Static Delay Model: Insights & Applications

• Static Delay Model is a fixed-delay representation capturing delay as a constant lag or bound, applied across diverse domains such as TSN, VLSI, and biological systems.
• It underpins methodologies like offline configuration in time-sensitive networking and LUT-based timing in VLSI, facilitating analytical and optimized delay estimations.
• By pre-characterizing delays with fixed lags or bounds, the model enables worst-case delay certification and design efficiency, though its validity depends on matching system dynamics.

Searching arXiv for papers on "static delay model" and closely related uses across domains. Search query: static delay model arXiv (Maile et al., 26 Aug 2025, Feng et al., 2016, Agarwal, 2014, Li et al., 2017, Li et al., 2017, Liu et al., 2020, Djeudjo et al., 12 May 2026) Across the cited literature, “static delay model” denotes fixed-delay, pre-characterized, or offline-optimized representations of latency rather than a single domain-independent formalism. In Time-Sensitive Networking, it appears as an offline configuration of per-hop service guarantees that yields a provable end-to-end worst-case delay bound (Maile et al., 26 Aug 2025). In delayed biological and collective-dynamics models, it denotes a constant time lag $\tau$ in delay-differential equations or interaction terms (Feng et al., 2016, Djeudjo et al., 12 May 2026). In VLSI timing, it refers both to analytic or piecewise timing abstractions used for LUT characterization and to reduced timing graphs that preserve input-to-output worst-case delays exactly (Agarwal, 2014, Li et al., 2017). Related uses also arise when geometry or operating conditions are static, as in directional cell search, static stall, and weak-field travel-time calculations (Li et al., 2017, Fouest et al., 2021, Liu et al., 2020). This suggests that the unifying feature is methodological: delay is encoded as a fixed lag, fixed bound, or fixed environment.

1. Deterministic fixed-delay formalisms

In the simplest abstract form, a static deterministic delay model is a delay-differential equation

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$

where $x(t)$ is the state variable at time $t$, $\tau$ is a fixed delay, and $f$ encodes production, degradation, and feedback (Feng et al., 2016). A concrete example is the self-activating gene circuit,

$$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$

with $\alpha,\beta,\gamma>0$, Hill exponent $b$, and half-maximal concentration $c$.

The same paper emphasizes that such a fixed-delay DDE is non-Markovian and implicitly assumes a delta-function delay distribution. To restore Markovianity, one may introduce explicit precursor states. For the self-activation example, one explicit model uses a precursor species $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$0 and active protein $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$1 with reactions

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$2

and mean-field equations

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$3

In that construction, the DDE corresponds to a delta-distributed delay, whereas the one-step explicit model produces an exponential delay distribution $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$4; with $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$5 identical sequential precursors, the distribution becomes Gamma with mean $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$6 (Feng et al., 2016).

The comparison between fixed-delay and explicit models is not merely formal. For bistable self-activation, increasing $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$7 in the delayed stochastic system shifts the stationary distribution from favoring the low-$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$8 state toward the high-$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$9 state, and the mean residence time in the low-$x(t)$0 basin grows rapidly with $x(t)$1. One explicit model can be tuned so that its mean residence time and equilibrium density function match the delayed stochastic system for small and moderate delays, while another explicit model with the same mean-field ODEs exhibits much shorter residence times and stationary distributions skewed toward low $x(t)$2 (Feng et al., 2016). For delayed degradation,

$x(t)$3

the fixed-delay formulation shows spectral peaks suggestive of quasi-oscillations, whereas the corresponding explicit precursor model has Jacobian eigenvalues $x(t)$4 and $x(t)$5, no Hopf bifurcation at the mean-field level, and no clear spectral peak in stochastic simulations. A plausible implication is that “static” fixed-delay models are phenomenological closures whose validity depends strongly on the true waiting-time structure.

2. Offline service-curve delay bounds in Time-Sensitive Networking

In Ethernet-based TSN with credit-based shaping queues, the static delay model in (Maile et al., 26 Aug 2025) is built from standard min-plus Network Calculus abstractions. Time-critical traffic traverses a sequence of bridges, one queue per priority, and each queue is abstracted as a server offering a guaranteed service curve to all flows mapped into that queue. Static flows are known before deployment and must meet hard deadlines; dynamic flows are admitted at runtime only if they can be shown to meet pre-provisioned per-hop delay bounds.

Each flow $x(t)$6 is upper-bounded by a leaky-bucket arrival curve

$x(t)$7

where $x(t)$8 is the long-term sending rate and $x(t)$9 the maximum burst. Each server $t$0 is lower-bounded by a rate-latency service curve

$t$1

where $t$2 is the reserved bandwidth in that queue and $t$3 is a latency parameter. The per-hop worst-case delay is the horizontal deviation between $t$4 and $t$5,

$t$6

which, for the chosen arrival and service families, simplifies to

$t$7

If flow $t$8 traverses $t$9 servers in sequence, the concatenated service curve is

$\tau$0

and the end-to-end bound is

$\tau$1

provided that each hop’s service-curve parameters have been chosen so as to avoid overtaking and ensure flow-aggregation constraints are met (Maile et al., 26 Aug 2025).

The static configuration problem is to choose the number of priority queues on each output port and, for each queue $\tau$2, the reserved rate $\tau$3 and latency $\tau$4, such that all static flows meet

$\tau$5

while sufficient slack remains on each link for future dynamic flows. Because this problem is combinatorial and non-convex, the paper uses Particle Swarm Optimization and a Genetic Algorithm. Their genes encode the vector of per-queue $\tau$6 or simply the $\tau$7 if $\tau$8 is inferred from a fixed idle-slope reservation, and the fitness function mixes three objectives: all static flows can be scheduled, a user-specified fraction of link bandwidth is reserved for future arrivals, and over-provisioning is avoided (Maile et al., 26 Aug 2025). The final static maximum end-to-end delay guarantee is

$\tau$9

3. Static timing abstractions in VLSI

In STA, cell delays are pre-characterized and stored in LUTs indexed by input rise-time $f$0 and output load $f$1, and the delay model used to fill an LUT governs both LUT-generation cost and STA accuracy (Agarwal, 2014). Within this setting, the Linear Delay Model exploits the observation that, for each fixed $f$2, the inverter delay $f$3 is nearly linear for small $f$4 up to a break point $f$5 and then becomes strongly nonlinear. For sufficiently small $f$6,

$f$7

where $f$8, $f$9, and $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$0 gather transistor- and load-dependent terms. Beyond $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$1, the NMOS falls out of saturation while the input is still rising, and the simple linear form no longer holds (Agarwal, 2014).

The break point is defined by the largest $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$2 for which the NMOS remains in saturation until the input reaches $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$3. From

$$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$4

and the condition $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$5, one obtains

$$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$6

The paper further models PVT dependence through $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$7, $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$8, and approximate relations in which $$\dot x(t)=\alpha+\beta \,\frac{x(t-\tau)^b}{c^b+x(t-\tau)^b}-\gamma x(t),$$9 scales with $\alpha,\beta,\gamma>0$0, $\alpha,\beta,\gamma>0$1, and $\alpha,\beta,\gamma>0$2 (Agarwal, 2014).

Validation in a $\alpha,\beta,\gamma>0$3 CMOS inverter uses Synopsys HSPICE with $\alpha,\beta,\gamma>0$4, $\alpha,\beta,\gamma>0$5, and $\alpha,\beta,\gamma>0$6. The extracted $\alpha,\beta,\gamma>0$7 is extremely linear in $\alpha,\beta,\gamma>0$8; the $\alpha,\beta,\gamma>0$9 exponent is approximately $b$0; the temperature exponent is approximately $b$1; and the typical fit error across PVT is less than $b$2 (Agarwal, 2014). In the example LUT with $b$3 input-rise points and $b$4 load points, naive characterization requires $b$5 SPICE runs, whereas the LDM plus analytic $b$6 requires about $b$7 runs, i.e. $b$8 fewer SPICE simulations. The stated trade-off is that power characterization in the linear region is forfeited.

A different but related static timing abstraction is the reduced timing graph for combinational modules (Li et al., 2017). A purely combinational module is modeled as a DAG $b$9 whose edge weights are maximum delays. For each input-output pair,

$c$0

A nonterminal vertex is redundant if it lies on no path achieving any $c$1, and an edge is redundant if

$c$2

for every input $c$3 and output $c$4, where $c$5 is the maximum delay from input $c$6 to $c$7 and $c$8 is the maximum delay from $c$9 to output $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$00 (Li et al., 2017). The pruning algorithm uses topological sorting, forward longest-path passes from each primary input, backward passes from each primary output, and then one pass each for redundant-vertex and redundant-edge removal. Its overall complexity is

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$01

and on $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$02 industrial combinational blocks it reduces edges by $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$03 and vertices by $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$04 on average while exactly preserving all $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$05 (Li et al., 2017). This use of a static delay model is therefore exact with respect to worst-case input-output timing, but compressed with respect to structural detail.

4. Delay under static geometry in wireless initial access

Directional cell search delay analysis for cellular networks with static users provides another use of static delay modeling (Li et al., 2017). Base stations are located according to a homogeneous planar PPP of intensity $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$06, and users are fixed or move so slowly that both the PPP realization and user locations can be treated as static over the timescale of interest. Time is divided into TDD initial-access cycles of duration $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$07, each cycle begins with a downlink cell search phase of $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$08 consecutive OFDM symbols of duration $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$09, and each base station sweeps synchronously through $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$10 sectorized beams with main-lobe gain $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$11 and beamwidth $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$12.

Conditioned on the static geometry $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$13, the indicator of success in each IA cycle is i.i.d. Bernoulli with success probability $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$14, so the number of cycles until success is geometric: $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$15 The spatially averaged mean number of IA cycles is

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$16

and the mean cell-search delay is

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$17

Because only fading refreshes across cycles while geometry remains fixed, the instantaneous SINR exhibits strong temporal correlation (Li et al., 2017).

The static geometry produces sharp finiteness results. In the noise-limited regime, if the NLOS path-loss exponent $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$18, then

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$19

for any finite $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$20 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$21; if $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$22, then

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$23

In the interference-limited, single-slope case with $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$24 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$25,

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$26

and if $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$27, the delay is always infinite (Li et al., 2017). The same work proves that if $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$28, then $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$29 for all $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$30. A plausible implication is that a static delay model under frozen geometry naturally generates heavy-tailed averages even when beam sweeping improves typical and percentile performance.

5. Static operating conditions in fluid, gravitational, and collective-dynamics models

For airfoil stall, the cited work argues that “static” stall is not phenomenologically different than dynamic stall and is merely a typical case of stall for low pitch rates (Fouest et al., 2021). The convective time scale is $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$31, the dimensionless time is $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$32, and the reduced pitch rate is

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$33

Using a slow continuous ramp with $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$34 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$35, and a step jump from $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$36 to $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$37 within $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$38, the reaction delay is defined as

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$39

where $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$40 is when $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$41 first exceeds $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$42 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$43 is determined from a $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$44 threshold in lift drop (Fouest et al., 2021). Across maneuvers with $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$45, the average stall reaction delay follows

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$46

with asymptotic minimum $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$47 convective times for $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$48. At very low $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$49, the static delay is approximately $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$50 convective times with normal distribution and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$51. The paper recommends continuous ramp-up motion at reduced frequency around $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$52 for static stall measurements, or waiting at least $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$53 convective times after a step jump before recording quasi-steady values (Fouest et al., 2021).

In weak-field static spherically symmetric spacetimes, the total travel time of null and timelike signals is expanded as a quasi-series in impact parameter $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$54 (Liu et al., 2020). With asymptotic metric expansions

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$55

the total time takes the form

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$56

and each coefficient $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$57 is determined by the metric coefficients up to order $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$58 (Liu et al., 2020). For a null ray, the gravitational time delay relative to flat space is

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$59

with

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$60

For two signals of speeds $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$61 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$62,

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$63

so the leading term is universal in $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$64 and the post-Newtonian parameter $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$65 (Liu et al., 2020).

In 1D swarmalators, a constant delay $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$66 can enter only the sine self-interaction channels,

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$67

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$68

For the static phase-wave state, generic Fourier modes have eigenvalues $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$69 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$70, hence these modes are stable if and only if $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$71; the mean mode requires $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$72 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$73 on the principal Lambert-$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$74 branch; and the special $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$75 and $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$76 modes obey a characteristic equation with a Hopf boundary determined by

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$77

and an explicit delay threshold $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$78 (Djeudjo et al., 12 May 2026). For the static $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$79-state, the linear stability condition is simply

$\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$80

The cited analysis states that increasing $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$81 contracts the static phase-wave domain, while in the asymmetric-sine-delay model it promotes the $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$82-state at the expense of phase-waves; in the symmetric-delay model, increasing $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$83 instead promotes unsteady states (Djeudjo et al., 12 May 2026). This use of “static” refers to stationary collective configurations coexisting with a fixed interaction lag.

6. Recurring assumptions, limitations, and points of interpretation

Several recurring assumptions delimit what these static delay models can guarantee. In the TSN formulation, end-to-end composition by summing per-hop horizontal deviations holds only provided that each hop’s service-curve parameters have been chosen so as to avoid overtaking and ensure flow-aggregation constraints are met (Maile et al., 26 Aug 2025). In fixed-delay DDEs, the delay is collapsed into a single uniform lag $\frac{dx}{dt} = f\bigl(x(t),x(t-\tau)\bigr),$84, which implicitly assumes a delta-function delay distribution and can lead to mis-predicted residence times, altered basin occupancies, spurious oscillations, or missed transitions if the real process is multi-step or broadly distributed (Feng et al., 2016). In LDM-based LUT characterization, delay accuracy remains within a few percent of a full NLDM LUT, but power characterization in the linear region is forfeited (Agarwal, 2014). In static-geometry wireless analysis, the spatial mean delay can be infinite even though percentile delays are numerically small and improve with beam sweeping (Li et al., 2017). In the swarmalator model, the results are restricted to identical agents, a single fixed delay, and linear stability analysis (Djeudjo et al., 12 May 2026).

A common misconception is to equate “static” with “steady” or “time-independent.” The static stall study explicitly states that the force and flow development of an airfoil undergoing static stall are highly unsteady (Fouest et al., 2021). The TSN framework distinguishes static flows from dynamic flows, yet dynamic flows are admitted online against static per-hop bounds (Maile et al., 26 Aug 2025). In cell search, the geometry is static but the fading is refreshed each IA cycle, creating strong temporal correlation rather than temporal constancy (Li et al., 2017). This suggests that “static” is best understood as describing what is held fixed in the model: a lag value, a service guarantee, a graph abstraction, a geometry, or an operating protocol.

Taken together, the cited works show that a static delay model can serve very different purposes: exact worst-case certification, compact pre-characterization, phenomenological closure, perturbative travel-time expansion, or stability analysis of delayed interactions. The technical value of such a model depends on whether the fixed or precomputed delay representation matches the structure of the underlying system.