Mode Selection: Principles and Applications

Updated 4 July 2026

Mode selection is the controlled preference of distinct intrinsic or operational modes, achieved by engineering thresholds, sparse reconstructions, or discrete associations.
It spans diverse applications, including laser cavity design, plasmonic antennas, fluid dynamics, and communication systems, reflecting its interdisciplinary impact.
Methodologies range from eigenmode analysis and greedy algorithm selection to graph-based and optically tuned approaches, optimizing parameters like stability, rate, and secrecy.

Searching arXiv for the supplied topic and papers to ground the article in current records. Search 1: general query for "mode selection". Search 2: verify the laser paper (Bitauld et al., 2010). Search 3: verify representative cross-domain papers used in the article. Mode selection is a domain-dependent term for the controlled preference of one mode, or one small subset of modes, among several admissible alternatives. In the surveyed literature, the term appears in two main senses. In one sense, it denotes the selection of intrinsic modes of a physical or mathematical system, such as longitudinal Fabry–Perot modes, plasmonic bonding and antibonding modes, dynamic modes in reduced-order modeling, aeroelastic vibration modes, or eigenmodes of a game-dynamical Jacobian (Bitauld et al., 2010, Huang et al., 2010, Ohmichi, 2017, Li et al., 2020, Wang et al., 2022). In the other sense, it denotes the choice among operating modes of a communication or coding system, such as direct versus cellular transmission, skip versus coded video reconstruction, or the number of simultaneously served users in MU-MIMO (Liu, 2016, Ladune et al., 2020, Chen et al., 2015). Across these settings, the common technical problem is to map a high-dimensional set of admissible behaviors to a smaller, controllable subset that optimizes threshold, rate, reconstruction error, secrecy, stability, or hardware cost.

1. Conceptual scope and mathematical formulations

In intrinsic-mode settings, mode selection is usually tied to an operator spectrum. A linearized dynamical system near a fixed point takes the form $\dot{y}=Jy$ , and the eigenpairs $(\lambda_i,\xi_i)$ of the Jacobian define eigenmodes. The resulting motion decomposes as $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ , so mode selection becomes the question of which eigenmodes dominate the observed trajectory (Wang et al., 2022). In DMD, the same logic appears in data-driven form: snapshots satisfy $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ , and selection means finding a small support set of amplitudes $\alpha_i$ that preserves reconstruction quality (Ohmichi, 2017). In laser physics, the relevant modes are cavity resonances, and selection is phrased in threshold terms: a cavity is engineered so that only a prescribed subset of modes has sufficiently low threshold gain to lase (Bitauld et al., 2010).

In operational-mode settings, the “mode” is a protocol or transmission state rather than an eigenfunction. In D2D-enabled cellular systems, a user chooses between direct D2D mode and cellular mode through a base station, often encoded by a discrete variable $x_{ij}\in\{0,1\}$ over the set $\mathcal{B}^+=\{0,1,\ldots,N\}$ , where $j=0$ denotes D2D and $j\in\mathcal{B}$ denotes association with BS $j$ (Liu, 2016). In learning-based video coding, the mode variable is a pixel-wise weighting matrix $(\lambda_i,\xi_i)$ 0 that arbitrates between skip/copy from motion-compensated prediction and transmission through a conditional codec (Ladune et al., 2020). In MU-MIMO secure downlink, the transmission mode is simply the number of beams $(\lambda_i,\xi_i)$ 1, so mode selection means choosing how many users are simultaneously served (Chen et al., 2015).

Domain	“Mode”	Selection variable
Laser cavities	Longitudinal cavity modes	Threshold-gain shaping (Bitauld et al., 2010)
Plasmonics	Bonding / antibonding eigenmodes	Excitation symmetry and spectral tuning (Huang et al., 2010)
DMD	Dynamic modes	Sparse support of $(\lambda_i,\xi_i)$ 2 (Ohmichi, 2017)
Video coding	Skip vs CodecNet	Pixel-wise $(\lambda_i,\xi_i)$ 3 map (Ladune et al., 2020)
D2D networks	Direct vs cellular	Association/mode variable (Liu, 2016)
MU-MIMO security	Number of active beams	Integer mode $(\lambda_i,\xi_i)$ 4 (Chen et al., 2015)

This split suggests that mode selection is less a single method than a recurrent design principle: identify a structured family of alternatives, then impose a selection rule that biases the system toward one element of that family or toward a sparse combination.

2. Optical cavities and active gain engineering

In semiconductor lasers, mode selection is most explicitly a threshold-engineering problem. A ridge-waveguide Fabry–Perot diode laser with cavity length $(\lambda_i,\xi_i)$ 5 was designed to support a discrete comb of modes by introducing 48 shallow-etched slots that pattern the effective index on one side of the cavity (Bitauld et al., 2010). The unperturbed Fabry–Perot spacing is about $(\lambda_i,\xi_i)$ 6, and the design selects six modes at $(\lambda_i,\xi_i)$ 7, spaced by twice the fundamental spacing. The resulting selected-mode spacing is about $(\lambda_i,\xi_i)$ 8, and passive mode locking with a $(\lambda_i,\xi_i)$ 9 saturable absorber adjacent to the HR-coated mirror yields pulses at $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 0 repetition rate. Experimentally, at a gain current of $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 1 and SA reverse bias near $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 2, the device produced a comb with Gaussian-fit FWHM spectral width $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 3, deconvolved pulsewidth $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 4, and time–bandwidth product $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 5, close to the transform-limited Gaussian value (Bitauld et al., 2010). In this setting, mode selection means that the cavity itself is engineered so that selected modes have lower threshold gain than all others.

A different laser-oriented formulation appears in active transformation optics. There, one begins from the SALT nonlinear dielectric with gain saturation and maps it to an equivalent unsaturated linear dielectric by constructing a transformed pump profile

$y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 6

with $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 7 (Ge, 2014). This establishes a correspondence between a saturated nonlinear lasing state and a linear threshold problem. The practical consequence is that a spatial pump profile can first equalize noninteracting thresholds of several competing modes and then be perturbed to lower the threshold of a chosen target mode. In a 1D slab laser, a higher-threshold mode that could not be selected by naive pumping proportional to $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 8 was made the first lasing mode; its actual threshold was reduced to $y(t)=\sum_i a_i e^{\lambda_i t}\xi_i$ 9, below the original lowest threshold (Ge, 2014). In a 2D diffusive random laser, the same construction selected one designated mode among six uniformly pumped lasing modes and produced single-mode emission. This suggests that in densely overlapping cavities, modal interactions themselves can be incorporated into the selection rule rather than treated merely as an obstacle.

3. Wave, plasmonic, and integrated-photonic realizations

In strongly coupled plasmonic nanoantennas, mode selection is governed by symmetry and detuning rather than lasing threshold. Symmetric gold dipole nanoantennas with 16 nm gaps support a bright bonding mode and a dark antibonding mode, analogous to coupled-oscillator supermodes (Huang et al., 2010). The bonding mode is even, has a field maximum at the gap, and couples efficiently to symmetric longitudinal excitation. The antibonding mode is odd, has a node at the gap, and is symmetry-forbidden for perfectly symmetric gap-centered excitation. By fixing the excitation wavelength at $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 0 and varying antenna length, the experiment tuned the laser spectrum first through the bonding resonance, then through a mixed regime, and finally predominantly through the antibonding resonance. TPPL maps changed correspondingly from a single centered spot to elongated profiles and then to a two-lobed pattern with a central node, directly imaging antibonding-mode selection (Huang et al., 2010). Simulations for a 236 nm antenna gave quality factors $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 1 and $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 2, making spectral discrimination between bonding and antibonding modes physically consequential.

In silicon photonics, the relevant operation is mode-selective phase shifting rather than mode conversion. A subwavelength-grating-based thermo-optic phase shifter on 220 nm SOI was designed so that TE $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 3 and TE $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 4 experience different effective thermo-optic coefficients because TE $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 5 remains concentrated in the central strip while TE $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 6 overlaps more strongly with the SWG side regions (Safaee et al., 2023). For an SWG duty cycle $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 7, the effective-medium model gives $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 8, and the measured mode selectivity was $\tilde{\bm x}_n=\sum_i \alpha_i \tilde{\bm\phi}_i \lambda_i^{n-1}$ 9, meaning the thermo-optic coefficient of TE $\alpha_i$ 0 was 44% larger than that of TE $\alpha_i$ 1 (Safaee et al., 2023). The phase shifter exhibited insertion loss at most $\alpha_i$ 2 and worst-case crosstalk of $\alpha_i$ 3 over a 40 nm wavelength range from 1520 to 1560 nm. Cascading this mode-selective phase shifter with a mode-insensitive thermo-optic phase shifter provides two degrees of freedom, allowing independent control of the relative phases of TE $\alpha_i$ 4 and TE $\alpha_i$ 5 (Safaee et al., 2023).

A more abstract spectral-control framework appears in non-Hermitian directed-graph networks. There, the graph connectivity condition $\alpha_i$ 6 supports a complete basis of geometry-protected pure decay modes with common exponential amplitude envelope and distinct phase windings (Liu et al., 15 May 2026). In fully connected graphs, one mode naturally emerges with a large tunable gap in the imaginary part of the spectrum. By adding synthetic gauge fields through phase-compensated non-reciprocal hopping, any chosen pure decay mode can be promoted to dominant status without changing its amplitude profile. The construction extends to paired-mode selection in half-connected graphs and to customizable multi-mode distributions through orthogonal folding in higher dimensions (Liu et al., 15 May 2026). This suggests a route to mode selection that is independent of gain/loss balancing and depends instead on graph geometry and gauge engineering.

4. Reduced-order modeling and coupled spatiotemporal dynamics

In large-scale data analysis, mode selection is a sparse model reduction problem. Preconditioned DMD with incremental POD first reduces high-dimensional snapshots to a low-dimensional basis $\alpha_i$ 7, forming $\alpha_i$ 8, then applies tlsDMD in the reduced space, and finally selects a sparse support of DMD amplitudes by approximately solving

$\alpha_i$ 9

where $x_{ij}\in\{0,1\}$ 0 (Ohmichi, 2017). The proposed greedy algorithm adds, at each step, the mode that most reduces the least-squares residual. In a 3D cylinder wake at $x_{ij}\in\{0,1\}$ 1, $x_{ij}\in\{0,1\}$ 2, $x_{ij}\in\{0,1\}$ 3, and reduced dimension $x_{ij}\in\{0,1\}$ 4, the method identified the mean flow $x_{ij}\in\{0,1\}$ 5, the primary shedding mode at $x_{ij}\in\{0,1\}$ 6, and lower-frequency oblique-vortex modes at $x_{ij}\in\{0,1\}$ 7 and $x_{ij}\in\{0,1\}$ 8 as the most important structures (Ohmichi, 2017). Reconstruction error decreased monotonically with the number of selected modes, and the greedy method achieved lower error than LASSO-based sparsity-promoting DMD for the same $x_{ij}\in\{0,1\}$ 9 on that dataset. Here, mode selection means choosing the subset of modes that best reconstructs the full time series under a cardinality constraint.

In fluid–structure interaction, mode selection emerges from lock-in between structural and flow-instability frequencies. A membrane wing at $\mathcal{B}^+=\{0,1,\ldots,N\}$ 0, $\mathcal{B}^+=\{0,1,\ldots,N\}$ 1, and angles of attack $\mathcal{B}^+=\{0,1,\ldots,N\}$ 2, $\mathcal{B}^+=\{0,1,\ldots,N\}$ 3, and $\mathcal{B}^+=\{0,1,\ldots,N\}$ 4 was analyzed with a global Fourier mode decomposition that stacked fluid vorticity, pressure, membrane displacement, and surface pressure difference into one joint state vector (Li et al., 2020). The dominant structural mode was consistently the chordwise second, spanwise first mode, and its structural natural frequency $\mathcal{B}^+=\{0,1,\ldots,N\}$ 5 was estimated with an approximate analytical formula including prestress, geometric nonlinearity, and added mass. The dominant coupled-system frequencies were $\mathcal{B}^+=\{0,1,\ldots,N\}$ 6 at $\mathcal{B}^+=\{0,1,\ldots,N\}$ 7, $\mathcal{B}^+=\{0,1,\ldots,N\}$ 8 at $\mathcal{B}^+=\{0,1,\ldots,N\}$ 9, and $j=0$ 0 at $j=0$ 1, with fluid and structural spectra showing synchronized peaks (Li et al., 2020). Lower non-integer frequency components such as $j=0$ 2 at $j=0$ 3 and $j=0$ 4 at $j=0$ 5 were associated with bluff-body-like vortex shedding rather than the lock-in branch. In this setting, mode selection means that one structural mode is preferentially amplified because its frequency and shape best match the forcing imposed by the separated shear layer.

5. Communications, coding, and network operating modes

In learning-based video coding, mode selection is a pixel-wise arbitration between copying a motion-compensated prediction and coding through a residual-generating network. MOFNet outputs optical flow $j=0$ 6, a mode-selection map $j=0$ 7, and a rate $j=0$ 8, and the reconstruction is

$j=0$ 9

where $j\in\mathcal{B}$ 0 and $j\in\mathcal{B}$ 1 is CodecNet (Ladune et al., 2020). The full system is trained with $j\in\mathcal{B}$ 2, using $j\in\mathcal{B}$ 3 as distortion. The mode map is soft rather than binary, which avoids blocking artifacts and encourages end-to-end learning of optical flow without a dedicated flow loss (Ladune et al., 2020). Ablations showed that forcing CodecNet everywhere increased rate by 53% BD-rate relative to the full system, while replacing conditional coding by residual coding increased rate by 32% BD-rate; the complete scheme performed on par with HEVC in the CLIC20 P-frame setting (Ladune et al., 2020). In this context, mode selection is not spectral but operational: each pixel chooses between skip and coded reconstruction.

In D2D-enabled cellular systems, mode selection usually means choosing direct communication or infrastructure-assisted transmission. In a multi-BS uplink formulation, each user selects $j\in\mathcal{B}$ 4, with $j\in\mathcal{B}$ 5 denoting D2D mode and $j\in\mathcal{B}$ 6 denoting association with BS $j\in\mathcal{B}$ 7 (Liu, 2016). The resulting joint problem of mode selection and user association is NP-complete in its original integer-program form, but it can be reformulated as a maximum weighted bipartite matching problem and solved centrally in polynomial time with the Hungarian algorithm (Liu, 2016). A distributed alternative uses dual prices $j\in\mathcal{B}$ 8 at BSs, so each user chooses

$j\in\mathcal{B}$ 9

with $j$ 0 for D2D mode (Liu, 2016). In stochastic-geometry analysis, a link-aware selection rule compares biased D2D and cellular path gains,

$j$ 1

so mode choice depends on both D2D distance and distance to the nearest BS, not only on $j$ 2 (ElSawy et al., 2014). That rule bounds D2D-to-BS interference by $j$ 3 under truncated channel inversion power control and yields lower cellular SINR outage and lower average D2D transmit power than distance-only selection at matched D2D intensity (ElSawy et al., 2014). In aerial networks, threshold D2D distance–based and RSS-based schemes similarly decide whether a Tx–Rx pair should use direct mode or aerial infrastructure; the D2D probability in the TDDS scheme depends non-monotonically on NFP altitude $j$ 4, while in the RSSS scheme it depends only on the D2D RSS threshold (1711.02220). From an effective-capacity perspective, a D2D link with fixed rate and fixed power plus direct/cellular mode selection becomes a Markov service process, and the effective capacity decays exponentially with increasing estimation error in the pathloss measurements used for mode selection (Shah et al., 2019).

Mode selection in secure MU-MIMO downlink has a different meaning: it is the choice of the number of simultaneous beams $j$ 5. In an OSDMA-style system with $j$ 6 transmit antennas, larger $j$ 7 increases inter-user interference at both legitimate users and the eavesdropper. That interference degrades legitimate channels but can also be exploited as anti-eavesdropping interference (Chen et al., 2015). The objective is to maximize the sum secrecy outage capacity $j$ 8 under a secrecy-outage constraint. Asymptotically, full mode $j$ 9 is optimal in the noise-limited regime, single-user mode $(\lambda_i,\xi_i)$ 00 is optimal in the interference-limited high-SNR regime, and full mode re-emerges as optimal when the number of secure users $(\lambda_i,\xi_i)$ 01 is large (Chen et al., 2015). In multi-mode pinching-antenna systems, the term again changes meaning: each pinching antenna either selects one guided mode for dominant radiation or combines several modes. Mode selection enforces $(\lambda_i,\xi_i)$ 02, while mode combining allows continuous tuning over $(\lambda_i,\xi_i)$ 03 (Xu et al., 9 Mar 2026). Joint optimization of digital beamforming, PA positions, and PA propagation constants showed that mode combining delivers the highest spectral efficiency, but mode selection approximates it with lower hardware complexity (Xu et al., 9 Mar 2026).

In pulsating stars, mode selection names an old and still unresolved nonlinear problem: of many linearly unstable normal modes predicted by stellar pulsation theory, only a fraction reaches observable amplitudes (Smolec, 2013). Linear growth rates are given by

$(\lambda_i,\xi_i)$ 04

with $(\lambda_i,\xi_i)$ 05, but growth rates do not determine final amplitudes or the winning mode set (Smolec, 2013). Non-resonant amplitude equations,

$(\lambda_i,\xi_i)$ 06

model self-saturation and cross-saturation, while resonant three-mode couplings account for parametric transfer of energy from unstable p-modes to damped g-modes (Smolec, 2013). In Cepheids and RR Lyrae stars, single-mode amplitude limitation is well captured by nonlinear hydrodynamic models, but robust physical explanations of F+1O and 1O+2O double-mode pulsation remain incomplete. In $(\lambda_i,\xi_i)$ 07 Sct and $(\lambda_i,\xi_i)$ 08 Cep stars, collective saturation and p–g resonant mode coupling reduce amplitudes, but neither mechanism alone yields a fully predictive selection theory (Smolec, 2013). Here, mode selection is inseparable from amplitude limitation and from nonlinear modal interactions.

Game dynamics provide a mathematically explicit social-science analogue. A five-strategy symmetric game was designed so that replicator dynamics around the interior Nash equilibrium has two complex eigenmodes, with eigenvalues $(\lambda_i,\xi_i)$ 09 and parameter-dependent imaginary parts controlled by a payoff parameter $(\lambda_i,\xi_i)$ 10 (Wang et al., 2022). Angular momentum measured in ten two-strategy subspaces yields an empirical vector $(\lambda_i,\xi_i)$ 11, and theory predicts that

$(\lambda_i,\xi_i)$ 12

where $(\lambda_i,\xi_i)$ 13 and $(\lambda_i,\xi_i)$ 14 are eigencycle sets induced by the two complex eigenvectors (Wang et al., 2022). In human-subject experiments with $(\lambda_i,\xi_i)$ 15, regression of $(\lambda_i,\xi_i)$ 16 on the two eigencycle sets produced very high fit quality, confirming the existence of two eigenmodes, and the estimated coefficients $(\lambda_i,\xi_i)$ 17 varied with $(\lambda_i,\xi_i)$ 18 in the direction predicted by the eigenvalue and myopic-response analysis (Wang et al., 2022). This suggests that mode selection is not confined to physical wave or oscillation systems; it can also describe the parameter-dependent dominance of one dynamical pattern over another in strategic behavior.

Across these natural and social examples, a common structure reappears. A system possesses multiple dynamically admissible modes, but only a subset becomes macroscopically visible. Whether the selector is nonlinear gain saturation, resonance locking, sparse reconstruction, a binary association rule, or a gauge-engineered phase pattern, mode selection is the mechanism that translates modal multiplicity into observed specificity.