Coordinated Harmonic Superposition (CHS)

• CHS is an additive, constraint-based approach that combines harmonic components via rules like gating, phase matching, and geometric coordination.
• It underpins architectures such as CHOIR in implicit neural representations and is applied in domains from turbulence modeling to nonlinear optics.
• Empirical validations demonstrate improved depth scalability, stable training dynamics, and efficient harmonic cancellation in vibration and wave-generation experiments.

Coordinated Harmonic Superposition (CHS) is an additive, constraint-structured view of harmonic synthesis in which multiple harmonic components are combined under a shared coordination rule so that the superposition realizes a target physical or computational effect. In the strictest usage presently represented in the literature, CHS is the architectural core of the Calibrated Harmonic Overlaid Implicit Neural Representation (CHOIR), where it replaces conventional deep function composition in periodic implicit neural representations by a residual-style harmonic overlay (Chen et al., 25 Jun 2026). A broader reading, explicitly suggested by several related works, is that CHS also describes a family of constructions in which harmonics are coordinated by geometry, symmetry, common complex dilatation, feedback control, or source-phase alignment rather than being superposed arbitrarily. This broader interpretation is not a single unified theory, but a cross-domain pattern visible in wall turbulence modeling, shaker excitation harmonization, harmonic surface theory, nonlinear potential theory, and microscopic harmonic generation (Liu, 2017, Hippold et al., 2024, Vasu, 4 May 2026, Brustad, 2016, Hardhienata, 2012).

1. Definition and conceptual scope

In CHOIR, CHS is introduced to replace the conventional function composition used in most periodic INRs, motivated by the claim that the signals being represented are more naturally understood as additive superpositions of harmonics than as deeply nested periodic nonlinearities (Chen et al., 25 Jun 2026). The core recurrence is

$$h_{l+1}=h_l+\beta_l F_l(h_l), \qquad l=0,1,\dots,L-1,$$

with initialization

$h_0(v)=W^{(0)}v+b^{(0)},$

and output

$P_\Theta(v)=W_{\text{out} h_L(v)$

or, after unrolling,

$P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$

Here the harmonic terms are not merely added; they are coordinated by learnable scalar gates $\beta_l$, initialized to zero, so that the network begins as a simple linear mapping and turns on harmonic modules progressively (Chen et al., 25 Jun 2026).

A broader, cross-domain interpretation is strongly supported by several papers that do not use the CHS term but instantiate comparable structures. In wall turbulence, a fixed-end, even-mode standing-wave family is coordinated by common boundary conditions, equal amplitudes, common sine phase, scale ordering, and envelope extraction (Liu, 2017). In shaker testing, higher harmonics are intentionally added to a shaker voltage command, and their complex coefficients are coordinated by harmonic-wise PI feedback until the applied excitation becomes purely sinusoidal at the drive point or base (Hippold et al., 2024). In harmonic surface theory, exact superposition is preserved when planar harmonic parts share the same complex dilatation $\nu$ (Vasu, 4 May 2026). In the $p$-Laplace equation, only a highly structured class of positive superpositions of translated fundamental solutions survives the nonlinearity, and then only as $p$-superharmonicity rather than exact $p$-harmonic closure (Brustad, 2016). In nonlinear optics, macroscopic harmonic fields are derived as the coherent sum of many microscopic anharmonic dipoles whose phases are coordinated by the pump and propagation (Hardhienata, 2012).

This suggests that CHS is best understood not as unrestricted Fourier addition, but as superposition under a compatibility condition. Depending on the problem, that condition may be gating, parity, common dilatation, nonnegative coefficients, harmonic-domain feedback, or phase matching.

2. CHS as an INR architecture

The most explicit formalization of CHS appears in CHOIR, which studies INR-based recovery of multidimensional data such as RGB images, multispectral images, hyperspectral images, and videos (Chen et al., 25 Jun 2026). The baseline periodic INR is written as

$$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$

while the harmonic target form is

$h_0(v)=W^{(0)}v+b^{(0)},$0

The paper’s thesis is that deep periodic networks and generalized Fourier series are fundamentally connected, yet the former are optimized through nested composition whereas the latter are additive expansions; CHS is proposed to reduce this mismatch (Chen et al., 25 Jun 2026).

The coordination mechanism is twofold. First, each harmonic module $h_0(v)=W^{(0)}v+b^{(0)},$1 is multiplied by a learnable scalar $h_0(v)=W^{(0)}v+b^{(0)},$2, initialized to zero. Second, the gate update is coupled to the downstream gradient through

$h_0(v)=W^{(0)}v+b^{(0)},$3

which the paper interprets as an implicit curriculum learning mechanism: a harmonic term activates only when it aligns with the useful descent direction (Chen et al., 25 Jun 2026). The paper explicitly states that at the beginning of training, the network degenerates into a simple linear mapping $h_0(v)=W^{(0)}v+b^{(0)},$4, and argues that this yields a highly stable Jacobian.

Within CHOIR, CHS is paired with Perceptual Spectrum Calibration (PSC), which defines the internal form of each $h_0(v)=W^{(0)}v+b^{(0)},$5. Frequencies are bounded by

$h_0(v)=W^{(0)}v+b^{(0)},$6

and

$h_0(v)=W^{(0)}v+b^{(0)},$7

with per-neuron frequencies assigned by

$h_0(v)=W^{(0)}v+b^{(0)},$8

The geometric mean frequency is

$h_0(v)=W^{(0)}v+b^{(0)},$9

and the $P_\Theta(v)=W_{\text{out} h_L(v)$0-th neuron of $P_\Theta(v)=W_{\text{out} h_L(v)$1 is defined as

$P_\Theta(v)=W_{\text{out} h_L(v)$2

The amplitude factor is motivated by the power-law spectrum prior

$P_\Theta(v)=W_{\text{out} h_L(v)$3

with $P_\Theta(v)=W_{\text{out} h_L(v)$4 initialized to $P_\Theta(v)=W_{\text{out} h_L(v)$5 (Chen et al., 25 Jun 2026).

The main reported effect of CHS is depth scalability. The paper states that methods such as SIREN, FINER, and FreSh peak at relatively shallow depth and degrade when made deeper, while CHOIR improves as depth increases and attains its best PSNR around 15 layers, with 12 layers chosen in experiments for a performance-efficiency trade-off (Chen et al., 25 Jun 2026). In the MSI Flowers ablation, CHS alone raises the Sine baseline from PSNR $P_\Theta(v)=W_{\text{out} h_L(v)$6, SSIM $P_\Theta(v)=W_{\text{out} h_L(v)$7, LPIPS $P_\Theta(v)=W_{\text{out} h_L(v)$8 to PSNR $P_\Theta(v)=W_{\text{out} h_L(v)$9, SSIM $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$0, LPIPS $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$1 for random missing completion at OR $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$2, and from PSNR $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$3, SSIM $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$4, LPIPS $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$5 to PSNR $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$6, SSIM $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$7, LPIPS $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$8 for mixed degradation on Scene 3 (Chen et al., 25 Jun 2026). The parameter increase is minimal: $P_\Theta(v)=W_{\text{out}\!\left(h_0(v)+\sum_l \beta_l F_l(h_l(v))\right).$9M for the Sine baseline, $\beta_l$0M for Sine+CHS, and $\beta_l$1M for CHOIR, with per-iteration times $\beta_l$2 s, $\beta_l$3 s, and $\beta_l$4 s respectively (Chen et al., 25 Jun 2026).

3. Geometry-, symmetry-, and envelope-coordinated superposition

A closely related construction appears in "Generating the Log Law of the Wall with Superposition of Standing Waves" (Liu, 2017). The paper does not use the term CHS, but its construction is explicitly a highly structured harmonic-superposition model in which fixed-end standing waves are treated as surrogates for multiscale turbulent shear stresses, and the envelope of their superposition is interpreted as the mean shear-stress profile that generates the logarithmic mean velocity profile.

The full standing-wave form is

$\beta_l$5

with

$\beta_l$6

and, in the implementation,

$\beta_l$7

For the main construction, only even harmonic modes

$\beta_l$8

are used (Liu, 2017). The temporal factor is then reduced phenomenologically so that the working superposition becomes purely spatial,

$\beta_l$9

The key inference layer is the envelope-based approximation

$\nu$0

where the envelope is determined by spline interpolation over local maxima and the maximum is taken over $\nu$1 (Liu, 2017).

The coordination rule here is geometric rather than dynamical. All modes satisfy common fixed-end boundary constraints; only even $\nu$2 are retained because they are symmetric with respect to $\nu$3; all harmonics have the same magnitude; all spatial modes share the same sine-phase convention; modes are ordered by $\nu$4; and the physically meaningful object is the smooth envelope rather than the raw oscillatory sum (Liu, 2017). The paper explicitly remarks: “Why nature favors only such even-mode harmonic waves is open for discovery elsewhere.”

The envelope is further adjusted to satisfy wall and centerline conditions. Near the centerline, the paper uses the sum of upper and lower envelopes so that

$\nu$5

and near the wall it clips the envelope to unity to impose

$\nu$6

Numerical integration then recovers the mean velocity profile by

$\nu$7

Using $\nu$8 in the main example, the resulting semi-log MVP exhibits a log-law region (Liu, 2017).

This standing-wave realization is also notable for finite-mode effects. The paper reports that for roughly $\nu$9, the MVP curvature profiles are self-similar; for $p$0, curvature changes; for $p$1, the curvature is “rather random”; at $p$2, a point of inflection appears; at $p$3, the profile becomes laminar-like with constant negative curvature; and for $p$4, interpretation is difficult, suggesting Eq. (7) may only be valid for $p$5 (Liu, 2017). A plausible implication is that, in this model class, regime structure depends on the number and ordering of coordinated harmonics rather than on broadband content alone.

4. Feedback-coordinated harmonic cancellation and waveform control

A different realization of CHS-like coordination is given by "An iteration-free approach to excitation harmonization" (Hippold et al., 2024). The paper addresses shaker-based nonlinear vibration testing, where a nominally sinusoidal command generates higher harmonics in the actually applied excitation because of nonlinear exciter-structure interaction. Its solution is to superpose compensation harmonics in the shaker voltage and coordinate their amplitudes and phases through harmonic-wise feedback so that the measured applied excitation becomes purely sinusoidal.

The command voltage is written as a multi-harmonic signal,

$p$6

where $p$7 is the complex Fourier coefficient of the fundamental command, $p$8 are the intentionally injected higher harmonics, $p$9 is the harmonization truncation order, and $p$0 is the fundamental phase with $p$1 (Hippold et al., 2024). For force excitation, each harmonic is regulated by

$p$2

for $p$3. The desired higher harmonics are zero, so the error is simply the negative of the measured harmonic coefficient (Hippold et al., 2024).

The harmonic coefficients are estimated in real time by a continuous-time LMS-type adaptive filter,

$p$4

for $p$5 (Hippold et al., 2024). The paper emphasizes that this makes the method iteration-free: instead of solving a sequence of Fourier-domain root-finding problems by Newton-type updates, it closes the loop directly in harmonic space.

The stability analysis begins from a coupled structure-exciter model. Under periodic-response and single-resonant-harmonic assumptions, the paper derives the simplified harmonic integrator dynamics

$p$6

where

$p$7

The fixed point corresponding to perfect suppression is $p$8, equivalently $p$9, which gives

$p$0

The asymptotic stability condition for the simplified linear case is

$p$1

In CHS terms, $p$2 is the coordinated compensation harmonic and $p$3 is the target cancellation condition (Hippold et al., 2024).

The reported validation is strong. In the virtual force-excitation experiment based on a beam with a cubic spring near a $p$4 internal resonance, the authors use $p$5, $p$6, and dimensionless gains $p$7, $p$8. Without harmonization, the third force harmonic reaches about 50% of the fundamental; with harmonization, higher harmonics are reduced by several orders of magnitude, largely below 0.01% of the target fundamental force, except in a regime where the orbit loses periodicity via a torus bifurcation (Hippold et al., 2024). In the real base-excitation experiment on a doubly clamped beam near a $p$9 internal resonance, using a dSpace MicroLabBox at 10 kHz with $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$0, $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$1, $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$2, $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$3, and $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$4, the non-fundamental Fourier coefficients of the base excitation are reduced to below 0.06% of the fundamental, effectively the noise floor (Hippold et al., 2024). Compared with an iterative Newton/Broyden method, the feedback method requires about 6 s per frequency point versus 37 s per point and achieves lower residual harmonic distortion (Hippold et al., 2024).

5. Exact, restricted, and geometric superposition principles

Several papers show that harmonic superposition survives outside linear Fourier settings only when compatibility conditions are imposed. In "Superposition in the $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$5-Laplace Equation," finite sums of translated fundamental solutions

$$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$6

satisfy, for $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$7,

$$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$8

so $$P_\Theta(v)=\phi_L \cdots \phi_1(v), \qquad \phi_l(x)=p^{(l)}(W^{(l)}x+b^{(l)}),$$9 wherever defined (Brustad, 2016). The sum is therefore generally $h_0(v)=W^{(0)}v+b^{(0)},$00-superharmonic rather than $h_0(v)=W^{(0)}v+b^{(0)},$01-harmonic. The coordination condition is stringent: the kernels must be translated copies of the fundamental solution, the coefficients must satisfy $h_0(v)=W^{(0)}v+b^{(0)},$02, and a concave term $h_0(v)=W^{(0)}v+b^{(0)},$03 may be added while preserving

$h_0(v)=W^{(0)}v+b^{(0)},$04

for $h_0(v)=W^{(0)}v+b^{(0)},$05 (Brustad, 2016). This is not arbitrary superposition; it is a sign-definite interaction mechanism controlled by kernel identity, coefficient sign, and the angular defect term $h_0(v)=W^{(0)}v+b^{(0)},$06.

An exact geometric superposition theorem is proved in "Superposition of Harmonic Surfaces: Helical Motifs in Lamellar Structures" (Vasu, 4 May 2026). A harmonic Enneper immersion is written as

$h_0(v)=W^{(0)}v+b^{(0)},$07

with

$h_0(v)=W^{(0)}v+b^{(0)},$08

If harmonic Enneper immersions $h_0(v)=W^{(0)}v+b^{(0)},$09 have planar parts with the same complex dilatation

$h_0(v)=W^{(0)}v+b^{(0)},$10

then for real constants $h_0(v)=W^{(0)}v+b^{(0)},$11,

$h_0(v)=W^{(0)}v+b^{(0)},$12

is again a harmonic Enneper immersion (Vasu, 4 May 2026). This is an exact superposition principle, but only under the common-dilatation coordination rule.

The same paper develops harmonic graph constructions for helical motifs,

$h_0(v)=W^{(0)}v+b^{(0)},$13

and finite superpositions

$h_0(v)=W^{(0)}v+b^{(0)},$14

For a finite collection of helical charges $h_0(v)=W^{(0)}v+b^{(0)},$15 at $h_0(v)=W^{(0)}v+b^{(0)},$16, the asymptotic expansion is

$h_0(v)=W^{(0)}v+b^{(0)},$17

and the paper explicitly concludes that a finite collection of helical motifs behaves asymptotically like a single helicoid whose pitch equals the sum of the individual pitches (Vasu, 4 May 2026). The same framework encodes infinite arrays through

$h_0(v)=W^{(0)}v+b^{(0)},$18

for a TGB phase, and

$h_0(v)=W^{(0)}v+b^{(0)},$19

for a $h_0(v)=W^{(0)}v+b^{(0)},$20 TGB configuration (Vasu, 4 May 2026). Here superposition remains exact at the harmonic-function level, while the geometric interpretation concerns layered structures with defect arrays.

Taken together, these works show three distinct superposition regimes: exact linear superposition under a common geometric invariant; restricted nonlinear superposition as a one-sided supersolution principle; and asymptotic aggregation in which many local harmonic motifs collapse to an effective far-field resultant.

6. Harmonic generation, selective enhancement, and source superposition

In wave- and quantum-generation problems, CHS-like behavior appears when harmonic outputs are built from coordinated emitters or coordinated quantum channels. In "Generation, reflection and transmission of nonlinear harmonic waves by direct superposition of anharmonic dipoles," the electric field radiated by many nonlinear dipoles in a medium is summed directly, yielding

$h_0(v)=W^{(0)}v+b^{(0)},$21

For transmission through a depth $h_0(v)=W^{(0)}v+b^{(0)},$22,

$h_0(v)=W^{(0)}v+b^{(0)},$23

The coherent buildup factor

$h_0(v)=W^{(0)}v+b^{(0)},$24

is the microscopic source-summation origin of phase matching and the familiar $h_0(v)=W^{(0)}v+b^{(0)},$25-type envelope (Hardhienata, 2012). In this formulation, source coordination comes from the common pump-induced phase $h_0(v)=W^{(0)}v+b^{(0)},$26, and constructive macroscopic harmonic buildup occurs only when the dipole phases remain aligned over depth.

A quantum-channel analogue appears in "Two states hydrogenlike model for High-Order Harmonic Generation and enhanced XUV Generation from a coherent superposition of bound states" (Batebi et al., 2011). The wavefunction ansatz is

$h_0(v)=W^{(0)}v+b^{(0)},$27

and the HHG dipole contains four bound-continuum-bound channels: ground $h_0(v)=W^{(0)}v+b^{(0)},$28 continuum $h_0(v)=W^{(0)}v+b^{(0)},$29 ground, excited $h_0(v)=W^{(0)}v+b^{(0)},$30 continuum $h_0(v)=W^{(0)}v+b^{(0)},$31 excited, ground $h_0(v)=W^{(0)}v+b^{(0)},$32 continuum $h_0(v)=W^{(0)}v+b^{(0)},$33 excited, and excited $h_0(v)=W^{(0)}v+b^{(0)},$34 continuum $h_0(v)=W^{(0)}v+b^{(0)},$35 ground (Batebi et al., 2011). The channel-resolved dipole terms are of the form

$h_0(v)=W^{(0)}v+b^{(0)},$36

and the paper states that the main contribution to high conversion efficiency is the transition from the excited to the ground state (Batebi et al., 2011). In CHS language, the emitted harmonic field is not a single-channel process but a coordinated coherent sum of quantum pathways, with enhancement arising from interference and from the asymmetric roles of launch and recombination states.

A control-theoretic version of selective harmonic coordination is given in "Optimal control theory of harmonic generation" (Schaefer et al., 2012). The frequency-domain objective is

$h_0(v)=W^{(0)}v+b^{(0)},$37

with spectral admissibility enforced by

$h_0(v)=W^{(0)}v+b^{(0)},$38

The field update is

$h_0(v)=W^{(0)}v+b^{(0)},$39

implemented by the relaxation step

$h_0(v)=W^{(0)}v+b^{(0)},$40

Although the paper focuses mainly on a specific harmonic order, it explicitly notes that the target filter $h_0(v)=W^{(0)}v+b^{(0)},$41 need not be a single narrow peak; a plausible implication is that multi-peak spectral weights implement coordinated enhancement of several harmonics from a band-limited drive (Schaefer et al., 2012).

Across these generation problems, the common structure is the same: harmonics are neither independent nor freely assigned. They are coordinated by phase matching, coherent channel interference, or spectral targeting, and the macroscopic effect depends on whether those constraints produce constructive rather than destructive superposition.

7. Interpretation, limitations, and recurrent misconceptions

A recurrent misconception is to treat CHS as synonymous with arbitrary Fourier decomposition. The literature summarized here does not support that interpretation. In every domain where a substantive result is obtained, the harmonic family is constrained. In CHOIR, the constraints are residual overlay, zero-initialized scalar gates, and PSC-defined modules (Chen et al., 25 Jun 2026). In the standing-wave turbulence model, they are fixed-end geometry, even-mode parity, equal amplitudes, common sine phase, and envelope extraction (Liu, 2017). In excitation harmonization, they are harmonic-wise feedback laws and measured residual cancellation (Hippold et al., 2024). In harmonic surfaces, they are common complex dilatation and immersion conditions (Vasu, 4 May 2026). In the $h_0(v)=W^{(0)}v+b^{(0)},$42-Laplace case, they are identical translated kernels with nonnegative weights and, optionally, a concave background (Brustad, 2016).

A second misconception is to assume that superposition always preserves the same kind of solution. The $h_0(v)=W^{(0)}v+b^{(0)},$43-Laplace result explicitly shows otherwise: positive superpositions of fundamental solutions are generally $h_0(v)=W^{(0)}v+b^{(0)},$44-superharmonic, not $h_0(v)=W^{(0)}v+b^{(0)},$45-harmonic (Brustad, 2016). The wall-turbulence standing-wave model similarly does not derive its envelope-to-shear identification from Reynolds averaging; the paper presents that step as an interpretive modeling assumption rather than a first-principles closure (Liu, 2017). The excitation-harmonization method assumes near-periodic operation and can degrade in strongly non-periodic regimes because the truncated Fourier basis and adaptive estimator then become inadequate (Hippold et al., 2024). CHOIR provides analytic and empirical support for CHS as an optimization-stabilizing architecture, but the main text does not present a theorem-level convergence or approximation guarantee (Chen et al., 25 Jun 2026).

A third misconception is to treat “coordination” as necessarily temporal phase locking. Several papers use a different notion. In the turbulence model, coordination is mainly geometric and parity-based (Liu, 2017). In harmonic surfaces, it is a compatibility invariant in complex dilatation (Vasu, 4 May 2026). In nonlinear optics, it is source-phase alignment under $h_0(v)=W^{(0)}v+b^{(0)},$46 (Hardhienata, 2012). In shaker testing, it is adaptive harmonic-domain error correction (Hippold et al., 2024). This suggests that CHS is not a single mechanism but a family resemblance centered on structured harmonic compatibility.

The most defensible generalization, therefore, is that CHS denotes coordinated additive harmonic synthesis in which the admissible harmonic family, coefficient law, and physical observable are specified together. Where those three ingredients are explicit, CHS can function as an architecture, a control law, a geometric theorem, a nonlinear supersolution principle, or a microscopic wave-generation model. Where they are absent, the term risks collapsing into a vague synonym for “using harmonics.”