Speed: Perspectives on Measurement & Control

• Speed is a context-dependent concept defined differently across disciplines, exemplified by geometry-dependent distinctions like expansion versus radial speed in astrophysics and propulsive versus instantaneous speed in active matter.
• Researchers derive speed indirectly using methodologies such as calibration, statistical inference, and geometric correction, highlighting the influence of measurement protocols and modeling techniques.
• Control and optimization applications leverage speed as a crucial variable for enhancing energy efficiency, traffic management, quantum dynamics, and computational architectures.

Speed is a context-dependent scientific quantity rather than a single universal observable. In the research literature surveyed here, it denotes physical velocity in astrophysical, vehicular, and active-matter systems; the rate of state change in open quantum dynamics; a control target in automation, communications, and energy systems; the asymptotic growth of a combinatorial class; and a family of acronymic systems named SPEED in computing, sensing, and molecular dynamics [(Gopalswamy et al., 2012); (Sekiguchi et al., 2024); (Wang et al., 2024); (Feng et al., 27 Feb 2026); (Wang et al., 2024)]. A recurring theme is that speed is often derived rather than directly observed: its meaning depends on geometry, state representation, measurement protocol, or optimization objective.

1. Physical motion and non-equivalent speeds

In heliophysics, speed is explicitly geometry-dependent. For Earth-directed coronal mass ejections (CMEs), a coronagraph on the Sun–Earth line does not directly observe the CME nose because of the occulting disk; it mainly observes the lateral spread of the halo, yielding an expansion speed rather than the true radial speed. Using quadrature observations of the 2011 February 15 CME, with SOHO/LASCO near the Sun–Earth line and STEREO-A and STEREO-B nearly $90^\circ$ away, the relation

$V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$

was confirmed for a full ice-cream-cone CME geometry. For the measured full width of $76^\circ$, hence $w=38^\circ$, the conversion becomes $V_{rad}=1.14\,V_{exp}$. With $V_{exp}=897\ \mathrm{km/s}$, the computed radial speed is $1023\ \mathrm{km/s}$, compared with direct STEREO measurements of $945\ \mathrm{km/s}$ and $1058\ \mathrm{km/s}$, differing by $7.6\%$ and $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$0 from the computed value (Gopalswamy et al., 2012). In this usage, speed is not a single measured scalar; it is a projection-sensitive quantity whose conversion depends on CME width.

Active-matter research makes a related distinction between several non-equivalent speeds. For catalytic nano- and micro-motors modeled as active Brownian particles,

$V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$1

the propulsive speed $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$2 is not identical to the finite-difference instantaneous speed extracted from image tracks. The standard mean-square displacement (MSD) is

$V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$3

with short-time approximation $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$4, but the paper emphasizes that instantaneous speed depends on frame interval, mixes propulsion with Brownian motion, and is therefore not a reliable proxy for true propulsive speed. It further argues that quadratic MSD fits can strongly bias both $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$5 and $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$6, whereas 3rd- and 4th-order Taylor fits are markedly more accurate; different formulas are also required for active angular speed, drift, and exponentially decaying propulsion (Mestre et al., 2020). Across both CME physics and active matter, speed is thus a model-indexed quantity rather than a measurement invariant.

2. Estimation, reconstruction, and intelligibility-constrained speed

In visual traffic surveillance, speed estimation is primarily a calibration problem. The monocular-camera benchmark for automatic vehicle speed measurement provides 18 full-HD videos, each around 1 hour long, captured at six different locations, with 20,865 vehicle passes carrying precise ground-truth speed from LiDAR and 2,779 additional instances marked invalid because of occlusion-related uncertainty. The camera model is written as

$V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$7

and the paper shows that errors in vanishing-point localization, especially for the second vanishing point $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$8, dominate downstream speed error. In the detailed evaluation of the fully automatic method of Dubská et al. (2014), the best-performing variant is OptCalib, while FullACC exhibits noticeably larger error; the dataset-wide tracking statistics are 9.745 false positives per minute of video and recall $V_{rad}=\frac{1}{2}(1+\cot w)V_{exp}$9 (Sochor et al., 2017). Here speed is reconstructed from image geometry and road-plane scale, so calibration error propagates directly into metric velocity bias.

A distinct reconstruction setting appears in smartphone inertial sensing. CarSpeedNet estimates car speed from three-axis accelerometer data alone, without CAN-bus access or gyroscopes. The dataset contains 13.2 hours of driving over 51 sessions, with accelerometer data at 500 Hz and GPS speed at 1 Hz. After low-pass filtering and downsampling to 20 Hz, the main model uses 4-second windows, producing $76^\circ$0 inputs, and maps them to scalar speed using a hybrid architecture comprising a bidirectional LSTM with 100 units, additional LSTM layers, Conv1D layers, and dense layers. On the held-out 30-minute test drive, the best reported result for the 4-second window is RMSE $76^\circ$1 and MAE $76^\circ$2 (Or, 2024). This usage treats speed as a latent state inferred statistically from motion history rather than integrated directly from acceleration.

Playback-speed optimization introduces a different criterion: intelligibility under deliberate time compression. AIx Speed uses a speech recognizer as a proxy for human comprehension and adjusts playback speed at units as small as phonemes. In the pilot study, human transcription accuracy and Wav2Vec2-based ASR accuracy at playback speeds above $76^\circ$3 show a reported correlation coefficient of $76^\circ$4, motivating recognizer score as the optimization signal. The system jointly optimizes a speed-adjuster and a CTC recognizer with

$76^\circ$5

and in evaluation it outperforms constant-speed playback at the same average rate: on LibriSpeech, AIx Speed reaches average speed 1.30 with CER 5.21 and WER 12.96, versus CER 5.61 and WER 14.58 for constant 1.30x; in blind listening tests, its mean opinion score is 0.5 points higher than baseline for LibriSpeech and 0.8 points higher for UME-ERJ (Kawamura et al., 2024). In this context, speed is no longer a physical state but a tunable compression factor constrained by recognition fidelity.

3. Speed as a control, conditioning, and planning variable

In variable-speed wind turbines, speed is the principal control target for maximum power extraction in region 2. The objective is to track

$76^\circ$6

with $76^\circ$7, so that the tip-speed ratio remains near its optimum and $76^\circ$8 stays near its maximum. The proposed controller combines a PI-type sliding surface $76^\circ$9 with an adaptive fuzzy disturbance observer, yielding a generator-torque law of the form $w=38^\circ$0. The Lyapunov analysis produces the robustness condition

$w=38^\circ$1

so the switching gain must dominate only the disturbance estimation error, not the full disturbance. In Simulink, the method reduces rotor-speed-tracking MSE from 15.6675 for PID to 4.3684, improves tip-speed-ratio MSE from 0.4667 to 0.1194, and raises energy-capture efficiency from 88.47% to 89.79% (Al-Jodah et al., 2021). Speed here is a regulated state linked directly to aerodynamic efficiency.

Mixed-autonomy traffic control uses speed as both a macroscopic field and a fleet-level command. The hierarchical Lagrangian variable speed limit framework places a server-side Speed Planner above vehicle-side controllers. The planner fuses coarse INRIX data, available about once per minute but delayed by about 3 minutes, with 1 Hz pings from controlled vehicles. A Transformer-like predictor uses the preceding 6 minutes of INRIX data to forecast congestion frontiers, then kernel smoothing and buffer design generate a target speed field that is broadcast to automated vehicles. In simulation on a single-lane 600 m bottleneck with 4% AV penetration, bottleneck throughput increases from 1492.4 veh/hr to 1567.18 veh/hr, a 5.01% increase, while speed standard deviation decreases from 20.81 mph to 13.66 mph, a 34.36% reduction. The field deployment used 100 vehicles in the 2022 MegaVanderTest on I-24 Westbound in Nashville (Wang et al., 2024). Speed is therefore both a control signal and an instrument for homogenizing flow.

Desired-speed conditioning in end-to-end autonomous driving makes speed an explicit user interface. Bench2Drive-Speed introduces a target-speed command and an overtake/follow command, a route-segment-wise speed profile $w=38^\circ$2, and evaluation metrics including Speed-Adherence Score and Overtake Score. The benchmark includes 48 evaluation routes and a Customized dataset of 2,100 clips. A central result is that speed supervision can be derived either from expert demonstrations or by re-annotating regular driving logs with a “Virtual Target Speed”; on the Drive benchmark, TCP without speed command attains Speed-Adherence 41.54, whereas TCP-Speed reaches 68.79 on Expert2.1k and 69.23 on Virtual2.1k. The paper further reports that target-speed following can be achieved without degrading regular driving performance, while overtaking remains challenging (Shao et al., 26 Mar 2026).

Other engineered systems also condition on speed. In vision-based multi-object tracking, SG-LKF treats ego-vehicle speed as a real-time signal for adaptive covariance modeling; its MotionScaleNet predicts $w=38^\circ$3, $w=38^\circ$4, and related covariance terms from speed and object scale, and the method achieves HOTA 79.59% on KITTI 2D MOT and AMOTA 69.00% on nuScenes 3D MOT (Gong et al., 1 Aug 2025). In small-cell networks, user speed enters directly into downlink power control: the optimal continuous-speed power law is affine in speed, so faster users receive more power and experience a larger “virtual” cell, with simulation gains reported up to 89% in some configurations relative to equal power (Kavitha et al., 2018). These cases treat speed as a conditioning variable for uncertainty modeling, handover mitigation, or behavior specification.

4. Speed limits, slowed dynamics, and accelerated simulation

Open quantum systems use speed in the sense of the rate of state conversion. For Lindblad/GKSL dynamics,

$w=38^\circ$5

the improved quantum speed-limit result separates a coherent contribution from a dissipative one: $w=38^\circ$6 The first term is governed by the SLD quantum Fisher information and represents quantum coherence or asymmetry; the second is governed by entropy production and an improved mobility term, representing classical bath-induced mixing. The paper’s central claim is that the quantum enhancement in speed is determined by the quantum Fisher information rather than the full energy variance, yielding tighter bounds on minimum evolution time and on observable currents (Sekiguchi et al., 2024).

Plasma simulation uses an apparently paradoxical strategy: slowing particles in order to accelerate computation. Speed-limited particle-in-cell simulation introduces a speed-limiting factor $w=38^\circ$7 through $w=38^\circ$8, leading to slowed equations of motion

$w=38^\circ$9

The resulting timestep is chosen roughly as $V_{rad}=1.14\,V_{exp}$0, where $V_{rad}=1.14\,V_{exp}$1 is the speed limit, rather than being constrained by the maximum physical particle speed or the electron plasma frequency. In a 1D electrostatic argon sheath problem, choosing $V_{rad}=1.14\,V_{exp}$2 enabled $V_{rad}=1.14\,V_{exp}$3, and the simulation reached the same steady state in about 600 timesteps and was about 160 times faster than PIC, without loss of accuracy (Werner et al., 2015). The method remains explicit and first-principles, but only for regimes where the suppressed fast motion is not itself the physics of interest.

A related acceleration strategy appears in mixed quantum-classical dynamics. Single potential evaluation Ehrenfest dynamics (SPEED) replaces trajectory-specific potential evaluations by a common local quadratic effective potential built around the ensemble-averaged nuclear position,

$V_{rad}=1.14\,V_{exp}$4

All trajectories are then propagated with a shared diabatic Taylor expansion, so one Hessian evaluation per step replaces one evaluation per trajectory. The method is exactly equivalent to standard Ehrenfest dynamics for diabatic potential energy surfaces and couplings that are at most quadratic, and this equivalence is demonstrated for a quadratic vibronic coupling model of pyrazine and for a model of atomic adsorption on a metallic surface. Combined with ALMO(MSDFT2), it qualitatively captures the temperature dependence of hole transfer in a furan dimer and accurately predicts the final 1:1 charge distribution, but it is insufficient for retinal photoisomerization because of strong anharmonicity (Scheidegger et al., 18 Mar 2025). In both SLPIC and Ehrenfest SPEED, “speed” is operationalized as the removal of stiff fast scales while preserving the slower dynamics of interest.

5. SPEED as architecture, processor, decoder, and sensor

Several recent systems use SPEED as a formal acronym. In edge AI hardware, SPEED is a scalable RISC-V vector processor for multi-precision DNN inference. It extends RVV v1.0 with customized instructions including $V_{rad}=1.14\,V_{exp}$5, $V_{rad}=1.14\,V_{exp}$6, $V_{rad}=1.14\,V_{exp}$7, and $V_{rad}=1.14\,V_{exp}$8, supports 4-, 8-, and 16-bit precision, and incorporates a parameterized multi-precision tensor unit with processing elements that can perform one 16-bit MAC, four 8-bit MACs, or sixteen 4-bit MACs. Synthesized in TSMC 28 nm, the processor achieves a peak throughput of 737.9 GOPS and peak energy efficiency of 1383.4 GOPS/W for 4-bit operators; in model-level evaluation it averages 4.88× speedup over Ara at 16-bit and 11.89× at 8-bit, and its design-space exploration identifies 4 lanes as the best balance between throughput and area utilization (Wang et al., 2024).

In language-model inference, Speculative Pipelined Execution for Efficient Decoding uses early-layer hidden states to predict future tokens and overlaps multiple token positions inside a cyclically parameter-shared decoder. Training uses a weighted multi-exit objective,

$V_{rad}=1.14\,V_{exp}$9

and inference tracks speculative predictions, invalidates downstream work when a token changes, and correspondingly invalidates KV-cache entries. The paper reports that for 2x6 configurations, after the first layer only about 13–17% of predictions need correction, and for 4x3 configurations about 6–14% do. The resulting accuracy-latency tradeoff is better than simply using shallow decoders, especially when output sequences are long enough to keep the pipeline occupied (Hooper et al., 2023).

In imaging, Single-pulse Photoacoustic Electromagnetic Detection proposes an image sensor that converts electromagnetic radiation into transient heating at absorbing pixels and reconstructs the scene from generated acoustic waves. The governing optoacoustic equation is

$V_{exp}=897\ \mathrm{km/s}$0

Simulation with a $V_{exp}=897\ \mathrm{km/s}$1 absorbing-pixel array shows correct reconstruction with no quality loss when the number of detectors is half or one-sixth of the number of absorbing pixels, but degradation when it is one-twentieth. A single-detector geometry fails in a homogeneous medium yet succeeds when an acoustic scatterer breaks symmetry. The paper discusses potential imaging rates around 50 Mfps in one geometry and around 1 Mfps in another, with a theoretical upper limit near $V_{exp}=897\ \mathrm{km/s}$2 fps set by non-radiative relaxation time (Aguirre, 2022). Across these systems, SPEED names architectures whose purpose is not merely high raw velocity but efficient reuse of shared structure—whether weights, tensor hardware, or acoustic channels.

6. Combinatorial speed and the interpretation of value

In graph theory, speed is a complexity function rather than a kinematic variable. For a graph class $V_{exp}=897\ \mathrm{km/s}$3,

$V_{exp}=897\ \mathrm{km/s}$4

and the speed is the map $V_{exp}=897\ \mathrm{km/s}$5, counting labeled graphs on vertex set $V_{exp}=897\ \mathrm{km/s}$6. The standard asymptotic regimes are factorial,

$V_{exp}=897\ \mathrm{km/s}$7

at most factorial, and exponential,

$V_{exp}=897\ \mathrm{km/s}$8

A main theorem states that if $V_{exp}=897\ \mathrm{km/s}$9 is finite, then the graph class $1023\ \mathrm{km/s}$0 represented by that finite binary language has at most factorial speed. The paper uses this criterion to classify many previously unknown graph classes as factorial, while proving that $1023\ \mathrm{km/s}$1-letter graphs have exponential speed (Feng et al., 27 Feb 2026). This is a fully non-kinematic use of the term: speed measures class size growth, not motion.

Transport economics treats speed as a welfare object and questions whether the conventional value of saved travel time is an adequate proxy. The paper argues that travel time is empirically rather stable—roughly around an hour per person per day in the studies it reviews—and that increases in transport speed are reflected more strongly in travel distance and land-use reorganization than in time savings. It distinguishes proximity, access, and accessibility, and reports that in a regression of travel time on speed, the speed coefficient is not significant ($1023\ \mathrm{km/s}$2), whereas the association between speed and travel distance is very strong ($1023\ \mathrm{km/s}$3). On this basis it concludes that the main effect of speed is better understood as a change in human wealth, not a direct time saving, and that the value of time is not a good proxy for the value of speed because value of speed is negatively correlated with prosperity while value of time is positively correlated (Goeverden, 2021).

Taken together, these usages suggest that “speed” is best understood as a field-specific rate concept whose interpretation depends on the state space being traversed. In some domains it is a projected velocity requiring geometric correction; in others it is a learned latent variable, a policy condition, a thermodynamic or quantum limit, a class-growth exponent, or the name of a system architecture. The literature repeatedly rejects naive identification of speed with a single scalar observable: expansion speed differs from radial speed in CMEs, instantaneous speed differs from propulsive speed in catalytic motors, playback speed is constrained by intelligibility rather than physics, and transport speed is not reducible to saved travel time [(Gopalswamy et al., 2012); (Mestre et al., 2020); (Kawamura et al., 2024); (Goeverden, 2021)].