Channel Multiplexing (CMX) Concepts

• Channel Multiplexing (CMX) is a design pattern that reorganizes multiple information channels into a shared carrier without losing data integrity.
• CMX is applied across domains such as audio generation, optical communications, quantum key distribution, and vehicular networking.
• Implementations leverage lossless reindexing, physical stream combining, and compression techniques to balance system complexity and performance.

Taken collectively, the research literature uses channel multiplexing (CMX) for a family of operations that place multiple information-bearing components into a shared carrier, representation, or infrastructure while preserving recoverability. The carrier may be a spectrogram tensor, a free-space optical beam, a few-mode fiber, a detector readout, or a coordinated radio stack. In recent audio generation work, CMX is a lossless spectrogram reshape that redistributes time–frequency samples into extra channels; in communications and sensing, it more often denotes the mapping of several logical channels onto one physical path or onto fewer hardware interfaces (Ristori et al., 30 Sep 2025, Turpin et al., 2013, Wang et al., 2024).

1. Conceptual scope

The literature suggests three recurrent meanings of channel multiplexing. First, CMX can denote a lossless redistribution across representation axes, where the number of spatial positions is reduced and the channel dimension is increased without discarding entries. Second, it can denote physical multiplexing, where several independent streams share one optical carrier, one beam, one fiber, or one radio infrastructure. Third, it can denote coordination or compression, where many detector or service channels are mapped onto fewer electronics channels or onto multiple managed radio channels.

Sense	Mechanism	Representative examples
Lossless reshaping	Space-to-depth or pixel-unshuffle–type remapping	MARS spectrogram CMX (Ristori et al., 30 Sep 2025)
Physical multiplexing	Polarization, subcarrier, OAM, or spatial-mode parallelization	CV-QKD SCM (Fang et al., 2013), twisted-light parity multiplexing (Rodenburg et al., 2015), FMF mode MUX (Zhou et al., 2023)
Coordination/compression	Graph-based readout encoding or multi-channel radio orchestration	MPGD Eulerian readout (Wang et al., 2024), ETSI MCO (Bazzi et al., 2023)

A plausible implication is that CMX is not a single operator but a design pattern: preserve decodability while shifting complexity from raw physical resources to structure, geometry, or inference.

2. Lossless tensor CMX in spectrogram autoregression

In "MARS: Audio Generation via Multi-Channel AutoRegression on Spectrograms" (Ristori et al., 30 Sep 2025), channel multiplexing is defined as “a reshaping technique that lowers height and width without discarding information.” The input waveform is transformed by STFT into a magnitude spectrogram, typically at 16 kHz with a 1024-point window and hop size 256. For an 8 s waveform this yields a $512 \times 512$ magnitude spectrogram, which the paper writes as

$$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$

CMX applies a spatial factor $m$ and transforms

$$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$

with

$$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$

Every non-overlapping $m\times m$ block is packed into the channel dimension. The mapping is a bijection, equivalent to a space-to-depth or pixel-unshuffle transform: $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$ Because this is pure reindexing, the inverse is exact and can be implemented with reshape-plus-permute operations.

Within MARS, CMX appears before tokenization and autoregression. The pipeline is waveform $\rightarrow$ STFT magnitude spectrogram $\rightarrow$ CMX-processed multi-channel image $\rightarrow$ shared tokenizer $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$0 VAR-style next-scale transformer $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$1 inverse CMX $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$2 Griffin–Lim waveform reconstruction. The tokenizer is adapted from ImageFolder and optimized with

$$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$3

combining $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$4 reconstruction, VQ codebook, and PatchGAN losses. The practical point is not compression in the lossy sense but bounded spatial size for vision-style transformers: a $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$5 or $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$6 token grid can carry the information of a much larger spectrogram by moving resolution into channels.

The paper gives both structural and empirical evidence for the usefulness of this formulation. A toy NSynth ablation compares truncation with CMX and reports that CMX makes training $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$7 faster, improves MSE by 0.002 and MAE by 0.022, and increases FAD by only 0.19. It also reports that simple frequency truncation produces a FAD gap of 2.16 and an MSE gap of 0.01 relative to the full-frequency setup. In the broader NSynth results, MARS reconstructs with NDB/$$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$8=0.19, PKID=0.0035, IKID=0.0015, MSE=0.0143, MAE=0.0915, and FAD=1.6429, and generated audio reaches FAD=1.8833. The same paper states that increasing input resolution in the original VAR model from $$X \in \mathbb{R}^{1 \times H \times W}, \quad H=W=512.$$9 to $m$0 raised parameter count from 310M to 2.3B, whereas CMX avoids that growth by keeping spatial size fixed and moving information into channels (Ristori et al., 30 Sep 2025).

3. Optical, photonic, and quantum realizations

In free-space optics, one realization of CMX uses conical refraction in biaxial crystals. A circularly polarized Gaussian beam is transformed into a ring whose polarization azimuth varies continuously with angle; diametrically opposite points are orthogonally polarized. With a degenerate two-cascade configuration, an angular amplitude mask selects sectors of that ring as distinct polarization–angular channels, and a second crystal remultiplexes them into one Gaussian-like beam. A third crystal at the receiver demultiplexes them again. The reported experiment demonstrates up to 12 independent channels from a single monochromatic beam over 4 m, with XT < 3% for the 12-channel mask and polarization crosstalk controlled below $m$1; alignment of the two transmitter crystals must be within 50 $m$2rad (Turpin et al., 2013).

In continuous-variable quantum key distribution, CMX appears as subcarrier multiplexing. A single optical carrier at $m$3 is phase-modulated by $m$4 RF subcarriers $m$5, producing $m$6 parallel Gaussian-modulated CV-QKD channels on sidebands $m$7. The central technical penalty is nonlinear intermodulation noise, summarized as

$m$8

For $m$9, the paper reports $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$0 for the noisiest channel and $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$1 for the cleanest one; with $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$2, the multi-channel gain $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$3 is approximately $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$4 at moderate distances (Fang et al., 2013).

In free-space twisted-light systems, CMX is implemented interferometrically by combining the even-$$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$5 OAM subspace from one input and the odd-$$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$6 OAM subspace from another input into a single output beam. The mechanism relies on a Mach–Zehnder interferometer with Dove prisms rotated by $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$7 relative to each other, so that a $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$8 rotation separates symmetric and antisymmetric components under OAM parity. The reported state tomography yields fidelity 0.9880, and the protocol supports either coherent superpositions or statistical mixtures, depending on source coherence (Rodenburg et al., 2015).

In integrated silicon photonics, CMX is realized as space-division and mode-division multiplexing for few-mode fibers. The reported integrated mode MUX combines a two-dimensional multimode grating coupler, mode size converters based on a subwavelength Mikaelian lens, adiabatic directional couplers, and phase shifters to launch up to eight spatial and polarization channels. Measured peak chip-to-fiber coupling efficiencies are $$X \in \mathbb{R}^{C \times H \times W} \longrightarrow X' \in \mathbb{R}^{C' \times H' \times W'}$$9 dB, $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$0 dB, $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$1 dB, and $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$2 dB for LP$$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$3, LP$$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$4, LP$$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$5, and LP$$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$6, respectively. The mode-independent field-size converter has loss $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$7 dB, and the combined MMGC-plus-MSC footprint is $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$8 (Zhou et al., 2023).

These cases differ in implementation, but they share a common structure: channels are distinguished by polarization angle, RF offset, OAM parity, or guided mode, then recombined into a single optical transport path.

4. Combinatorial and sparse multiplexing in sensing and waveform design

In micro-pattern gaseous detectors, CMX is a readout-compression method built on Eulerian circuits. Electronics channels are treated as graph vertices and detector strips as edges; contiguous strip clusters can then be decoded from which electronics channels fire. For even $$H'=\frac{H}{m},\quad W'=\frac{W}{m},\quad C'=C\cdot m^2.$$9 electronics channels, the paper gives the maximal readable detector-channel count as

$m\times m$0

Three hardware realizations are reported, including 8-to-1 and 16-to-1 multiplexing with 64 electronics channels. The abstract states that with a multiplexing factor of 8, spatial resolution remains comparable to direct readout and detection efficiency exceeds 94\%; with a factor of 16, spatial resolution degrades slightly but detection efficiency remains above 93.6\% (Wang et al., 2024).

In wireless ad-hoc networks, "Combinatorial Channel Signature Modulation" uses sparse codeword dictionaries, user-specific multipath channel signatures, and compressed sensing recovery so that all users may transmit and receive in the same time interval and on the same carrier frequency. Messages are encoded as $m\times m$1-out-of-$m\times m$2 sparse combinations, and the receiver solves a joint sparse recovery problem with the proposed Group Subspace Pursuit algorithm. The paper’s abstract states that this scheme “at least doubles the throughput when compared to the state-of-the art” (Piechocki et al., 2012).

In "Meta-Multiplexing" (Ji et al., 2017), multiplexing is achieved by splitting an ordinary modulated stream into $m\times m$3 accelerated streams, delaying them by $m\times m$4, and superposing them on a single channel. The effective linear model

$m\times m$5

is diagonalized by SVD into artificial parallel Gaussian sub-channels. The paper claims a capacity region in which, for sufficiently large $m\times m$6, capacity grows linearly with SNR rather than following the conventional logarithmic dependence, and reports a proof-of-concept hardware experiment with 81.7 bits/s/Hz spectral efficiency over a single channel (Ji et al., 2017).

A plausible implication is that these systems replace strict orthogonality with structure: graph incidence in detector readout, sparse recovery in ad-hoc access, or designed intersymbol overlap in waveform multiplexing.

5. Cooperative and geometric spatial multiplexing in wireless systems

In the half-duplex relay channel, CMX denotes cooperative multiplexing gain. For a source with one antenna, a half-duplex relay with one antenna, and a destination with two antennas, the paper characterizes diversity–multiplexing performance when the source–relay link scales as $m\times m$7. The maximum multiplexing gain is

$m\times m$8

Thus $m\times m$9 when $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$0, and when $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$1 a blind schedule with the relay listening for $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$2 of the time and transmitting for $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$3 of the time achieves the $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$4 MIMO DMT for multiplexing gains $$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$5 (0901.2164).

In IRS-aided LoS MIMO, CMX is recast as a geometric property of the cascaded channel

$$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$6

Here, full multiplexing depends on Tx–IRS and IRS–Rx distances, array orientations, IRS size, and a passive beamforming law called reflective focusing. The paper introduces the full multiplexing region (FMR) as the set of distance pairs for which the cascaded LoS channel can support full multiplexing, and derives an inner bound through generalized Rayleigh distances under reflective focusing (Zhang et al., 2021).

These two lines of work interpret multiplexing not as frequency allocation but as the creation of additional usable spatial degrees of freedom. In the relay case, the extra degree of freedom is distributed across time and terminals; in the IRS case, it is synthesized by spherical-wave phase diversity over a programmable aperture.

6. Protocol-level CMX in vehicular systems

In ETSI Cooperative Intelligent Transport Systems Release 2, channel multiplexing appears as multi-channel operation (MCO). The architecture manages multiple 10 MHz ITS-band channels labeled SCH0–SCH7 and may coordinate multiple radio technologies through Access Layer Instances (ALIs) and ALI groups. At the Facilities layer, the Bandwidth Management Entity (BME) negotiates Functional Configuration Profiles (FCPs) and Functional Configuration Limits (FCLs), while the Message Handling Entity (MHE) routes or offloads packets across channels. The standard supports sequential filling, load balancing, and elastic channel usage, as well as predefined and flexible channel association policies (Bazzi et al., 2023).

A more specific vehicular realization is CMD, a coordination scheme for emergency dissemination in IEEE 1609.4. The legacy standard uses a 100 ms synchronization interval with 50 ms CCHI and 50 ms SCHI, separated by 4 ms guard intervals. CMD selects channel coordinators by Least Average Distance (LAD) during CCHI and forwards emergency messages across SCHs during SCHI using multiple coordinators rather than one sequential switcher. Its analytical delay model gives

$$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$7

for the WSD baseline, but

$$X'_{[c',i,j]} = X_{[c,\, mi+a,\, mj+b]}, \quad c' = c\cdot m^2 + a\cdot m + b.$$8

for CMD. The paper also derives an optimal CCHI operation interval of approximately 8.38 ms for the coordination phases under its parameterization (Eyobu et al., 2018).

This suggests a clear systems distinction: some CMX mechanisms are waveform- or channel-matrix constructions, whereas others are architectural control planes that orchestrate many pre-existing channels.

7. Limits, trade-offs, and nomenclature

The principal technical misconception is that CMX is always either lossless or always capacity-increasing. The literature shows both regimes. In MARS, CMX is an exactly invertible permutation/reshape with no learnable parameters and no numerical loss (Ristori et al., 30 Sep 2025). By contrast, subcarrier-multiplexed CV-QKD introduces non-Gaussian extra source noise through intermodulation (Fang et al., 2013). Eulerian detector CMX depends on low occupancy, contiguous multi-strip clusters, and degrades on ambiguous single-strip events (Wang et al., 2024). Optical CMX can be limited by mask diffraction, beam quality, or alignment tolerances, as in conical-refraction multiplexing (Turpin et al., 2013).

A second source of confusion is nomenclature. The acronym CMX is not semantically uniform across fields. In RGB–X semantic segmentation, for example, "CMX" names a unified cross-modal fusion framework built from a Cross-Modal Feature Rectification Module (CM-FRM) and a Feature Fusion Module (FFM); that usage is terminologically separate from channel multiplexing proper (Zhang et al., 2022). The broader literature therefore suggests that “CMX” should be interpreted from context: in some papers it is an operator on tensor geometry, in others a physical multiplexing primitive, in others a readout-compression method, and in still others a protocol stack for multi-channel orchestration.

Across these meanings, one invariant remains. CMX denotes the deliberate restructuring of channel degrees of freedom—spatial, spectral, polarization, temporal, combinatorial, or tensorial—so that a constrained system can carry, process, or expose more useful information than a direct one-to-one mapping would permit.