ScaleDisturb Mechanisms

• ScaleDisturb is a recurrent scaling principle where disturbances vary with system, observation, or control scale, influencing outcomes in diverse fields.
• It categorizes mechanisms where scale enters through forcing laws, response kernels, or measurement operators, thereby altering propagation and detectability.
• In DRAM, ScaleDisturb exploits temporal asymmetry to lower activation thresholds, challenging traditional ECC and mitigation strategies.

to=arxiv_search ＿影音先锋 期六合:{"query":"ScaleDisturb arXiv (Wang et al., 5 Jun 2026) temporal asymmetry DRAM read disturbance", "max_results": 5} ScaleDisturb denotes a family of scale-dependent disturbance mechanisms in which the magnitude, propagation, detectability, or mitigation of a perturbation depends explicitly on a system scale, an observation scale, or a scale-coupled control variable. Across the cited literature, this dependence appears as horizon scale in cosmology, curve length in stochastic geometry, local Stokes number in turbulence, Hurst exponent in landscapes, field-of-view in avalanche imaging, platoon length in networked control, and aggressor-row open-time asymmetry in DRAM read disturbance (Amin et al., 2014, Yan, 26 Nov 2025, Hartlep et al., 2017, Fehr et al., 2011, Chen et al., 2011, Herman et al., 2016, Wang et al., 5 Jun 2026). A plausible implication is that ScaleDisturb is best understood not as a single domain-specific object, but as a recurrent scaling principle: disturbances become qualitatively different once their driving, transport, or measurement is tied to the relevant scale of the underlying system.

1. Conceptual scope

Across these works, the controlling scale is not uniform. In some settings it is a physical size, such as the instantaneous curve length $L(t)$ in stochastic curve shortening flow or the core-scale wavelength $\lambda_{\rm MAX}$ in prestellar collapse. In others it is a spectral or observational variable, such as wavenumber $k$ in neutrino-induced growth, the Hurst exponent $H$ in watershed displacement, the local Stokes number $\mathrm{St}_r$ in inertial-range turbulence, the window width $W$ in avalanche imaging, the Lundquist number $S$ in current-sheet disruption, or the platoon length $N$ in bidirectional vehicle strings (Yan, 26 Nov 2025, Hernández, 2016, Fehr et al., 2011, Hartlep et al., 2017, Chen et al., 2011, Huang et al., 2019, Herman et al., 2016).

Domain	Scale variable	Disturbance observable
Stochastic curve shortening	$L(t)$	$\sigma L(t)\circ dW_t$
Cosmology and inflation	$\lambda_{\rm MAX}$0, $\lambda_{\rm MAX}$1, horizon entry	$\lambda_{\rm MAX}$2, $\lambda_{\rm MAX}$3, $\lambda_{\rm MAX}$4, $\lambda_{\rm MAX}$5
Landscapes and turbulence	$\lambda_{\rm MAX}$6, $\lambda_{\rm MAX}$7, $\lambda_{\rm MAX}$8	$\lambda_{\rm MAX}$9, multiplier PDFs, fragmentation
Avalanche imaging	$k$0	windowed scaling functions
Plasma and platoons	$k$1, $k$2	$k$3, $k$4, $k$5
DRAM	OTB, $k$6, $k$7	$k$8, bitflips

This taxonomy also separates three recurrent mechanisms. First, scale can enter the forcing law itself, as in $k$9. Second, scale can enter the response kernel or transfer function, as in $H$0, $H$1, or plasmoid growth rates. Third, scale can enter the measurement operator, as in field-of-view windowing and surface-brightness-limited cluster catalogs. The literature repeatedly shows that these cases should not be conflated: identical microscopic disturbances can yield different macroscopic laws once the relevant scale variable changes.

2. Cosmological and gravitational contexts

In early-universe cosmology, scale-dependent disturbance appears most directly in perturbations sourced by scaling seeds. Global, symmetry-breaking phase transitions generate horizon-scale field gradients that act as active gravitational seeds; the resulting seed-induced acoustic waves in the photon–baryon plasma are Silk damped and produce CMB spectral distortions. In the large-$H$2 $H$3 nonlinear $H$4-model, the seed potentials at horizon crossing scale as $H$5, and solving the sourced Boltzmann–Einstein system yields

$H$6

Saturating Planck temperature-anisotropy bounds gives the paper’s headline prediction, $H$7 and $H$8 up to decoupling (Amin et al., 2014).

Massive neutrinos furnish a distinct scale-dependent effect: free streaming suppresses clustering below a characteristic scale and induces a $H$9-dependent linear growth rate $\mathrm{St}_r$0. The reported signature is not a large tilt, but a percent-level running: typically $\mathrm{St}_r$1 varies between $\mathrm{St}_r$2–$\mathrm{St}_r$3 from low $\mathrm{St}_r$4 to high $\mathrm{St}_r$5 over the modeled range. In the linear Kaiser description,

$\mathrm{St}_r$6

so the anisotropy retains the usual $\mathrm{St}_r$7-dependence while acquiring a $\mathrm{St}_r$8-dependent amplitude. The Fisher forecasts emphasize that detecting this effect is volume-limited rather than $\mathrm{St}_r$9-limited: a survey of about $W$0 at $W$1 is needed to detect the scale-dependent growth for all considered neutrino masses, while $W$2 is needed to reach $W$3 across the allowed $W$4 range (Hernández, 2016).

A useful counterexample appears in DHOST inflation. Derivative scordatura, despite modifying the higher-derivative sector, does not generate scale dependence in de Sitter: in all derivative scordatura cases studied, the scalar power spectrum remains exactly scale-invariant and $W$5. Nontrivial ScaleDisturb arises only when shift symmetry is broken by the axion-like potential

$W$6

which induces explicit time dependence in the perturbation coefficients and yields Planck-compatible values such as $W$7, $W$8, $W$9, $S$0 in one benchmark, and $S$1, $S$2, $S$3, $S$4 in another (Brax et al., 2021). This directly rebuts the common misconception that any higher-derivative detuning automatically produces scale dependence.

Wave transport from a uniform distribution provides a further cosmological-scale disturbance mechanism. For an infinite lattice of incoherent emitters, exact compensation between oscillator energy loss and far-field energy transport gives $S$5 at leading order, consistent with the Zel’dovich bound. At next order, however, near fields $S$6 yield a finite positive large-scale power, with $S$7 at late times, while the two-point function still vanishes for $S$8 once emission is truncated. The result is therefore scale-disturbing without being acausal (Lieu, 2017).

In a different gravitational setting, a globally scale-invariant scalar–tensor model admits a de Sitter background in which scalar, vector, and tensor perturbations all decay. The scalar mode satisfies a damped wave equation,

$S$9

and the exact solution decays as $N$0 near the future boundary. Here scale invariance of the action is compatible with perturbative stability, rather than with amplified scale dependence (Jain et al., 2011).

3. Hierarchical media and cascade disturbances

In geomorphology, the disturbance of a watershed by a local perturbation is explicitly scale-free. For real landscapes, the enclosed-area distribution follows

$N$1

while the outlet-separation distribution obeys

$N$2

The conditional distribution satisfies $N$3 and $N$4. In artificial fractional Brownian surfaces, the exponents vary approximately linearly with the Hurst exponent $N$5, and the paper stresses that the power laws are independent of perturbation magnitude: even infinitesimal $N$6 can trigger large, nonlocal watershed shifts (Fehr et al., 2011).

Turbulent particle clustering displays an analogous but dynamically local notion of scale dependence. Cascade multiplier PDFs for particle concentration, dissipation, and enstrophy are all scale dependent, but the particle multipliers collapse when parameterized by the local Stokes number

$N$7

The collapsed $N$8 curve is U-shaped, with strongest intermittency near $N$9. By contrast, enstrophy and dissipation multipliers approach scale-independent asymptotes at sufficiently small scales, with $L(t)$0 and $L(t)$1 in the reported DNS. This scale-local parameterization then supports inertial-range cascade predictions for the radial distribution function at Reynolds numbers inaccessible to DNS (Hartlep et al., 2017).

In weakly turbulent prestellar cores, the critical scale is not the spectral slope $L(t)$2 itself, but the maximum turbulent wavelength $L(t)$3, provided $L(t)$4. Because the largest mode carries most of the turbulent energy, $L(t)$5 controls both the amount and coherence of angular momentum, the formation of dense filaments, and the fragmentation pathway. The study reports that the core only has a high probability of fragmenting if $L(t)$6, with fragmentation common for $L(t)$7–$L(t)$8 and absent in the sampled realizations at $L(t)$9 or $\sigma L(t)\circ dW_t$0. Small disks with $\sigma L(t)\circ dW_t$1 form routinely, whereas large fragmenting disks are rare because early filament fragmentation dominates (Walch et al., 2011).

Hierarchical structure also governs star-cluster disruption. In the hierarchical-ISM picture, clusters drift away from dense birth sites through a cloud complex whose density falls with radius, so the harassment or collision hazard declines with age. When the effective hazard takes the form

$\sigma L(t)\circ dW_t$2

the survival probability becomes $\sigma L(t)\circ dW_t$3 and the observed mass–age distribution is reproduced as

$\sigma L(t)\circ dW_t$4

The model was constructed precisely to explain the empirical $\sigma L(t)\circ dW_t$5 decline with $\sigma L(t)\circ dW_t$6–$\sigma L(t)\circ dW_t$7 without stitching together unrelated fixed-rate mechanisms (Elmegreen et al., 2010).

4. Geometric, kinetic, and observational formulations

In stochastic geometry, scale-dependent disturbance is literal: the noise amplitude is proportional to the curve length. The stochastic curve shortening flow is

$\sigma L(t)\circ dW_t$8

so larger interfaces experience stronger perturbations. Rewriting the problem in curvature–length variables yields a stochastic one-phase Stefan problem; after transforming to a fixed domain and converting to Itô form, the system becomes a quasilinear SPDE–SDE. Using the Agresti–Veraar framework for quasilinear stochastic evolution equations, the paper proves a unique $\sigma L(t)\circ dW_t$9-maximal local strong solution for sufficiently small $\lambda_{\rm MAX}$00, with blow-up characterized by either curvature divergence or collapse of the length $\lambda_{\rm MAX}$01 to zero (Yan, 26 Nov 2025).

Collisionless stellar dynamics presents a different kind of scale disturbance: a neutral, scale-invariant mode that stretches or shrinks a spherical equilibrium while preserving total mass. For ergodic models with a single length parameter $\lambda_{\rm MAX}$02, the equilibrium family satisfies

$\lambda_{\rm MAX}$03

and the perturbation $\lambda_{\rm MAX}$04 is an exact stationary solution of the perturbed Vlasov–Poisson system. The paper derives explicit first- and second-order fields and shows that the true second-order perturbation energy and the familiar bilinear pseudoenergy are both integrals of motion but differ by a constant (Polyachenko et al., 2023).

Observation itself can be the source of ScaleDisturb. In avalanche imaging, limited field of view produces systematic distortions that reorganize universal scaling functions into three classes: internal avalanches $\lambda_{\rm MAX}$05, split avalanches $\lambda_{\rm MAX}$06, and spanning avalanches $\lambda_{\rm MAX}$07. The qKPZ study develops multivariable scaling forms in the two control parameters $\lambda_{\rm MAX}$08 and $\lambda_{\rm MAX}$09, with fitted full-system exponents

$\lambda_{\rm MAX}$10

The main encyclopedic point is methodological: apparent scale dependence may be induced by finite observation windows rather than by the underlying avalanche dynamics (Chen et al., 2011).

5. Collective dynamics, transport, and system-size amplification

Current-sheet disruption by the plasmoid instability is a scale-dependent disturbance problem in which thinning, resistive growth, and reconnection outflow compete. For a current sheet of half-length $\lambda_{\rm MAX}$11 and half-width $\lambda_{\rm MAX}$12, with $\lambda_{\rm MAX}$13, the dominant tearing mode at disruption satisfies

$\lambda_{\rm MAX}$14

while the disruption width has the characteristic form of a power law multiplied by a logarithmic factor,

$\lambda_{\rm MAX}$15

The paper distinguishes two seeding scenarios—an injected initial perturbation and system noise—and shows that reconnection outflow changes the effective scaling because initial noise can decay while system noise acts as a floor (Huang et al., 2019).

In bidirectional vehicle platoons, the scale variable is the string length $\lambda_{\rm MAX}$16, and the disturbance metric is the growth of control errors under additive disturbances. The sharp result is structural. With symmetric position coupling and symmetric velocity coupling, the energy-like bound scales quadratically, $\lambda_{\rm MAX}$17. With symmetric position coupling and asymmetric velocity coupling, linear scaling $\lambda_{\rm MAX}$18 is achieved. With asymmetric position coupling, exponential scaling may occur, and the system may even become unstable. In the paper’s summary notation,

• SPSV: $\lambda_{\rm MAX}$19,
• SPAV: $\lambda_{\rm MAX}$20,
• APAV: $\lambda_{\rm MAX}$21 with $\lambda_{\rm MAX}$22. The design lesson is correspondingly precise: symmetry in the position coupling and asymmetry in the velocity coupling qualitatively improves string performance (Herman et al., 2016).

These two cases illustrate a broader distinction. In plasmas, the decisive scale law is set by competition between amplification and advection. In platoons, it is set by how the inter-agent coupling geometry interacts with system size. Both are ScaleDisturb problems, but only the latter makes asymmetry itself the control knob.

6. ScaleDisturb in DRAM read disturbance

In modern memory systems, ScaleDisturb is the proper name of a DRAM access pattern that amplifies read disturbance by exploiting temporal asymmetry across two aggressor rows flanking a victim row. The mechanism is defined at fixed total open-time budget,

$\lambda_{\rm MAX}$23

so the effect does not arise from simply increasing total aggressor open time. The two limiting cases are explicit: $\lambda_{\rm MAX}$24 reduces to double-sided RowHammer, and $\lambda_{\rm MAX}$25 reduces to double-sided RowPress. ScaleDisturb instead sweeps asymmetric splits along the constant-OTB isocontour, breaking the partial cancellation of electric fields that can occur when both adjacent wordlines are held open for equal durations (Wang et al., 5 Jun 2026).

The command-level loop alternates the upper and lower aggressor rows. One loop consists of ACT Row$\lambda_{\rm MAX}$26 and hold for $\lambda_{\rm MAX}$27, PRE and wait $\lambda_{\rm MAX}$28, ACT Row$\lambda_{\rm MAX}$29 and hold for $\lambda_{\rm MAX}$30, PRE and wait $\lambda_{\rm MAX}$31, repeated until a bitflip is observed or a search bound is reached. The principal metric is $\lambda_{\rm MAX}$32, the minimum total number of aggressor activations needed to induce at least one victim-row bitflip. Lower $\lambda_{\rm MAX}$33 therefore means higher vulnerability.

The reported characterization covers 196 DDR4 chips and 3 HBM2 chips. Relative to double-sided RowPress at the same OTB, ScaleDisturb reduces $\lambda_{\rm MAX}$34 by $\lambda_{\rm MAX}$35 on average across rows and by up to $\lambda_{\rm MAX}$36 in individual rows; at the module level, reductions average $\lambda_{\rm MAX}$37 at smaller OTBs and reach $\lambda_{\rm MAX}$38 at $\lambda_{\rm MAX}$39. The reductions are reported across all three major vendors—$\lambda_{\rm MAX}$40 for Samsung, $\lambda_{\rm MAX}$41 for SK Hynix, and $\lambda_{\rm MAX}$42 for Micron on average—and vulnerability increases with technology scaling, with representative average-$\lambda_{\rm MAX}$43 reductions at $\lambda_{\rm MAX}$44 of $\lambda_{\rm MAX}$45 for Samsung $\lambda_{\rm MAX}$46 A$\lambda_{\rm MAX}$47B, $\lambda_{\rm MAX}$48 for SK Hynix $\lambda_{\rm MAX}$49 C$\lambda_{\rm MAX}$50D, and $\lambda_{\rm MAX}$51 for Micron $\lambda_{\rm MAX}$52 B$\lambda_{\rm MAX}$53F. The same trend generalizes to HBM2, where average reductions across the six tested OTBs were $\lambda_{\rm MAX}$54, $\lambda_{\rm MAX}$55, $\lambda_{\rm MAX}$56, $\lambda_{\rm MAX}$57, $\lambda_{\rm MAX}$58, and $\lambda_{\rm MAX}$59.

The per-row response is not uniform. Sweeping $\lambda_{\rm MAX}$60 at fixed OTB reveals three classes: L-type rows, whose minimum $\lambda_{\rm MAX}$61 occurs when Row$\lambda_{\rm MAX}$62 has shorter open time; R-type rows, whose minimum occurs when Row$\lambda_{\rm MAX}$63 has longer open time; and Flat-type rows, with less than $\lambda_{\rm MAX}$64 reduction. L-type and R-type together account for $\lambda_{\rm MAX}$65 of rows on average, with nearly equal rates, which the paper interprets as consistent with device-level asymmetry in upper versus lower aggressor coupling.

The software-level exploitability is demonstrated on a real Intel-based system with in-DRAM TRR. A user-level program alternates between the two aggressor rows, prolongs each row’s effective open time by issuing multiple cache-line reads and clflushopt operations, uses mfence to serialize, and accesses 16 dummy rows four times each to exhaust TRR tracking capacity. On an Ubuntu 18.04 system with an Intel i5-10400 and a 16 GB dual-rank DDR4 DIMM, the attack flipped 33 of the top-50 rows under low dummy-row activation frequency ($\lambda_{\rm MAX}$66), versus only 3 for double-sided RowPress; for rows flipped by both, ScaleDisturb induced up to 34 bitflips where double-sided RowPress induced at most 5. Aggregate bitflip counts also favored ScaleDisturb strongly, for example 289 versus 8 at $\lambda_{\rm MAX}$67.

The mitigation analysis is correspondingly severe. ECC alone is inadequate because ScaleDisturb can induce up to 40 bitflips in a single 64-bit word, beyond SECDED and ChipKill capability, and large fractions of words per victim row have at least 3 bitflips. Lowering protection thresholds in Graphene, Hydra, PRAC, PRFM, and PARA by safety margins improves security but incurs nontrivial performance and energy cost: at threshold 128 and 60% margin, the performance overheads relative to the same mechanisms without margin were $\lambda_{\rm MAX}$68, $\lambda_{\rm MAX}$69, $\lambda_{\rm MAX}$70, $\lambda_{\rm MAX}$71, and $\lambda_{\rm MAX}$72, with energy overhead reaching $\lambda_{\rm MAX}$73 for PARA. Adapting RowPress mitigation via open-time-aware counting also creates asymmetry-induced over-refresh. The proposed TeACUp mechanism addresses this by scaling the faster row’s counter increments using

$\lambda_{\rm MAX}$74

and updating $\lambda_{\rm MAX}$75, $\lambda_{\rm MAX}$76. In the reported workloads, TeACUp improved normalized performance by $\lambda_{\rm MAX}$77 on average, up to $\lambda_{\rm MAX}$78, relative to ImPress.

The DRAM case crystallizes the strongest form of the ScaleDisturb concept. A disturbance need not increase by adding more energy or more activations; it can increase because the same budget is repartitioned asymmetrically across the relevant scale variable. In this sense, temporal asymmetry is the hardware-security analogue of the broader literature’s central lesson: scale-coupled structure changes what a disturbance does.