Spectrum: Scientific and Mathematical Perspectives

• Spectrum is a multifaceted concept encompassing radio frequencies, optical transfer functions, and mathematical spectral spaces, defining core signal and data properties.
• In wireless communications, spectrum management employs dynamic access methods such as SURs, database-assisted trading, and fine-granular scheduling to mitigate interference.
• Advanced sensing and prediction techniques use persistent measurements and algorithmic models to achieve high accuracy in wideband recovery under challenging conditions.

Spectrum denotes several technically distinct objects across the sciences. In wireless communications it refers to the set of radio frequency bands over which electromagnetic signals are transmitted and received, spanning roughly from $8.3\ \mathrm{kHz}$ to $275\ \mathrm{GHz}$ in the International Telecommunication Union framework (Webb et al., 8 Mar 2025). In other research traditions the same term denotes spectral spaces associated with rings and lattice-ordered groups, spectral sets carrying exponential orthonormal bases, optical frequency responses, and wavenumber- or frequency-resolved descriptions of natural phenomena such as solar convection and the cosmic background radiation (Wehrung, 2017, Lai et al., 2019, Agin et al., 2024, Luo et al., 2023, Hathaway et al., 2015, Hill et al., 2018).

1. Radio spectrum as an engineered and regulated resource

In radio engineering and policy, spectrum is a finite shared national resource used by broadcasting, mobile broadband, satellite links, radar, and public safety communications (Webb et al., 8 Mar 2025). Spectrum management exists “to avoid interference between different users,” and the stakeholder set includes national regulatory authorities, the ITU and regional bodies such as CEPT and CITEL, mobile network operators, broadcasters, unlicensed device ecosystems such as Wi‑Fi, government users, and cloud or platform providers entering “Private 5G” (Webb et al., 8 Mar 2025). Traditional “command-and-control” management allocates bands to services under highly specific technical assumptions, including maximum transmit power, receiver sensitivity, target signal-to-noise ratio at service-area edges, channelization, duplexing, static guard bands, and minimum geographic separation distances; the paper describing this regime characterizes regulatory change as slow, often requiring $5$–$10$ years including international negotiations (Webb et al., 8 Mar 2025).

Recent work emphasizes more flexible regimes. “Spectrum Usage Rights” define property-like rights through limits on aggregate received power, typically expressed in Power-Flux Density units, rather than through transmitter-specific equipment rules (Webb et al., 8 Mar 2025). In that formulation, interference control is expressed through boundary constraints such as

$$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$

together with PFD relations such as

$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$

The same literature places liberalized trading and full shared spectrum access beside SURs as the other main mechanisms of reform, with CBRS in $3.5\ \mathrm{GHz}$, Licensed Shared Access in Europe, and $6\ \mathrm{GHz}$ coexistence between unlicensed Wi‑Fi and cellular networks treated as mature or emerging exemplars of database-assisted and dynamic access (Webb et al., 8 Mar 2025).

2. Measurement, occupancy, and event formation

Operational spectrum analysis now depends on persistent sensing pipelines rather than isolated measurements. A deployed U.S. multi-site framework uses remote antennas and Signal Hound BB60C spectrum scanners, local Python control, thresholded $24/7$ recording, minute-partitioned Parquet files, hourly JSON metadata, daily $7$-ZIP/LZMA compression, FTP transfer, centralized MATLAB/Python analytics, MySQL storage, and a Flask dashboard (Gandotra et al., 16 Jan 2026). In that framework, one month of compressed data across seven sites is approximately $275\ \mathrm{GHz}$0, and post-processing produces channel occupancy and airtime utilization with one-hour temporal granularity over $275\ \mathrm{GHz}$1 channels (Gandotra et al., 16 Jan 2026).

Occupancy is commonly reduced to binary labels. In a $275\ \mathrm{GHz}$2-day mid-band dataset built from continuous U.S. measurements, the monitored span was divided into $275\ \mathrm{GHz}$3 bins of $275\ \mathrm{GHz}$4, occupancy was computed on one-minute intervals, and each bin label was set by thresholding received power so that a channel was considered occupied if at least one sample within the minute exceeded a predefined power threshold (Mao et al., 16 Jan 2026). The resulting binary process

$275\ \mathrm{GHz}$5

supports sliding-window prediction, while the deployed framework above defines hourly airtime utilization for channel $275\ \mathrm{GHz}$6 by

$275\ \mathrm{GHz}$7

where $275\ \mathrm{GHz}$8 if $275\ \mathrm{GHz}$9 and $5$0 otherwise (Gandotra et al., 16 Jan 2026). This threshold-based construction is intentionally deployment-oriented, but it also implies sensitivity to the threshold rule and to minute-level maxima rather than average power (Mao et al., 16 Jan 2026).

A complementary line of work converts streaming power spectral density into event objects. “Spectrum Streamer” defines a transmission event as a time-frequency tuple $5$1, computes per-bin recent and historical energy histograms, applies a Chi-square test to detect activity, and then groups adjacent active bins in frequency and time into full transmission blocks (Fortuna et al., 2018). This produces real-time event streams, historical query interfaces, and statistical reports without protocol-specific labels. Airborne measurements extend the same occupancy logic to altitude-dependent sensing: Helikite campaigns over urban and rural sites up to $5$2 and $5$3 found that the mean measured power generally increases with altitude as line-of-sight links to nearby base stations become more available, while downlink bands are more crowded than uplink bands across the monitored LTE, 5G NR, and CBRS allocations (Raouf et al., 2023).

3. Algorithmic sensing, reconstruction, and prediction

Short-horizon occupancy prediction treats spectrum as a multi-output temporal classification problem. Using the $5$4-day, $5$5 mid-band dataset described above, the next-minute occupancy task takes the previous $5$6 minutes of binary occupancy across $5$7 bins as input and predicts the $5$8-dimensional occupancy vector one minute ahead (Mao et al., 16 Jan 2026). The study compares a first-order two-state Markov chain baseline with Random Forest, XGBoost, and LSTM models. With $5$9, the Markov baseline achieved Accuracy $10$0, Balanced Accuracy $10$1, $10$2, and $10$3; Random Forest achieved $10$4, $10$5, $10$6, and $10$7; XGBoost achieved $10$8, $10$9, $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$0, and $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$1; and LSTM achieved $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$2, $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$3, $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$4, and $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$5 (Mao et al., 16 Jan 2026). The gains concentrate on dynamic channels with higher transition rates, whereas static channels remain near-perfect for both statistical and learning-based methods (Mao et al., 16 Jan 2026).

Wideband sensing also admits sub-Nyquist formulations. In compressive wideband frequency sensing, the monitored signal is modeled as sparse in frequency and acquired through an analog-to-information converter, with measurements

$$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$6

When primary allocations induce known spectral group boundaries, recovery can be posed through mixed group sparsity, and the EVLBS-CWSS method uses an iteratively reweighted $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$7 program to better approximate an $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$8 objective (Liu et al., 2010). Simulations on a $$I_k = \sum_{i \in \mathcal{L}} P_i G_{ik} \le I_k^{\max},$$9–$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$0 band with active subbands at $\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$1–$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$2, $\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$3–$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$4, $\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$5–$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$6, and $\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$7–$\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$8, SNR $\mathrm{PFD}(R) = \frac{P_t G_t}{4 \pi R^2}.$9, and subsampling ratios $3.5\ \mathrm{GHz}$0–$3.5\ \mathrm{GHz}$1 showed cleaner reconstruction, sharper edge localization, and much smaller inactive-band residual energy than plain $3.5\ \mathrm{GHz}$2 or unweighted group lasso, with convergence typically within about $3.5\ \mathrm{GHz}$3–$3.5\ \mathrm{GHz}$4 iterations (Liu et al., 2010).

A different trend replaces spectrogram object detection with direct I/Q-domain segmentation. Semantic spectrum segmentation uses a $3.5\ \mathrm{GHz}$5 U-Net-like encoder-decoder with a non-local block, processes each time window as a $3.5\ \mathrm{GHz}$6 slice of real and imaginary frequency-domain samples, and outputs multi-label per-bin protocol occupancy (Uvaydov et al., 2024). On a stitched wideband dataset spanning WiFi, LTE, BLE, LoRa, and ZigBee, the method reported mean IoU $3.5\ \mathrm{GHz}$7 across five protocols, latency $3.5\ \mathrm{GHz}$8, and approximately $3.5\ \mathrm{GHz}$9 better accuracy on the most challenging wideband signals than a standard U-Net (Uvaydov et al., 2024). This suggests a spectrum representation in which occupancy, localization, and protocol identity are inferred jointly at bin level rather than through image-like bounding boxes.

4. Quantified sharing, coexistence, and market mechanisms

Dynamic sharing requires more than binary occupancy: it requires an accounting of what transmitters use, what receivers forbid, and what remains available. In discretized spectrum-space models, each space-time-frequency cell carries three complementary quantities: spectrum-occupancy $6\ \mathrm{GHz}$0, spectrum-opportunity $6\ \mathrm{GHz}$1, and receiver-liability $6\ \mathrm{GHz}$2, with the conservation law

$6\ \mathrm{GHz}$3

in one formulation (Khambekar et al., 2014), or

$6\ \mathrm{GHz}$4

in the MUSE framework (Khambekar et al., 2015). In both cases, receiver consumption is explicit: receivers deny additional interference power through their minimum SINR requirements, so the available spectrum is not simply the complement of transmitter emissions.

Scheduling under these constraints is NP-hard. A quantified sharing study therefore proposes Network Spectrum Consumption based Coexistence, a greedy method that computes each request’s minimal network spectrum consumption in isolation, orders requests by ascending cost, and admits each request only if all receiver SINR constraints remain satisfied after admission (Khambekar et al., 2014). That work emphasizes the significance of the active role of incumbents, the benefits of fine granular spectrum access, and the need for transceiver standards, and it reports that small-cell ranges dramatically increase the number of schedulable secondary networks while reducing harmful interference (Khambekar et al., 2014). A plausible implication is that spectrum efficiency in shared bands is governed at least as much by receiver tolerance, geometry, and footprint granularity as by raw vacancy.

Allocation can also be implemented through explicit economic mechanisms. MTSSA studies multi-tier federal-commercial sharing in which a broker runs secure spectrum auctions at the base-station level under conflict-graph interference constraints, using VCG-style allocation and Paillier homomorphic encryption to prevent fraud and bid-rigging (Abdelhadi et al., 2015). Revenue-maximizing DSA mechanisms further distinguish frequency-division and spread-spectrum sharing, derive incentive-compatible and individually rational payments through Myerson-style virtual values

$6\ \mathrm{GHz}$5

and show that revenue maximization reduces to expected virtual surplus maximization, convex in the frequency-division setting and generally non-convex under interference-coupled spread spectrum (Kakhbod et al., 2011). These approaches complement policy frameworks based on SURs, database-driven access, CBRS-like SAS control, and liberalized trading (Webb et al., 8 Mar 2025).

5. Frequency-domain analysis in photonics, cosmology, and solar physics

In photonics, spectrum refers not only to intensity versus optical frequency but to the full complex transfer function

$6\ \mathrm{GHz}$6

A wideband vector spectrum analyzer based on chirped external-cavity diode lasers and a calibrated fiber cavity measures loss, phase, and dispersion over $6\ \mathrm{GHz}$7 from $6\ \mathrm{GHz}$8 to $6\ \mathrm{GHz}$9, with $24/7$0 frequency resolution and $24/7$1 dynamic range (Luo et al., 2023). The retrieved phase supports direct computation of group delay $24/7$2, group index, and higher-order dispersion, and the same platform is used for passive device characterization, frequency-comb mapping, and FMCW LiDAR (Luo et al., 2023).

In observational cosmology, the spectrum of the Universe is the sky-averaged specific intensity $24/7$3 of the cosmic background radiation from radio to $24/7$4-rays (Hill et al., 2018). The CMB monopole follows the Planck spectrum

$24/7$5

with best-fit $24/7$6 (Hill et al., 2018). The same synthesis identifies major peaks associated with the CMB, the cosmic infrared background, the cosmic optical background, the cosmic X-ray background, and the cosmic $24/7$7-ray background, while also discussing line perturbations such as global $24/7$8, [C II], CO, Ly$24/7$9, and Fe K$7$0 signals (Hill et al., 2018). Directional dependence is represented by

$7$1

so that a full set of spectra for the multipole moments would encode the statistical history of nuclear, atomic, and molecular processes (Hill et al., 2018).

Solar physics uses spectrum in yet another sense, as a wavenumber-resolved velocity distribution. Full-disk HMI/SDO Doppler data yield a photospheric convection spectrum

$7$2

which rises to a peak near $7$3 ($7$4), levels to about $7$5, and rises again to a peak near $7$6 ($7$7) (Hathaway et al., 2015). The decomposition into radial, poloidal, and toroidal components shows toroidal dominance at $7$8, poloidal dominance above that scale, and a radial contribution that increases from about $7$9 of the total velocity at the lowest wavenumbers to about $275\ \mathrm{GHz}$00 near $275\ \mathrm{GHz}$01 (Hathaway et al., 2015).

6. Spectra in algebra, Diophantine approximation, and harmonic analysis

In algebraic and order-theoretic settings, spectrum denotes a spectral space of prime-type objects. For a commutative unital ring $275\ \mathrm{GHz}$02, the real spectrum $275\ \mathrm{GHz}$03 is the space of prime cones; for an Abelian lattice-ordered group $275\ \mathrm{GHz}$04, the $275\ \mathrm{GHz}$05-spectrum $275\ \mathrm{GHz}$06 is the space of prime $275\ \mathrm{GHz}$07-ideals; and for an $275\ \mathrm{GHz}$08-ring, the Brumfiel spectrum $275\ \mathrm{GHz}$09 consists of prime $275\ \mathrm{GHz}$10-ideals (Wehrung, 2017). Wehrung shows that every real spectrum can be embedded, as a spectral subspace, into some $275\ \mathrm{GHz}$11-spectrum, but not every real spectrum is an $275\ \mathrm{GHz}$12-spectrum, a spectral subspace of a real spectrum may fail to be a real spectrum, not every $275\ \mathrm{GHz}$13-spectrum embeds into a real spectrum, and there exists a completely normal spectral space that cannot be embedded as a spectral subspace into any $275\ \mathrm{GHz}$14-spectrum (Wehrung, 2017).

In Diophantine approximation, the Dirichlet spectrum records limsup values of rescaled approximation error. For $275\ \mathrm{GHz}$15, arbitrary norms on $275\ \mathrm{GHz}$16 and $275\ \mathrm{GHz}$17, and

$275\ \mathrm{GHz}$18

the Dirichlet spectrum is

$275\ \mathrm{GHz}$19

For $275\ \mathrm{GHz}$20, the result is

$275\ \mathrm{GHz}$21

where $275\ \mathrm{GHz}$22 is the sharp Minkowski-Dirichlet constant determined by the norms; related $275\ \mathrm{GHz}$23-Dirichlet spectra satisfy $275\ \mathrm{GHz}$24 under the paper’s stated hypotheses (Agin et al., 2024). Special cases include $275\ \mathrm{GHz}$25 for max norms in $275\ \mathrm{GHz}$26 or $275\ \mathrm{GHz}$27, and

$275\ \mathrm{GHz}$28

for $275\ \mathrm{GHz}$29 with the Euclidean norm on $275\ \mathrm{GHz}$30 (Agin et al., 2024).

Harmonic analysis uses spectrum differently again. A bounded measurable set $275\ \mathrm{GHz}$31 is spectral if $275\ \mathrm{GHz}$32 admits an exponential orthonormal basis $275\ \mathrm{GHz}$33, and in one dimension any spectrum $275\ \mathrm{GHz}$34 with $275\ \mathrm{GHz}$35 of a spectral set $275\ \mathrm{GHz}$36 with $275\ \mathrm{GHz}$37 must be rational, meaning $275\ \mathrm{GHz}$38 for some $275\ \mathrm{GHz}$39 (Lai et al., 2019). Combined with previously established periodicity, that result makes Fuglede’s conjecture on $275\ \mathrm{GHz}$40 equivalent to the corresponding conjecture on finite cyclic groups $275\ \mathrm{GHz}$41 (Lai et al., 2019). The contrast with higher-dimensional radio, optical, and physical frequency spectra is structural rather than terminological: here spectrum is a set of admissible frequencies or prime-like points, not a measured intensity distribution.