betaFeedback

WorksEngineering

Low-Latency Joint Source-Channel Coding with Multiple Constant Curvature Curves

Abstract

We propose constant-curvature curve multicode (C3M), a multi-branch extension of constant-curvature curve tube (C3T) analog joint source-channel coding for one-shot transmission over fixed isotropic AWGN channels. Instead of mapping a scalar source to a single curve on the power sphere, C3M partitions the source interval across a union of constant-speed branch curves in even-dimensional Euclidean space. This construction increases the available geometric degrees of freedom while preserving a low block length analog encoding structure. We derive the induced mixture output distribution, posterior representation, MAP and MMSE estimators, and mutual-information decomposition into branch-identification and along-curve components. We then introduce geometry- and classification-aware branch allocation rules based on tube density, curve length, and branch-classification reliability. The resulting design procedure fixes the optimal C3T curve as a starting branch and sequentially selects additional branches, thereby reducing the combinatorial cost of exhaustive multi-curve search. We use NORM-inspired branch-conditioned spectral decoders based on Fourier-feature lifting of the received ambient vector. Numerical experiments in four and six dimensions show that adding candidate branch curves improves the finite-dimensional end-to-end signal-to-distortion ratio, with selected two- and three-candidate C3M systems outperforming the one-curve baseline at +10 dB SNR. The results show that multi-curve analog mappings provide a practical finite-dimensional design route for low block length joint source-channel coding.

Introduction

Sixth-generation wireless technology is expected to combine communication, sensing, distributed learning, and non-terrestrial connectivity in a single networked infrastructure [1, 2, 3]. Integrated sensing and communication, edge intelligence, radio-access-network slicing, and large numbers of heterogeneous IoT devices will create many short-packet, low-latency, and energy-constrained links [4, 5, 6]. In such settings, the main difficulty is not only to increase

This work was supported by the Research Council of Norway.

Anders Buvarp and Stefan Werner are with the Department of Electronic Systems, Norwegian University of Science and Technology, 7491 Trondheim, Norway (e-mail: {anders.buvarp,stefan.werner}@ntnu.no). Stefan Werner is also with the Department of Information and Communications Engineering, Aalto University, 00076, Finland. data rate, but also to transmit analog-valued sensor information with small delay, limited complexity, and robustness to interference, impulsive disturbances, and model mismatch [7, 8, 9].

Shannon’s separation theorem states that, in the infinite-block-length limit, a source can be compressed to a rate below the channel capacity and then transmitted reliably using a channel code [10]. This separated architecture, commonly referred to as separate source–channel coding (SSCC), is asymptotically optimal under the classical assumptions, but it can require long source and channel blocks to approach the asymptotic limit [11]. The resulting buffering and computational delay are poorly matched to ultra-reliable low-latency communication and real-time sensing applications. At finite block length and under implementation constraints, joint source–channel coding (JSCC) provides an alternative in which the source is mapped directly to channel inputs.

Analog JSCC, also known in this context as Shannon–Kotel’nikov (SK) mapping, encodes continuous-alphabet source samples directly into channel vectors without a separate quantization and channel-coding stage [12]. For scalar sources with bandwidth expansion by a factor of nn, a common construction maps one source sample to a point on a one-dimensional curve embedded in Rn\mathbb{R}^{n}. The curve geometry determines the tradeoff between local resolution and noise robustness. Longer curves improve tangential resolution, whereas excessive folding or small separation between curve segments increases the risk of threshold effects. Curve-based analog JSCC mappings on spheres, tori, and related manifolds therefore provide a natural finite-dimensional design problem for one-shot and low-blocklength communication [13, 14, 15, 16]. Related nonlinear analog JSCC schemes with MMSE decoding have also been studied as practical finite-dimensional reference points [17].

The present work is also adjacent to finite-blocklength JSCC, hybrid digital–analog JSCC, and learned JSCC. Finite-blocklength JSCC characterizes distortion under short-block constraints where asymptotic separation arguments are not directly decisive [11]. Learned JSCC instead optimizes encoder–decoder pairs end to end for structured sources, especially images [18]. C3M differs from these lines by focusing on one-shot scalar analog mappings whose signal sets are explicit unions of constant-curvature curves, so that branch geometry, source allocation, and branch identification can be studied directly.

The present paper builds on constant-curvature curve tube (C3T) mappings, where a scalar source is transmitted through a single constant-speed curve embedded in an even-dimensional Euclidean space [15]. C3T provides a geometrically structured one-shot analog code, but a single curve also limits the available design degrees of freedom. This motivates a multi-branch extension in which the source interval is partitioned across several constant-curvature branch curves. The resulting construction can be viewed as a discrete-index/continuous-position analog code: the branch index selects a geometric component, while the position along the selected branch carries the analog refinement.

Multi-curve analog mappings have appeared previously in related forms. Hekland et al. studied a two-curve analog code based on a double Archimedes spiral in a 2:1 bandwidth-reducing JSCC scheme [19]. In that setting, two intertwined spiral curves form the codebook for mapping pairs of i.i.d. Gaussian source samples to a single channel input. Campello, Torrezzan, and Costa proposed another multi-curve construction based on a set of mm curves foliating the surface of an nn-dimensional sphere [20]. These works show that multi-curve mappings can increase geometric design flexibility, but they also introduce inter-curve separation and branch-identification constraints that are absent in the single-curve case.

These observations motivate the constant-curvature curve multicode (C3M) developed in this paper. C3M partitions the source interval across multiple constant-speed branch curves and uses a branch-conditioned decoder to estimate the transmitted local coordinate. The resulting design problem combines intra-branch resolution, inter-branch separation, source allocation across branches, and receiver-side branch classification.

This paper derives the induced mixture output distribution, posterior representation, MAP and MMSE estimators, mutual-information expressions, and end-to-end and branch-known SDR references for the C3M construction. It then evaluates learned branch-conditioned decoders under nominal AWGN and non-Gaussian robustness tests. The learned receiver includes a NORM-inspired branch-conditioned spectral decoder that applies Fourier-feature lifting to the received ambient vector and uses circular Huber training for local-coordinate regression.

Prior work on analog curve codes suggests that performance is governed by the geometry of the signal locus, including curve length, local stretch, and minimum separation between distinct curve segments [14]. In multi-curve analog mappings, these geometric constraints become central because intra-curve reconstruction and inter-curve branch identification must be handled simultaneously [20].

Our previous work developed constant-curvature curve tube (C3T) mappings for 1:n1:n bandwidth-expansion analog JSCC using tube-packing and ropelength ideas [15, 21]. In C3T, a uniformly distributed scalar source is mapped

to a single constant-curvature curve in Rn\mathbb{R}^n. Figure 1 shows an example of such a C3T trajectory projected into three dimensions. The corresponding receivers used geometric feature extraction, including averaged torus projections, together with learned MLP decoders. A subsequent extension studied complex-valued and quaternion-valued neural decoders for related curve-based analog mappings [16].

Three-dimensional projection of the C3T manifold for dimensions 2, 3, and 4 with n = 4

Figure 1. Three-dimensional projection of the C3T manifold for dimensions 2, 3, and 4 with n=4n = 4. The colors denote the source parameter α\alpha.

The present work extends this single-curve construction to a multi-branch constant-curvature code. Related multi-curve analog mappings were studied by Campello et al. [20], who used multiple curves on spherical geometries and length-based allocation principles for non-overlapping curves. Here, the branch curves are derived from the C3T geometry, and the receiver is explicitly branch conditioned. The source interval is therefore not allocated solely according to geometric length. Instead, the allocation also accounts for receiver-side branch-identification reliability. The relevant design question is whether the receiver can first identify the correct branch and then accurately estimate the local coordinate along that branch.

In this work, our main contributions are as follows:

  1. Multi-branch C3M construction: We extend C3T from a single constant-curvature curve to a multi-branch analog JSCC construction, denoted C3M. Each branch is a C3T-type constant-speed constant-curvature curve embedded in an even-dimensional Euclidean space and associated with a flat-torus parameterization [22]. The resulting encoder is a piecewise constant-curvature map in which each source symbol activates one branch curve and one local coordinate on that branch.

  1. Analytical model and SDR references: We derive the induced mixture output distribution, posterior representation, MAP and MMSE estimators, and mutual-information expressions for the C3M construction. We also distinguish between the ordinary end-to-end Gaussian SDR reference and the genie-aided branch-known Gaussian SDR reference. This distinction is essential because the practical receiver must infer the branch index from the noisy channel output.

  1. Interval-allocation rules and diagnostics: We formulate equal-splitting, length-only, tube-density-only, and density–reliability allocation rules for partitioning the U[0,1)\mathcal{U}[0,1) source interval across branch curves. The density–reliability rule accounts for both the local geometry of each branch and the probability that the receiver selects the correct branch before performing along-curve decoding. The allocation ablation separates the effect of source-interval allocation from the effect of branch-geometry selection.

  1. NORM-inspired branch-conditioned spectral receiver: We propose a hierarchical receiver that first estimates the branch index and then applies a branch-conditioned decoder. Motivated by neural operators on Riemannian manifolds (NORM) [23], the implemented decoder applies Fourier-feature lifting to the received ambient vector and then uses residual and dense layers to estimate the local branch coordinate. A separate decoder is trained for each branch. A circular Huber loss is used for periodicity-aware local-coordinate regression under Gaussian-mixture impulsive noise, SαSS\alpha S perturbations, and branch-confusion events.

  1. Numerical evaluation and design tradeoff: We evaluate C3M in four and six dimensions under fixed isotropic AWGN and compare the learned decoders with the appropriate SDR references. The results show that additional branches can improve finite-dimensional end-to-end SDR, but they also introduce branch-confusion errors. The resulting tradeoff links curve geometry, source allocation, branch priors, and classifier reliability.

The remainder of this paper is organized as follows. § 2 reviews the background material underlying the proposed C3M construction, including C3T, flat tori, and geometry-based learning methods such as neural operators. § 3 introduces the system model, including the source, encoder, channel, noise, and receiver models. Section IV presents the proposed multi-curve design and develops equal-splitting, length-only, tube-density-only, and density–reliability allocation rules. § 5 presents the NORM-inspired branch-conditioned spectral decoder, including the Fourier-feature lifting, decoder architecture, and circular Huber-loss training objective. Section VI develops the analytical model, posterior estimators, mutual-information expressions, and SDR references. § 9 presents the numerical results, and § 11 concludes the paper.

Notation: O(⋅)\mathcal{O}(\cdot) denotes Bachmann–Landau big-O notation, and ∘\circ denotes degrees. The set of real numbers is denoted by R\mathbb{R}. The unit sphere in Rn\mathbb{R}^{n} is denoted by Sn−1={x∈Rn:∥x∥2=1}\mathbb{S}^{n-1} = \{x \in\mathbb{R}^{n} : \lVert x\rVert_{2} = 1\}. The operators E[⋅]\mathbb{E}[\cdot] and Pr⁡(⋅)\Pr(\cdot) denote expectation and probability, and ×\times denotes the Cartesian product. The Euclidean norm is denoted by ∥⋅∥\lVert\cdot\rVert, and In\mathbf{I}_{n} denotes the n×nn \times n identity matrix. The notation x∼U[a,b)x \sim\mathcal{U}[a,b) denotes a continuous uniform random variable on [a,b)[a,b), and x∼N(μ,Σ)x \sim\mathcal{N}(\mu,\boldsymbol{\Sigma}) denotes a Gaussian random vector with mean μ\mu and covariance matrix Σ\boldsymbol{\Sigma}. A symmetric alpha-stable random variable is denoted by x∼SαS(α,c,μ)x \sim\mathcal{S}\alpha\mathcal{S}(\alpha,c,\mu), where α\alpha is the stability parameter, cc is the scale, and μ\mu is the location. Uppercase letters denote random variables, lowercase letters denote their realizations, and hatted variables denote estimates. For example, SS, KK, and ZZ denote random variables, ss, kk, and zz denote realizations, and S^\widehat{S}, K^\widehat{K}, and Z^\widehat{Z} denote estimates. The transmitted and received vectors are denoted by X\mathbf{X} and Y\mathbf{Y}, respectively. The integer mm denotes the number of branch curves in the C3M construction. The interval endpoints are denoted by a0,a1,…,ama_{0},a_{1},\ldots,a_{m}, the corresponding subintervals by IkI_{k}, and their lengths by pkp_{k}. The random variable KK denotes the transmitted branch index. The kkth branch curve is denoted by γk(⋅)\gamma_{k}(\cdot), where γk:[0,1)→Mk⊂Tk⊂Rn\gamma_{k} : [0,1) \to\mathcal{M}_{k} \subset\mathcal{T}_{k} \subset\mathbb{R}^{n}. Here, Mk\mathcal{M}_{k} denotes the associated one-dimensional branch manifold, Tk\mathcal{T}_{k} the associated flat torus, and θk=(rk,ωk)\boldsymbol{\theta}_{k} = (\mathbf{r}_{k},\boldsymbol{\omega}_{k}) the corresponding parameter vector, with rk\mathbf{r}_{k} and ωk\boldsymbol{\omega}_{k} denoting the radii and frequency vectors, respectively. The signal set is X≜⋃k=1mγk([0,1))\mathcal{X} \triangleq\bigcup_{k=1}^{m} \gamma_{k}([0,1)). The length and protection radius of the kkth curve are denoted by LkL_{k} and ρk\rho_{k}, and the minimum separation between curves kk and ℓ\ell is denoted by dkℓd_{k\ell}.

Background

Section II-A reviews constant-curvature curves and geodesics on flat tori, which form the geometric basis for the C3T and C3M constructions. Section II-B discusses geometric principles for multi-curve code design, including curve length, separation, packing, and Grassmannian-frame viewpoints. Section II-C reviews feature extraction on non-Euclidean data structures and motivates the NORM-inspired branch-conditioned spectral decoder used later in the paper.

Constant-curvature curves and geodesics on flat tori

A manifold is a space that is locally homeomorphic to Euclidean space. In particular, a one-dimensional manifold M\mathcal{M} may be represented by a parametric curve embedded in Rn\mathbb{R}^{n}. Thus, for a suitable fundamental parameter interval I⊂RI \subset\mathbb{R}, we view a branch curve as a bijective map γ:I→M⊂Rn\gamma: I \to\mathcal{M} \subset\mathbb{R}^{n}, where γ\gamma is a smooth parametric curve and M\mathcal{M} is the associated embedded one-dimensional manifold.

In our setting, the relevant one-dimensional manifolds are the images of parametric constant-curvature curves embedded in Rn\mathbb{R}^{n}. The C3T and C3M curves considered in this paper are constructed as geodesics on flat tori. In even dimensions, these constant-curvature curves have constant generalized curvatures and are geodesics on flat tori Tn/2\mathbb{T}^{n/2} embedded in Sn−1⊂Rn\mathbb{S}^{n-1} \subset\mathbb{R}^{n}. These curves are represented by the vector [22]

x(α)=[r1cos⁡(ω1α),r1sin⁡(ω1α),r2cos⁡(ω2α),r2sin⁡(ω2α),…,rn/2cos⁡(ωn/2α),rn/2sin⁡(ωn/2α)]T.(1)\mathbf{x}(\alpha) = [r_1\cos(\omega_1\alpha), r_1\sin(\omega_1\alpha), r_2\cos(\omega_2\alpha), r_2\sin(\omega_2\alpha), \ldots, r_{n/2}\cos(\omega_{n/2}\alpha), r_{n/2}\sin(\omega_{n/2}\alpha)]^{T}. \tag*{(1)}

where r1,r2,…,rn/2r_1,r_2,\ldots,r_{n/2} are the radii and ω1,ω2,…,ωn/2\omega_1,\omega_2,\ldots,\omega_{n/2} are the frequency parameters. These harmonic coordinates satisfy a number of useful structural properties, summarized in the following lemma.

Lemma 1. Consider the even-dimensional parametric curve in Eq. (1), with α∈[0,2π)\alpha\in[0,2\pi), radii rj≥0r_j \ge0, and frequencies ωj∈R\omega_j \in\mathbb{R} for j=1,…,n/2j = 1,\ldots,n/2. Then:

  1. The squared norm is constant and satisfies ∥x(α)∥2=∑j=1n/2rj2\|\mathbf{x}(\alpha)\|^2 = \sum_{j=1}^{n/2} r_j^2. Hence, under the power constraint ∑j=1n/2rj2=P\sum_{j=1}^{n/2} r_j^2 = P, the curve lies on the sphere Sn−1(P)\mathbb{S}^{n-1}(\sqrt{P}).

  1. The squared speed is constant and satisfies ∥x′(α)∥2=∑j=1n/2ωj2rj2\|\mathbf{x}'(\alpha)\|^2 = \sum_{j=1}^{n/2}\omega_j^2r_j^2. Therefore the natural parametrization is constant-speed,

v=∥x′(α)∥=∑j=1n/2ωj2rj2.(2)v = \|\mathbf{x}'(\alpha)\| = \sqrt{\sum_{j=1}^{n/2}\omega_j^2r_j^2}. \tag*{(2)}
  1. The curve length over the parameter interval α∈[0,2π)\alpha\in[0,2\pi) is

L=2πv=2π∑j=1n/2ωj2rj2.L = 2\pi v = 2\pi\sqrt{\sum_{j=1}^{n/2}\omega_j^2r_j^2}.

When the frequencies are compatible with a 2π2\pi-periodic closed parametrization, this is the length over one period.

Proof. Using cos⁡2(θ)+sin⁡2(θ)=1\cos^2(\theta) + \sin^2(\theta) = 1 for each harmonic pair gives

∥x(α)∥2=∑j=1n/2rj2(cos⁡2(ωjα)+sin⁡2(ωjα))=∑j=1n/2rj2,\|\mathbf{x}(\alpha)\|^2 = \sum_{j=1}^{n/2}r_j^2\left(\cos^2(\omega_j\alpha) + \sin^2(\omega_j\alpha)\right) = \sum_{j=1}^{n/2}r_j^2,

which proves the first claim. Differentiating component-wise yields

x′(α)=[−r1ω1sin⁡(ω1α),r1ω1cos⁡(ω1α),…,−rn/2ωn/2sin⁡(ωn/2α),rn/2ωn/2cos⁡(ωn/2α)]T.(3)\mathbf{x}'(\alpha) = [-r_1\omega_1\sin(\omega_1\alpha), r_1\omega_1\cos(\omega_1\alpha), \ldots, -r_{n/2}\omega_{n/2}\sin(\omega_{n/2}\alpha), r_{n/2}\omega_{n/2}\cos(\omega_{n/2}\alpha)]^{T}. \tag*{(3)}

Hence

∥x′(α)∥2=∑j=1n/2ωj2rj2(sin⁡2(ωjα)+cos⁡2(ωjα))=∑j=1n/2ωj2rj2,\|\mathbf{x}'(\alpha)\|^2 = \sum_{j=1}^{n/2}\omega_j^2r_j^2\left(\sin^2(\omega_j\alpha) + \cos^2(\omega_j\alpha)\right) = \sum_{j=1}^{n/2}\omega_j^2r_j^2,

which is independent of α\alpha. This proves the constant-speed claim. Since the speed is constant, the length over α∈[0,2π)\alpha\in[0,2\pi) is

L=∫02π∥x′(α)∥ dα=2π∥x′(α)∥=2π∑j=1n/2ωj2rj2.L = \int_{0}^{2\pi}\|\mathbf{x}'(\alpha)\|\,\mathrm{d}\alpha= 2\pi\|\mathbf{x}'(\alpha)\| = 2\pi\sqrt{\sum_{j=1}^{n/2}\omega_j^2r_j^2}.

This proves the third claim.

In the present C3T/C3M construction, we typically choose integer frequencies, and in particular harmonic choices such as ωj=j\omega_j=j, so that the resulting torus geodesic is closed over one 2π2\pi parameter interval. Each coordinate pair in Eq. (1) lies on a radius-rir_i circle Si1(ri)S_i^1(r_i), so the curve lies on the flat torus

Trn/2=S11(r1)×S21(r2)×⋯×Sn/21(rn/2).T_{\mathbf r}^{n/2}=S_1^1(r_1)\times S_2^1(r_2)\times\cdots\times S_{n/2}^1(r_{n/2}).

Defining the angular coordinates θi=ωiα\theta_i=\omega_i\alpha gives (θ1,…,θn/2)∈Tn/2(\theta_1,\ldots,\theta_{n/2})\in T^{n/2}. Thus, the C3T/C3M branch curve corresponds to the linear flow θ(α)=(ω1α,…,ωn/2α)\boldsymbol{\theta}(\alpha)=(\omega_1\alpha,\ldots,\omega_{n/2}\alpha) on the flat torus, where ω=(ω1,…,ωn/2)\boldsymbol{\omega}=(\omega_1,\ldots,\omega_{n/2}) is the slope vector of the geodesic. A flat torus can be represented as Tn/2=Rn/2/(2πZn/2)T^{n/2}=\mathbb{R}^{n/2}/(2\pi\mathbb{Z}^{n/2}). Under the Euclidean embedding with radii r\mathbf r, the induced metric is the constant diagonal metric ds2=∑i=1n/2ri2 dθi2ds^2=\sum_{i=1}^{n/2}r_i^2\,d\theta_i^2. For i=1,…,n/2i=1,\ldots,n/2, geodesics of this flat metric satisfy d2θidα2=0\frac{d^2\theta_i}{d\alpha^2}=0. Since the C3T/C3M flow is given by θi=ωiα\theta_i=\omega_i\alpha, it follows that the curve is a geodesic on the corresponding flat torus. Mapping these torus coordinates into the Euclidean embedding via (ricos⁡θi,risin⁡θi)(r_i\cos\theta_i,r_i\sin\theta_i) yields the harmonic coordinate representation in Eq. (1).

It is known that packing density is important for the performance of continuous codes [24], and the radii are therefore selected to increase the usable tube density within the ambient power sphere. Consider the curve length in Lemma 1, the tube cross-sectional volume Vcs=κn−1ρn−1V_{cs}=\kappa_{n-1}\rho^{n-1}, the bounding nn-ball volume Vb=κn(1+ρ)nV_b=\kappa_n(1+\rho)^n, and the radii vector r\mathbf{r}. The tube density is defined as

D(r)=LVcs(ρ)Vb(ρ),(4)D(\mathbf{r}) = L\frac{V_{\mathrm{cs}}(\rho)}{V_{\mathrm{b}}(\rho)}, \tag*{(4)}

where κn\kappa_n denotes the volume of the unit nn-ball,

κn=πn/2Γ(n2+1).(5)\kappa_n = \frac{\pi^{n/2}}{\Gamma\left(\frac{n}{2}+1\right)}. \tag*{(5)}

Here Γ(t)=∫0∞xt−1e−x dx\Gamma(t) = \int_{0}^{\infty} x^{t-1}e^{-x}\,\mathrm{d}x is the Gamma function [21]. For even nn, this reduces to κn=πn/2/(n/2)!\kappa_n = \pi^{n/2}/(n/2)!. The protection radius ρ\rho is obtained from the global radius of curvature [21]. For two parameter values α1\alpha_1 and α2\alpha_2 on the curve in Eq. (1), let Δ=α1−α2\Delta= \alpha_1-\alpha_2 and define

ρ2(Δ)=[∑i=1n/2ri2(1−cos⁡(ωiΔ))]22∑i=1n/2ri2(1−cos⁡(ωiΔ))−[2π∑i=1n/2ωiri2sin⁡(ωiΔ)]2L2.(6)\rho^2(\Delta) = \frac{\left[\sum_{i=1}^{n/2}r_i^2(1-\cos(\omega_i\Delta))\right]^2}{2\sum_{i=1}^{n/2}r_i^2(1-\cos(\omega_i\Delta))-\frac{\left[2\pi\sum_{i=1}^{n/2}\omega_i r_i^2\sin(\omega_i\Delta)\right]^2}{L^2}}. \tag*{(6)}

The global protection radius used in the density calculation is then obtained by minimizing ρ(Δ)\rho(\Delta) over the admissible nonzero parameter separations. For a C3T curve on S3S^3, Figure 2 shows tube densities and curve lengths for a range of (r1,r2)(r_1,r_2) pairs, where all pairs satisfy the same power constraint ∥x(α)∥2=P\lVert\mathbf{x}(\alpha)\rVert^2=P.

Tube density and curve length versus radius angle

Figure 2. Tube density (left axis) and curve length (right axis) versus the radius angle ϕ=atan2⁡(r2,r1)\phi= \operatorname{atan2}(r_2,r_1) for the n=4n=4 radii parameters in Eq. (1). The SPSA-optimized radii reported in [15 Table I], (r1,r2)≈(0.8165,0.5774)(r_1,r_2) \approx(0.8165,0.5774), correspond to ϕ=35.27∘\phi= 35.27^\circ, which coincides with the maximum tube density. For ω=(1,2)\omega=(1,2) and r1=cos⁡ϕr_1=\cos\phi, r2=sin⁡ϕr_2=\sin\phi, the curve length is L(ϕ)=2πcos⁡2ϕ+4sin⁡2ϕL(\phi)=2\pi\sqrt{\cos^2\phi+4\sin^2\phi}, ranging from 2π2\pi to 4π4\pi.

In our C3T construction, we used harmonic integer frequencies, ωi=i\omega_i=i, so that the parametrized curve is closed over one 2π2\pi interval and may be interpreted as a torus-knot-type curve on the associated flat torus. In knot theory and ropelength analysis, the global radius of curvature determines the radius of an embedded tube around the curve centerline [21]. It is defined as the minimum circumradius over all triplets of points on the curve and therefore determines the admissible tube radius, see Figure 3 for an illustration. This tube enforces a minimum separation between distinct points on the curve, thereby reducing confusion between nearby source values at the receiver. The corresponding global protection radius is ρG=min⁡Δ∈[0,π]ρ2(Δ)\rho_G=\sqrt{\min_{\Delta\in[0,\pi]}\rho^2(\Delta)}.

Schematic illustration of a protection tube around a constant-curvature curve

Figure 3. Schematic illustration of a protection tube around a constant-curvature curve. The tube radius is determined by the global protection radius ρG\rho_G, while the minimum separation dmin⁡≥2ρGd_{\min} \ge2\rho_G between distinct folds controls the onset of threshold errors when noise moves a received point toward a neighboring fold.

Geometric principles for multi-curve code design

The design of multi-curve analog codes can be viewed as a geometric optimization problem. The relevant design variables include curve shape, orientation, spacing, and source allocation across branches. These variables affect both along-curve resolution and the probability of confusing nearby curve segments or nearby branch curves.

Consequently, ideas from manifold geometry, Riemannian optimization, frame design, and packing theory provide useful guiding principles for C3M.

Vaishampayan and Costa showed that the performance of analog codes is governed by the geometry of the signal locus [14]. In their curved modulation construction, they derived expressions for arc length and minimum distance along the curve and related these quantities to distortion performance. This reinforces the central design tradeoff in curve-based analog JSCC: increasing curve length improves local resolution, while maintaining sufficient separation between distinct curve segments reduces threshold errors at high SNR. In the multi-curve setting, the same principle applies both within each branch and across different branches.

A related viewpoint is provided by Grassmannian frames. Strohmer and Heath studied frames that minimize the maximum pairwise correlation, thereby linking frame design to packings in Grassmannian spaces and spherical codes [25]. Although their focus was on discrete vector sets rather than continuous curves, the underlying design principle is relevant here: different signaling components should be spread apart in the ambient space so that mutual interference is reduced. For C3M, this suggests choosing branch curves whose dominant directions, occupied regions, or local tangent structures are sufficiently separated. This is analogous to placing lines or subspaces far apart in a Grassmannian packing, but applied here to continuous branch manifolds rather than to finite frame elements.

These geometric viewpoints motivate the C3M design problem considered in this paper. The branch curves should provide sufficient along-curve resolution while maintaining enough inter-branch separation for reliable classifier-assisted decoding. Thus, the relevant design criteria combine packing geometry, manifold structure, source allocation, and receiver-side branch identification.

Feature extraction on non-Euclidean data structures

The geometric viewpoint above also motivates learning architectures that respect manifold structure. The receiver problem in C3M is not merely a Euclidean regression problem: the transmitted point lies on one of several structured branch manifolds, and the receiver must first identify the active branch and then estimate a local coordinate on that branch. This motivates feature-extraction methods that exploit non-Euclidean structure rather than treating the received vector only as an unstructured point in Rn\mathbb{R}^{n}.

Masci et al. generalized convolutional neural networks (CNNs) to non-Euclidean manifolds by introducing geodesic CNNs (GCNNs) [26]. Their construction defines convolutions through local geodesic coordinate patches on curved surfaces, thereby learning filters that exploit intrinsic manifold geometry. This illustrates how neural networks can be adapted to data that lie on curved domains and provides part of the motivation for decoder architectures that are adapted to branch geometry.

In our previous work on C3T, decoding was performed using a conventional multilayer perceptron (MLP). Although effective, such a model does not explicitly exploit the geometry of the underlying branch manifold. A broader class of approaches for non-Euclidean feature extraction is provided by geometric deep learning. One representative example is MoNet [27], which introduces mixture-model convolution kernels for graphs and manifolds. MoNet defines convolution as a template-matching operation on local intrinsic patches using learnable pseudo-coordinates and parametric kernels, typically based on Gaussian mixtures. In this way, MoNet unifies several earlier geometric deep-learning constructions, including geodesic CNNs and anisotropic CNNs [26, 28]. The main relevance for C3M is that such methods show how feature extraction can be made geometry-aware by operating on local manifold structure rather than on a fixed Euclidean grid.

A related line of work moves from discrete graph processing toward continuous-domain models. Continuous graph neural networks (CGNNs) [29, 30] extend geometric learning from finite graph neighborhoods to continuous domains, where learned kernels can be evaluated at arbitrary points on the underlying space. This is appealing in the present setting because the signal geometry is fundamentally continuous, even when the receiver is implemented through sampled or discretized representations.

Neural operators provide a further step in this direction by learning mappings between function spaces rather than only mappings between finite-dimensional vectors. Instead of approximating a continuous function through a fixed discrete graph or mesh, a neural operator learns the underlying continuous operator. In Euclidean domains, the Fourier neural operator realizes this idea by parameterizing operator kernels in the Fourier domain, which gives a resolution-independent representation of the learned map.

Neural operators have subsequently been extended to Riemannian manifolds through NORM [23]. NORM represents functions in truncated subspaces spanned by Laplace–Beltrami eigenfunctions. Since the Laplace–Beltrami operator plays a role on manifolds analogous to the Fourier transform in Euclidean space, it provides a natural spectral representation for learning maps on curved domains. This viewpoint motivates the decoder used in this paper: each C3M branch is associated with an embedded one-dimensional manifold contained in a flat torus, and the implemented receiver uses branch-conditioned spectral features to estimate the corresponding local coordinate.

This progression motivates the NORM-inspired receiver adopted in this work. The geometric-design viewpoints explain why the branch manifolds and their relative placement matter, while geometric deep-learning methods explain how feature extraction can be adapted to non-Euclidean domains. NORM provides a suitable endpoint for the present setting because it learns maps on Riemannian manifolds using spectral structure, allowing each branch-conditioned decoder to exploit the geometry associated with the selected C3M branch. § 3-C introduces the receiver model, and § 5 presents the corresponding NORM-inspired decoder and circular Huber training objective.

System Model

Our system model is illustrated in Figure 4 for several additive-noise settings, including symmetric alpha-stable noise, Gaussian mixture noise, and Gaussian noise, with additive white Gaussian noise (AWGN) arising as a special case when ϵ=0\epsilon= 0 or α=2\alpha= 2. The proposed C3M framework extends the single-curve C3T construction to a multi-branch setting in which the source interval is partitioned across mm branches, each associated with its own constant-curvature curve on a nearby flat torus. In this way, the encoder operates as a piecewise constant-curvature map from the scalar source domain to an nn-dimensional signal set. For each realization of the scalar memoryless source, the encoder first determines the active branch and a normalized local coordinate within that branch, and then maps this pair to a point on the corresponding branch curve. The resulting channel input is transmitted over an nn-dimensional additive-noise channel. At the receiver, the noisy observation is first processed by a curve classifier that estimates the transmitted branch index, after which a branch-conditioned decoder reconstructs the source symbol. The system performance is measured through the end-to-end mean-squared distortion D≜E[(S−S^)2]D \triangleq\mathbb{E}\left[(S-\widehat{S})^{2}\right], which is used throughout the paper to compare interval allocations, branch geometries, and decoder configurations.

Block diagram of the proposed C3M system

Figure 4. Block diagram of the proposed C3M system. The source symbol satisfies S∼U[0,1)S \sim\mathcal{U}[0,1) and is first mapped to a branch index and a normalized local coordinate, which determine the active constant-curvature branch curve for transmission. The nominal channel model is fixed isotropic AWGN. Gaussian-mixture impulsive noise and symmetric alpha-stable noise are used as robustness test cases. At the receiver, a curve classifier estimates the transmitted branch, after which a branch-conditioned decoder reconstructs the source symbol.

Source and encoder model

Consider a scalar memoryless source S∼U(0,1)S \sim\mathcal{U}(0,1), which is partitioned into mm disjoint source intervals

Ik=[ak−1,ak),k=1,…,m,(7)I_k = [a_{k-1},a_k), \qquad k=1,\ldots,m, \tag*{(7)}

with

0=a0≤a1≤⋯≤am=1,pk≜ak−ak−1≥0.(8)0 = a_0 \le a_1 \le\cdots\le a_m = 1, \qquad p_k \triangleq a_k-a_{k-1} \ge0. \tag*{(8)}

Branches with pk>0p_k > 0 are active source-carrying branches. Let A={k:pk>0}\mathcal{A} = \{k : p_k > 0\} denote the active branch set. All interval maps below are understood to be restricted to branches in A\mathcal{A}. A candidate branch with pk=0p_k = 0 is interpreted as pruned by the allocation rule and does not carry source mass in the final encoder. The corresponding branch-index random variable is defined by

K=k⟺S∈Ik,(9)K = k \quad\Longleftrightarrow\quad S \in I_k, \tag*{(9)}

so that, for the uniform source, Pr⁡(K=k)=pk\Pr(K = k) = p_k.

For an active branch with pk>0p_k > 0 and a realization s∈Iks \in I_k, define the normalized local coordinate

z=s−ak−1pk∈[0,1).(10)z = \frac{s-a_{k-1}}{p_k} \in[0,1). \tag*{(10)}

Accordingly, the corresponding local-coordinate random variable is

Z≜S−aK−1pK∈[0,1).(11)Z \triangleq\frac{S-a_{K-1}}{p_K} \in[0,1). \tag*{(11)}

The first stage of the encoder is the interval map

f:[0,1)→{1,…,m}×[0,1),f(s)=(k,z).(12)f : [0,1) \to\{1,\ldots,m\} \times[0,1), \qquad f(s) = (k,z). \tag*{(12)}

or, equivalently,

f(s)=(k,s−ak−1pk),s∈Ik.(13)f(s) = \left(k,\frac{s-a_{k-1}}{p_k}\right), \qquad s \in I_k. \tag*{(13)}

At the random-variable level, this becomes (K,Z)=f(S)(K,Z) = f(S). Because each source symbol belongs to exactly one active interval and can be recovered uniquely from the pair (k,z)(k,z) with k∈Ak \in\mathcal{A}, the map ff is bijective on the active branch set, with inverse f−1(k,z)=ak−1+pkzf^{-1}(k,z) = a_{k-1} + p_k z.

For each branch k∈{1,…,m}k \in\{1,\ldots,m\}, let

γk:[0,1)→Rn(14)\gamma_k : [0,1) \to\mathbb{R}^{n} \tag*{(14)}

denote a constant-curvature branch curve of the form reviewed in § 2-A, with branch-specific parameter vector θk=(rk,ωk)\theta_k = (r_k,\omega_k). Define the branch signal set Xk≜γk([0,1))\mathcal{X}_k \triangleq\gamma_k([0,1)) and the full signal set

X≜⋃k=1mXk=⋃k=1mγk([0,1)).(15)\mathcal{X} \triangleq\bigcup_{k=1}^{m} \mathcal{X}_k = \bigcup_{k=1}^{m} \gamma_k([0,1)). \tag*{(15)}

The second stage of the encoder is geometric. Given a branch index kk and local coordinate zz, the map gg selects the corresponding branch curve and generates the transmitted point on that curve:

g:{1,…,m}×[0,1)→X,g(k,z)=γk(z).(16)g : \{1,\ldots,m\} \times[0,1) \to\mathcal{X}, \qquad g(k,z) = \gamma_k(z). \tag*{(16)}

Since the branch curves are pairwise disjoint and each parameterization z↦γk(z)z \mapsto\gamma_k(z) is one-to-one on [0,1)[0,1), the map gg is bijective onto X\mathcal{X}. Consequently, the overall noiseless encoder

e≜g∘f:[0,1)→X(17)e \triangleq g \circ f : [0,1) \to\mathcal{X} \tag*{(17)}

is also bijective. The transmitted vector random variable is therefore

X=e(S)=g(f(S))=γK(Z)∈Rn.(18)\mathbf{X} = e(S) = g(f(S)) = \gamma_K(Z) \in\mathbb{R}^{n}. \tag*{(18)}

Figure 4 illustrates the resulting C3M transmitter, channel, classifier, and decoder. Throughout, the branch curves are normalized to satisfy the transmit-power constraint E[∥X∥2]=P\mathbb{E}[\lVert\mathbf{X}\rVert^2] = P and, unless otherwise stated, we use the normalization P=1P = 1.

Channel model and noise assumptions

Typically, wireless channels are modeled using AWGN together with some form of fading. However, the Gaussian assumption can be inadequate in practical implementations, where the channel may be perturbed by impulsive or heavy-tailed disturbances [31, 32]. In such settings, the noise distribution is no longer well described by a single Gaussian law. Common non-Gaussian models include Gaussian-mixture impulsive-noise models [31, 32, 33] and symmetric alpha-stable models [33, 34], which provide useful test cases for robustness evaluation.

The Gaussian-mixture impulsive-noise model is particularly attractive because it connects naturally to robust estimation theory. In this interpretation, most samples follow a nominal Gaussian distribution, while a fraction of the observations behave as outliers. This viewpoint is well aligned with impulsive-noise channels, in which rare but large perturbations contaminate an otherwise nominal background-noise process. To also capture heavier-tailed behavior, we consider symmetric alpha-stable noise [33, 34].1 Stable distributions are heavy-tailed and leptokurtotic, except for the Gaussian case obtained at α=2\alpha= 2. For β=0\beta= 0, the distribution is symmetric about its location parameter and is referred to as a symmetric alpha-stable (SαSS\alpha S) distribution. Thus, the two non-Gaussian noise models considered here provide complementary robustness test cases: Gaussian-mixture impulsive noise captures finite-variance contamination by rare large perturbations, while SαSS\alpha S noise captures more general heavy-tailed behavior.

After encoding, the transmitted vector is corrupted by additive channel noise. Figure 4 illustrates the proposed C3M system. The nominal channel model is fixed isotropic AWGN, and the Gaussian-mixture impulsive-noise and symmetric alpha-stable (SαSS\alpha S) models are used as robustness test cases. Accordingly, the channel is modeled as

Y=X+N,\mathbf{Y} = \mathbf{X} + \mathbf{N},

where the noise vector N\mathbf{N} is independent of the source S\mathbf{S}. Each source symbol is transmitted using one nn-dimensional channel use, corresponding to a bandwidth-expansion factor of nn. In the Gaussian case, we write N∼N(0,Σn)\mathbf{N} \sim\mathcal{N}(\mathbf{0},\boldsymbol{\Sigma}_n), where Σn\boldsymbol{\Sigma}_n is a positive-definite covariance matrix. The isotropic AWGN case is recovered as the special case

Σn=σn2In.\boldsymbol{\Sigma}_n = \sigma_n^2\mathbf{I}_n.

For impulsive Gaussian-mixture noise, we write the noise vector N\mathbf{N} as

N∼(1−ϵ)N(0,Σw)+ϵN(0,Σv),(19)\mathbf{N} \sim(1-\epsilon)\mathcal{N}(\mathbf{0},\boldsymbol{\Sigma}_w) + \epsilon\mathcal{N}(\mathbf{0},\boldsymbol{\Sigma}_v), \tag*{(19)}

where ϵ∈[0,1]\epsilon\in[0,1] is the contamination probability and Σw,Σv∈Rn×n\boldsymbol{\Sigma}_w,\boldsymbol{\Sigma}_v \in\mathbb{R}^{n \times n} are covariance matrices. The AWGN case is recovered when ϵ=0\epsilon= 0 and Σw=σn2In\boldsymbol{\Sigma}_w = \sigma_n^2\mathbf{I}_n. For symmetric alpha-stable noise, we write the jjth component of N\mathbf{N} as

Nj∼i.i.d.SαS(α,c,0),j=1,…,n,(20)N_j \overset{\mathrm{i.i.d.}}{\sim} S\alpha S(\alpha,c,0), \qquad j = 1,\ldots,n, \tag*{(20)}

where α∈(0,2]\alpha\in(0,2] is the stability parameter, and c>0c > 0 is the scale parameter. The Gaussian case is recovered at α=2\alpha= 2. Unless otherwise stated, we normalize the transmit power to

P=E[∥X∥2]=1,(21)P = \mathbb{E}\left[\|\mathbf{X}\|^2\right] = 1, \tag*{(21)}

and hence in the isotropic AWGN case, i.e., when N∼N(0,σn2In)\mathbf{N} \sim\mathcal{N}(0,\sigma_n^2 I_n), we define the total signal-to-noise ratio as

SNR≜Pσn2=1σn2.(22)\mathrm{SNR} \triangleq\frac{P}{\sigma_n^2} = \frac{1}{\sigma_n^2}. \tag*{(22)}

Unless explicitly stated otherwise, the analytical posterior expressions, MAP and MMSE estimators, mutual-information calculations, capacity references, and SDR reference curves in Sections VI–VII are evaluated under the fixed isotropic AWGN model. In Figure 9, the one-curve panels use the end-to-end Gaussian SDR reference, while the multi-curve panels use the branch-known Gaussian SDR reference. The Gaussian-mixture and SαSS\alpha S models are used only as robustness test cases for the Huber-trained branch-conditioned decoder and for evaluating deviations from the nominal Gaussian channel.

Six-panel plots of SDR and decoder-inference accuracy versus SNR for C3M systems

Figure 9. SDR and decoder-inference accuracy versus SNR for C3M systems with different ambient dimensions and numbers of branch curves. The top row shows the four-dimensional systems and the bottom row shows the six-dimensional systems. The columns correspond to m=1m = 1, m=2m = 2, and m=3m = 3 branch curves, respectively. In each panel, the upper plot reports SDR together with decoder-inference accuracy, while the lower plot reports the corresponding bootstrap standard deviation. The one-curve panels show the end-to-end Gaussian SDR reference from Eq. (69) and include the C3T baseline [15]. The multi-curve panels show the branch-known Gaussian SDR reference from Eq. (71) and compare NORM-inspired C3M decoding with the available MLP baselines. These reference curves are included for interpretation and are not finite-dimensional upper bounds on the learned decoders.

Receiver model

The receiver is hierarchical. The first stage estimates the transmitted branch index from the noisy observation, and the second stage performs branch-conditioned decoding within the selected branch. This structure mirrors the encoder: first a coarse branch decision, then a fine along-curve reconstruction.

The first stage is a curve classifier

K^=C(Y),C:Rn→{1,…,m},(23)\widehat{K} = C(\mathbf{Y}), \qquad C : \mathbb{R}^{n} \to\{1,\ldots,m\}, \tag*{(23)}

which estimates the transmitted branch index.

Equivalently, the receiver may be viewed as a bank of mm branch-conditioned decoders. Each decoder observes the same channel output Y∈Rn\mathbf{Y} \in\mathbb{R}^{n} and is associated with one candidate branch. The classifier estimate K^\widehat{K} selects the active decoder, which then produces the local-coordinate estimate Z^\widehat{Z} and, through the inverse interval map, the final scalar estimate S^\widehat{S}. In the absence of side information, the receiver must infer the transmitted branch identity from the channel output. From a detection perspective, this can be formulated as a maximum a posteriori (MAP) branch-selection problem, followed by branch-conditioned decoding within the selected branch.

Conditioned on K^\widehat{K}, the decoder produces the reconstructed source symbol

h:Rn×{1,…,m}→[0,1),S^=h(Y,K^).(24)h : \mathbb{R}^{n} \times\{1,\ldots,m\} \to[0,1), \qquad\widehat{S} = h(\mathbf{Y},\widehat{K}). \tag*{(24)}

In the implementation considered in this paper, h(⋅,⋅)h(\cdot,\cdot) is realized by a NORM-inspired branch-conditioned spectral decoder, see Section V. Conceptually, the decoder may be viewed as first estimating a branch-local coordinate and then mapping it back to the source interval through the inverse interval map.

Multi-Curve Design

This section describes how the branch curves used in C3M are constructed once a baseline C3T design has been fixed. The main idea is to retain the constant-curvature branch structure while introducing additional nearby branches through new radii vectors. Throughout this section, the frequency parameters are kept fixed, and the branch design is performed through the choice of the radii parameters.

Because the measured SDR depends on the noise normalization, all main C3M design sweeps are evaluated using a fixed channel-noise covariance that is independent of the candidate curve radii. This distinguishes the communication-channel model from radius-dependent normalizations, where coordinate pairs with small radii also receive proportionally smaller noise. The latter can mask near-degenerate radius choices and is therefore treated only as an auxiliary normalization rather than as the main C3M channel model. Unless otherwise stated, all C3M design sweeps are performed under fixed isotropic AWGN, N∼N(0,σn2In)\mathbf{N} \sim\mathcal{N}(0,\sigma_n^2\mathbf{I}_n), where the noise variance is independent of the branch radii and candidate curve parameters.

The decoders are trained using a mixed-SNR data set with samples drawn from 0, +3, +6, and +10 dB. Unless otherwise stated, the SNRs are sampled uniformly, so that the learned receiver is exposed to a range of moderate-noise conditions rather than a single operating point. Under the power normalization used in this work, the radii vector lies on the positive orthant of the unit sphere, i.e., ∑i=1n/2ri2=1\sum_{i=1}^{n/2} r_i^2 = 1, ri≥0r_i \ge0. Thus, the multi-curve design problem may be viewed as the selection of multiple radii vectors, each defining a constant-curvature branch curve on its own nearby flat torus.

For n=4n=4, let ϕ\phi denote the angle associated with the radii pair (r1,r2)(r_1,r_2) in Eq. (1), i.e., ϕ=atan2⁡(r2,r1)\phi= \operatorname{atan2}(r_2,r_1). The optimal pair (r1,opt,r2,opt)(r_{1,\mathrm{opt}},r_{2,\mathrm{opt}}) maximizing the tube density in Eq. (4) is reported in [15 Table I]. Note that this optimal pair satisfies r1,opt≈45r_{1,\mathrm{opt}} \approx\frac{4}{5}, r2,opt≈35r_{2,\mathrm{opt}} \approx\frac{3}{5}, corresponding approximately to the Pythagorean triplet (3,4,5)(3,4,5).

Given this optimal radii pair, and keeping the frequency parameters fixed, the multi-curve design objective is to identify additional branch curves (r1,r2)(r_1,r_2) that maximize the resulting SDR while remaining geometrically distinguishable from the baseline curve. For n=4n = 4, the candidate radii pairs are parameterized as

\begin{cases} r_{1} = \cos\phi, \\ r_{2} = \sin\phi, \end{cases} \qquad0^\circ\le\phi\le90^\circ. \tag*{(25)}\end{cases}
QuantityStarting C3T Curvea^{a}Second CurveThird Curve
ϕ=atan2⁡(r2,r1)\phi= \operatorname{atan2}(r_{2}, r_{1}) [deg]35.2747.0044.00
r1r_{1}0.81650.68200.7193
r2r_{2}0.57740.73140.6947
Length8.886010.14049.8300
Tube density0.37670.34110.3554

Table 1. Four-dimensional C3M candidate-branch parameters at +10 dB SNR.

Thus, the search is restricted to the first quadrant. For the case n=4n = 4 and m=2m = 2, Figure 5a illustrates the dependence of both the system SDR and the tube density of the second curve on ϕ\phi, when the first curve is fixed to the optimal C3T design from [15 Table I]. The resulting multi-curve design details, including radii, tube densities, curve lengths, interval fractions, and classification accuracies for n=4n = 4, are reported in Table I.

Sequential branch-selection results for the $n=4$ C3M design at +10 dB SNR

Figure 5. Sequential branch-selection results for the n=4n=4 C3M design at +10 dB SNR. (a) System SDR (left axis) and system tube density (right axis) versus ϕ\phi for the one-, two-, and three-candidate systems. (b) Classification accuracy (left axis) and interval fraction Pi(m)(ϕ)P_i^{(m)}(\phi) (right axis) versus ϕ\phi for the corresponding m=2m=2 and m=3m=3 sweeps.

For n=6n = 6, the candidate radii triples are parameterized by two angles as

r1=sin⁡θcos⁡ϕ,r2=sin⁡θsin⁡ϕ,0∘≤θ≤90∘,0∘≤ϕ≤90∘,r3=cos⁡θ.(26)\begin{aligned} r_1 &= \sin\theta\cos\phi,\\ r_2 &= \sin\theta\sin\phi,\qquad0^\circ\le\theta\le90^\circ,\quad0^\circ\le\phi\le90^\circ,\\ r_3 &= \cos\theta. \tag*{(26)} \end{aligned}

Again, the search is restricted to the first quadrant of the corresponding angular parameterization. The radii vector r=(r1,r2,r3)\mathbf{r} = (r_1,r_2,r_3) satisfies the power constraint r12+r22+r32=1r_1^2+r_2^2+r_3^2=1, so the admissible radii lie on the unit sphere S2⊂R3S^2 \subset\mathbb{R}^3. Since the harmonic pairs occupy mutually orthogonal two-dimensional subspaces of R6\mathbb{R}^6, the corresponding encoder curve lies on the unit sphere S5⊂R6S^5 \subset\mathbb{R}^6, while its geometry is determined by the choice of r\mathbf{r}. Thus, once the frequencies are fixed, designing an additional branch in R6\mathbb{R}^6 may be viewed as selecting an additional point on the positive orthant of S2S^2. The multi-curve design details for n=6n = 6 are shown in Table II and the corresponding two-curve and three-curve sequential branch-selection procedures are illustrated in Figure 6. For n=8n = 8, the candidate radii quadruples (r1,r2,r3,r4)(r_1,r_2,r_3,r_4) are parameterized analogously by three angles in the first quadrant. More generally, for an nn-dimensional space with even nn, the radii design involves n/2−1n/2 - 1 angular parameters. Thus, the branch-design problem reduces to a search over the positive orthant of the unit sphere in Rn/2\mathbb{R}^{n/2}.

Figure 6. Six-dimensional C3M candidate sweeps at +10 dB SNR, using the optimal C3T curve from [15 Table I] as the starting branch. Panels (a)–(c) show the 15 × 15 second-curve sweep for the m=2m = 2 system: (a) classification accuracy, (b) system SDR, and (c) system tube density. The selected m=2m = 2 point is (θ,ϕ)=(50∘,60∘)(\theta,\phi) = (50^\circ,60^\circ). Panels (d)–(e) show the corresponding 15 × 15 third-curve sweep for the m=3m = 3 system: (d) classification accuracy and (e) system SDR. The selected third-curve candidate is (θ,ϕ)=(50∘,55∘)(\theta,\phi) = (50^\circ,55^\circ).

QuantityStarting C3T Curve2Second CurveThird Curve
θ\theta [deg]60.5450.0050.00
ϕ\phi [deg]47.9260.0055.00
r1r_10.58350.38300.4394
r2r_20.64630.66340.6275
r3r_30.49180.64280.6428
Length12.8614.9014.72
Tube density0.10140.08250.0848

Table II. Six-dimensional C3M candidate-branch parameters at +10 dB SNR.

tbl:3 shows that the selected six-dimensional three-candidate configuration contains one candidate branch with zero allocated source mass. Thus, this case should be interpreted as a three-candidate branch set in which the allocation rule effectively prunes the second candidate branch. The reported overall branch-classification accuracy is the interval-weighted value ∑kpkqk\sum_k p_kq_k, not the unweighted average over candidate branches. This distinction is important for the n=6n=6, m=3m=3 case because the second candidate branch has p2=0p_2=0.

For even ambient dimension nn, the branch geometry admits a convenient radius-vector interpretation. Writing N=n/2N=n/2, the radii vector of branch kk is rk=(r1,k,…,rN,k)∈RN\mathbf{r}_k=(r_{1,k},\ldots,r_{N,k})\in\mathbb{R}^{N}, ∥rk∥2=∑i=1Nri,k2=1\lVert\mathbf{r}_k\rVert^2=\sum_{i=1}^{N}r_{i,k}^2=1, so that rk\mathbf{r}_k lies on the unit sphere SN−1\mathbb{S}^{N-1}. Thus, once the frequency parameters are fixed, selecting multiple branch curves may be viewed as selecting multiple points on SN−1\mathbb{S}^{N-1}, restricted to the positive orthant by the nonnegativity of the radii.

In this sense, the multi-curve design problem reduces to a spherical-code-like packing problem in radius-vector space: one seeks branch radii vectors that are sufficiently well separated to support reliable classification while still yielding geometrically efficient curves.

This viewpoint is particularly useful for interpreting the low-dimensional cases considered in this paper. For n=4n=4, one has N=2N=2, so the admissible radii vectors lie on S1\mathbb{S}^{1} and may be parameterized by a single angle ϕ\phi. For n=6n=6, one has N=3N=3, so the admissible radii vectors lie on S2\mathbb{S}^{2} and may be parameterized by the angular pair (θ,ϕ)(\theta,\phi). More generally, for even nn, the branch-design problem may be viewed as the selection of points on Sn/2−1\mathbb{S}^{n/2-1}, followed by evaluation of the resulting geometric metrics, classification reliability, and end-to-end SDR.

It is useful to separate multi-curve design into two stages. The first stage selects the branch curves. The second stage determines how the source interval is divided among the selected branches. In the geometry stage, the objective is to find a branch set with favorable geometric scores and sufficient separation. In the allocation stage, the source interval is assigned according to the resulting per-branch utilities. Sections IV-A, IV-B, and IV-C focus on the allocation stage, where the selected branches are already fixed and the remaining task is to divide the source interval across them.

Classification-aware interval allocation

Once the branch curves have been selected, the next step is to determine how the source interval should be divided among them. Consider mm candidate branch curves in Rn\mathbb{R}^{n}. The source interval [0,1)[0,1) is partitioned into subintervals of lengths p1,p2,…,pmp_1,p_2,\ldots,p_m, with pk≥0p_k\geq0 and ∑k=1mpk=1\sum_{k=1}^{m}p_k=1. Branches with pk>0p_k>0 are active source-carrying branches, while branches with pk=0p_k=0 are pruned by the allocation rule. The kkth active subinterval is mapped to branch kk. Let δk\delta_k denote the tube-density measure associated with branch kk, and let qk≜Pr⁡(K^=k∣K=k)=1−ϵkq_k\triangleq\Pr(\widehat{K}=k\mid K=k)=1-\epsilon_k

denote the corresponding probability of correct branch classification, where ϵk\epsilon_k is the misclassification probability for branch kk, that is, ϵk=Pr⁡(K^≠k∣K=k)\epsilon_k = \Pr(\widehat{K} \ne k \mid K = k).

To account jointly for geometric efficiency and branch-identification reliability, we allocate the interval lengths proportionally to the effective branch weights δkqk\delta_k q_k. Under the adopted approximation that branch utility is proportional to the product of tube density and correct-classification probability, the resulting interval allocation is

pk=δkqk∑ℓ=1mδℓqℓ=δk(1−ϵk)∑ℓ=1mδℓ(1−ϵℓ),k=1,…,m.(27)p_k = \frac{\delta_k q_k}{\sum_{\ell=1}^{m}\delta_\ell q_\ell} = \frac{\delta_k(1-\epsilon_k)}{\sum_{\ell=1}^{m}\delta_\ell(1-\epsilon_\ell)}, \qquad k=1,\ldots,m. \tag*{(27)}

Thus, each branch is weighted by both its tube density and its probability of correct classification. In the special case of two branches, the allocation rule becomes

p1=δ1(1−ϵ1)δ1(1−ϵ1)+δ2(1−ϵ2),p2=1−p1.p_1 = \frac{\delta_1(1-\epsilon_1)}{\delta_1(1-\epsilon_1)+\delta_2(1-\epsilon_2)}, \qquad p_2 = 1-p_1.

If the source is parameterized over [0,2π)[0,2\pi), the corresponding split point is θ=2πp1\theta=2\pi p_1.

The rule in Eq. (27) is a classification-aware extension of density-based source allocation. A branch with larger tube density can support finer along-curve resolution and improved noise tolerance. In a multi-curve receiver, however, this geometric advantage is useful only if the branch can also be identified reliably before along-curve decoding. The factor δk\delta_k captures the geometric contribution, and qk=1−ϵkq_k=1-\epsilon_k captures branch-identification reliability. Their product defines the effective branch utility used for proportional allocation. The rule should therefore be viewed as a geometry- and reliability-aware allocation surrogate, not as a finite-SNR optimality claim.

For the case n=4n=4, Figure 5b visualizes the sequential branch-selection procedure. In the m=2m=2 sweep, the second-curve candidate is varied over ϕ\phi, and the plotted interval fraction is P2(2)(ϕ)P_2^{(2)}(\phi); the corresponding starting-branch fraction is P1(2)(ϕ)=1−P2(2)(ϕ)P_1^{(2)}(\phi)=1-P_2^{(2)}(\phi). In the m=3m=3 sweep, the second curve is fixed while the third-curve candidate is varied. Since the interval allocation is recomputed for each candidate, the figure shows both the fixed second-branch fraction P2(3)(ϕ)P_2^{(3)}(\phi) and the third-candidate fraction P3(3)(ϕ)P_3^{(3)}(\phi), with P1(3)(ϕ)=1−P2(3)(ϕ)−P3(3)(ϕ)P_1^{(3)}(\phi)=1-P_2^{(3)}(\phi)-P_3^{(3)}(\phi). Together, these curves illustrate the tradeoff between system SDR, geometric admissibility, and branch-identification reliability that underlies the proposed allocation rule.

Density-only allocation rule

A useful special case of Eq. (27) arises when the branch-classification reliabilities are identical, i.e., ϵ1=ϵ2=⋯=ϵm\epsilon_1=\epsilon_2=\cdots=\epsilon_m. Then the interval allocation reduces to pk=δk∑ℓ=1mδℓp_k=\frac{\delta_k}{\sum_{\ell=1}^{m}\delta_\ell}, k=1,…,mk=1,\ldots,m, which is the density-only allocation rule. In this regime, the interval division depends only on the geometric efficiency of the branch curves, as measured by their tube densities, and no longer distinguishes among branches through classification reliability.

Equal-splitting allocation rule

A useful baseline is obtained by dividing the source interval uniformly across all branches. For a source S∼U(0,1)S\sim U(0,1), the source variance is σS2=1/12\sigma_S^2=1/12. If the source interval is partitioned uniformly, then

Ik=[k−1m,km),k=1,…,m,I_k=\left[\frac{k-1}{m},\frac{k}{m}\right), \qquad k=1,\ldots,m,
Algorithm 1 Sequential C3M candidate-branch selection for $n \in \{4,6\}$ and $m \in \{2,3\}$.
Input: Design case $(n,m)$ and the optimal C3T curve from [15, Table I]
Output: Selected candidate branches $\mathcal{B}_m = \{\text{Curve 1},\ldots,\text{Curve }m\}$, selected parameters $\{\psi_j^\star\}_{j=2}^{m}$, interval allocation $\{p_k\}_{k=1}^{m}$, and final performance metrics

1: Initialize $\mathcal{B}_1$ with the optimal C3T branch
2: Compute $\delta_1$ and $L_1$ for the starting branch
3: for $j = 2$ to $m$ do
4:     $SDR_{\max} \leftarrow -\infty$
5:     if $n = 4$ then
6:         $\Psi \leftarrow$ finite sweep grid for $\phi \in [0^\circ,90^\circ]$
7:     else
8:         $\Psi \leftarrow$ finite sweep grid for $(\theta,\phi) \in [0^\circ,90^\circ]^2$
9:     end if
10:     for each $\psi \in \Psi$ do
11:         Form the candidate branch set $\mathcal{B}_j(\psi) = \mathcal{B}_{j-1} \cup \{\text{Curve }j(\psi)\}$
12:         Compute $\delta_k(\psi)$ and $L_k(\psi)$ for $k = 1,\ldots,j$
13:         Estimate branch reliabilities $q_k(\psi) = 1-\epsilon_k(\psi)$ for $k = 1,\ldots,j$
14:         Compute interval fractions $p_k(\psi) \propto \delta_k(\psi)q_k(\psi)$ for $k = 1,\ldots,j$
15:         Train and evaluate the classifier and branch-conditioned decoders for $\mathcal{B}_j(\psi)$
16:         Compute the end-to-end $SDR(\psi)$
17:         if $SDR(\psi) > SDR_{\max}$ then
18:             $SDR_{\max} \leftarrow SDR(\psi)$
19:             $\psi_j^\star \leftarrow \psi$
20:             $\mathcal{B}_j^\star \leftarrow \mathcal{B}_j(\psi)$
21:             Store the corresponding allocation and performance metrics
22:         end if
23:     end for
24:     $\mathcal{B}_j \leftarrow \mathcal{B}_j^\star$
25: end for
26: Re-evaluate the final selected system $\mathcal{B}_m$ and record the reported SDR, inference accuracy, interval allocation, and branch reliabilities
27: return $\mathcal{B}_m$, $\{\psi_j^\star\}_{j=2}^{m}$, $\{p_k\}_{k=1}^{m}$, and the final performance metrics

so that pk=1/mp_k = 1/m. Conditioned on S∈IkS \in I_k, the source is uniform on IkI_k, and the conditional variance within each interval is

Var⁡(S∣S∈Ik)=112m2.\operatorname{Var}(S \mid S \in I_k) = \frac{1}{12m^2}.

This conditional variance reduction is a property of the source partition and does not by itself imply an end-to-end m2m^2 information-theoretic improvement, since the branch index must still be inferred at the receiver. This equal-splitting construction provides a natural allocation baseline against which the classification-aware allocation rule in Section IV-A can be compared.

Equal splitting also serves as the natural allocation rule when the branches are geometrically identical. In particular, if all branches have the same geometric score and the same branch-reliability score, then a geometry-and classification-aware allocation assigns equal mass to each branch. Thus, in the absence of branch-dependent geometric or classification advantages, the uniform choice pk=1/mp_k = 1/m is the baseline allocation.

Algorithmic selection procedure

The proposed C3M multi-curve selection procedure combines geometric branch construction, classification-aware interval allocation, and branch-conditioned decoder evaluation, see Figure 7 and Algorithm 1. The first branch is fixed to the optimal single-curve C3T design from [15 Table II], and the remaining branches are selected sequentially. For each branch-selection stage, a candidate parameter set Ψ\Psi is defined. For n=4n = 4, Ψ\Psi consists of values of ϕ\phi according to Eq. (25), while for n=6n = 6, Ψ\Psi consists of angular pairs (θ,ϕ)(\theta,\phi) according to Eq. (26). For each candidate parameter choice in Ψ\Psi, the geometric branch quantities and classification reliability are evaluated, the interval fractions are computed, and the decoder performance is assessed. Among the geometrically admissible candidates, the parameter choice yielding the largest end-to-end SDR is selected as the next branch. This procedure is repeated until the desired multi-curve design with mm branches is obtained.

Figure 7. Multi-curve selection procedure for (n,m)∈{(4,2),(4,3),(6,2),(6,3)}(n,m) \in\{(4,2),(4,3),(6,2),(6,3)\}. At each branch-selection stage, a candidate parameter set Ψ\Psi is defined, where Ψ={ϕ}\Psi= \{\phi\} for n=4n = 4 and Ψ={(θ,ϕ)}\Psi= \{(\theta,\phi)\} for n=6n = 6. Geometric quantities and branch-classification reliability determine the interval allocation, while decoder performance is evaluated in parallel. These quantities are then combined to select the parameter choice that maximizes the end-to-end SDR.

For the n=4n = 4 case in Figure 5, the second-curve candidate is swept over ϕ\phi, yielding the selected second branch at ϕ=47∘\phi= 47^\circ for the final NORM C3M configuration. This second curve is then fixed, and a third-curve candidate is swept over ϕ\phi; among the admissible third-curve candidates, the final branch is selected according to the highest system SDR. Figure 5a shows the resulting system SDR and tube-density trends, while Figure 5b shows the corresponding

classification accuracy and interval-fraction curves. The interval-fraction curves are denoted by Pi(m)(ϕ)P_i^{(m)}(\phi), where ii is the candidate-branch index and mm is the number of candidate branches in the sweep. Thus, the m=2m=2 sweep shows P2(2)(ϕ)P_2^{(2)}(\phi), while the m=3m=3 sweep shows both the recomputed second-branch fraction P2(3)(ϕ)P_2^{(3)}(\phi) and the third-candidate fraction P3(3)(ϕ)P_3^{(3)}(\phi).

Although the above description focuses on the cases n∈{4,6}n \in\{4,6\} for concreteness, the same workflow extends directly to higher even dimensions. In general, the radii vector in Rn/2\mathbb{R}^{n/2} is parameterized by n/2−1n/2-1 angular variables in the positive orthant of the unit sphere, and the corresponding candidate set is denoted by Ψ\Psi. For example, the case n=8n=8 requires three angular parameters, so that each candidate branch corresponds to one element of a three-dimensional parameter set Ψ\Psi. Once Ψ\Psi is specified, the remainder of the procedure is unchanged: for each candidate ψ∈Ψ\psi\in\Psi, the geometric branch quantities and classification reliability are evaluated, the interval fractions are computed via the classification-aware allocation rule, the branch-conditioned decoder performance is assessed, and the final branch is selected by maximizing the resulting end-to-end SDR. In this way, the proposed selection procedure provides a unified framework for C3M design in arbitrary even dimensions.

The ambient dimension nn alone does not impose a hard upper bound on the number of admissible branch curves. Without robustness constraints, arbitrarily many distinct curves can be placed in the ambient sphere by allowing the corresponding tubes or inter-curve separations to become sufficiently small. A meaningful design limit appears only after imposing geometric robustness requirements, including a minimum tube radius, a minimum inter-branch separation, or a target branch-classification reliability. The practical multi-curve design problem is therefore a constrained packing problem. The number of branches must be balanced against tube robustness, branch separability, and decoding performance.

Remark (Sequential search complexity and design rationale). One could in principle perform a full unordered two-curve search over ϕ1,ϕ2∈{1∘,…,89∘}\phi_1,\phi_2 \in\{1^\circ,\ldots,89^\circ\} in the four-dimensional case. Excluding identical pairs, this gives the binomial coefficient (892)=89⋅882=3916\binom{89}{2}=\frac{89\cdot88}{2}=3916, i.e., the number of unordered distinct pairs. By contrast, the sequential procedure fixes the first branch to the optimal C3T curve and sweeps only the second branch, reducing the second-curve selection stage to 89 candidates, a factor of approximately 44. This reduction is significant because each candidate is not only a geometric lookup because it requires interval allocation, branch-classifier evaluation, decoder training/evaluation, and end-to-end SDR estimation under Monte Carlo sampling.

Moreover, the cost grows combinatorially when adding further branches. On the same four-dimensional grid, an unordered three-candidate exhaustive search would require (893)=113564\binom{89}{3}=113564 distinct triples before accounting for training seeds, SNR sweeps, and decoder hyperparameter choices.

The sequential strategy is therefore not merely a computational convenience. It is a scalable code-design rule that anchors the construction at the known optimal single-curve C3T design and then measures the marginal benefit of adding each new branch. This makes the procedure reproducible, extensible to higher dimensions and larger branch counts, and aligned with the practical objective of constructing finite-dimensional multi-curve analog codes without an intractable global combinatorial search.

Norm-Inspired Branch-Conditioned Decoder

As motivated in § 2-C, decoding on curved signal geometries should ideally exploit the geometric structure of the underlying branch set. The receiver described in § 3 is hierarchical: after branch classification, decoding is performed conditionally on the estimated branch index. This motivates a branch-dependent decoder rather than a single regressor shared across all branch curves.

The decoder used in this work is motivated by the spectral viewpoint underlying neural operators on Riemannian manifolds, but it is not a full Laplace–Beltrami neural-operator implementation. The implemented decoder does not construct a sampled function on the branch manifold, compute a nearest-point projection onto the curve, or form an explicit Laplace–Beltrami coefficient field. Instead, after the classifier selects branch K^=k\widehat{K}=k, the corresponding branch-conditioned decoder takes the ambient noisy channel vector Y∈Rn\mathbf{Y}\in\mathbb{R}^{n} as its input and estimates the local coordinate on that branch:

Z^=dK^(Y),dk:Rn→[0,1).\widehat{Z}=d_{\widehat{K}}(\mathbf{Y}),\qquad d_k:\mathbb{R}^{n}\to[0,1).

The overall decoder structure is illustrated in Figure 8. The estimated branch K^\widehat{K} selects one decoder from a bank of branch-conditioned decoders. Each decoder applies a Fourier-feature lifting of the received vector, followed by residual and dense layers, and produces a scalar local-coordinate estimate. During training, the estimate is compared with the true local coordinate using the circular Huber loss described in § 5-B.

NORM-inspired branch-conditioned spectral decoder

Figure 8. NORM-inspired branch-conditioned spectral decoder used in the proposed C3M receiver. The curve classifier first estimates the transmitted branch K^=C(Y)\widehat{K}=C(\mathbf{Y}). The estimated branch selects one decoder from a bank of branch-conditioned decoders. The selected decoder takes the received ambient vector Y∈Rn\mathbf{Y}\in\mathbb{R}^{n} as input, applies a Fourier-feature lifting ΓK^(Y)\Gamma_{\widehat{K}}(\mathbf{Y}), processes the lifted representation through residual and dense layers, and outputs a normalized local-coordinate estimate Z^∈[0,1)\widehat{Z}\in[0,1). The source estimate is then obtained using the inverse interval map. During training, the local-coordinate error is converted to an angular residual, wrapped onto the principal interval, and penalized using a robust Huber loss.

Branch-conditioned decoder architecture

For each branch k∈{1,…,m}k\in\{1,\ldots,m\}, training samples are generated according to

Y=γk(z)+N,z∼U[0,1),(28)\mathbf{Y}=\gamma_k(z)+\mathbf{N},\qquad z\sim\mathcal{U}[0,1), \tag*{(28)}

where γk(z)\gamma_k(z) is the kkth branch embedding and N\mathbf{N} is the channel noise. The branch-conditioned decoder is therefore a regression map

dk:Rn→[0,1),Z^=dk(Y).(29)d_k:\mathbb{R}^n \to[0,1), \qquad\widehat{Z}=d_k(\mathbf{Y}). \tag*{(29)}

The term branch-conditioned means that a separate decoder is trained for each candidate branch and that the decoder indexed by the detected branch K^\widehat{K} is used at inference time. Thus, the overall receiver may be viewed as a bank of decoders {d1,d2,…,dm}\{d_1,d_2,\ldots,d_m\}, together with the branch classifier from § 3-C. Given K^=k\widehat{K}=k, the receiver applies dkd_k to the same received vector Y\mathbf{Y} and then maps the estimated local coordinate back to the corresponding source interval.

In the implementation used here, each branch decoder begins with a Fourier-feature lifting of the ambient input. Specifically, for branch kk, the received vector is mapped to

Γk(Y)=[sin⁡(2πYBk),cos⁡(2πYBk)],(30)\Gamma_k(\mathbf{Y})=\left[\sin(2\pi\mathbf{Y}\mathbf{B}_k),\cos(2\pi\mathbf{Y}\mathbf{B}_k)\right], \tag*{(30)}

where Bk\mathbf{B}_k is the Fourier-feature projection matrix used by the branch-conditioned decoder. The lifted representation is then processed by a neural network FkF_k consisting of residual and dense layers, followed by a scalar readout:

Z^=dk(Y)=qk(Fk(Γk(Y))).(31)\widehat{Z}=d_k(\mathbf{Y})=q_k\left(F_k\left(\Gamma_k(\mathbf{Y})\right)\right). \tag*{(31)}

The readout qkq_k produces the normalized local-coordinate estimate Z^∈[0,1)\widehat{Z}\in[0,1). Thus, decoding is performed directly from the ambient received vector through a branch-conditioned Fourier-feature representation, rather than through a discretized likelihood field or an explicit Laplace–Beltrami expansion.

This architecture is NORM-inspired because it uses spectral features to learn branch-dependent maps associated with geometric signal sets [23]. In the present implementation, the spectral representation is obtained by Fourier-feature lifting of the received ambient vector and by training a separate decoder for each branch. After branch classification, the selected decoder only has to learn the inverse map associated with one branch geometry. Thus, the method is a branch-conditioned spectral decoder inspired by NORM, rather than a full Laplace–Beltrami neural-operator implementation. The implemented receiver architecture is kept fixed across all reported C3M experiments: the classifier maps y∈Rn\mathbf{y}\in\mathbb{R}^n to a candidate branch, and the selected branch-conditioned decoder applies the Fourier-feature lifting in Eq. (30) before residual and dense layers estimate the local coordinate. The decoder hyperparameters are summarized in Table IV, and training uses the circular Huber loss introduced in § 5-B.

HyperparameterValues
Epochs100
Batch size128
Learning rate (LR)1e−41\mathrm{e}{-4}
Minimum LR1e−61\mathrm{e}{-6}
LR schedulerReduceLROnPlateau
LR patience70
LR factor0.5
Loss functionHuber
OptimizerAdamW
Weight Decay1e−41\mathrm{e}{-4}
Number of NORM training examples per curve200,000
Number of NORM evaluation examples per curve50,000
Number of training examples for curve classifier15,000
Number of evaluation examples for curve classifier15,000

Table IV. HYPERPARAMETERS OF THE NORM-INSPIRED BRANCH-CONDITIONED SPECTRAL DECODER.

Huber loss on the angular error

For neural receivers trained for regression, the mean-squared error (MSE), or ℓ2\ell_2 loss, is often the default choice because it is smooth and performs well when residuals are approximately Gaussian. In the present setting, however, large residuals can be produced by Gaussian-mixture impulses, heavy-tailed perturbations, branch-confusion events, or difficult regions of the branch manifold. Since MSE grows quadratically with residual magnitude, such samples can dominate the optimization and destabilize training.

Robust alternatives such as the mean absolute error (MAE), or ℓ1\ell_1 loss, reduce the influence of large residuals, but their nonsmooth behavior at the origin is less attractive for gradient-based optimization. The Huber loss combines a quadratic penalty for small errors with a linear penalty for large errors. It therefore preserves local smoothness near the optimum while reducing sensitivity to outliers. This makes it well matched to branch-conditioned local-coordinate regression under impulsive or heavy-tailed noise.

The decoder estimates a normalized local coordinate in [0,1)[0,1), which is mapped to an angular coordinate on the selected branch. The regression residual should therefore respect the periodic geometry of the branch parameterization. Let Δ\Delta denote the difference between the predicted and target angular coordinates. We define the circular residual by projecting this difference onto the unit circle and taking the principal angle,

Δcirc=atan2⁡(sin⁡(Δ),cos⁡(Δ)).(32)\Delta_{\mathrm{circ}} = \operatorname{atan2}(\sin(\Delta), \cos(\Delta)). \tag*{(32)}

This is equivalent, up to endpoint convention, to wrapping Δ\Delta into the principal interval (−π,π](-\pi,\pi]. The angular error magnitude is then

δ=∣Δcirc∣.\delta= \lvert\Delta_{\mathrm{circ}}\rvert.

We define the Huber loss on this circular residual as

Lκ(δ)={12δ2,δ≤κ,κ(δ−12κ),δ>κ,(33)L_{\kappa}(\delta) = \begin{cases} \frac{1}{2}\delta^2, & \delta\le\kappa, \\ \kappa\left(\delta- \frac{1}{2}\kappa\right), & \delta> \kappa, \end{cases} \tag*{(33)}

where κ>0\kappa> 0 is the transition threshold. In the experiments, we use κ=1\kappa= 1. Equivalently, this special case can be implemented through q=min⁡(δ,1)q = \min(\delta,1), ℓ=δ−q\ell= \delta- q, and

L1(δ)=12q2+ℓ.L_1(\delta) = \frac{1}{2}q^2 + \ell.

The constant offset in Eq. (33) ensures continuity at the transition point, and the loss is averaged over the training batch.

Together, circular residual wrapping and Huber regression define a robust training objective for branch-conditioned local-coordinate estimation. This objective is well matched to C3M because large residuals can arise from branch-confusion events as well as from impulsive or heavy-tailed noise realizations. The Gaussian-mixture noise model in Eq. (19) also connects directly to Huber’s robust-estimation viewpoint [35]. Most observations follow a nominal distribution, while a small fraction behave as outliers. This motivates the use of robust loss functions such as the Huber loss [36].

Analysis

This section develops the analytical model and reference quantities used for the proposed C3M construction. We first characterize the induced output distribution under fixed isotropic AWGN and use the resulting posterior representation to define MAP and MMSE estimators. We then express the mutual information induced by the C3M input distribution and relate it to rate–distortion theory and channel-capacity references.

The remainder of the section studies the main multi-curve effects. These include branch-classification reliability, capacity-surrogate allocation, geometry-only allocation, end-to-end and branch-known SDR references, non-uniform source allocation, geometric branch-selection behavior, and heuristic upper bounds on the number of robustly admissible branch curves.

Induced output distribution

Under the C3M encoder, the channel input is a random point on one of the mm branch curves. In the main Gaussian-noise model considered in this paper, the channel output is Y=X+N\mathbf{Y}=\mathbf{X}+\mathbf{N}, where N∼N(0,σn2In)\mathbf{N}\sim\mathcal{N}(0,\sigma_n^2\mathbf{I}_n), with σn2\sigma_n^2 fixed independently of the branch index, radius parameters, and candidate curve. Since the encoder is memoryless and generally nonlinear, the proposed C3M construction induces a non-Gaussian mixture distribution at the receiver even under isotropic Gaussian channel noise.

Let

ϕn(y;μ,σn2In)=1(2πσn2)n/2exp⁡(−∥y−μ∥22σn2)(34)\phi_n(\mathbf{y};\boldsymbol{\mu},\sigma_n^2\mathbf{I}_n)=\frac{1}{(2\pi\sigma_n^2)^{n/2}}\exp\left(-\frac{\lVert\mathbf{y}-\boldsymbol{\mu}\rVert^2}{2\sigma_n^2}\right) \tag*{(34)}

denote the nn-dimensional isotropic Gaussian density with mean μ\boldsymbol{\mu} and covariance σn2In\sigma_n^2\mathbf{I}_n. The channel law conditioned on the branch index K=kK=k and local coordinate Z=zZ=z is therefore pY∣K,Z(y∣k,z)=ϕn(y;γk(z),σn2In)p_{Y\mid K,Z}(\mathbf{y}\mid k,z)=\phi_n(\mathbf{y};\boldsymbol{\gamma}_k(z),\sigma_n^2\mathbf{I}_n). Because S∼U(0,1)S\sim\mathcal{U}(0,1) and the local normalization on each branch interval is affine, the conditional distribution of the local coordinate satisfies pZ∣K(z∣k)=1p_{Z\mid K}(z\mid k)=1, z∈[0,1)z\in[0,1), for every k=1,…,mk=1,\ldots,m and zero otherwise.

For example, in the n=4n=4 case, a branch with two radius pairs may be written as

γk(z)=[rk,1cos⁡(2πωk,1z)rk,1sin⁡(2πωk,1z)rk,2cos⁡(2πωk,2z)rk,2sin⁡(2πωk,2z)],z∈[0,1).(35)\boldsymbol{\gamma}_k(z)= \begin{bmatrix} r_{k,1}\cos(2\pi\omega_{k,1}z)\\ r_{k,1}\sin(2\pi\omega_{k,1}z)\\ r_{k,2}\cos(2\pi\omega_{k,2}z)\\ r_{k,2}\sin(2\pi\omega_{k,2}z) \end{bmatrix}, \qquad z\in[0,1). \tag*{(35)}

Writing y=[y1,y2,y3,y4]T\mathbf{y}=[y_1,y_2,y_3,y_4]^{T}, the fixed isotropic AWGN model implies that the received coordinates are conditionally independent given (K,Z)=(k,z)(K,Z)=(k,z). Hence,

pY∣K,Z(y∣k,z)=∏i=1412πσn2exp⁡(−(yi−γk,i(z))22σn2)=1(2πσn2)2exp⁡(−∥y−γk(z)∥22σn2).(36)p_{Y|K,Z}(\mathbf{y}|k,z)=\prod_{i=1}^{4}\frac{1}{\sqrt{2\pi\sigma_n^2}}\exp\left(-\frac{(y_i-\gamma_{k,i}(z))^2}{2\sigma_n^2}\right)=\frac{1}{(2\pi\sigma_n^2)^2}\exp\left(-\frac{\lVert\mathbf{y}-\boldsymbol{\gamma}_k(z)\rVert^2}{2\sigma_n^2}\right). \tag*{(36)}

After marginalizing over the latent coordinate ZZ, however, the received coordinates are generally dependent because they share the same underlying curve parameter.

Thus, conditioned on K=kK=k, the source induces a uniform distribution on the normalized local parameter ZZ along branch kk. This does not imply that the ambient coordinates of γk(Z)\boldsymbol{\gamma}_k(Z) are uniformly distributed; rather, the uniformity holds in the branch parameter itself, and for constant-speed parametrizations equivalently in normalized arc length. In particular, the coordinate marginals are generally sinusoidal transformations of a uniform parameter and are therefore non-uniform.

Conditioning only on the branch index gives pY∣K(y∣k)=∫01ϕn(y;γk(u),σn2In) dup_{Y|K}(\mathbf{y}|k)=\int_0^1\phi_n(\mathbf{y};\boldsymbol{\gamma}_k(u),\sigma_n^2\mathbf{I}_n)\,du, where u∈[0,1)u\in[0,1) is the normalized local coordinate on branch kk. Averaging over the branch index yields the unconditional output density

pY(y)=∑k=1mpkpY∣K(y∣k)=∑k=1mpk∫01ϕn(y;γk(u),σn2In) du.(37)p_Y(\mathbf{y})=\sum_{k=1}^{m}p_kp_{Y|K}(\mathbf{y}|k)=\sum_{k=1}^{m}p_k\int_0^1\phi_n(\mathbf{y};\boldsymbol{\gamma}_k(u),\sigma_n^2\mathbf{I}_n)\,du. \tag*{(37)}

Thus, the receiver distribution is a mixture over the mm branch curves, with weights given by the interval-allocation probabilities pkp_k. This representation makes explicit the probabilistic structure seen by the curve classifier and provides the basis for the information-theoretic analysis that follows. A corresponding joint representation is pK,Z,Y(k,z,y)=pkpZ∣K(z∣k)pY∣K,Z(y∣k,z)=pkϕn(y;γk(z),σn2In)p_{K,Z,Y}(k,z,\mathbf{y})=p_kp_{Z|K}(z|k)p_{Y|K,Z}(\mathbf{y}|k,z)=p_k\phi_n(\mathbf{y};\boldsymbol{\gamma}_k(z),\sigma_n^2\mathbf{I}_n), for k=1,…,mk=1,\ldots,m and z∈[0,1)z\in[0,1), where we used the fact that Z∣K=kZ|K=k is uniform on [0,1)[0,1). For a single branch, the induced output density reduces to the corresponding one-manifold Gaussian mixture obtained by averaging the Gaussian channel law along that branch.

Remark (Unit-circle output law). As a simple special case, consider a single unit-circle branch in R2\mathbb{R}^2 with γ(a)=[cos⁡a,sin⁡a]T\boldsymbol{\gamma}(a)=[\cos a,\sin a]^T, with angular parameter a∼U[0,2π)a\sim U[0,2\pi), and isotropic AWGN with variance σn2\sigma_n^2 per coordinate. Writing y=[y1,y2]T\mathbf{y}=[y_1,y_2]^T, the induced output density is

pY(y)=12π∫02π12πσn2exp⁡(−∥y−γ(a)∥22σn2) da=12πσn2exp⁡(−∥y∥2+12σn2)I0(∥y∥σn2),(38)p_Y(\mathbf{y})=\frac{1}{2\pi}\int_0^{2\pi}\frac{1}{2\pi\sigma_n^2}\exp\left(-\frac{\lVert\mathbf{y}-\boldsymbol{\gamma}(a)\rVert^2}{2\sigma_n^2}\right)\,da=\frac{1}{2\pi\sigma_n^2}\exp\left(-\frac{\lVert\mathbf{y}\rVert^2+1}{2\sigma_n^2}\right)I_0\left(\frac{\lVert\mathbf{y}\rVert}{\sigma_n^2}\right), \tag*{(38)}

where I0(⋅)I_0(\cdot) is the modified Bessel function of the first kind. This example illustrates that even a single smooth analog curve induces a non-Gaussian output law. Moreover, although the two output coordinates are uncorrelated by symmetry in this example, they are not independent, since both are coupled through the same latent curve parameter. ■

Posterior representation and Bayesian estimators

By Bayes’ rule, the posterior density of the branch index and local coordinate given the observation y\mathbf{y} is

pK,Z∣Y(k,z∣y)=pkϕn(y;γk(z),σn2In)pY(y),(39)p_{K,Z|Y}(k,z|\mathbf{y})=\frac{p_k\phi_n(\mathbf{y};\boldsymbol{\gamma}_k(z),\sigma_n^2\mathbf{I}_n)}{p_Y(\mathbf{y})}, \tag*{(39)}

k=1,…,m,z∈[0,1)k=1,\ldots,m,\quad z\in[0,1). Here pY(y)p_{\mathbf{Y}}(\mathbf{y}) is the marginal likelihood of the observation y\mathbf{y}, i.e., the normalizing constant that ensures the posterior density integrates to one over all admissible branch-index and local-coordinate pairs. Equivalently, for fixed y\mathbf{y},

pK,Z∣Y(k,z∣y)∝pkpY∣K,Z(y∣k,z),(40)p_{K,Z\mid Y}(k,z\mid\mathbf{y}) \propto p_k p_{Y\mid K,Z}(\mathbf{y}\mid k,z), \tag*{(40)}

so the posterior is proportional to likelihood times prior, with the branch probabilities {pk}\{p_k\} serving as priors over the discrete branch index. The normalization is over both the discrete branch index and the continuous local coordinate,

pY(y)=∑k=1mpk∫01pY∣K,Z(y∣k,z) dz.(41)p_Y(\mathbf{y}) = \sum_{k=1}^{m} p_k \int_0^1 p_{Y\mid K,Z}(\mathbf{y}\mid k,z)\,dz. \tag*{(41)}

This posterior representation makes explicit the Bayesian inference problem induced by the C3M encoder. In particular, the MAP estimator is

(K^MAP,Z^MAP)=arg⁡max⁡k, z∈[0,1) pK,Z∣Y(k,z∣y).(42)(\widehat{K}_{\mathrm{MAP}},\widehat{Z}_{\mathrm{MAP}}) = \underset{k,\ z\in[0,1)}{\arg\max}\ p_{K,Z\mid Y}(k,z\mid\mathbf{y}). \tag*{(42)}

For fixed isotropic AWGN, this is equivalent to minimizing a branch-dependent Euclidean-distance criterion with the branch probabilities {pk}\{p_k\} entering as prior weights:

(K^MAP,Z^MAP)=arg⁡min⁡k, z∈[0,1) {∥y−γk(z)∥2−2σn2log⁡pk}.(43)(\widehat{K}_{\mathrm{MAP}},\widehat{Z}_{\mathrm{MAP}}) = \underset{k,\ z\in[0,1)}{\arg\min}\ \left\{\lVert\mathbf{y}-\gamma_k(z)\rVert^2 - 2\sigma_n^2\log p_k\right\}. \tag*{(43)}

When the branch probabilities are equal, the prior term is constant across branches and the MAP rule reduces to nearest-curve decoding in Euclidean distance.

The MMSE source estimate is

S^MMSE(y)=E[S∣Y=y].(44)\widehat{S}_{\mathrm{MMSE}}(\mathbf{y}) = \mathbb{E}[S\mid Y=\mathbf{y}]. \tag*{(44)}

The estimator in Eq. (44) is optimal in the mean-squared-error sense, i.e., it minimizes E[(S−g(Y))2]\mathbb{E}[(S-g(Y))^2] over all measurable estimators g(Y)g(Y). Its optimality under mean-squared error is recalled in Appendix A.

Since SS is in one-to-one correspondence with (K,Z)(K,Z) through the interval map ff, both estimators may be interpreted equivalently in the source domain or in the branch-coordinate domain. For a Gaussian model with general covariance Σn\boldsymbol{\Sigma}_n, the same posterior representation applies after replacing σn2In\sigma_n^2\mathbf{I}_n by Σn\boldsymbol{\Sigma}_n, and thus the MAP distance term becomes the Mahalanobis distance (y−γk(z))TΣn−1(y−γk(z))(\mathbf{y}-\gamma_k(z))^{\mathsf{T}}\boldsymbol{\Sigma}_n^{-1}(\mathbf{y}-\gamma_k(z)).

Mutual information induced by C3M

Since the source symbol SS is in one-to-one correspondence with the pair (K,Z)(K,Z) through the bijective interval map ff, the mutual information between the source and the received vector satisfies

I(S;Y)=I(K,Z;Y)=I(K;Y)+I(Z;Y∣K).(45)I(S;Y) = I(K,Z;Y) = I(K;Y) + I(Z;Y\mid K). \tag*{(45)}

The first term, I(K;Y)I(K;Y), quantifies the coarse branch-identification information, whereas the second term, I(Z;Y∣K)I(Z;Y\mid K), quantifies the fine along-curve information once the transmitted branch is known. This decomposition mirrors the proposed receiver structure, which first performs curve classification and then applies a branch-conditioned decoder.

For the fixed isotropic AWGN channel, the mutual information induced by the C3M encoder may also be written as I(S;Y)=h(Y)−h(N)I(S;Y)=h(Y)-h(N), where h(⋅)h(\cdot) denotes differential entropy. Since N∼N(0,σn2In)N\sim\mathcal{N}(0,\sigma_n^2 I_n), the noise entropy is

h(N)=n2log⁡(2πeσn2).h(N)=\frac{n}{2}\log(2\pi e\sigma_n^2).

Remark (Diagonal Gaussian extension). As a Gaussian extension, if the n=4n=4 noise covariance is diagonal with unequal variances across coordinate pairs, Σn=diag⁡(σ12,σ12,σ22,σ22)\Sigma_n=\operatorname{diag}(\sigma_1^2,\sigma_1^2,\sigma_2^2,\sigma_2^2), then the noise entropy becomes

h(N)=12log⁡((2πe)4det⁡Σn)=2log⁡(2πe)+2log⁡σ1+2log⁡σ2.(46)h(N)=\frac{1}{2}\log\left((2\pi e)^4\det\Sigma_n\right)=2\log(2\pi e)+2\log\sigma_1+2\log\sigma_2. \tag*{(46)}

Equivalently, h(N)=log⁡(2πeσ12)+log⁡(2πeσ22)h(N)=\log(2\pi e\sigma_1^2)+\log(2\pi e\sigma_2^2). ■

In either case, the remaining task is to evaluate the output entropy induced by the mixture density in Eq. (37).

Equivalently,

I(S;Y)=∑k=1mpk∫01∫RnpY∣K,Z(y∣k,z)log⁡pY∣K,Z(y∣k,z)pY(y) dy dz.(47)I(S;Y)=\sum_{k=1}^{m}p_k\int_0^1\int_{\mathbb{R}^n}p_{Y\mid K,Z}(y\mid k,z)\log\frac{p_{Y\mid K,Z}(y\mid k,z)}{p_Y(y)}\,dy\,dz. \tag*{(47)}

Unless otherwise stated, logarithms in this subsection are natural and the resulting mutual information is measured in nats; conversion to bits is obtained by dividing by log⁡2\log2.

For numerical evaluation, the mutual information in Eq. (47) may be approximated by Monte Carlo sampling from the joint distribution pK,Z,Yp_{K,Z,Y}. Specifically, one may draw samples (Ki,Zi,Yi)(K_i,Z_i,Y_i) from the encoder–channel model and estimate I(S;Y)I(S;Y) by averaging the corresponding log-likelihood ratios log⁡pY∣K,Z(Yi∣Ki,Zi)−log⁡pY(Yi)\log p_{Y\mid K,Z}(Y_i\mid K_i,Z_i)-\log p_Y(Y_i).

Remark (Single-branch continuous-parameter interpretation). Conditioned on branch K=kK=k, the mutual information between the local coordinate and the received vector can also be written in entropy form as

I(Z;Y∣K=k)=h(Y∣K=k)−h(N),(48)I(Z;Y\mid K=k)=h(Y\mid K=k)-h(N), \tag*{(48)}

where

h(Y∣K=k)=−∫RnpY∣K(y∣k)log⁡pY∣K(y∣k) dy.(49)h(Y\mid K=k)=-\int_{\mathbb{R}^n}p_{Y\mid K}(y\mid k)\log p_{Y\mid K}(y\mid k)\,dy. \tag*{(49)}

For fixed isotropic AWGN, h(N)=n2log⁡(2πeσn2)h(N)=\frac{n}{2}\log(2\pi e\sigma_n^2), so for n=4n=4 this becomes h(N)=2log⁡(2πeσn2)h(N)=2\log(2\pi e\sigma_n^2). The output entropy term must generally be evaluated numerically from the branch-induced mixture density. In the small-noise limit, the posterior over the continuous curve parameter concentrates and I(Z;Y∣K=k)I(Z;Y\mid K=k) can grow without bound; in the high-noise limit, the observation becomes independent of ZZ and I(Z;Y∣K=k)→0I(Z;Y\mid K=k)\to0. ■

Remark (Scalar projection bound). As a scalar projection example, let Ysc=X+NY_{\mathrm{sc}}=X+N, where X=cos⁡(2πS)X=\cos(2\pi S), S∼U(0,1)S\sim\mathcal{U}(0,1), and N∼N(0,σ2)N\sim\mathcal{N}(0,\sigma^2) is independent of SS. Equivalently, with A=2πSA=2\pi S, one has A∼U[0,2π)A\sim\mathcal{U}[0,2\pi) and X=cos⁡AX=\cos A. Since YscY_{\mathrm{sc}} depends on SS only through XX, I(S;Ysc)=I(X;Ysc)=h(Ysc)−12log⁡(2πeσ2)I(S;Y_{\mathrm{sc}})=I(X;Y_{\mathrm{sc}})=h(Y_{\mathrm{sc}})-\frac{1}{2}\log(2\pi e\sigma^2). This remark uses natural logarithms; conversion to bits is obtained by dividing by log⁡2\log2.

The induced scalar output density is

pYsc(y)=12π∫02πϕ1(y;cos⁡a,σ2) da.(50)p_{Y_{\mathrm{sc}}}(y)=\frac{1}{2\pi}\int_0^{2\pi}\phi_1(y;\cos a,\sigma^2)\,da. \tag*{(50)}

Thus, equivalently,

I(S;Ysc)=−∫RpYsc(y)log⁡pYsc(y) dy−12log⁡(2πeσ2),(51)I(S;Y_{\mathrm{sc}})=-\int_{\mathbb{R}}p_{Y_{\mathrm{sc}}}(y)\log p_{Y_{\mathrm{sc}}}(y)\,dy-\frac{1}{2}\log(2\pi e\sigma^2), \tag*{(51)}

where the entropy integral generally has to be evaluated numerically. The scalar entropy power inequality [37, 38, 39] gives

e2h(Ysc)≥e2h(X)+e2h(N).(52)e^{2h(Y_{\mathrm{sc}})} \ge e^{2h(X)} + e^{2h(N)}. \tag*{(52)}

The random variable X=cos⁡(2πS)X=\cos(2\pi S) has the arcsine distribution on [−1,1][-1,1] and differential entropy h(X)=log⁡(π/2)h(X)=\log(\pi/2), while h(N)=12log⁡(2πeσ2)h(N)=\frac{1}{2}\log(2\pi e\sigma^2). Substituting these quantities into Eq. (52) yields

I(S;Ysc)≥12log⁡(1+π8eσ2).(53)I(S;Y_{\mathrm{sc}}) \ge\frac{1}{2}\log\left(1+\frac{\pi}{8e\sigma^2}\right). \tag*{(53)}

In this scalar example, the limiting behavior is also explicit. As σ2→0\sigma^2\to0, the output density converges to the arcsine law induced by X=cos⁡(2πS)X=\cos(2\pi S), whose differential entropy is h(X)=log⁡(π/2)h(X)=\log(\pi/2). Hence

I(S;Ysc)=h(Ysc)−h(N)≈log⁡(π2)−12log⁡(2πeσ2),(54)I(S;Y_{\mathrm{sc}})=h(Y_{\mathrm{sc}})-h(N)\approx\log\left(\frac{\pi}{2}\right)-\frac{1}{2}\log(2\pi e\sigma^2), \tag*{(54)}

which diverges as −12log⁡σ2-\frac{1}{2}\log\sigma^2 in the vanishing-noise limit. This does not contradict the two-to-one nature of the cosine map: the noiseless observation does not fully determine SS, but it does determine the continuous variable X=cos⁡(2πS)X=\cos(2\pi S) exactly. In the high-noise limit, the mutual information vanishes, with first-order behavior

I(S;Ysc)=Var⁡(X)2σ2+o(σ−2)=14σ2+o(σ−2),(55)I(S;Y_{\mathrm{sc}})=\frac{\operatorname{Var}(X)}{2\sigma^2}+o(\sigma^{-2})=\frac{1}{4\sigma^2}+o(\sigma^{-2}), \tag*{(55)}

since Var⁡(cos⁡(2πS))=1/2\operatorname{Var}(\cos(2\pi S))=1/2. This bound and these limiting expressions are only scalar projection examples, while the C3M analysis uses the full vector output distribution in Eq. (37).

It is important to distinguish this quantity from channel capacity. The capacity of the Gaussian channel depends only on the channel law and the input power constraint, whereas Eqs. (45)–(47) characterize the mutual information induced by the specific distribution generated by the proposed C3M encoder. In what follows, we therefore refer to this quantity as the mutual information induced by the C3M encoder, or equivalently, the achievable information rate under the C3M input distribution.

Rate–distortion and channel-capacity references

The information-theoretic reference for distortion-limited communication is obtained by combining the source rate–distortion function with the channel capacity constraint [40, 41]. Since one source symbol is transmitted using one nn-dimensional channel use, Shannon separation implies that reliable communication with distortion DD is achievable whenever R(D)≤CnR(D)\le C_n, where R(D)R(D) is measured in bits per source symbol and CnC_n is the channel capacity per nn-dimensional channel use.

For the fixed isotropic AWGN channel with total input power constraint P=E[∥X∥2]=1P=\mathbb{E}[\lVert X\rVert^2]=1 and noise covariance σn2In\sigma_n^2 I_n, the capacity of one nn-dimensional channel use is

Cn=n2log⁡2(1+Pnσn2)=n2log⁡2(1+SNRn),(56)C_n=\frac{n}{2}\log_2\left(1+\frac{P}{n\sigma_n^2}\right)=\frac{n}{2}\log_2\left(1+\frac{\mathrm{SNR}}{n}\right), \tag*{(56)}

where, under the normalization adopted in § 3, SNR⁡=P/σn2=1/σn2\operatorname{SNR}=P/\sigma_n^2=1/\sigma_n^2. This capacity serves as an AWGN capacity reference against which the mutual information induced by the C3M input distribution can be compared. For the four-dimensional case, the fixed isotropic AWGN capacity reference specializes to

C4=2log⁡2(1+P4σn2).C_4=2\log_2\left(1+\frac{P}{4\sigma_n^2}\right).

Under the normalization P=1P=1, this becomes

C4=2log⁡2(1+14σn2)=2log⁡2(1+SNR⁡4).C_4=2\log_2\left(1+\frac{1}{4\sigma_n^2}\right)=2\log_2\left(1+\frac{\operatorname{SNR}}{4}\right).

Equivalently, if σ2=4σn2\sigma^2=4\sigma_n^2 denotes the total noise variance across the four channel dimensions, then C4=2log⁡2(1+1/σ2)C_4=2\log_2(1+1/\sigma^2).

Because the C3M encoder e≜g∘fe\triangleq g\circ f is deterministic and bijective between the source interval [0,1)[0,1) and the admissible code set X\mathcal{X}, the source symbol SS and the transmitted vector X=e(S)\mathbf{X}=e(S) are in one-to-one correspondence. Consequently, I(S;Y)=I(X;Y)I(S;\mathbf{Y})=I(\mathbf{X};\mathbf{Y}), since X=e(S)\mathbf{X}=e(S) is an invertible reparameterization of SS on the C3M codebook. This distinction matters because nonlinearity itself does not reduce mutual information. Rather, the induced input distribution and its support determine how close the mapping can come to the AWGN capacity reference. Under an average-power constraint, the capacity-achieving input for the fixed isotropic AWGN channel is an nn-dimensional Gaussian distribution, whereas the C3M input distribution is supported on a union of one-dimensional branch curves in Rn\mathbb{R}^n. Thus, CnC_n should be interpreted as an AWGN capacity reference, while I(S;Y)I(S;\mathbf{Y}) is the achievable information rate induced by the specific C3M source–channel mapping.

For a Gaussian source with variance σS2\sigma_S^2, the rate–distortion function under MSE distortion is

RG(D)=12log⁡2(σS2D).(57)R_G(D)=\frac{1}{2}\log_2\left(\frac{\sigma_S^2}{D}\right). \tag*{(57)}

Equating Eq. (57) with the channel-capacity reference CnC_n yields the Gaussian-source reference distortion

DG∗=σS22−2Cn.(58)D_G^*=\sigma_S^2 2^{-2C_n}. \tag*{(58)}

For a uniform source, the true rate–distortion function is not given by Eq. (57). At high rate, however, the uniform-source distortion reference is lower than the Gaussian-source reference by the asymptotic shaping factor 6/(πe)6/(\pi e):

DUasy=DG∗6πe=σS22−2Cn6πe.(59)D_U^{\mathrm{asy}}=D_G^*\frac{6}{\pi e}=\sigma_S^2 2^{-2C_n}\frac{6}{\pi e}. \tag*{(59)}

Equivalently, the corresponding high-rate SDR reference is larger by the factor πe/6\pi e/6, or approximately 1.53 dB. This refined reference is therefore best interpreted as a high-rate uniform-source reference, not as an exact low-SNR rate–distortion limit.

The branch decomposition can also be used to define conditional diagnostic references. If the true branch index KK were provided to the decoder, then a conditional branch-known rate constraint could be written as

∑k=1mpkRk(Dk)≤Cn.\sum_{k=1}^{m}p_kR_k(D_k)\le C_n.

where Rk(Dk)R_k(D_k) is the rate–distortion function associated with the source interval assigned to branch kk. This is a branch-known reference, not an ordinary end-to-end reference for the original source. In the practical C3M receiver, the branch index must be inferred from YY, and any information needed to identify the branch is part of the end-to-end source description. The corresponding end-to-end and branch-known SDR references are separated explicitly in Section VI-H.

Structured-code constructions provide an important asymptotic reference for Gaussian-source settings, showing that separation-based references can be approached arbitrarily closely in high-dimensional regimes [42]. The C3M construction studied here is instead finite-dimensional and low-blocklength. Its performance is governed by curve geometry, branch separation, branch-classification reliability, and decoder accuracy rather than by claims of asymptotic optimality.

It should be emphasized that the separation-based rate–distortion and capacity references are asymptotic information-theoretic limits. In practical finite-blocklength and low-SNR regimes, separated systems may operate far from these limits [11], whereas analog joint source–channel mappings can provide graceful degradation under changing channel conditions [12]. Throughout the numerical comparisons in this paper, these references are evaluated under the fixed isotropic AWGN model. For a Gaussian channel model with general covariance Σn\Sigma_n, the corresponding capacity and entropy expressions are obtained by replacing σn2In\sigma_n^2\mathbf{I}_n with Σn\Sigma_n.

Classification reliability

To interpret the branch-classification reliabilities appearing in the interval-allocation rule in Eq. (27), it is useful to relate them to the local separation geometry of the branch set. Under fixed isotropic AWGN, a convenient nearest-neighbor approximation is obtained from the Euclidean inter-branch distances

dkk′≜min⁡z,z′∈[0,1)∥γk(z)−γk′(z′)∥.(60)d_{kk'} \triangleq\min_{z,z'\in[0,1)} \left\lVert\gamma_k(z)-\gamma_{k'}(z') \right\rVert. \tag*{(60)}

In the moderate- to high-SNR regime, the conditional error probability for branch kk is then dominated by nearby competing branches and may be approximated as

Pr⁡(K^≠k∣K=k)≈∑k′≠kQ(dkk′2σn).(61)\Pr(\widehat{K}\ne k\mid K=k) \approx\sum_{k'\ne k} Q\left(\frac{d_{kk'}}{2\sigma_n}\right). \tag*{(61)}

Thus, the quantities qk=1−ϵkq_k=1-\epsilon_k used in the interval-allocation rule are closely tied to the local separation structure of the branch set: larger inter-branch distances improve classification reliability, whereas closely spaced branches become the dominant source of confusion. For a Gaussian extension with general covariance Σn\Sigma_n, the Euclidean distance in Eq. (60) is replaced by the corresponding Mahalanobis distance. This reliability viewpoint will also be useful in interpreting the SDR behavior of multi-curve designs under different interval-allocation rules.

The nearest-neighbor approximation further suggests that branch-classification reliability may be strongly asymmetric across the branch set. Although the multi-curve design is chosen to maximize global separation, the conditional error for branch kk is controlled mainly by its closest competitors. Hence branches with multiple nearby neighbors, or with smaller admissible tube radii, may exhibit substantially lower values of

qk=Pr⁡(K^=k∣K=k)q_k=\Pr(\widehat{K}=k\mid K=k)

than more isolated branches. This phenomenon is particularly relevant in low-dimensional settings, where geometric packing constraints limit the number of strongly separated branches. In such cases, one or more curves may become bottlenecks for classification performance, even when the overall design remains near-optimal in a global sense.

A useful conceptual decomposition of end-to-end distortion separates branch-index errors from within-branch reconstruction error. In particular, let EKE_K denote the event that the branch index is decoded incorrectly. Then the total mean-squared error may be decomposed as

E[(S−S^)2]=Pr⁡(EK)E[(S−S^)2∣EK]+(1−Pr⁡(EK))E[(S−S^)2∣EKc].(62)\mathbb{E}\left[(S-\hat{S})^2\right] = \Pr(E_K)\mathbb{E}\left[(S-\hat{S})^2 \mid E_K\right] + \left(1-\Pr(E_K)\right)\mathbb{E}\left[(S-\hat{S})^2 \mid E_K^c\right]. \tag*{(62)}

This viewpoint clarifies why both inter-branch separation and along-branch resolution matter simultaneously in the multi-curve setting.

A more refined decomposition is obtained by conditioning on the transmitted branch. Let DkcorrD_k^{\mathrm{corr}} denote the conditional distortion when branch kk is correctly identified, and let DkerrD_k^{\mathrm{err}} denote the corresponding distortion under branch misclassification. Then

Dtotal=∑k=1mpk[(1−ϵk)Dkcorr+ϵkDkerr].(63)D_{\mathrm{total}} = \sum_{k=1}^{m} p_k\left[(1-\epsilon_k)D_k^{\mathrm{corr}}+\epsilon_kD_k^{\mathrm{err}}\right]. \tag*{(63)}

where ϵk=Pr⁡(K^≠k∣K=k)\epsilon_k=\Pr(\hat{K}\ne k\mid K=k). This form makes explicit that end-to-end performance depends jointly on within-branch reconstruction quality and branch-classification reliability.

The decomposition in Eq. (63) also highlights a useful optimistic reference obtained by assuming perfect branch identification, in which case only the within-branch reconstruction error remains. Practical receivers must infer the transmitted branch from noisy observations, so the gap to this reference measures the additional distortion associated with branch-identification errors.

In multi-branch systems, branch misclassification can induce a large distortion contribution because the reconstructed source value may be forced into an interval associated with a different branch. Consequently, even modest classification error probabilities may significantly affect end-to-end SDR when the branch intervals are well separated in source space.

Capacity-surrogate viewpoint

A useful high-SNR perspective is obtained by viewing multi-curve design through a geometric capacity surrogate. Let GkG_k denote the geometric score associated with branch kk. In a small-noise approximation, the induced mutual information increases with the aggregate geometric score ∑k=1mGk\sum_{k=1}^{m}G_k, so that the corresponding objective may be viewed as increasing monotonically with

I⋆∝log⁡(∑k=1mGk).I^\star\propto\log\left(\sum_{k=1}^{m}G_k\right).

From this perspective, maximizing the sum of the per-branch geometric scores provides a natural high-SNR design principle for the branch set.

A related geometric viewpoint is to balance total usable branch length against effective union thickness, which captures the tradeoff between resolution and robustness. This viewpoint separates the C3M design problem into two stages. The geometry stage seeks a branch set with a large aggregate geometric score while maintaining sufficient robustness and inter-branch separation. The allocation stage divides the source interval among the selected branches according to the resulting per-branch utilities. For constant-speed branches, the geometric scores satisfy Gk=Lk/(2π)G_k=L_k/(2\pi), so the capacity-surrogate viewpoint connects directly to a geometry-only, length-based allocation rule.

Remark. In the constant-speed even-dimensional case, the capacity-surrogate viewpoint also suggests a direct algorithmic route to branch-set design. A natural formulation is

max⁡{rk}log⁡(∑k=1mGk)−λ[ρmin⁡−ρU]+2(64)\max_{\{r_k\}} \log\left(\sum_{k=1}^{m}G_k\right)-\lambda\left[\rho_{\min}-\rho_U\right]_+^2 \tag*{(64)}

where ρU\rho_U denotes an overall robustness measure for the multi-curve construction, ρmin⁡\rho_{\min} is a target robustness threshold, λ≥0\lambda\ge0 is a penalty parameter, and [x]+≜max⁡{x,0}[x]_+\triangleq\max\{x,0\}. Here the first term promotes a large aggregate geometric score, while the penalty term discourages branch sets whose usable tube radius falls below the target level. Writing

Gk=∑j=1n/2j2rj,k2,S=∑ℓ=1mGℓ,G_k=\sqrt{\sum_{j=1}^{n/2}j^2r_{j,k}^2},\qquad S=\sum_{\ell=1}^{m}G_\ell,

the gradient of the objective term log⁡S\log S with respect to the radii parameters is

∂∂rj,klog⁡(∑ℓ=1mGℓ)=1Sj2rj,kGk.(65)\frac{\partial}{\partial r_{j,k}}\log\left(\sum_{\ell=1}^{m}G_\ell\right)=\frac{1}{S}\frac{j^2r_{j,k}}{G_k}. \tag*{(65)}

Hence the geometry stage may be treated algorithmically, for example through projected-gradient or stochastic-approximation methods, with robustness and inter-branch separation imposed as additional constraints or penalties.

■

Geometry-only allocation perspective

A complementary perspective is obtained by considering source allocation based solely on branch geometry. In a high-SNR small-noise approximation, a natural geometric score for branch kk is

Gk=exp⁡(12π∫02πlog⁡vk(θ) dθ).(66)G_k=\exp\left(\frac{1}{2\pi}\int_0^{2\pi}\log v_k(\theta)\,\mathrm{d}\theta\right). \tag*{(66)}

where vk(θ)v_k(\theta) denotes the local branch speed, as in Eq. (2). This quantity is the geometric mean of the local speed over one period. A geometry-only allocation rule would then assign interval fractions proportionally to these scores,

pk=Gk∑ℓ=1mGℓ,k=1,…,m.(67)p_k=\frac{G_k}{\sum_{\ell=1}^{m}G_\ell},\qquad k=1,\ldots,m. \tag*{(67)}

Remark (Local Fisher-information interpretation). *Another local interpretation is obtained from the Fisher information of the curve parameter. For a single branch observed through fixed isotropic AWGN, Y=γk(z)+N\mathbf{Y}=\gamma_k(z)+\mathbf{N}, N∼N(0,σn2In)\mathbf{N}\sim\mathcal{N}(0,\sigma_n^2\mathbf{I}_n), the Fisher information for the normalized local coordinate zz is Jk(z)=1σn2∥γk′(z)∥2J_k(z)=\frac{1}{\sigma_n^2}\lVert\gamma_k'(z)\rVert^2. Thus, branches with larger local speed provide greater local sensitivity of the observation to changes in the source coordinate. For constant-speed branches, Jk(z)J_k(z) is constant along the branch, which supports the use of curve length or speed-based geometric gains in high-SNR allocation heuristics. This Fisher-information viewpoint is local and does not by itself provide a global mutual-information, rate–distortion, or branch-identification bound.

Other geometry-only surrogates, such as the tube-density factors {δk}\{\delta_k\}, may be used in the same way. The classification-aware rule in Eq. (27) extends this geometry-only viewpoint by weighting each branch by its reliability qk=1−ϵkq_k = 1 - \epsilon_k, thereby reducing the allocation to branches that are geometrically efficient but unreliable to classify.

A branch parametrization is said to be constant-speed if the local speed vk(θ)=∥γk′(θ)∥v_k(\theta) = \lVert\gamma_k'(\theta) \rVert does not depend on the parameter θ\theta. In that case, the branch parameter is proportional to arc length, so the natural parametrization is already compatible with normalized arc-length coordinates. This simplification is useful in the present setting because it allows the geometric score to be expressed directly in terms of branch length and avoids additional Jacobian factors in the corresponding decoder expressions. For the branch families considered here, the natural parametrizations are constant-speed. Hence Gk=vk=Lk2πG_k = v_k = \frac{L_k}{2\pi}, since the branch length over one period is Lk=2πvkL_k = 2\pi v_k. Thus, for constant-speed branches, the geometry-only rule reduces to an allocation proportional to branch length. For non-constant-speed branch parametrizations, related decoder formulations may also include a Jacobian compensation term based on the local branch speed.

Since the full branch of length LkL_k is traversed over a source interval of length pkp_k, the effective encoder slope scales like Lk/pkL_k/p_k. Hence, in the high-SNR regime, the local tangential distortion is expected to scale like σ2pk2/Lk2\sigma^2 p_k^2/L_k^2, and after averaging over the interval mass this yields a contribution of order σ2pk3/Lk2\sigma^2 p_k^3/L_k^2.

Remark (Heuristic high-SNR distortion decomposition). A useful high-SNR interpretation of multi-curve performance is obtained by separating tangential distortion from tube-escape effects. Heuristically, the mean-squared error may be approximated as

D(σ2)≈σ2∑k=1mckpk3Lk2+C∑k=1mpkexp⁡(−rk22σ2),(68)D(\sigma^2) \approx\sigma^2 \sum_{k=1}^{m} c_k \frac{p_k^3}{L_k^2} + C \sum_{k=1}^{m} p_k \exp\left(-\frac{r_k^2}{2\sigma^2}\right), \tag*{(68)}

where LkL_k denotes the length of branch kk, rkr_k its admissible tube radius, and ck,C>0c_k,C > 0 are branch-dependent and global constants, respectively. The first term represents tangential distortion when the received point remains in the correct tube, whereas the second term captures threshold effects associated with tube escapes. This decomposition helps explain why geometry-only allocation tends to favor long branches in the high-SNR regime, while branches with larger admissible tube radii become increasingly important when tube escapes are non-negligible.

When the admissible tube radii vary significantly across branches and the noise level is such that tube escapes are no longer negligible, the escape term penalizes branches with small rkr_k. This motivates radius-aware geometric weighting in addition to length-based allocation. When tube escapes are negligible, the leading term in (68) decreases as the branch length increases, which provides a high-SNR justification for geometry-only allocation rules that favor longer branches. This viewpoint is useful because it clarifies the role of geometry in source allocation. However, in the present multi-curve setting, end-to-end performance is influenced not only by geometric efficiency but also by branch-classification reliability. At low or moderate SNR, performance may be dominated not by along-curve resolution alone, but by tube escapes and branch-classification errors. This provides further motivation for the classification-aware rule in Section IV-A, where the geometric contribution is weighted by the branch-reliability factors {qk}k=1m\{q_k\}_{k=1}^{m}.

End-to-end and branch-known SDR references

We distinguish between an end-to-end SDR reference, in which the branch index is not available as side information, and a branch-known reference, in which the decoder is given the true branch index. This distinction is important because a multi-curve code represents the source through both a discrete branch index and a continuous coordinate along the selected branch. In the practical C3M receiver, the branch index must be inferred from the received channel output and is therefore not free side information.

Let S∼U[0,1)S \sim\mathcal{U}[0,1), so that σS2=1/12\sigma_S^2 = 1/12, and let CnC_n denote the channel-capacity reference per nn-dimensional channel use from Eq. (56). For the ordinary end-to-end reference, the rate–distortion calculation is applied to the original scalar source using the classical Gaussian rate–distortion and AWGN capacity formulas [43, 44]. The Gaussian-source distortion and SDR references are

DG,end=1122−2Cn,SDRG,end=22Cn.(69)D_{\mathrm{G,end}} = \frac{1}{12}2^{-2C_n}, \qquad\mathrm{SDR}_{\mathrm{G,end}} = 2^{2C_n}. \tag*{(69)}

Using the high-rate uniform-source correction gives

DU,endasy=12πe2−2Cn,SDRU,endasy=22Cnπe6.(70)D^{\mathrm{asy}}_{\mathrm{U,end}} = \frac{1}{2\pi e}2^{-2C_n}, \qquad\mathrm{SDR}^{\mathrm{asy}}_{\mathrm{U,end}} = 2^{2C_n}\frac{\pi e}{6}. \tag*{(70)}

These end-to-end references do not contain an m2m^2 factor, because any information needed to identify the branch is part of the transmitted source description.

For comparison, we also define a genie-aided branch-known reference. In this reference, the decoder is given the true branch index KK and only needs to estimate the local coordinate within the corresponding source interval. Under equal splitting into mm subintervals, each interval has length 1/m1/m and conditional variance 1/(12m2)1/(12m^2). The branch-known Gaussian distortion and SDR references are therefore

DG,bk=112m22−2Cn,SDRG,bk=m222Cn.(71)D_{\mathrm{G,bk}} = \frac{1}{12m^2}2^{-2C_n}, \qquad\mathrm{SDR}_{\mathrm{G,bk}} = m^2 2^{2C_n}. \tag*{(71)}

Similarly, the high-rate uniform-source branch-known reference is

DU,bkasy=12πem22−2Cn,SDRU,bkasy=m222Cnπe6.(72)D^{\mathrm{asy}}_{\mathrm{U,bk}} = \frac{1}{2\pi e m^2}2^{-2C_n}, \qquad\mathrm{SDR}^{\mathrm{asy}}_{\mathrm{U,bk}} = m^2 2^{2C_n}\frac{\pi e}{6}. \tag*{(72)}

Thus, the familiar m2m^2 factor belongs to the branch-known reference. It is not an ordinary end-to-end rate–distortion gain for the original source.

For non-uniform branch probabilities p1,…,pmp_1,\ldots,p_m with ∑k=1mpk=1\sum_{k=1}^{m} p_k = 1, we also use a same-rate branch-known Gaussian diagnostic reference. If branch kk is assigned interval fraction pkp_k, then its conditional source variance is pk2/12p_k^2/12. Applying the same Gaussian capacity reference CnC_n to each active branch gives

DG,bk=1122−2Cn∑k=1mpk3,(73)D_{\mathrm{G,bk}} = \frac{1}{12}2^{-2C_n}\sum_{k=1}^{m}p_k^3, \tag*{(73)}

and the corresponding branch-known Gaussian SDR reference is

SDRG,bk=22Cn∑k=1mpk3.(74)\mathrm{SDR}_{\mathrm{G,bk}} = \frac{2^{2C_n}}{\sum_{k=1}^{m}p_k^3}. \tag*{(74)}

Equal splitting gives ∑k=1mpk3=1/m2\sum_{k=1}^{m}p_k^3=1/m^2, which recovers Eq. (71). This expression is used as a diagnostic branch-known reference. It assumes the same effective conditional rate reference for each active branch and does not optimize a separate rate allocation across branches.

Using the fixed isotropic AWGN capacity expression in Eq. (56),

22Cn=(1+Pnσn2)n=(1+SNRn)n.(75)2^{2C_n}=\left(1+\frac{P}{n\sigma_n^2}\right)^n=\left(1+\frac{\mathrm{SNR}}{n}\right)^n . \tag*{(75)}

Hence the end-to-end references can be written as

SDRG,end=(1+SNRn)n,SDRU,endasy=(1+SNRn)nnπe6,(76)\mathrm{SDR}_{\mathrm{G,end}}=\left(1+\frac{\mathrm{SNR}}{n}\right)^n,\qquad\mathrm{SDR}^{\mathrm{asy}}_{\mathrm{U,end}}=\left(1+\frac{\mathrm{SNR}}{n}\right)^n\frac{n\pi e}{6}, \tag*{(76)}

while the equal-splitting branch-known references are

SDRG,bk=m2(1+SNRn)n,SDRU,bkasy=m2(1+SNRn)nnπe6.(77)\mathrm{SDR}_{\mathrm{G,bk}}=m^2\left(1+\frac{\mathrm{SNR}}{n}\right)^n,\qquad\mathrm{SDR}^{\mathrm{asy}}_{\mathrm{U,bk}}=m^2\left(1+\frac{\mathrm{SNR}}{n}\right)^n\frac{n\pi e}{6}. \tag*{(77)}

The practical C3M performance reported in § 9 is neither of these idealized references. The receiver observes only Y\mathbf{Y}, estimates the branch index, and then estimates the source coordinate on the selected branch. Its end-to-end distortion and SDR are

DC3M=E[(S−S^)2],SDRC3M=σS2DC3M.(78)D_{\mathrm{C3M}}=\mathbb{E}\left[(S-\hat{S})^2\right],\qquad\mathrm{SDR}_{\mathrm{C3M}}=\frac{\sigma_S^2}{D_{\mathrm{C3M}}}. \tag*{(78)}

This quantity includes both along-curve estimation error and branch-identification error. The gap between a branch-known decoder and the practical decoder therefore measures the penalty due to branch confusion. Table V summarizes the distinction between the end-to-end Gaussian reference, the branch-known Gaussian reference, and the practical C3M SDR used in the numerical evaluation.

QuantityExpressionInterpretation
End-to-end Gaussian reference(1+SNRn)n\left(1+\frac{\mathrm{SNR}}{n}\right)^nNo branch side information
Branch-known Gaussian referencem2(1+SNRn)nm^2\left(1+\frac{\mathrm{SNR}}{n}\right)^nGenie-aided diagnostic reference
Practical C3M SDRσS2E[(S−S^)2]\frac{\sigma_S^2}{\mathbb{E}[(S-\hat{S})^2]}Learned-decoder end-to-end performance

Table V. Interpretation of the SDR quantities used in the C3M analysis and evaluation.

A useful high-SNR interpretation of multi-curve performance is obtained by separating along-curve distortion from tube-escape effects. At a heuristic level, the mean-squared error may be decomposed into a tangential term, which decreases with increasing effective curve length, and a threshold term associated with tube escapes or branch confusions. This viewpoint helps explain why geometry-only allocation tends to favor long curves in the high-SNR regime, whereas more robust branches become increasingly important when the noise level is high enough for tube escapes or branch confusions to play a significant role.

These analytical references do not include finite-dimensional geometric admissibility constraints, branch-overlap constraints, decoder training effects, or classification errors. They should therefore be interpreted as idealized information-theoretic references rather than finite-dimensional performance guarantees. The Gaussian reference is not the exact Shannon limit for a uniform source. It is a convenient reference obtained from the Gaussian rate–distortion function. The expression SDRUasy\mathrm{SDR}_{U}^{\mathrm{asy}} is best interpreted as a high-rate uniform-source reference, not as an exact low-SNR rate–distortion limit.

Non-uniform source allocation

Finally, when the source allocation is highly non-uniform, a more refined design criterion is obtained by protecting the branches with largest probabilities {pk}\{p_k\}. In that case, the standard max–min inter-branch separation heuristic may be replaced by a weighted geometric surrogate max⁡{rk}min⁡k≠k′wkk′∥rk−rk′∥\max_{\{r_k\}}\min_{k\ne k'} w_{kk'}\lVert r_k-r_{k'}\rVert, where wkk′≥0w_{kk'}\ge0 are design weights. By assigning larger values of wkk′w_{kk'} to pairs involving dominant branches, the resulting curve set encourages larger separation margins around the branches that contribute most strongly to the overall distortion. This is especially relevant in low-dimensional settings, where geometric packing constraints limit the number of well-separated branches and one or more curves may otherwise become bottlenecks for classification performance.

It is useful to distinguish non-uniform branch probabilities from unequal conditional variances. If two branches have probabilities p1p_1 and p2p_2 but identical conditional source variances and are assigned the same target distortion, then the weighted rate constraint p1R1(D)+p2R2(D)≤Cnp_1R_1(D)+p_2R_2(D)\le C_n reduces to R(D)≤CnR(D)\le C_n whenever R1(D)=R2(D)R_1(D)=R_2(D) and p1+p2=1p_1+p_2=1. In that special symmetric case, changing the probabilities alone does not change the per-branch reference distortion. For a uniform source partitioned into intervals, however, the branch probability is tied to the interval length. Thus, non-uniform interval fractions generally imply different conditional variances; for a contiguous interval of length pkp_k, σS∣K=k2=pk212\sigma^2_{S|K=k}=\frac{p_k^2}{12}, and the allocation problem becomes genuinely weighted.

For Gaussian rate–distortion references, one may write the weighted constraint as

∑k=1mpkRk(Dk)≤Cn,Rk(Dk)=12log⁡2(σS∣K=k2Dk).(79)\sum_{k=1}^{m}p_kR_k(D_k)\le C_n,\qquad R_k(D_k)=\frac{1}{2}\log_2\left(\frac{\sigma^2_{S|K=k}}{D_k}\right). \tag*{(79)}

This form makes explicit that non-uniform source allocation affects the branch-known reference through both the probabilities pkp_k and the conditional variances σS∣K=k2\sigma^2_{S|K=k}.

A complementary way to interpret non-uniform allocation is through a decoder-switching model. For simplicity, consider two branches with probabilities p1p_1 and p2p_2, and let α=Pr⁡(K^=2∣K=1)\alpha=\Pr(\hat{K}=2|K=1) and β=Pr⁡(K^=1∣K=2)\beta=\Pr(\hat{K}=1|K=2) denote the two branch-confusion probabilities. If Dk→ℓD_{k\to\ell} denotes the conditional distortion incurred when branch kk was transmitted but the branch-conditioned decoder for branch ℓ\ell is used, then the total distortion can be written as

Dtotal=p1[(1−α)D1→1+αD1→2]+p2[βD2→1+(1−β)D2→2].(80)D_{\mathrm{total}}=p_1\left[(1-\alpha)D_{1\to1}+\alpha D_{1\to2}\right]+p_2\left[\beta D_{2\to1}+(1-\beta)D_{2\to2}\right]. \tag*{(80)}

This expression separates the within-branch reconstruction distortion from the penalty associated with decoder-switching errors. In particular, when the branch probabilities are non-uniform, switching errors involving high-probability branches contribute more strongly to the total distortion and should therefore be weighted more heavily in both branch placement and receiver training.

Remark (Switching-aware numerical design objective). More generally, one can define a switching-aware numerical design objective by optimizing over branch curves, a branch-selection rule, and branch-conditioned decoders,

min⁡{γk}, h, {dk}E[(S−dh(Y)(Y))2]s.t.\underset{\{\gamma_k\},\,h,\,\{d_k\}}{\min}\quad\mathbb{E}\left[(S-d_{h(\mathbf{Y})}(\mathbf{Y}))^2\right] \quad\text{s.t.}
∑k=1mpk E[∥γk(Z)∥2∣K=k]≤P.(81)\sum_{k=1}^{m}p_k\,\mathbb{E}\left[\|\gamma_k(Z)\|^2 \mid K=k\right]\le P. \tag*{(81)}

Here h(Y)∈{1,…,m}h(\mathbf{Y})\in\{1,\ldots,m\} denotes the branch-selection rule, and dk(⋅)d_k(\cdot) denotes the decoder used when branch kk is selected. If the functions {γk}\{\gamma_k\}, hh, and {dk}\{d_k\} are restricted to a finite-dimensional parameterized family, such as C3M curves with a learned branch classifier and branch-conditioned neural decoders, Eq. (81) yields a numerical design objective rather than a closed-form analytical performance limit. ■

When the branch-selection probabilities {pk}\{p_k\} are non-uniform, these priors should also be reflected in the decoder. In a MAP formulation under fixed isotropic AWGN, the branch decision is no longer determined solely by Euclidean proximity, but must include the prior contribution associated with each branch. Thus, for an observation y\mathbf{y}, a branch-conditioned MAP rule takes the form (k^,z^)=arg⁡min⁡k, z∈[0,1){∥y−γk(z)∥2−2σn2log⁡pk}(\hat{k},\hat{z})=\arg\min_{k,\,z\in[0,1)}\{\|\mathbf{y}-\gamma_k(z)\|^2-2\sigma_n^2\log p_k\}. Hence, highly probable branches receive the appropriate prior weighting through the term −2σn2log⁡pk-2\sigma_n^2\log p_k, while the geometric mismatch is measured by Euclidean distance under the fixed isotropic AWGN model. For a Gaussian extension with general covariance Σn\boldsymbol{\Sigma}_n, the Euclidean distance is replaced by the corresponding Mahalanobis distance.

The same principle also applies to learning-based receivers. When {pk}\{p_k\} are non-uniform, classifier and decoder training should account for the resulting class imbalance, for example through class-weighted losses or probability-aware sampling. Otherwise, dominant branches may be underprotected during training even though they contribute most strongly to the overall distortion. In this sense, the source allocation {pk}\{p_k\} influences not only the encoder-side interval division, but also the statistically consistent design of the receiver.

Geometric interpretation of multi-curve branch selection

A useful geometric interpretation is obtained by viewing multi-curve branch selection in terms of spherical codes in radius-vector space. For even ambient dimension nn, one has N=n/2N=n/2, so the admissible radii vectors lie on SN−1S^{N-1}. In this way, multi-curve branch selection may be viewed as a packing problem on SN−1S^{N-1}, where one seeks sufficiently well-separated radii vectors to obtain robust and reliably distinguishable branch curves. In low dimensions, this immediately explains some of the observed branch-count limitations. For n=4n=4, the radii vectors lie on S1S^1, where only a very small number of strongly separated points can be placed; this is consistent with the practical difficulty of using many robustly distinguishable curves in four dimensions. For n=6n=6, the design problem moves to S2S^2, where larger and more symmetric point sets become feasible, for example six points arranged like an octahedron. For n=8n=8, the corresponding design space is S3S^3, where substantially richer spherical configurations are possible, for example the 24 vertices of the 24-cell. These examples should be interpreted as geometric illustrations rather than strict optimality statements, but they provide useful intuition for why the admissible number of robust branch curves tends to increase with dimension.

Heuristic upper bounds on the number of branch curves

A further geometric question is whether one can derive explicit upper bounds on the number of robustly admissible branch curves. Dimension alone does not impose such a bound, since arbitrarily many distinct curves may in principle be placed on the ambient sphere if tube thickness and inter-branch separation are allowed to shrink. A meaningful finite bound arises only after geometric robustness constraints are imposed, such as a minimum admissible tube radius or a minimum inter-branch separation. A natural robustness metric for the multi-curve construction combines intra-branch tube admissibility with inter-branch separation, so that the overall usable tube radius is limited by the smaller of the individual branch thicknesses and half the minimum inter-branch distance.

At a heuristic level, let Uk(rmin⁡)U_k(r_{\min}) denote the radius-rmin⁡r_{\min} tube around branch kk, i.e., the tubular neighborhood generated by all points within Euclidean distance rmin⁡r_{\min} of the curve γk\gamma_k. If each branch admits a disjoint tube of radius at least rmin⁡r_{\min}, then the total volume occupied by the branch tubes must remain within the available bounding region. For a one-dimensional curve of length LkL_k embedded in Rn\mathbb{R}^n, the corresponding tube volume scales locally like Vol⁡(Uk(rmin⁡))≈κn−1Lkrmin⁡n−1\operatorname{Vol}(U_k(r_{\min})) \approx\kappa_{n-1} L_k r_{\min}^{n-1}, where κn−1\kappa_{n-1} is the volume of the unit ball in the normal (n−1)(n-1)-dimensional cross-section. Consequently, if all admissible branch tubes are pairwise disjoint and satisfy a common lower bound Lk≥Lmin⁡L_k \ge L_{\min}, one obtains the heuristic packing estimate

M≲Vol⁡(bounding region)κn−1Lmin⁡rmin⁡n−1.(82)M \lesssim\frac{\operatorname{Vol}(\text{bounding region})}{\kappa_{n-1}L_{\min}r_{\min}^{n-1}}. \tag*{(82)}

Although the precise constants and admissibility conditions depend on the detailed geometry of the construction, this scaling suggests that the maximum number of robustly usable branches is governed jointly by ambient dimension, minimum tube radius, and curve length, rather than by dimension alone.

Numerical Experiments

We evaluate the proposed C3M designs under the fixed isotropic AWGN model, N∼N(0,σ2In)N \sim\mathcal{N}(0,\sigma^2 I_n), using the power normalization in Eq. (21) and the SNR convention in Eq. (22). When useful for interpretation, the corresponding per-dimension SNR is SNRdim=P/nσn2=SNRn\mathrm{SNR}_{\mathrm{dim}} = \frac{P/n}{\sigma_n^2} = \frac{\mathrm{SNR}}{n}. Unless otherwise stated, the reported SDR values are evaluated under this nominal AWGN setting. In Figure 9, the one-curve panels use the end-to-end Gaussian SDR reference in Eq. (69). The multi-curve panels use the branch-known Gaussian SDR reference in Eq. (71), which includes the equal-splitting m2m^2 factor and is used as a diagnostic reference. The hyperparameters of the NORM-inspired decoder are shown in Table IV.

Table VI and Figure 9 summarize the main AWGN performance results. The figure compares SDR, decoder-inference accuracy, and bootstrap variability across n∈{4,6}n \in\{4,6\} and m∈{1,2,3}m \in\{1,2,3\}, while the table reports the selected or newly added curve parameters, SDR values, decoder-inference accuracies, and bootstrap statistics. In Figure 9, the one-curve panels use the end-to-end Gaussian SDR reference, whereas the multi-curve panels use the branch-known Gaussian SDR reference. The results show an increase in SDR as candidate branches are added. At +10 dB SNR, the selected m=3m = 3 NORM-inspired systems reach 31.6 dB for n=4n = 4 and 37.5 dB for n=6n = 6. The n=6n = 6, m=3m = 3 case is a three-candidate branch set with one candidate effectively pruned by the allocation rule. These multi-curve gains should be interpreted as finite-dimensional C3M performance gains, with the branch-known reference used only as a diagnostic comparison. They are not end-to-end rate–distortion gains for the original scalar source. The MLP baselines trained for comparison follow the same increasing-SDR trend and provide learned-decoder reference points. Since the MLP rows are slightly higher than the NORM-inspired rows in some cases, the main performance claim is the multi-branch C3M construction rather than architectural superiority of the NORM-inspired decoder.

Table VII compares equal splitting, length-only allocation, tube-density-only allocation, and density–reliability allocation on the same selected C3M branch geometries. This ablation separates the effect of interval allocation from the effect of branch-geometry selection. For the two-branch cases, the allocation rule has only a small effect on SDR. For the three-candidate cases, length-only allocation gives the largest mean SDR among the tested rules. The density–reliability rule instead favors branches with higher classifier reliability; in the n=6n = 6, m=3m = 3 case, it prunes the second candidate branch because its measured reliability is zero. Thus, the density–reliability allocation should be interpreted as a reliability-aware allocation diagnostic rather than as an SDR-optimal rule for these fixed trained receivers. Across all tested allocations, the main multi-branch performance trend remains unchanged.

ConfigurationEqual splitLength-onlyTube-density-onlyDensity–reliability δkqk\delta_kq_k
n=4,m=2n=4,m=228.03 ± 0.0728.08 ± 0.0728.01 ± 0.0628.02 ± 0.07
n=4,m=3n=4,m=331.72 ± 0.0531.75 ± 0.0731.70 ± 0.0631.60 ± 0.06
n=6,m=2n=6,m=234.06 ± 0.0234.09 ± 0.0234.02 ± 0.0334.02 ± 0.03
n=6,m=3n=6,m=337.68 ± 0.0337.71 ± 0.0337.65 ± 0.0337.52 ± 0.03

Table VII. Allocation-rule ablation at +10 dB SNR for the selected C3M candidate branch sets. All rules are evaluated on the same selected branch geometries. Entries report the mean SDR over 100 Monte Carlo trials, with one standard deviation.

For the selected NORM-inspired C3M systems, the total system curve length and density increase from 19.23

DecodernnmmSelected/new curvef^{f}SDR [dB]Inference accuracy [%]g^{g}σc\sigma^{c}MADd^{d}
C3Ta^{a}41–20.896.3N/AeN/A^{e}N/AeN/A^{e}
MLPb^{b}41ϕ1=47∘\phi_{1}=47^\circ22.997.30.007880.00550
MLPb^{b}42ϕ2=46∘\phi_{2}=46^\circ28.597.20.008840.00594
MLPb^{b}43ϕ3=44∘\phi_{3}=44^\circ32.197.20.01800.0117
NORM41ϕ1=47∘\phi_{1}=47^\circ22.697.30.01640.0110
NORM42ϕ2=47∘\phi_{2}=47^\circ28.097.10.01820.0118
NORM43ϕ3=44∘\phi_{3}=44^\circ31.697.10.008680.00594
C3Ta^{a}61–25.297.6N/AeN/A^{e}N/AeN/A^{e}
MLPb^{b}61(45∘,45∘)(45^\circ,45^\circ)28.998.60.009470.00609
MLPb^{b}62(50∘,65∘)(50^\circ,65^\circ)34.498.40.004140.00339
MLPb^{b}63(50∘,50∘)(50^\circ,50^\circ)37.898.40.009970.00687
NORM61(50∘,60∘)(50^\circ,60^\circ)28.598.50.004630.00317
NORM62(50∘,60∘)(50^\circ,60^\circ)34.098.40.005740.00342
NORM63(50∘,55∘)(50^\circ,55^\circ)37.598.40.005100.00345

Table VI. Decoder SDR and inference-accuracy statistics at +10 dB SNR.

nnmmp1p_1p2p_2p3p_3q1q_1 [%]q2q_2 [%]q3q_3 [%]Overall [%]
420.52070.4793–73.1674.23–73.67
430.47290.32340.203767.1950.6930.8854.45
620.54830.4517–82.3883.33–82.81
630.54130.00000.458777.670.0078.7778.17

Table III. Final interval allocations and branch-classification reliabilities at +10 dB SNR for the selected C3M configurations. The reliabilities qkq_k are the diagonal entries of the row-normalized branch-confusion matrix, and the overall branch-classification accuracy is ∑kpkqk\sum_k p_kq_k.

and 0.7062 for the four-dimensional m=2m=2 design to 28.95 and 1.0656 for the four-dimensional m=3m=3 design. In six dimensions, the corresponding values increase from 27.58 and 0.1862 for m=2m=2 to 41.88 and 0.2733 for

m=3m = 3. These trends show that adding branch curves increases the usable geometric resolution, while larger ambient dimension provides additional degrees of freedom for branch placement and along-curve reconstruction.

MAP and MMSE estimators are characterized analytically in Section VI-B; numerical MAP/MMSE comparisons for the single-curve C3T case were studied in [15]. The present experiments therefore focus on scalable learned decoders for the multi-curve setting.

Computational complexity

The proposed C3M design has two distinct computational regimes: (a) offline code design and training, and (b) online inference. The offline stage includes candidate branch selection, classifier training, branch-conditioned decoder training, and Monte Carlo evaluation. The online stage consists only of applying the trained classifier and the selected branch decoder to each received vector.

Let GG denote the number of candidate branches considered at each sequential selection step. In the four-dimensional one-parameter sweep, G=∣Φ∣G = |\Phi|, while in the six-dimensional two-parameter sweep, G=∣Θ∣∣Φ∣G = |\Theta||\Phi|. For a final mm-branch design initialized by a fixed C3T branch, the sequential search requires approximately (m−1)G(m - 1)G candidate evaluations. By contrast, an unordered exhaustive search over GG possible branches would require (Gm)\binom{G}{m} candidate sets. Thus, the sequential design replaces a combinatorial global search with a linear sequence of marginal branch-selection problems.

For each candidate, let NtrN_{\mathrm{tr}} denote the number of training samples, ECE_C and EhE_h the numbers of classifier and decoder training epochs, and let PC(m)P_C(m) and PhP_h denote the numbers of trainable parameters in the branch classifier and in one branch-conditioned decoder, respectively. Up to architecture-dependent constants, the offline training cost per candidate scales as O(ECNtrPC(m)+mEhNtrPh)O(E_C N_{\mathrm{tr}} P_C(m) + m E_h N_{\mathrm{tr}} P_h), with an additional Monte Carlo evaluation cost proportional to NMC(PC(m)+Ph)N_{\mathrm{MC}}(P_C(m) + P_h) for NMCN_{\mathrm{MC}} test samples when one selected decoder is evaluated per received vector. This cost is incurred offline and is therefore not part of the online inference cost. At inference time, the trained receiver first estimates the branch index and then applies the corresponding branch-conditioned decoder. Hence the per-symbol online complexity is O(PC(m)+Ph)O(P_C(m) + P_h), or simply O(Ph)O(P_h) in the single-branch case. The dependence on the number of active branches enters primarily through the classifier output dimension and through the number of decoders stored in the decoder bank. Importantly, the learned receiver does not perform an online search over the continuous curve parameter.

This distinguishes the learned C3M receiver from a discretized MAP or MMSE implementation. If each branch were discretized using NzN_z points, then a nearest-curve MAP approximation would require O(mNzn)O(mN_z n) distance evaluations per received vector, and an MMSE approximation would require a comparable summation over the branch-coordinate grid. The NORM-inspired branch-conditioned spectral receiver replaces this online grid search with fixed forward passes through the classifier and the selected branch decoder. Thus, the online cost is determined by the classifier and selected-decoder network sizes rather than by the branch-coordinate grid resolution.

Single-curve decoder robustness under impulsive non-Gaussian noise

To isolate decoder robustness from branch-classification effects, we evaluate the single-curve setting under the non-Gaussian noise models in Eqs. (19) and (20). For the Gaussian-mixture experiments, we set Σw=σw2In\boldsymbol{\Sigma}_w = \sigma_w^2\mathbf{I}_n and Σv=σv2In\boldsymbol{\Sigma}_v = \sigma_v^2\mathbf{I}_n, and fix the impulsive-to-nominal variance ratio to σv2/σw2=100\sigma_v^2/\sigma_w^2 = 100. At the nominal +10 dB SNR operating point, the corresponding linear SNR is SNRlin=10\mathrm{SNR}_{\mathrm{lin}} = 10. With P=1P = 1, this gives σw2=P/SNRlin=0.1\sigma_w^2 = P/\mathrm{SNR}_{\mathrm{lin}} = 0.1. Hence, σv2=10\sigma_v^2 = 10, or σv=10≈3.1623\sigma_v = \sqrt{10} \approx3.1623, so impulsive samples have ten times larger standard deviation than the nominal AWGN component. We then sweep the contamination probability over the range ϵ∈[0,0.5]\epsilon\in[0,0.5]. For the symmetric alpha-stable model, we decrease the stability parameter from α=2\alpha= 2 to α=1\alpha= 1, where α=2\alpha= 2 corresponds to the Gaussian endpoint and α=1\alpha= 1 corresponds to the Cauchy endpoint of the sweep. These single-curve experiments assess how the circular Huber loss used by the NORM-inspired spectral decoder affects stability as the noise departs from the nominal AWGN model. They do not evaluate multi-branch operation or branch-classification robustness under non-Gaussian noise.

We report robustness using the normalized SDR loss relative to the nominal AWGN case,

ΔrelSDR=SDRAWGN−SDRnonGaussianSDRAWGN,(83)\Delta_{\mathrm{rel}}\mathrm{SDR} = \frac{\mathrm{SDR}_{\mathrm{AWGN}} - \mathrm{SDR}_{\mathrm{nonGaussian}}}{\mathrm{SDR}_{\mathrm{AWGN}}}, \tag*{(83)}

where the SDR values in Eq. (83) are evaluated on a linear scale. Thus, smaller values of ΔrelSDR\Delta_{\mathrm{rel}}\mathrm{SDR} indicate greater robustness, with ΔrelSDR=0\Delta_{\mathrm{rel}}\mathrm{SDR} = 0 corresponding to no loss relative to the nominal AWGN setting. In Figure 10, this quantity is plotted in percent. The curves in Figure 10 are estimated from 100 Monte Carlo runs. For Gaussian-mixture noise at ϵ=0.10\epsilon= 0.10, the NORM-inspired spectral decoder has a smaller normalized SDR loss than the MLP baseline by 1.64 percentage points for n=4n = 4 and 0.06 percentage points for n=6n = 6. For Sα\alphaS noise at α=1.8\alpha= 1.8, the corresponding reductions are 5.45 percentage points for n=4n = 4 and 1.47 percentage points for n=6n = 6. In these single-curve tests, the Huber-trained decoder is therefore competitive with the MLP baseline and gives lower normalized SDR loss in several impulsive and heavy-tailed regimes. In the single-curve setting, the Gaussian-mixture results are consistent with the bounded-influence behavior of the Huber objective. The NORM-inspired decoder generally exhibits smaller normalized SDR loss than the MLP baseline under low-to-moderate contamination.

Gaussian-mixture noise and SαS noise robustness plots

Figure 10. (a) Gaussian-mixture noise. (b) SαS noise. Robustness of the single-curve decoders under non-Gaussian noise. Panel (a) sweeps the Gaussian-mixture contamination probability ϵ\epsilon in Eq. (19), and panel (b) sweeps the symmetric alpha-stable stability parameter α\alpha in Eq. (20). Both panels report the AWGN-relative normalized SDR loss ΔrelSDR\Delta_{\mathrm{rel}}\mathrm{SDR} in percent, where smaller values indicate greater robustness. The curves compare the n=4n = 4 and n=6n = 6 single-curve systems for the NORM-inspired decoder with circular Huber loss and the MLP baseline. Curves are estimated from 100 Monte Carlo runs.

Conclusion and Future Work

This paper introduced constant-curvature curve multicode (C3M), a multi-branch extension of C3T for finite-dimensional analog joint source–channel coding over fixed isotropic AWGN channels. The construction partitions a scalar source interval across multiple constant-speed branch curves in even-dimensional Euclidean space, thereby increasing the available geometric design degrees of freedom while preserving a one-shot analog encoding structure. We derived the induced mixture output distribution, posterior representation, MAP and MMSE estimators, mutual-information expressions, and end-to-end and branch-known SDR reference curves. The analysis separates along-curve reconstruction from branch-identification effects, which is essential for interpreting multi-curve performance.

The numerical results show that adding candidate branches increases the finite-dimensional SDR of the learned C3M receivers under the nominal AWGN model. At +10 dB SNR, the selected m=3m=3 NORM-inspired systems reach 31.6 dB for n=4n=4 and 37.5 dB for n=6n=6. The n=6n=6, m=3m=3 case is a three-candidate branch set with one candidate effectively pruned by the allocation rule. These gains should be interpreted as finite-dimensional C3M performance gains, not as rate–distortion gains for the original scalar source. The branch-known Gaussian reference is used only as a diagnostic comparison for the multi-curve panels.

Several extensions follow directly from the present formulation. The first is automated branch-set design. In the present work, branch candidates are selected by finite sweeps over radii parameters. A natural next step is to treat the radii vectors and branch orientations as variables in a constrained geometric optimization problem. For constant-speed even-dimensional constructions, the geometry-only objectives admit simple derivatives with respect to the radii parameters, which suggests projected-gradient, stochastic-approximation, and simultaneous-perturbation methods for automated C3M branch design [45, 46, 47]. This direction would also allow branch geometry and interval allocation to be optimized jointly, rather than sequentially. The fixed-geometry ablation in Table VII provides a first comparison of allocation rules; a natural extension is to include the allocation variables directly in the branch-selection optimization.

A second direction is to develop receivers that more directly exploit the intrinsic geometry of the C3M signal set. One approach is to lift each noisy channel observation to a likelihood or log-likelihood function over the circular branch coordinate and then process this intrinsic signal using S¹-equivariant or manifold-aware neural architectures [26, 27, 48, 49]. Such decoders could exploit the harmonic structure of the C3T/C3M encoder, refine multimodal posteriors caused by curve folds, and produce calibrated directional uncertainty through a von Mises or related circular-statistics readout [50]. Kolmogorov–Arnold networks provide another possible class of decoders and branch classifiers and should be compared with MLP and Fourier-feature baselines on the same selected branch geometries [51].

A third direction is to separate branch identification from along-curve estimation more explicitly. A hybrid receiver could first preselect a small set of candidate branches using coarse geometry-aware scores and then apply refined branch-conditioned projection, MAP estimation, MMSE estimation, or learned local-coordinate decoding only to the selected candidates. This coarse-to-fine structure may reduce online complexity when the number of branches is large. Low-complexity branch classifiers are also natural in this setting. In four dimensions, the features u1(y)=y12+y22u_1(\mathbf{y}) = \sqrt{y_1^2 + y_2^2} and u2(y)=y32+y42u_2(\mathbf{y}) = \sqrt{y_3^2 + y_4^2} are invariant to phase within each harmonic pair. In the noiseless limit, these features coincide with the corresponding branch radii and therefore provide a low-dimensional representation for coarse branch classification in radius-vector space. Such magnitude-based classifiers may be useful as front ends for hybrid receivers, followed by refined branch-conditioned projection, MAP estimation, MMSE estimation, or learned local-coordinate decoding. Gray-style assignment of neighboring source intervals to geometrically nearby branches is a related design option, since it can reduce the source-domain distortion caused by occasional branch-index errors [52]. For non-uniform branch probabilities, receiver training should also incorporate source priors through class-weighted losses, probability-aware sampling, or prior-weighted decision rules so that high-probability branches are not underprotected [53].

Finally, algebraic constructions may provide systematic tools for generating structured branch families. One exploratory direction is to use number-field embeddings to generate torus curves whose winding directions and relative separations are controlled algebraically. CM subfields of K=Q(i,−2)K = \mathbb{Q}(i,\sqrt{-2}) may provide candidate curve families with distinct torus directions, while related structures such as Minkowski embeddings, Eisenstein-integer lattices [54], and Hurwitz-quaternion constructions may provide additional tools for multi-curve design [55, 56]. These algebraic constructions could complement numerical branch optimization by providing structured candidate families with analyzable separation and symmetry properties. A longer-term direction is to use C3T/C3M mappings as one-shot analog source–channel interfaces in distributed sensing, edge inference, or federated-learning systems, connecting the finite-dimensional geometric approach studied here with learned JSCC systems [18].

Proof of Optimal Bayesian Estimator

Proposition 1. The MMSE estimator S^MMSE(Y)=E[S∣Y]\widehat{S}_{\mathrm{MMSE}}(\mathbf{Y}) = \mathbb{E}[S \mid\mathbf{Y}] minimizes E[(S−g(Y))2]\mathbb{E}[(S - g(\mathbf{Y}))^2] over all measurable estimators g(Y)g(\mathbf{Y}).

Proof. For any estimator g(Y)g(\mathbf{Y}),

E[(S−g(Y))2]=E[(S−E[S∣Y]+E[S∣Y]−g(Y))2]=E[(S−E[S∣Y])2]+E[(E[S∣Y]−g(Y))2]+2E[(S−E[S∣Y])(E[S∣Y]−g(Y))].\begin{aligned} \mathbb{E}\left[(S-g(\mathbf{Y}))^2\right] = \mathbb{E}\left[\left(S-\mathbb{E}[S\mid\mathbf{Y}]+\mathbb{E}[S\mid\mathbf{Y}]-g(\mathbf{Y})\right)^2\right] \\ &= \mathbb{E}\left[\left(S-\mathbb{E}[S\mid\mathbf{Y}]\right)^2\right] + \mathbb{E}\left[\left(\mathbb{E}[S\mid\mathbf{Y}]-g(\mathbf{Y})\right)^2\right] \\ &+ 2\mathbb{E}\left[(S-\mathbb{E}[S\mid\mathbf{Y}])(\mathbb{E}[S\mid\mathbf{Y}]-g(\mathbf{Y}))\right]. \end{aligned}

The cross-term vanishes by the tower property: E[(S−E[S∣Y])(E[S∣Y]−g(Y))]=0\mathbb{E}[(S-\mathbb{E}[S\mid\mathbf{Y}])(\mathbb{E}[S\mid\mathbf{Y}]-g(\mathbf{Y}))]=0. Hence E[(S−g(Y))2]=E[(S−E[S∣Y])2]+E[(E[S∣Y]−g(Y))2]\mathbb{E}[(S-g(\mathbf{Y}))^2] = \mathbb{E}[(S-\mathbb{E}[S\mid\mathbf{Y}])^2] + \mathbb{E}[(\mathbb{E}[S\mid\mathbf{Y}]-g(\mathbf{Y}))^2], and the second term is nonnegative, with equality if and only if g(Y)=E[S∣Y]g(\mathbf{Y})=\mathbb{E}[S\mid\mathbf{Y}] almost surely.

References

  1. [1]K. B. Letaief, Y. Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 1, pp. 5–36, 2022.arxiv.org/abs/2111.12444
  2. [2]K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y. J. A. Zhang, “The roadmap to 6G: AI empowered wireless networks,” IEEE Communications Magazine, vol. 57, no. 8, pp. 84–90, 2019.arxiv.org/abs/1904.11686
  3. [3]Y. Liu, X. Yuan, Z. Xiong, J. Kang, X. Wang, and D. Niyato, “Federated learning for 6G communications: Challenges, methods, and future directions,” China Communications, vol. 17, no. 9, pp. 105–118, 2020.DOI
  4. [4]P. Pulkkinen, M. Esfandiari, and V. Koivunen, “Beamspace and frequency domain ISAC resource allocation using reinforcement learning,” in Asilomar Conference on Signals, Systems, and Computers, 2024, pp. 443–449.
  5. [5]D. A. Ravi, V. K. Shah, C. Li, Y. T. Hou, and J. H. Reed, “RAN slicing in multi-MVNO environment under dynamic channel conditions,” IEEE Internet of Things Journal, vol. 9, no. 6, pp. 4748–4757, 2022.arxiv.org/abs/2104.04900
  6. [6]V. C. Cognini, S. Werner, F. Gauthier, Y.-F. Huang, and A. Kuh, “Personalized online federated learning for IoT/CPS: Challenges and future directions,” IEEE Internet of Things Magazine, vol. 5, no. 4, pp. 78–84, 2022.
  7. [7]M. Matthaiou, O. Yurduseven, H. Q. Ngo, D. Morales-Jimenez, S. L. Cotton, and V. F. Fusco, “The road to 6G: Ten physical layer challenges for communications engineers,” IEEE Communications Magazine, vol. 59, no. 1, pp. 64–69, 2021.arxiv.org/abs/2004.07130
  8. [8]L. Clavier, T. Pedersen, I. R. Larrad, and M. Egan, “Alpha-stable model for interference in IoT networks,” in IEEE Conference on Antenna Measurements and Applications, 2021, pp. 575–578.DOI
  9. [9]A. Buvarp and S. Werner, “Robust complex-valued federated learning for secure 6G mobile communications,” TechRxiv, Jul. 2025, 175339544.49904893/v1.DOI
  10. [10]C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423, 1948.
  11. [11]V. Kostina and S. Verdú, “Lossy joint source-channel coding in the finite blocklength regime,” in IEEE International Symposium on Information Theory Proceedings, 2012, pp. 1553–1557.
  12. [12]F. Hekland, P. Floor, and T. Ramstad, “Shannon-Kotelnikov mappings in joint source-channel coding,” IEEE Transactions on Communications, vol. 57, no. 1, pp. 94–105, 2009.DOI
  13. [13]V. A. Vaishampayan and S. R. Costa, “Curves on a sphere, shift-map dynamics, and error control for continuous alphabet sources,” IEEE Transactions on Information Theory, vol. 45, no. 5, pp. 1652–1666, Jul. 1999.
  14. [14]V. Vaishampayan and S. Costa, “Curves on a sphere, shift-map dynamics, and error control for continuous alphabet sources,” IEEE Transactions on Information Theory, vol. 49, no. 7, pp. 1658–1672, 2003.DOI
  15. [15]A. M. Buvarp, R. M. Taylor, K. M. Mishra, L. Mili, and A. Zaghloul, “Constant curvature curve tube codes for low-latency analog error correction,” IEEE Transactions on Information Theory, vol. 69, no. 12, pp. 7738–7754, 2023.arxiv.org/abs/2205.12332
  16. [16]A. M. Buvarp, L. Mili, and A. I. Zaghloul, “Robust constant curvature curve communications with complex and quaternion neural networks,” IEEE Transactions on Communications, vol. 72, no. 12, pp. 7787–7797, 2024.DOI
  17. [17]Y. Hu, J. Garcia-Frias, and M. Lamarca, “Analog joint source-channel coding using non-linear curves and mmse decoding,” IEEE Transactions on Communications, vol. 59, no. 11, pp. 3016–3026, 2011.DOI
  18. [18]E. Bourtsoulaze, D. B. Kurka, and D. Gündüz, “Deep joint source-channel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking, vol. 5, no. 3, pp. 567–579, 2019.
  19. [19]F. Hekland, G. Oien, and T. Ramstad, “Using 2:1 Shannon mapping for joint source-channel coding,” in Data Compression Conference, 2005, pp. 223–232.DOI
  20. [20]A. Campello, C. Torrezzan, and S. I. R. Costa, “Curves on torus layers and coding for continuous alphabet sources,” in IEEE International Symposium on Information Theory, 2012, pp. 2127–2131.DOI
  21. [21]O. Gonzalez and J. H. Maddocks, “Global curvature, thickness and the ideal shapes of knots,” Proceedings of the National Academy of Sciences, vol. 96, no. 9, pp. 4769–4773, 1999.pubmed.ncbi.nlm.nih.gov/10220368
  22. [22]S. I. R. Costa, “On closed twisted curves,” Proceedings of the American Mathematical Society, vol. 109, no. 1, pp. 205–214, 1990.
  23. [23]G. Chen, X. Liu, Q. Meng, L. Chen, C. Liu, and Y. Li, “Learning neural operators on Riemannian manifolds,” National Science Open, vol. 3, no. 6, p. 20240001, 2024.
  24. [24]B. Kravtsov and D. Raphaeli, “Multidimensional analog modulation and optimal mappings,” in IEEE Global Communications Conference, 2007, pp. 1426–1430.DOI
  25. [25]T. Strohmer and R. Heath, “Grassmannian frames with applications to coding and communication,” Applied and Computational Harmonic Analysis, vol. 14, pp. 257–275, 05 2003.arxiv.org/abs/math/0301135
  26. [26]J. Masci, D. Boscaini, M. M. Bronstein, and P. Vандергheynst, “Geodesic convolutional neural networks on riemannian manifolds,” in IEEE International Conference on Computer Vision Workshop, 2015, pp. 832–840.arxiv.org/abs/1501.06297
  27. [27]F. Monti, D. Boscaini, J. Masci, E. Rodolà, J. Svoboda, and M. M. Bronstein, “Geometric deep learning on graphs and manifolds using mixture model CNNs,” in IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 5425–5434.arxiv.org/abs/1611.08402
  28. [28]F. Yu and V. Koltun, “Multi-scale context aggregation by dilated convolutions,” in International Conference on Learning Representations, 05 2016.arxiv.org/abs/1511.07122
  29. [29]M. Poli, S. Massaroli, C. M. Rabideau, J. Park, A. Yamashita, H. Asama, and J. Park, “Continuous-Depth Neural Models for Dynamic Graph Prediction,” preprint arXiv:2106.11581, 2021.arxiv.org/abs/2106.11581
  30. [30]L. Ma, S. Pal, Y. Zhang, J. Zhou, Y. Zhang, and M. Coates, “Ckgconv: General graph convolution with continuous kernels,” in International Conference on Machine Learning, 07 2024.arxiv.org/abs/2404.13604
  31. [31]L. M. H. Shahb, A. Rizaner, A. H. Ulusoy, and H. Amca, “Impact of impulsive noise on millimeter wave cellular systems performance,” in UK-Europe-China Workshop on Millimetre Waves and Terahertz Technologies, 2017, pp. 1–4.
  32. [32]L. Clavier, G. Peters, F. Septier, and I. Nevat, “Impulsive noise modeling and robust receiver design,” EURASIP Journal on Wireless Communications and Networking, no. 13, 2021.DOI
  33. [33]C. L. Nikias and M. Shao, Signal Processing with Alpha-Stable Distributions and Applications. New York, NY, USA: Wiley-Interscience, 1995.
  34. [34]J. P. Nolan, Univariate Stable Distributions: Models for Heavy Tailed Data. Cham, Switzerland: Springer, 2020.DOI
  35. [35]P. J. Huber, “Robust estimation of a location parameter,” The Annals of Mathematical Statistics, vol. 35, no. 1, pp. 73–101, 1964.
  36. [36]J. Zhao, M. Netto, and L. Mili, Robust Dynamic State Estimation of Power Systems, 1st ed. New York, NY: Elsevier, 2024.
  37. [37]A. J. Stam, “Some inequalities satisfied by the quantities of information of fisher and shannon,” Information and Control, vol. 2, no. 2, pp. 101–112, 1959.
  38. [38]A. Dembo, T. M. Cover, and J. A. Thomas, “Information theoretic inequalities,” IEEE Transactions on Information Theory, vol. 37, no. 6, pp. 1501–1518, Nov. 1991.DOI
  39. [39]S. Verdu and D. Guo, “A simple proof of the entropy-power inequality,” IEEE Transactions on Information Theory, vol. 52, no. 5, pp. 2165–2166, 2006.DOI
  40. [40]R. G. Gallager, Information Theory and Reliable Communication. New York: John Wiley & Sons, 1968.
  41. [41]T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley, 2006.
  42. [42]U. Erez and R. Zamir, “Achieving 1/2 log (1 + SNR) on the AWGN channel with lattice encoding and decoding,” IEEE Transactions on Information Theory, vol. 50, no. 10, pp. 2293–2314, 2004.
  43. [43]T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. Hoboken, NJ, USA: Wiley-Interscience, 2006.
  44. [44]T. Berger, Rate Distortion Theory: A Mathematical Basis for Data Compression. Englewood Cliffs, NJ, USA: Prentice-Hall, 1971.
  45. [45]H. Robbins and S. Monro, “A stochastic approximation method,” The Annals of Mathematical Statistics, vol. 22, no. 3, pp. 400–407, 1951.
  46. [46]J. Kiefer and J. Wolfowitz, “Stochastic estimation of the maximum of a regression function,” The Annals of Mathematical Statistics, vol. 23, no. 3, pp. 462–466, 1952.
  47. [47]J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,” IEEE Transactions on Automatic Control, vol. 37, no. 3, pp. 332–341, 1992.DOI
  48. [48]T. S. Cohen and M. Welling, “Group equivariant convolutional networks,” in Proceedings of the International Conference on Machine Learning, 2016.arxiv.org/abs/1602.07576
  49. [49]T. S. Cohen, M. Weiler, B. Kicanaoglu, and M. Welling, “Gauge equivariant convolutional networks and the icosahedral cnn,” in Proceedings of the International Conference on Machine Learning, 2019.
  50. [50]K. V. Mardia and P. E. Jupp, Directional Statistics. John Wiley & Sons, 2000.DOI
  51. [51]Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljacic, T. Y. Hou, and M. Tegmark, “KAN: Kolmogorov–Arnold networks,” in The Thirteenth International Conference on Learning Representations, 2025.DOI
  52. [52]F. Gray, “Pulse code communication,” U.S. Patent 2,632,058, Mar. 1953, issued Mar. 17, 1953.
  53. [53]Y. Cui, M. Jia, T.-Y. Lin, Y. Song, and S. Belongie, “Class-balanced loss based on effective number of samples,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 9268–9277.arxiv.org/abs/1901.05555
  54. [54]A. M. Buvarp, L. Mili, and J. Fishbone, “Robust media-based modulation with an Einsenstein constellation generated by a reconfigurable intelligent surface with blind equalization and complex-valued neural receivers,” IEEE Transactions Wireless Communications, vol. 24, no. 5, pp. 3794–3804, 2025.DOI
  55. [55]J. G. F. Souza, S. I. R. Costa, and C. Ling, “Multilevel lattice codes from hurwitz quaternion integers,” IEEE Transactions on Information Theory, vol. 71, no. 11, pp. 8280–8293, 2025.arxiv.org/abs/2401.10773
  56. [56]L. Natarajan, Y. Hong, and E. Viterbo, “Lattice index coding,” IEEE Transactions on Information Theory, vol. 61, no. 12, pp. 6505–6525, 2015.DOI

Paper details

vlnc:W_hg7bvwf9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
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