placement, the normalized array gain (11) is the Dirichlet kernel in frequency
where is the angle between the waveguide axis and the line from the cluster to the user. Its two-sided 3-dB coherence bandwidth is
where is the unique root of in . At the broadside operating point the factor reduces to , and in a non-dispersive guide to unity.
Proof. By Lemma 2, for a small aperture the pinch-user angle is nearly constant, so and . Substituting into (17) gives the residual phase with as stated. With , (11) becomes a geometric series,
which is (19). Over the main lobe , so with the condition reads , giving . Hence the half-power half-width is and is (20).
(19) is the array factor of a uniform linear array in the frequency variable : the pinching array scans in frequency and its peak walks off as grows, exactly beam squint [2, 5], a phenomenon also motivating flexible-antenna remedies in phase-shifter arrays [19].
Remark 1. In conventional wideband arrays, squint is a harmful artifact of frequency-flat phase shifters, cured by TTD [2, 5]. Squint in a PA is equally harmful, collapsing the gain away from , but its cause differs: alignment is set by physical path length, correct only modulo a wavelength (Lemma 1), not by an electronic phase correct only at one frequency. This is consequential: because co-phasing is an integer condition on path length, the admissible positions form an exact lattice (Theorem 4), which makes the placement problem exactly solvable and lets a few feed delays reach the dispersion floor with rather than elements. Squint is thus the same effect in consequence, but a more structured object in mechanism.
Remark 2 (Frequency-selective coherence loss vs. angular squint). With a single fixed user and no beam scanning there is no angular walk-off: the effect here is a frequency-selective loss of coherent combining gain, not a spatial beam that swings with frequency. It is, however, the same mechanism: if the array were steered, the coherent focal point would drift with frequency, so the Dirichlet-kernel description is exact rather than analogical, now in the frequency variable in place of a spatial angle. We retain the term “beam squint” for continuity with the wideband-array literature [5], with this distinction understood.
Remark 3 (Coherent aperture vs. placement range). The coherent cluster spans only , which is centimetre-scale at THz (e.g. cm for at 300 GHz, ). The metres-long waveguide is therefore used not to form a large coherent aperture but to position this small cluster at the pathloss-optimal site near the user, so the long guide provides placement freedom, not aperture. This smallness validates the constant- linearization used throughout. A small simultaneously widens the coherence bandwidth () and the spatial robustness () at the price of a proportionally smaller gain , the law of Proposition 1.
Because the coherent array gain equals while the fractional coherence bandwidth is , their product
is independent of : array gain can be traded for coherence bandwidth only at a fixed product set by the guide dispersion and the geometry. This invariance is one instance of a single law that also governs the dispersion floor of § 3-B and the spatial robustness of Section VI.
Proposition 1 (Generalized gain–coherence law). Let a size- consecutive-order cluster have, about a design point, a dominant residual per-pinch phase of the single-term form in a scalar perturbation , with constant and integer order . Then the normalized gain is the Dirichlet kernel , its two-sided 3-dB half-width obeys , and the gain–coherence product scales as .
Proof. With , the summed geometric series gives the stated kernel; the main-lobe half-power condition yields , i.e. , so and .
Proposition 1 unifies the three trade-offs through a common exponent governing their gain–bandwidth scaling. TTD correction thus does not merely displace the trade-off; it changes its order (), converting the -invariant ceiling into one that improves with array size, the sense in which the uncorrected array is -limited, whereas the corrected array is only -limited.
The Irreducible Dispersion Floor
Theorem 1 retained only the first order in , where the group index merely rescaled the coherence bandwidth. That first-order squint is a pure delay effect, and we show below that it is removable by TTD compensation. The genuinely dispersive effect appears at second order, through group-velocity dispersion, and it cannot be removed by any delay, so it sets a fundamental floor on the usable bandwidth. This section isolates and quantifies that floor.
We expand the propagation constant about the carrier angular frequency to second order. Since the free-space phase is exactly linear in and carries no curvature, only the guided term contributes, and [22, 23]
where is the group delay per unit length, is the group velocity, and is the group-velocity-dispersion parameter, with units of . A non-dispersive guide has .
Lemma 3. Retaining (23), the residual phase of pinch relative to pinch 1 decomposes into a linear group-delay term and a quadratic dispersion term,
with as in Lemma 2 and .
Proof. The total phase is . Only the guided term carries curvature, so with (23) the difference relative to pinch 1, after subtracting the common phase constant , is . The first bracket equals since , and with the two terms become (24).
The two terms are physically distinct. A frequency-flat feed delay adds a phase that is linear in , so it cancels the group-delay term exactly when . No delay can reproduce a term in , so the dispersion term is irreducible. After ideal first-order compensation the residual phase is purely
Theorem 2. Under the canonical placement, for which , and with ideal cancellation of the
first-order squint, the residual array gain is a Dirichlet kernel quadratic in frequency,
and the two-sided 3-dB dispersion-limited bandwidth is
Proof. Ideal compensation leaves (25). Substituting gives , and summing the geometric series as in the proof of Theorem 1 yields (26). The main-lobe condition solves to , and with is (27). ■
Remark 4 (Interpreting the dispersion floor). Two features of (27) are important. First, the floor scales as , whereas the uncorrected coherence bandwidth of (20) scales as . TTD correction therefore improves the gain-bandwidth scaling from to , and the floor widens as for a less dispersive guide. Second, the dispersion term is masked by the first-order squint until the latter is removed. Equating the two terms of (24) at broadside gives the crossover offset , below which the linear squint dominates and dispersion is negligible. Since the uncorrected coherence bandwidth is in practice far smaller than , waveguide dispersion is immaterial to the uncorrected array and becomes the binding limit only once the first-order squint has been compensated, which is the subject of Section V. Finally, combines material dispersion, from the frequency dependence of the dielectric index, and modal dispersion, from the guided-mode geometry, the latter dominating near modal cutoff [22, 23]. Both enter (25) and (27) identically, and is precisely the quantity discarded by the lossless narrowband models, so by (27) it alone sets the ultimate wideband limit of a pinching array.
Wideband Achievable Rate under Random User Location
We now relate the squint and dispersion results to the ergodic spectral efficiency of (13), averaged over the band and over a uniformly random user location with and . The main finding is structural. The user-position statistics and the wideband squint separate, and the squint enters the rate only through a single deterministic scalar.
A first observation reduces the wideband rate to known narrowband quantities. By the linearity of expectation applied to (13),
where each is the narrowband single-pinch ergodic rate evaluated at with the received SNR scaled by the squint factor . The wideband rate is thus the subcarrier average of the closed-form narrowband expression derived for loss-aware pinching links in [18], and it inherits that machinery term by term. Equation (28) is exact at every SNR. The low-SNR regime, which is the relevant one for power- or distance-limited THz links [20], admits a fully decoupled closed form.
Theorem 3. In the low-SNR regime , the wideband ergodic rate factors as
where the wideband efficiency
is deterministic and captures jointly the roll-off and the squint, and the position-averaged reference SNR is available in closed form as
Proof. For , , so . Write with reference SNR and . The attenuation dispersion over the band is negligible, so is independent of position, and . Summing gives (29) with . Since and are independent, . The elementary integrals and give (31). ■
Remark 5 (Validity of the attenuation decoupling). The step drops the attenuation dispersion by assuming . From (4), the excess exponent across the band is , so the decoupling is accurate when . For large loss exponent , large fractional bandwidth, or long guide runs this fails; the factor then reacquires a deterministic, position-coupled dependence, the amplitude-dispersion analogue of the molecular-absorption taper of Section IV-A, and the factorization (29) holds only with a position-averaged correction.
Equation (29) isolates two independent effects. The user-position statistics are carried by , whose closed form in (31) reduces to the narrowband result of [18] when evaluated at a single frequency. The wideband squint is carried by the scalar . As the bandwidth shrinks relative to the coherence bandwidth of Theorem 1, every and , so the narrowband rate is recovered. As the bandwidth grows past , an increasing fraction of subcarriers fall below the 3-dB point and decreases, so the fractional loss is a direct closed-form measure of the squint penalty. The deterministic effect of molecular absorption, significant over THz distances [2], is developed next; at general SNR, (28) provides the exact rate as the subcarrier average of the narrowband closed form.
Molecular Absorption and the Squint
With molecular absorption the free-space factor becomes , where [Np/m] is the power absorption coefficient and the free-space hop. The pinch amplitude in (6) acquires the extra factor . For the small cluster of Remark 3 (, so ), this factor is common-mode: it leaves the array factor in (11) unchanged and instead reweights each subcarrier. Hence the wideband efficiency (30) generalizes to
Absorption and squint therefore multiply: absorption is a deterministic amplitude taper over the subcarriers, the squint a phase-driven coherence loss. Absorption can depress but does not interact with the squint beyond reweighting.
Inside a transmission window (between water-vapour lines) is smooth; expanding about , , the slope produces an asymmetric spectral tilt (shifting the effective passband centroid and lowering ) and the curvature a symmetric roll-off . Defining the two-sided 3-dB absorption-limited bandwidth from the symmetric term,
the usable band is the intersection of the squint- and absorption-limits, mirroring (38):
Which limit binds is distance-dependent: equating (33) with gives the crossover hop
so squint binds for and absorption for . Over the short LOS hops that PASS creates, absorption typically does not dominate; over longer hops or near a window edge it does, and (35) marks the boundary. Near a water absorption line at , varies sharply across the band, causing to suppress the affected subcarriers. Consequently, must be evaluated numerically using spectroscopic line databases. As in practical THz systems, the design response is to confine the transmission band to a low-absorption window, where (33)–(35) apply.
Rate-Optimal Pinch Count
Theorem 1 shows coherence shrinks as while the coherent SNR grows as by (12), so band-integrated rate has an interior sweet spot. Let denote the coherent band-centre SNR, where is the single-pinch (full-power) reference SNR; the out-of-band subcarriers combine incoherently ( on average) and fall back to this same , independent of . Adopt the sharp-coherence idealization .
Proposition 2 (Rate-optimal pinch count). Under the sharp-coherence idealization and neglecting the slow roll-off, the band-averaged rate is strictly increasing for and strictly decreasing for , with the unique maximizer
The optimum places the coherence bandwidth exactly at the signal bandwidth.
Proof. If then , the whole band is coherent, every subcarrier carries SNR , and , strictly increasing. If then a fraction of the band is coherent and the rest contributes zero under the idealization, so with . Then , where satisfies and , hence and for . Continuity at completes the proof. ■
Corollary 1 (TTD-corrected sweet spot). Replacing by the dispersion floor of Theorem 2, the diminishing-returns knee moves to (from ), with : TTD correction lifts the useful pinch count, hence attainable array gain, by orders of magnitude.
Remark 6 (Low-SNR knee vs. high-SNR peak). At strict low SNR, the decrease beyond is , so remains essentially flat, yielding a diminishing-returns knee rather than a sharp peak. Restoring the incoherent floor shifts the maximizer modestly above . Nevertheless, marks the onset of diminishing returns and remains the appropriate design target.
At broadside , so (36) specializes to , scaling inversely with the signal bandwidth. With the dispersion-limited count of Corollary 1 is orders of magnitude larger, so post-TTD the binding constraint is spacing/aperture feasibility, not dispersion.
Delay-Phase Squint Mitigation
Theorem 2 showed that ideal TTD correction leaves only the dispersion floor . We now give a realizable architecture that approaches this floor with few delay elements, by letting pinch positioning rather than radio-frequency hardware perform the pathloss minimization and the fine alignment, while a small number of feed delays remove the coarse squint. The design is the pinching counterpart of sub-connected delay-phase precoding for wideband THz arrays [5].
The array uses the multi-waveguide model of Section II-A. Waveguide carries pinches clustered about a site and is driven through a feed delay , so the combined channel is (9). The per-pinch group delay and the cluster-centroid delay are
Proposition 3. The causal feed delays with cancel the inter-cluster group-delay differences to first order in , leaving only the intra-cluster group-delay spread and the common dispersion curvature.
Proof. The feed delay adds the frequency-flat phase , a pure delay common to the cluster. The effective group delay from is . The global constant does not affect relative phases, and the centroid delays , which carry the coarse differences, are removed. Only the intra-cluster deviations remain, and by Lemma 3 the dispersion phase is unchanged by a frequency-flat delay. ■
:::
With equal clusters , each co-phased in the canonical sense with , two residuals remain after Proposition 3. The intra-cluster first-order squint of a size- cluster contributes a 3-dB bandwidth by Theorem 1, and the dispersion term, unaffected by the feed delays and acting across the full -pinch aperture, contributes the floor of Theorem 2. The gain falls below once either residual reaches its 3-dB point, so the usable bandwidth is
More feed delays widen the first term, which scales as , until it meets the dispersion floor, beyond which further delays bring no gain. Equating the two terms of (38) and substituting from (27) at broadside gives , which simplifies to
valid when the result satisfies . The dispersion floor is therefore attained with feed delays rather than one per pinch, because pinch positioning performs the fine alignment and the pathloss minimization at no delay cost. This contrasts with the delay elements of a conventional delay-phase array. Setting , one pinch per waveguide, makes the first term vanish and recovers the pure floor , so (39) shows that choice to be wasteful. The feed delays follow in closed form from Proposition 3, while the pinch positions are obtained from the placement problem of Section VI.
Optimal and Robust Pinch Placement
Sections III–V assumed the canonical placement. We now solve the placement problem (14), prove the canonical choice optimal, analyze robustness to an uncertain user position, and assemble these facts into the proposed algorithm with its benchmarks.
Problem and Exact Integer Reformulation
The placement problem maximizes the band-averaged received power (14) over pinch positions and feed delays,
By Theorem 3, placement enters the ergodic rate only through the band-averaged gain, and (14) is exactly that quantity ( under Assumption 1). Maximizing is thus no surrogate: the rate analysis sets this objective and fixes the array size (Proposition 2) that placement realizes. This problem is severely non-convex. Along any single coordinate , the cross terms of oscillate with the guided phase , which advances by cycles as sweeps the region. Each cycle creates a maximum and a minimum, so the objective has an order stationary points per coordinate and up to joint local optima. At GHz with m this is about stationary points per coordinate, the pinching analogue of the oscillatory landscape that forces grid search in wideband array design [15]. Gradient and Newton methods are therefore unreliable, and we instead exploit an exact integer structure.
The structure follows from the monotonicity of the electrical path. Since has derivative , it is a strictly increasing bijection on . Fixing a reference path , the co-phased lattice points defined by for are the positions at which a pinch is co-phased with the reference at . Every position decomposes uniquely as
into an integer order and a sub-wavelength offset , where is the local cell width.
Theorem 4. The placement problem (40) is equivalent, through the change of variables (41), to the mixed-integer program
The equivalence is exact. The combinatorial difficulty resides entirely in the integer orders , which set the wideband array factor of Theorem 1, while for fixed orders the objective is smooth in within each cell, with order-one stationary points per coordinate.
Proof. Monotonicity of makes the map a bijection between and the lattice cells, so the change of variables (41) preserves the feasible set and the objective without approximation. Fixing fixes the integer part of every electrical path, hence the co-phasing pattern that generates the array factor, while the dependence on enters only through the amplitude and a sub-wavelength phase that sweeps less than one cycle over a cell, which is smooth. ■
This reformulation is specific to pinching. It follows from Lemma 1, where co-phasing is an equality on path length modulo , which makes the admissible positions an exact integer lattice. A conventional phased array, whose element phases are set by electronic shifters rather than by physical path, has no such lattice. Moreover, with the optimal feed delays of Proposition 3 the objective separates across clusters, since distinct clusters are separated by many wavelengths and their cross terms average to zero over the band, so the order and offset selection is solved independently within each cluster.
Tractable Solution
The integer layer, which appears combinatorial, has a closed-form solution in the operating regime.
Proposition 4. For , that is throughout the coherent regime, the band-averaged gain of a size- cluster is maximized over all feasible integer order sets (adjacent order gaps at least of Section VI-C) by the minimal-step progression , uniquely up to a global shift. When the spacing-feasibility condition (46) holds, and the canonical placement of Sections III–IV is optimal.
Proof. Neglecting the slow roll-off, the band-averaged gain is proportional to with and . Expanding the square,
with ; maximizing the band-averaged gain is equivalent to maximizing over feasible integer sets. Order with gaps . The regime condition keeps every pairwise argument of the minimal-step configuration in , where sinc is positive and strictly decreasing. Suppose some gap ; form by decrementing every , , by one. Pairs on the same side of the cut keep their differences; each crossing pair has its difference reduced by exactly one, remaining . Since the reduction only decreases pairwise differences toward those of the minimal-step configuration, all arguments stay in , so every crossing term strictly increases and none decreases: . Iterating removes all excess gaps, so the minimal-step progression, the unique gap-floored configuration up to a global shift, is the unique maximizer. ■
With the orders fixed to the minimal-step progression, only the sub-wavelength offsets remain, and the objective is smooth in them by Theorem 4. Under the equal-amplitude approximation of Assumption 1 these offsets vanish at the optimum. They do work only when the amplitude varies across the cluster, as for a large aperture or a strong attenuation gradient, where, with the orders fixed, a successive convex approximation adapted from the movable-antenna capacity framework [8] refines them through the update
a quadratic program with a projected closed-form solution [24] whose surrogate lower-bounds the objective and is tight at the iterate, so the iterates increase monotonically to a stationary point [25]. The per-cluster cost is with the iteration count and the interior-point quadratic program, and across clusters the subproblems are independent with closed-form feed delays, giving an overall , polynomial in , , and . The exponential landscape of (P) is thus eliminated, the integer layer being closed-form and the continuous layer smooth.
Spacing Feasibility and Order Skipping
The minimal-step progression of Proposition 4 assumes consecutive orders are feasible. With consecutive orders, adjacent pinches sit one cell apart,
At GHz, broadside, : mm, which is sub-wavelength. If the coupling-limited spacing of (40) exceeds , consecutive placement is infeasible. The feasibility condition is
When (46) fails, co-phasing must skip orders: use step
giving adjacent spacing . The residual phase of Theorem 1 becomes , i.e. , so the coherence bandwidth and the spatial robustness of Theorem 5 both shrink by :
Both are aperture effects: order-skipping stretches the aperture to , and a larger aperture means more delay spread (narrower ) and tighter focus (narrower ), consistent with Proposition 1. When (46) holds, and Theorems 1 and 5 hold verbatim.
Robust Placement under Position Uncertainty
The previous subsections optimized placement for a known user position. In practice the position is known only to lie in an uncertainty set , such as a localization cell. Whereas Section IV averaged over the user distribution, we seek here a worst-case guarantee,
Robustness is governed by a spatial coherence region that mirrors the coherence bandwidth of Theorem 1.
Theorem 5. For a size- consecutive-order cluster co-phased for a reference user at distance , the array gain toward a user displaced axially by is the spatial Dirichlet kernel
whose two-sided 3-dB position-coherence width is
with the cluster aperture.
Proof. Let the pinches sit at with , co-phased for the reference user at distance . For an axial displacement , the distance perturbs as to first order. The co-phasing cancels the reference phase, so the array sum is with . The per-index phase is , a geometric series whose squared magnitude is (50). The main-lobe condition gives the half-width, and doubling it yields (51). ■
Because clusters are millimeter-scale, is of order decimeters to meters, so a pinching cluster is intrinsically robust to position uncertainty within its focal spot. The gain-robustness product is independent of , so array gain trades against position robustness exactly as it trades against bandwidth in (22), the first-order () case of Proposition 1. The robust problem then has a simple structure.
Proposition 5. Let be a convex cell with axial extent . The worst-case user (49) lies on the boundary of . If , the focal-spot-centered placement is robust, with worst-case gain within 3 dB of nominal. Otherwise the robust cluster center solves the one-dimensional problem
with the worst-case transverse term, which is quasi-concave in and balances the Chebyshev center of the cell against the attenuation gradient.
Algorithm 1. Structured Wideband PASS Placement
Require: user (or cell ), , , , subcarriers , guide
Ensure: pinch positions , feed delays
partition the pinches into clusters of size ; site each centroid at the pathloss optimum near the user (tile if , Prop. 5)
for each cluster do
;
feasible step, (47)
optimal orders, Prop. 4
by bisection monotone, Thm. 4
end for
causal feed delays, Prop. 3
if amplitude taper negligible (Assumption 1) then
else
refine by the SCA update (44) to convergence
end if
return ,
Proof. For a fixed placement, as a function of user position is the product of a unimodal focusing factor and a monotone pathloss factor , hence quasi-concave on the convex cell, so its minimum is attained on the boundary. If the axial diameter does not exceed , then throughout by Theorem 5, so the focal-centered placement loses at most 3 dB. Beyond the focal spot the binding term is pathloss, and for fixed the worst user maximizes distance, giving the inner value . Maximizing over yields (52), whose logarithm is concave in with convex and piecewise linear, so the maximizer trades the Chebyshev center against a smaller . ■
For a cell much larger than the focal spot, multiple co-phased clusters tile , each covering one focal spot. In all cases the feed delays remain closed-form, since they depend on cluster geometry rather than on the user, and the intra-cluster orders remain minimal-step by Proposition 4, so robustness is carried entirely by the cluster center and the number of tiling clusters.
The results above assemble into a solver that avoids any search over the oscillatory landscape. Each cluster centroid is sited at its pathloss optimum (or the cell is tiled by focal spots, Proposition 5); the feasible step and the optimal orders are set in closed form; the lattice positions follow from the monotone electrical path by one-dimensional root finding; the feed delays are closed-form; and the sub-wavelength offsets are either zero under Assumption 1, or refined by the SCA update (44) when the amplitude taper is non-negligible. The procedure is Algorithm 1, with cost when offsets are refined and when they vanish, in either case polynomial in , , , against the local optima of (P) and the delay elements of a conventional delay-phase array [5].
Optimization Benchmarks
Two baselines validate Algorithm 1. The first is an exhaustive lattice search that, made finite by Theorem 4, certifies global optimality. Since every candidate optimum lies on the co-phasing lattice and the feed delays of Proposition 3 decouple the clusters, for a cluster of pinches one enumerates all feasible order sets (strictly increasing integer vectors with adjacent gaps at least ) within a window of lattice cells and returns the maximizer of . This visits candidates and finds the global optimum exactly, a certificate against which any heuristic can be checked and one with no counterpart in a continuous-phase array. The enumeration is tractable because wideband coherence forces the cluster to be small (Remark 3).
The second baseline applies particle swarm optimization (PSO) directly to the continuous, non-convex (P), representative of the metaheuristics usually entrusted with oscillatory wideband-array objectives [15]. Each particle encodes a full placement , scored by the penalized fitness
with the sorted positions and , updated by the canonical velocity rule. PSO needs no structural knowledge but offers no guarantee and, as Section VII shows, stalls well below the certified optimum because the -per-coordinate local optima of (P) trap the swarm. Table II contrasts the three strategies; the quantitative comparison is deferred to Section VII.
| Strategy | Complexity | Optimality |
| Proposed (Alg. 1) | lattice-optimal (Prop. 4) | |
| Exhaustive lattice | global (certificate) | |
| PSO on (P) | none |
Table II. Solution strategies for wideband placement.
Numerical Results
This section validates the theoretical analysis through extensive Monte Carlo simulations. Throughout the numerical results, the analytical predictions are compared against independent simulations of the considered wideband channel model, demonstrating the accuracy of the derived expressions and the validity of the underlying theoretical approximations. Unless stated otherwise we adopt THz parameters drawn from the 6G and pinching-antenna literature. The system operates in the sub-THz 6G band at GHz over GHz [1, 26]–[28], on a low-loss dielectric guide with , [14, 29], and GVD parameter ps/m referenced and swept toward larger near-cutoff values [22, 23]. The user lies in a m service region at array height m [18], with minimum pinch spacing [14]. Ergodic quantities are averaged over user drops with , [18].
Inherent Squint and Coherence Bandwidth
Figure 2 overlays the Dirichlet closed form (19) on the full model, which coincides to marker width for every , so the kernel is exact and the array scans in rather than angle. Figure 3 turns this into a design law: panel (a) confirms and its slope, so coherent combining over needs , with dispersion costing a factor at broadside; panel (b) shows the product flat at (22), the case of Proposition 1, so gain and coherence bandwidth are one operating point on a fixed hyperbola that only TTD correction leaves.
Figure 2. Normalized array gain vs. fractional offset: full model (markers) vs. Dirichlet closed form (dashed).
Figure 3. Coherence bandwidth and its gain–bandwidth invariance.
b) The Irreducible Dispersion Floor: Figure 4(a) places the uncorrected and the post-TTD floor on log–log axes, with slopes and : TTD correction changes the exponent of the gain–bandwidth trade, not merely its constant, so the product grows as . At the nominal near-lossless guide the floor sits in the THz range (the crossover exceeds ), so an enlarged near-cutoff is used to bring it onto the axes; dispersion is then not the operative constraint, and after TTD the pinch count is spacing-limited, not dispersion-limited. Figure 4(b) confirms over a decade.

Figure 4. Dispersion floor and the order-changing effect of TTD correction.
c) Delay-Phase Mitigation with Few Delays: Figure 5(a) tracks the usable bandwidth as feed delays are added. Two regimes meet at a sharp knee. For small , grows linearly: each delay halves the sub-cluster size and, since the size- squint scales as , doubles the coherent band. Past the knee saturates at the dispersion floor , where the full-aperture GVD curvature takes over and further delays are wasted; the full-model markers follow the envelope of (38). The knee is the point of diminishing returns on delay hardware. Figure 5(b) shows the delays needed to reach the floor, , following the law (39) against the line of a conventional delay-phase array: the delayed fraction vanishes with array size. Pinch positioning provides per-element fine alignment, leaving the feed delays responsible only for the inter-cluster offsets. As in Figure 4, an enlarged near-cutoff brings the floor into the plotted range; the rise and scaling are independent of , which determines only the floor height.

Figure 5. Delay-phase mitigation: usable band and the delays needed to reach it.
d) Wideband Ergodic Rate and Molecular Absorption: Figure 6(a) validates the structural claim of Theorem 3: the user-position statistics and the wideband squint separate. The full Monte-Carlo matches the factorization through the low-to-moderate SNR range, departing only above 0 dB where the logarithm compresses. Since THz links are power- and distance-limited, this is exactly the operative regime: the wideband rate collapses to a product of a geometry-only functional (31) and a deterministic squint scalar , so array sizing and site planning decouple. Figure 6(b) plots the penalty against : small and flat below , rising steadily beyond it as subcarriers spill outside the main lobe. This identifies as the economical operating point and motivates the pinch-count optimization of Section VII-Oe.

Figure 6. Wideband ergodic rate and the squint penalty.
Figure 7 adds molecular absorption. Panel (a) shows for several hop distances: as grows the taper pulls efficiency down faster in but never reshapes the squint roll-off, confirming that absorption multiplies the squint kernel as an independent factor. Panel (b) makes the trade explicit through : is -independent while falls with distance, and their crossover (35) splits the deployment into a squint-limited regime () and an absorption-limited one (). Over the short LOS hops PASS is built to create, the operating point is squint-limited, so dispersion-driven squint is the wideband bottleneck; absorption dominates only on long hops or near a window edge, where is large and contracts toward the metre scale shown.
Figure 7. Molecular absorption as a deterministic taper multiplying the squint.
e) Rate-Optimal Pinch Count: Because the coherent band-centre SNR grows as while , the band-averaged rate has an interior optimum. Figure 8(a) shows the rate rising while the band stays coherent, then flattening into a knee at the closed-form (36), where , for both and GHz, confirming Proposition 2: size the array so coherence bandwidth just meets signal bandwidth. Figure 8(b) distinguishes two regimes. At low SNR the curve is a knee, over-provisioning costs only second order in the per-pinch SNR, so a power-limited THz array is forgiving. At high SNR it is a genuine peak, where over-provisioning wastes array gain on incoherent subcarriers; such links must size precisely at . The panel also shows that replacing by moves the knee to , orders of magnitude higher: after TTD the useful pinch count is limited by the spacing and aperture feasibility of Section VII-0f, not dispersion.

Figure 8. Rate-optimal pinch count and its TTD lift (Proposition 2, Corollary 1).
f) Spatial Robustness and Order Skipping: Figure 9(a) plots the carrier-frequency gain as the user is displaced axially from the co-phasing position. The full-model markers fall on the spatial Dirichlet kernel of Theorem 5 for every , the exact counterpart of the frequency kernel with in place of : spatial robustness and coherence bandwidth are one phenomenon in conjugate variables. The main-lobe width is of order decimetres, so a cluster tolerates position errors of that scale within 3 dB without tracking, and larger cells are tiled by several co-phased clusters. Figure 9(b) shows the gain–robustness product flat at , the mirror of the gain–bandwidth invariance and the same case of Proposition 1, so gain, coherence bandwidth, and position robustness are three readings of one design choice. Figure 10 treats the coupling constraint. The co-phasing cell is sub-wavelength ( at broadside), so when the array must skip orders by . Figure 10(a) shows this step as a staircase with the feasibility boundary at separating the consecutive regime (, where Sections VII-0a–VII-0e hold verbatim) from the order-skipped regime. Figure 10(b) shows both and contracting as , the full-model points tracking the law across . The mechanism is a single aperture effect, skipping orders stretches the cluster to , so element coupling, by setting , imposes a minimum aperture that debits both the usable bandwidth and the robustness budget.
Figure 9. Spatial robustness and its gain–robustness invariance.
Figure 10. Order skipping under coupling-limited spacing.
g) Placement Optimization: Proposed vs. Exhaustive vs. PSO: Figure 11(a) shows why (40) resists generic optimization and why its integer structure is the way out. Sweeping one pinch across the region, oscillates times, giving as many stationary per-point coordinate and joint local optima. Yet the maxima land exactly on the co-phasing lattice points of (41), so the continuous search reduces without approximation to a choice of integer orders (Theorem 4), and one should snap positions to the lattice and optimize the orders.

Figure 11. Global-Optimality of the Proposed Placement Algorithm
Figure 11(b) compares the three strategies. The proposed algorithm places pinches at minimal-step consecutive orders in closed form (Proposition 4) and attains an objective indistinguishable from the exhaustive certificate, so its optimality gap is numerically zero (0.00 dB) at sub-millisecond cost for every . PSO on the raw continuous problem stalls below the certificate, trapped by the per-coordinate optima, and its gap widens with array size from dB at to dB at and dB at . Meanwhile the exhaustive runtime climbs steeply, from 9 ms at to 0.6 s at , whereas the proposed runtime stays near 0.2 ms throughout. The structured method is thus the only one that is both optimal and scalable.
Conclusion
In this paper, we developed the first wideband, dispersion-aware theory of pinching-antenna systems, and showed that the pinching mechanism carries an inherent limitation. Because single-waveguide co-phasing equalizes path length only modulo a wavelength, a pinching array is intrinsically beam-squint limited, with a Dirichlet-kernel response, a closed-form coherence bandwidth, and an irreducible dispersion floor that even TTD correction cannot remove. A unifying law ties the behavior together, in which array gain trades against coherence bandwidth and against spatial robustness through one common invariant that TTD correction breaks, reaching the floor with far fewer delay elements than a conventional array. The law is actionable, yielding a rate-optimal pinch count where coherence meets signal bandwidth, a molecular absorption that multiplies the squint as a deterministic taper, and a placement problem that the integer structure of co-phasing renders an exactly solvable mixed-integer program with a provably optimal, polynomial-time solution. The analysis is foundational, isolating the dispersion–squint mechanism through a single-user LOS link. The multi-user case, where per-user co-phasing lattices conflict, is the natural next step, alongside multipath channels, hardware impairments, and experimental validation.
References
- [1]H. Sarieddeen, M.-S. Alouini, and T. Y. Al-Naffouri, “An overview of signal processing techniques for terahertz communications,” Proc. IEEE, vol. 109, no. 10, pp. 1628–1665, 2021.arxiv.org/abs/2005.13176
- [2]S. Tarboush et al., “TeraMIMO: A channel simulator for wideband ultra-massive MIMO terahertz communications,” IEEE Trans. Veh. Technol., vol. 70, no. 12, pp. 12 325–12 341, 2021.arxiv.org/abs/2104.11054
- [3]A. B. Enad et al., “Performance analysis of data detection in the THz-band under channel-correlated noise,” IEEE Commun. Lett., vol. 29, no. 9, pp. 2098–2102, 2025.arxiv.org/abs/2507.03702
- [4]A. B. Enad and C. Tellambura, “Exact performance analysis for correlated Rayleigh MIMO channels,” IEEE Commun. Lett., pp. 1–1, 2026.
- [5]L. Dai et al., “Delay-phase precoding for wideband THz massive MIMO,” IEEE Trans. Wireless Commun., vol. 21, no. 9, pp. 7271–7286, 2022.DOI
- [6]C. Huang et al., “Holographic MIMO surfaces for 6G wireless networks: Opportunities, challenges, and trends,” IEEE Wireless Commun., vol. 27, no. 5, pp. 118–125, 2020.DOI
- [7]M. Di Renzo et al., “Smart radio environments empowered by Reconfigurable Intelligent Surfaces: How it works, state of research, and the road ahead,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, 2020.
- [8]W. Ma, L. Zhu, and R. Zhang, “MIMO capacity characterization for movable antenna systems,” IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 3392–3407, 2024.arxiv.org/abs/2210.05396
- [9]A. Dkhan, S. Tarboush, H. Sarieddeen et al., “Physically consistent channel modeling and signal processing for reconfigurable wireless systems,” 2026.arxiv.org/abs/2608.17122
- [10]Y. Liu et al., “Pinching-antenna systems: Architecture designs, opportunities, and outlook,” IEEE Commun. Mag., vol. 64, no. 1, pp. 190–196, 2026.DOI
- [11]Z. Ding et al., “Flexible-antenna systems: A Pinching-Antenna perspective,” IEEE Trans. Commun., vol. 73, no. 10, pp. 9236–9253, 2025.arxiv.org/abs/2412.02376
- [12]S. A. Tegos et al., “Minimum data rate maximization for uplink Pinching-Antenna systems,” IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1516–1520, 2025.
- [13]K. Wang, Z. Ding, and R. Schober, “Antenna activation for NOMA assisted Pinching-Antenna systems,” IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1526–1530, 2025.arxiv.org/abs/2412.13969
- [14]C. Ouyang et al., “Array gain for pinching-antenna systems (PASS),” IEEE Commun. Lett., vol. 29, no. 6, pp. 1471–1475, 2025.arxiv.org/abs/2501.05657
- [15]A. Bereyhi et al., “MIMO-PASS: Uplink and downlink transmission via MIMO Pinching-Antenna systems,” IEEE Trans. Commun., vol. 74, pp. 5701–5716, 2026.arxiv.org/abs/2503.03117
- [16]C. Ouyang et al., “Capacity characterization of Pinching-Antenna systems,” IEEE Trans. Wireless Commun., vol. 25, pp. 10387–10404, 2026.arxiv.org/abs/2506.14298
- [17]Z. Ding et al., “Design of DoF-robust MIMO via generalized pinching antennas,” 2026. [Online]. Available: https://arxiv.org/abs/2610.36670
- [18]D. Tyrovolas et al., “Performance analysis of Pinching-Antenna systems,” IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 590–601, 2026.arxiv.org/abs/2502.06701
- [19]Y. Xie et al., “Terahertz beam squint mitigation via six-dimensional Movable Antennas,” 2026. [Online]. Available: https://arxiv.org/abs/2603.23859
- [20]T. S. Rappaport et al., “Wireless communications and applications above 100 GHz: Opportunities and challenges for 6G and beyond,” IEEE Access, vol. 7, pp. 78729–78757, 2019.DOI
- [21]A. B. Enad et al., “Diversity analysis for Terahertz communication systems under small-scale fading,” IEEE Open J. Commun. Soc., vol. 7, pp. 954–964, 2026.arxiv.org/abs/2504.09986
- [22]C. Yeh and F. I. Shimabukuro, The Essence of Dielectric Waveguides. Springer, 2008.
- [23]C. A. Balanis, Advanced Engineering Electromagnetics. John Wiley & Sons, Hoboken, NJ, 2024.
- [24]S. Boyd et al., Convex Optimization. Cambridge Univ. Press, 2004.
- [25]M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis of block successive minimization methods for nonsmooth optimization,” SIAM J. Optim., vol. 23, no. 2, pp. 1126–1153, 2013.arxiv.org/abs/1209.2385
- [28]A. B. Enad et al., “Symbol error analysis of linear receivers in Terahertz channels under channel-noise dependence,” 2026. [Online]. Available: https://arxiv.org/abs/2606.15006
- [29]J. R. Rodrigues et al., “All-dielectric scale invariant waveguide,” Nature Commun., vol. 14, no. 1, p. 6675, Oct. 2023.DOI





