Graded Bandwidth in Adaptive Networks

• Graded bandwidth is a systems concept that decomposes available bandwidth into discrete, tunable grades for controlled allocation and performance optimization.
• It is applied across wireless, multimedia, IP/MPLS, quantum, optical, and speech systems to balance resource isolation with opportunistic sharing.
• Studies indicate that intermediate, graded allocations can outperform full-spectrum use by enhancing SIR, reducing congestion, and boosting throughput.

Searching arXiv for the cited papers on graded/adaptive bandwidth to ground the article. arXiv search query: "(Baccelli et al., 2021) graded bandwidth adaptive bandwidth (Chowdhury et al., 2018, Lingaraju et al., 2020, Du et al., 2017, 0711.0277, Reale et al., 2018, Valin et al., 2016, Chen et al., 2016, 0710.4046)" Graded bandwidth denotes a class of allocation and representation schemes in which bandwidth is not treated as a fixed, binary resource but as a multi-level quantity that can be partitioned, assigned, or consumed in discrete grades or tunable portions. In the cited literature, the grading variable is context-dependent: the number of orthogonal chunks used by a transmitter in D2D wireless networks (Baccelli et al., 2021), the number of active enhancement layers in scalable video sessions (Chowdhury et al., 2018), the amount of private and loanable bandwidth attached to a Traffic Class in DS-TE (Reale et al., 2018), the spectral width and position of flex-grid slices in quantum networks (Lingaraju et al., 2020), the number of frequency slots in decentralized wireless partitioning (0711.0277), or the bandwidth selected to keep a mmWave link near its rate-maximizing SNR (Du et al., 2017). A related but distinct usage appears in graded-index optical interconnects, where refractive-index grading is used to sustain high bandwidth over a broad alignment window (Chen et al., 2016), and in speech coding, where a narrowband core is augmented by a low-rate enhancement layer to reconstruct wideband audio (Valin et al., 2016).

1. Core meanings and representational forms

Across the literature, graded bandwidth is consistently associated with an intermediate design space between rigid fixed allocation and unconstrained sharing. In one line of work, a user’s “grade” is the number of frequency chunks it occupies: with total system bandwidth normalized to $W=1$, split into $K$ orthogonal equal-width chunks, a type-$i$ user uses $i$ chunks and therefore has effective bandwidth $B_i=i/K$ (Baccelli et al., 2021). In another line, a video session’s grade is the number of active enhancement layers between a minimum and a maximum bit rate (Chowdhury et al., 2018). In DS-TE, the grade of a class is determined by its Bandwidth Constraint together with the extent to which that bandwidth is private, High-To-Low loanable, or Low-To-High loanable (Reale et al., 2018).

A broader implication is that graded bandwidth is not a single mechanism but a family of resource abstractions. The common structure is a finite or parameterized set of admissible bandwidth states, plus rules that govern transitions among those states. This suggests that “graded bandwidth” is best understood as a systems concept rather than a domain-specific protocol.

Domain	Grading unit	Primary purpose
D2D wireless	Number of chunks $i$ with $B_i=i/K$	Service differentiation and stochastic interference analysis (Baccelli et al., 2021)
Scalable video / cellular	Base layer plus enhancement layers	Congestion response and call-level QoS protection (Chowdhury et al., 2018)
IP/MPLS/DS-TE	$BC_i$, $HTL_i$, $LTH_i$, $K$0	Class isolation and controlled sharing (Reale et al., 2018)
Quantum networking	Width and placement of spectral slices	Coincidence-rate equalization and QoS (Lingaraju et al., 2020)
Decentralized wireless	Number of sub-bands $K$1	Maximize simultaneous links under outage (0711.0277)
mmWave access	Selected occupied bandwidth $K$2	Maximize pilot-limited achievable rate (Du et al., 2017)
Speech coding	NB core plus low-rate enhancement	Wideband reconstruction over legacy NB infrastructure (Valin et al., 2016)

2. Chunk-based graded bandwidth in wireless networks

A mathematically explicit model of graded bandwidth appears in D2D wireless networks inspired by 5G NR bandwidth parts (Baccelli et al., 2021). The total bandwidth is partitioned into $K$3 orthogonal chunks, and the user type is defined solely by how many chunks it uses. Each transmitter independently chooses type $K$4 with probability $K$5, where $K$6. This traffic mix encodes the fractions of “small” and “large” users, and the heterogeneity of bandwidth demand is captured entirely by random type assignment together with random chunk selection within each type.

Two allocation variants are analyzed. In random BA, a type-$K$7 user selects uniformly at random one subset of size $K$8 from $K$9, with no contiguity constraint, matching generalized LTE or carrier-aggregation style behavior. In contiguous BA, a type-$i$0 user chooses uniformly at random one contiguous interval of length $i$1, matching 5G NR BWP-like behavior (Baccelli et al., 2021). For a typical type-$i$2 user and a type-$i$3 interferer, the overlap cardinality

$i$4

determines the interference weight, and the overlap probabilities $i$5 are hypergeometric in the random BA case and piecewise-defined in the contiguous BA case.

The network model is a homogeneous PPP $i$6 of intensity $i$7, with each transmitter paired to a dedicated receiver at deterministic distance $i$8 in a random direction. With path loss $i$9, $i$0, i.i.d. Rayleigh fading, and power density $i$1 per chunk, the type-$i$2 signal power is

$i$3

while the interference is the chunk-overlap-weighted aggregate

$i$4

The corresponding SIR is

$i$5

The factor $i$6 cancels, so the analysis is scale-invariant in transmit power (Baccelli et al., 2021).

The model yields exact type-conditioned performance metrics. For success probability with target threshold $i$7,

$i$8

where $i$9 and $B_i=i/K$0 (Baccelli et al., 2021). The paper also gives moments of the SIR meta distribution and the type-$B_i=i/K$1 Shannon throughput

$B_i=i/K$2

Two results are central to the encyclopedia treatment of graded bandwidth. First, for random BA, $B_i=i/K$3, which makes the mean signal and the mean interference both proportional to $B_i=i/K$4; this yields the proposition that random BA is roughly egalitarian in Shannon throughput per Hertz and produces a linear service differentiation in aggregate Shannon throughput (Baccelli et al., 2021). Second, the paper compares networks with the same mean signal and the same mean interference powers but different variance in chunk usage, and reports that the more variable network performs better for all SIR thresholds in overall success probability, and also in overall Shannon throughput and Shannon throughput per Joule (Baccelli et al., 2021). This directly contradicts the naive expectation that added traffic variability is always harmful.

3. Congestion-driven and class-based graded allocation

In cellular multimedia systems, graded bandwidth appears as a discrete-step adaptation mechanism for scalable traffic. A single-cell downlink model with total capacity

$B_i=i/K$5

and $B_i=i/K$6 MBS sessions treats scalable video as a base layer plus enhancement layers, where the active number of layers determines the current grade of service (Chowdhury et al., 2018). For unicast video, the allocated bandwidth $B_i=i/K$7 lies between

$B_i=i/K$8

and for each MBS session $B_i=i/K$9,

$i$0

Voice is non-adaptive with

$i$1

while background traffic can be degraded from $i$2 to $i$3 (Chowdhury et al., 2018).

Under low traffic, all MBS sessions operate at maximum grade,

$i$4

whereas under congestion the system removes enhancement layers almost equally across MBS sessions: the maximum difference in removed layers between any two sessions is at most one. If $i$5 is the minimum number of layers removed from every active MBS session and $i$6 is the number of sessions that lose exactly $i$7 layers, then the remaining $i$8 sessions lose $i$9 layers (Chowdhury et al., 2018). Handover calls are admitted if the system can provide at least $B_i=i/K$0 after degradation, while new calls are admitted only if $B_i=i/K$1. Unicast video is degraded only to accept handover calls, background traffic may be degraded for both handover and some higher-priority new calls, and MBS sessions can be degraded in both circumstances (Chowdhury et al., 2018).

The reported outcome is that releasing MBS bandwidth in steps during congestion keeps handover dropping probability negligible even at high new call arrival rates and keeps forced call termination much lower than two non-adaptive fixed-MBS schemes in which MBS always occupies either $B_i=i/K$2 Mbps or $B_i=i/K$3 Mbps (Chowdhury et al., 2018). The mechanism is therefore not merely adaptive coding but an explicit graded bandwidth policy in which discrete quality grades are traded against call-level QoS.

A different but related formalization appears in IP/MPLS/DS-TE. G-BAM defines, for each Traffic Class $B_i=i/K$4, a Bandwidth Constraint $B_i=i/K$5, High-To-Low loan $B_i=i/K$6, Low-To-High loan $B_i=i/K$7, and private bandwidth

$B_i=i/K$8

subject to $B_i=i/K$9 and $BC_i$0, $BC_i$1 (Reale et al., 2018). The maximum possible occupancy of class $BC_i$2 is

$BC_i$3

and dynamic loan availability is determined by the currently allocated bandwidth $BC_i$4 through $BC_i$5 and $BC_i$6 (Reale et al., 2018).

This model reproduces existing BAMs by parameter choice. Setting all $BC_i$7 yields MAM behavior, in which all bandwidth is private. Setting $BC_i$8 and $BC_i$9 yields RDM behavior, in which higher-class unused capacity is loanable downward. Setting both $HTL_i$0 and $HTL_i$1 yields AllocTC-Sharing behavior, in which capacity is bidirectionally loanable and private bandwidth vanishes (Reale et al., 2018). A crucial extension is that G-BAM can also integrate private resources with both HTL and LTH loans, allowing intermediate grades that are not expressible by MAM, RDM, G-RDM, or AllocTC-Sharing alone (Reale et al., 2018). This suggests that graded bandwidth in packet networks is most naturally viewed as a policy continuum between isolation and opportunistic sharing.

4. Spectrum partitioning and optimal partial bandwidth use

A distinct research thread studies graded bandwidth not as user classes or service layers but as optimal partitioning of spectrum itself. In decentralized wireless networks with total system bandwidth $HTL_i$2, fixed per-link rate $HTL_i$3, and PPP-distributed transmitters, the band is partitioned into $HTL_i$4 equal sub-bands of width $HTL_i$5. Each transmitter independently and randomly chooses one sub-band (0711.0277). The required SINR threshold becomes

$HTL_i$6

or equivalently $HTL_i$7 if $HTL_i$8 is the spectral efficiency (0711.0277). Increasing $HTL_i$9 lowers interferer density per sub-band but raises the required SINR exponentially, and the total outage-constrained transmission density is maximized at an intermediate $LTH_i$0 satisfying a one-dimensional fixed-point equation (0711.0277). In the interference-limited regime, $LTH_i$1 depends only on the path-loss exponent $LTH_i$2; for example, $LTH_i$3 gives $LTH_i$4 bps/Hz and $LTH_i$5 gives $LTH_i$6 bps/Hz (0711.0277). In the power-limited regime, $LTH_i$7, where $LTH_i$8 is the interference-free AWGN spectral efficiency (0711.0277).

A related but more explicit “too much spectrum can hurt” result appears in mmWave access. With pilot-based MMSE channel estimation over an i.i.d. block-fading channel of coherence length

$LTH_i$9

the achievable rate is

$K$00

with effective SNR

$K$01

(Du et al., 2017). The paper shows that the optimal, rate-maximizing SNR is a constant determined only by channel coherence, and for large $K$02,

$K$03

so the optimal bandwidth is

$K$04

The immediate consequence is that optimal bandwidth scales linearly with received power (Du et al., 2017).

The numerical findings are operationally important. Under 3GPP UMi-NLOS with $K$05 ms and $K$06 MHz, EIRP $K$07 dBm and $K$08 dBi receive gain at $K$09 GHz, the maximum beneficial bandwidth is $K$10 GHz beyond about $K$11 m and $K$12 MHz beyond about $K$13 m; the corresponding max throughput is about $K$14 Mbps and about $K$15 Mbps, respectively (Du et al., 2017). The same study states that, under some typical deployment scenarios with both transmit and receive side beamforming, $K$16 GHz bandwidth can be too much (Du et al., 2017). This is a direct rebuttal to the common assumption that the optimal policy is always to occupy the entire available mmWave band.

The wideband-regime analysis of BICM adds a different but compatible viewpoint. At low SNR,

$K$17

which yields

$K$18

(0710.4046). For Gray-mapped $K$19-PAM and $K$20-QAM,

$K$21

and therefore

$K$22

as $K$23 (0710.4046). The paper then derives an explicit power–bandwidth trade-off showing how a power loss can be traded for a bandwidth gain in the wideband regime (0710.4046). A plausible implication is that graded bandwidth can also be interpreted as gradual bandwidth expansion used to compensate for modulation or receiver suboptimality.

5. Optical, quantum, and signal-processing interpretations

In quantum networking, graded bandwidth is realized as adaptive spectral slicing. A broadband entangled-photon source is partitioned by a wavelength-selective switch into energy-matched spectral slices that can be routed to different user pairs in a four-user network with links $K$24 (Lingaraju et al., 2020). The spectrum is carved into $K$25 spectral slices, each $K$26 GHz wide, yielding $K$27 channels across the biphoton bandwidth (Lingaraju et al., 2020). Because the coincidence rate is proportional to the integral of joint spectral intensity over the allocated spectral support, more optical spectral bandwidth corresponds directly to more entangled pairs per second (Lingaraju et al., 2020).

The experiment uses bandwidth grading to equalize strongly heterogeneous links. Under fixed $K$28 GHz grid, alphabetical assignment yields an AB/CD coincidence-rate ratio of about $K$29. Reassigning which $K$30 GHz pairs go to which link reduces that disparity to about $K$31. Using full flex-grid with individual $K$32 GHz slices, dropping link $K$33, and interleaving slices across a $K$34-link subgraph brings coincidence rates for the included links within a factor of about $K$35 (Lingaraju et al., 2020). The Letter further emphasizes that the WSS-based architecture introduces insertion loss that is roughly constant with the number of users, about $K$36 dB in their device, whereas a nested DWDM architecture requires $K$37 filters for full connectivity and has worst-case loss scaling roughly as $K$38 dB, reaching about $K$39 dB at $K$40 users (Lingaraju et al., 2020). In this setting, graded bandwidth is simultaneously a fairness tool and a scalability mechanism.

In board-level optical interconnects, the phrase “graded bandwidth” is not used as a resource-allocation policy but arises from graded-index engineering. A siloxane multimode polymer waveguide with a measured triangle-like refractive-index profile and $K$41 is designed to reduce multimode dispersion and maintain high bandwidth over practical launch offsets (Chen et al., 2016). The reported bandwidth-length products exceed $K$42 GHz·m and are about $K$43 GHz·m for a $K$44 m spiral under $K$45m MMF launch without and with a mode mixer, respectively, for $K$46m offsets corresponding to the $K$47 dB alignment tolerance (Chen et al., 2016). Under a restricted $K$48 objective launch into a $K$49 m reference waveguide, BLP exceeds $K$50 GHz·m over an area of approximately $K$51, entirely within the $K$52 dB coupling-loss contour (Chen et al., 2016). Here the bandwidth is “graded” by the index profile rather than by dynamic control, but the shared theme is explicit engineering of bandwidth performance across multiple admissible operating states.

Speech coding provides yet another interpretation. Starting from a narrowband $K$53–$K$54 Hz telephone signal, wideband $K$55–$K$56 Hz speech is reconstructed using a zero-bit low-band extension plus a very low-rate high-band enhancement layer (Valin et al., 2016). The frame size is $K$57 samples at $K$58 kHz, or $K$59 ms, and the transmitted side information is an $K$60-bit VQ index per frame, yielding

$K$61

The low band $K$62–$K$63 Hz is modeled by up to two sinusoids, whose amplitudes are predicted from $K$64 MFCCs plus pitch gain and pitch delay by a two-hidden-layer MLP with $K$65 units per layer and $K$66 activation (Valin et al., 2016). The high band $K$67–$K$68 Hz uses a source–filter model in which excitation is extrapolated locally and the spectral envelope is transmitted by $K$69-bit vector quantization (Valin et al., 2016). The result is an explicitly layered architecture: a narrowband core plus an enhancement layer of about $K$70 bit/s. This is graded bandwidth in the coding sense rather than in the spectrum-allocation sense.

6. Cross-cutting properties, common misconceptions, and limitations

Several recurring properties emerge across these otherwise heterogeneous formulations. First, graded bandwidth is frequently discrete. In the D2D model, the grade is the integer chunk count $K$71 (Baccelli et al., 2021). In scalable video, it is the integer number of enhancement layers removed or activated (Chowdhury et al., 2018). In DS-TE, although $K$72 and $K$73 are configurable as proportions of $K$74, operational behavior is still mediated through class-level admission and loan rules (Reale et al., 2018). In quantum networking, the admissible resources are finite spectral slices of $K$75 GHz (Lingaraju et al., 2020). This suggests that a central structural feature of graded bandwidth is quantization of a resource that would otherwise be treated as continuous.

Second, the literature repeatedly shows that full bandwidth occupancy is not universally optimal. In dense D2D networks, assigning everyone all $K$76 chunks corresponds to a non-adaptive baseline, and adaptive graded BA improves success probability and throughput (Baccelli et al., 2021). In mmWave access, the rate-maximizing bandwidth is generally finite once pilot overhead is accounted for, and can be far below the total available spectrum (Du et al., 2017). In decentralized wireless partitioning, excessive sub-band splitting becomes counterproductive because the SINR target rises exponentially with spectral efficiency (0711.0277). A common misconception is therefore that maximizing used bandwidth automatically maximizes performance.

Third, heterogeneity can be beneficial. The D2D analysis shows that, for fixed mean signal and mean interference, higher traffic variability improves success probability, SIR meta distribution behavior, and Shannon throughput (Baccelli et al., 2021). In the quantum case, strongly unequal detector efficiencies and spectral flux are partially neutralized by nonuniform slice assignment (Lingaraju et al., 2020). In DS-TE, G-BAM exists precisely because no single fixed sharing model is suitable for all traffic profiles (Reale et al., 2018). This suggests that graded bandwidth is often valuable not because it equalizes allocations, but because it makes structured inequality controllable.

The limitations are equally consistent. The stochastic-geometry D2D model assumes a homogeneous PPP, fixed link distance $K$77, Rayleigh fading, no noise, homogeneous power density per chunk, and no temporal scheduling (Baccelli et al., 2021). The scalable-video cellular model assumes a single wireless cell, Poisson call arrivals, exponentially distributed dwell times, and focuses on numerical results rather than a fully spelled-out Markov-chain solution (Chowdhury et al., 2018). The mmWave result assumes pilot-based MMSE estimation, i.i.d. block fading, and idealized beam-switching abstractions (Du et al., 2017). The quantum Letter is experimental and does not specify an explicit optimization solver for slice assignment (Lingaraju et al., 2020). The speech system is speech-specific and exhibits low-band artifacts, especially under pitch errors (Valin et al., 2016). The polymer-waveguide model neglects mode coupling along bent spirals and approximates launch mode distributions (Chen et al., 2016). G-BAM focuses on reproducibility of existing behaviors and does not fully resolve dynamic preemption strategy under parameter changes (Reale et al., 2018).

Taken together, these works support a precise encyclopedia-level conclusion. Graded bandwidth is a general systems principle in which bandwidth is decomposed into ordered or configurable grades, and performance is shaped by how those grades are exposed to allocation, adaptation, sharing, or physical propagation. In wireless access it appears as chunk counts, sub-band counts, or finite beneficial occupied bandwidth [(Baccelli et al., 2021); (0711.0277); (Du et al., 2017)]. In multimedia and packet networks it appears as layer counts, class constraints, and loan policies (Chowdhury et al., 2018, Reale et al., 2018). In quantum and optical systems it appears as spectral slicing and refractive-index engineering (Lingaraju et al., 2020, Chen et al., 2016). In coding, it appears as enhancement layers and explicit power–bandwidth trade-offs [(Valin et al., 2016); (0710.4046)]. A plausible implication is that graded bandwidth is best regarded as a unifying abstraction for controlled bandwidth differentiation under resource scarcity, rather than as a single algorithmic family.