ColumnDisturb: Column-Based Disturbance Mechanisms

• ColumnDisturb is a multifaceted term characterizing disturbances organized along column structures in domains such as fluid dynamics, soft matter, control systems, and memory architecture.
• It encompasses phenomena like the formation and breakdown of columnar vortices, instabilities in active media, and cellwise perturbations that preserve statistical properties.
• Mitigation strategies range from adaptive control in networked systems and robust statistical adjustments to hardware-level refresh schemes in DRAM to counteract column-based failures.

Searching arXiv for the term to ground the article in the cited papers. arxiv_search("ColumnDisturb", max_results=10, sort_by="relevance") ColumnDisturb is not a single canonical construct. In the cited literature, the label denotes multiple domain-specific phenomena and methods whose common organizing motif is the column: columnar vortices in rapidly rotating flows, disturbance rejection in the column space of a graph Laplacian, column-wise perturbation or diagnosis of data matrices, disturbances and instabilities of physical columns in colloidal, active, liquid-crystalline, sedimentation, and granular systems, and column-based read disturbance in DRAM. This suggests that the term functions as a recurring descriptor for column-centered disturbance, instability, control, or masking mechanisms rather than a unified theory (Whitehead et al., 2012, Yucelen et al., 2013, Maruyama et al., 2015, Toledo-Zucco et al., 2023, Yüksel et al., 16 Oct 2025).

1. Terminological scope and recurring motif

Across the cited papers, the word “column” refers to different structural objects: vertically coherent vortices in rotating Boussinesq flow, columns in ordered soft matter and granular media, columns of variables in data matrices, columns of agents’ disagreement dynamics through $\mathrm{col}(\mathcal{L}(\mathcal{G}))$, physical sedimentation columns, and DRAM columns implemented as bitlines. The accompanying “disturb” component is likewise heterogeneous: it may denote stochastic forcing, persistent exogenous disturbance, privacy-preserving perturbation, cellwise deviation, hydrodynamic or morphological instability, or read disturbance in memory hardware (Swan et al., 2013, Kole et al., 2023, Belli et al., 2010, Rousseeuw et al., 2016, Heussinger, 2023, Das et al., 25 Aug 2025, Kakolyris et al., 21 Jun 2026).

A common misconception would be to assume that ColumnDisturb names one standardized method or one physical effect. The cited usage indicates otherwise. What is shared is not ontology but structure: a disturbance process is organized by a columnar degree of freedom, or a mitigation acts specifically on a column-aligned subspace, geometry, or hardware path.

2. Rotating and stratified flows

In the fluid-dynamical usage introduced by Whitehead and Wingate, a rotating, weakly stratified Boussinesq fluid is driven from rest by stochastic “white-noise” momentum forcing at an intermediate wavenumber and, under rapid rotation, spontaneously organizes into vertically coherent columnar vortices. The DNS solves

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$

$$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$

with momentum forcing centered spectrally at $k \approx 3$, zero initial conditions, $Ro = 0.05$, $Fr = 1$, and a $256^3$ grid. The stated mechanism combines geostrophic balance, Taylor–Proudman columnarity at small Rossby number, weak stratification with finite Froude number, and near-conservation of Ertel potential vorticity $q = (\omega + f \hat z)\cdot \nabla b$. The video shows emergence, persistence, advection, interaction, and merging of vertically coherent vortex columns whose horizontal scale is near the forcing wavelength (Whitehead et al., 2012).

A distinct but related rotating-turbulence usage studies the destruction rather than the formation of columns. In direct numerical simulations of forced rotating turbulence in a three-dimensional periodic box, intermittent bursting of cyclonic columnar structures is observed only when the random isotropic forcing is applied at low wavenumber, $k_f=[3,4]$, and not at $k_f=[6,7]$, for the same rotation rate and initial conditions. The reported diagnostics include energy spectra, ring spectra, time series of modal kinetic energy for modes with $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$0, and mode-to-mode energy transfer

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$1

The reported interpretation is that a circular cyclonic vortex becomes elliptical because of cyclone–anticyclone interactions, leading to elliptical instability, while energy transfer to higher-$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$2 modes triggers the Crow instability; the burst is then followed by re-formation of the column due to external rotation. Growth rates are reported for pre-burst modes, for example $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$3, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$4, and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$5, and the representative case exhibits bursts at $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$6 (Das et al., 25 Aug 2025).

Taken together, these two uses define a rotating-flow motif for ColumnDisturb. At low Rossby number, stochastic forcing can self-organize vertical vorticity into Taylor–Proudman-like columns, yet the same rotationally constrained state can later undergo burst-like breakdown when triadic transfer feeds higher-$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$7 content. This suggests that “disturbance” in rotating columns can mean either spontaneous organization under broadband forcing or intermittent loss of that organization under instability-mediated forward transfer.

3. Columnar order, instability, and collapse in soft and granular matter

In a magnetically annealed colloidal dispersion, ColumnDisturb refers to the disturbance and decomposition of particle-rich columns formed under a uniform pulsed magnetic field. Superparamagnetic beads with radius $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$8 form dense threads aligned with the field during the field-on phase. At low pulse frequency, especially $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$9, these columns undergo a Rayleigh–Plateau instability: sinusoidal undulations grow, necks form, and pinch-off produces droplets. The instability criterion is stated as $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$0, equivalently $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$1, with the inviscid fastest mode at $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$2 and $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$3. As pulse frequency increases to $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$4 the instability becomes coarser and less pronounced, and at $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$5 and $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$6 the system instead forms a slowly relaxing, system-spanning network. The paper interprets this as degeneration of the Rayleigh–Plateau instability as frequent re-magnetization shortens the off-time available for capillary-like growth and increases the effective viscosity contrast (Swan et al., 2013).

In ordered active matter, the relevant object is the active columnar phase: a three-dimensional fluid with a two-dimensional translationally ordered lattice of columns perpendicular to a mean column axis. The hydrodynamic theory formulated within an active Model H$$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$7 framework yields an in-plane displacement field $$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$8, passive elastic free energy

$$\partial_t b + u\cdot\nabla b + N^2 w = \kappa \nabla^2 b + \Phi(x,t), \qquad \nabla\cdot u = 0,$$9

and three active stresses: achiral force-dipole stress, apolar chiral stress, and polar chiral stress producing odd elasticity. Reported consequences include two-dimensional odd elasticity from three-dimensional plasmon-like oscillations, an active Helfrich–Hurault buckling instability with threshold

$k \approx 3$0

and handed helical undulation selection by apolar chirality. Here, ColumnDisturb is a disturbance spectrum of relaxational, chiral, and oscillatory modes supported by ordered columns (Kole et al., 2023).

In equilibrium simulations of columnar liquid crystals of perfectly aligned hard spherocylinders, the disturbance mechanism is heterogeneous hopping between columns. The in-plane broken translational symmetry generates a periodic potential of mean force $k \approx 3$1, with barrier heights that grow with packing fraction and rod anisotropy. For long rods at $k \approx 3$2 and $k \approx 3$3, the barriers are reported as “close to and even higher than” $k \approx 3$4. The consequences are intermittent inter-column jumps, non-Gaussian transverse diffusion, a plateau in the transverse mean-squared displacement, a peak in the transverse non-Gaussian parameter, and two-step relaxation in the self-intermediate scattering function. The longitudinal direction remains more liquid-like but still shows slight deviations from Gaussian behavior (Belli et al., 2010).

In columns of granular rods, the disturbance problem becomes one of collapse versus stability. Three-dimensional DEM simulations with spherocylinders show that free-standing columns are possible because rod contacts can persist while sliding tangentially along rod axes, mobilizing what the paper calls “frictional cohesion.” The contact law uses a Cundall–Strack force with Coulomb cap $k \approx 3$5, plus viscous rolling and twisting torques. Stability is controlled primarily by friction coefficient $k \approx 3$6 and aspect ratio $k \approx 3$7. Representative thresholds reported are $k \approx 3$8–$k \approx 3$9 and $Ro = 0.05$0 for $Ro = 0.05$1, whereas nearly spherical grains always collapse in the explored range. The paper interprets stability through the emergence of a Mohr–Coulomb-like cohesion term $Ro = 0.05$2 with $Ro = 0.05$3 for rods and $Ro = 0.05$4 for spheres (Heussinger, 2023).

These soft- and granular-matter uses share a common structure: columnar order produces a distinct failure or transport pathway not present in a simple isotropic medium. In colloids the column is destabilized by Rayleigh–Plateau breakup, in active matter by chiral and odd-elastic modes, in hard-rod liquid crystals by entropic hopping across inter-column barriers, and in granular rods by gravitational collapse resisted by rod-specific frictional contacts.

4. Distributed control in networked systems

In distributed control, ColumnDisturb denotes a disturbance-rejection architecture for consensus and formation of single-integrator multiagent systems under unknown persistent disturbances. The agent dynamics are

$Ro = 0.05$5

on a static, connected, undirected graph with Laplacian $Ro = 0.05$6. The baseline consensus and formation terms are $Ro = 0.05$7 and $Ro = 0.05$8. The new element is an adaptive integral term $Ro = 0.05$9 driven by a predictor error $Fr = 1$0, where

$Fr = 1$1

with $Fr = 1$2, $Fr = 1$3, and

$Fr = 1$4

where $Fr = 1$5 and $Fr = 1$6. The paper’s central interpretation is that $Fr = 1$7 is a localized surrogate of the orthogonal projector onto $Fr = 1$8, so the integral action lives in the disagreement subspace and annihilates the consensus direction (Yucelen et al., 2013).

The constant-disturbance analysis shows that $Fr = 1$9 is Hurwitz and that $256^3$0 has $256^3$1 positive eigenvalues and one zero eigenvalue. The resulting closed-loop error system is Lyapunov-stable, drives the $256^3$2 component to zero, and achieves consensus or formation in the disagreement subspace. Without an additional local term, the consensus value may drift along $256^3$3; adding $256^3$4 to the integral update yields $256^3$5 and $256^3$6, removing drift and recovering convergence to a constant consensus or formation point. For time-varying disturbances with bounds $256^3$7 and $256^3$8, adding $256^3$9 produces uniform ultimate boundedness under the stated Assumption 1. The reported examples use a cycle graph with $q = (\omega + f \hat z)\cdot \nabla b$0, $q = (\omega + f \hat z)\cdot \nabla b$1, $q = (\omega + f \hat z)\cdot \nabla b$2, and, when needed, $q = (\omega + f \hat z)\cdot \nabla b$3 or $q = (\omega + f \hat z)\cdot \nabla b$4 (Yucelen et al., 2013).

This usage is mathematically precise and differs sharply from the physical-column papers. Here “column” refers to the column space of the Laplacian, and disturbance handling is achieved by projected integral action rather than by geometric or hydrodynamic mechanisms.

5. Statistical data perturbation and cellwise diagnostics

One statistical usage of ColumnDisturb is a column-wise perturbation of a dependent variable designed to preserve ordinary least squares outputs exactly. Let $q = (\omega + f \hat z)\cdot \nabla b$5 include an intercept and have full column rank, and let $q = (\omega + f \hat z)\cdot \nabla b$6 be the dependent variable. The perturbation is $q = (\omega + f \hat z)\cdot \nabla b$7 with $q = (\omega + f \hat z)\cdot \nabla b$8, so $q = (\omega + f \hat z)\cdot \nabla b$9 and $k_f=[3,4]$0. The constructive procedure computes the residuals $k_f=[3,4]$1, draws a random vector $k_f=[3,4]$2, forms

$k_f=[3,4]$3

and sets

$k_f=[3,4]$4

With the recommended choice $k_f=[3,4]$5 and $k_f=[3,4]$6, the paper states exact invariance of OLS coefficients, fitted values, residual sum of squares, $k_f=[3,4]$7, estimated noise variance, covariance matrix of $k_f=[3,4]$8, standard errors, $k_f=[3,4]$9-statistics, and $k_f=[6,7]$0-values. In the real-estate case study of $k_f=[6,7]$1 cases of newly built detached houses in Setagaya Ward, $k_f=[6,7]$2 is recommended as a practical default and $k_f=[6,7]$3 gives very high robustness of model equivalence in the reported subsampling tests (Maruyama et al., 2015).

A second statistical usage concerns diagnosis rather than masking. The DetectDeviatingCells method is designed for multivariate data with cellwise outliers, including situations in which many rows have a few contaminated cells. The data matrix is robustly standardized columnwise, univariate outliers are capped at

$k_f=[6,7]$4

robust pairwise correlations and slopes are estimated, and each cell is predicted from correlated variables. Standardized residuals

$k_f=[6,7]$5

are then thresholded at the same $k_f=[6,7]$6. The method exploits the fact that a column disturbance may be invisible marginally but detectable conditionally through violated inter-variable relations. The paper emphasizes that if the per-cell contamination rate is $k_f=[6,7]$7, the expected contaminated-row fraction is $k_f=[6,7]$8, so rowwise robust methods quickly fail as dimension grows, whereas cellwise methods remain viable. The default correlation threshold is $k_f=[6,7]$9, and the algorithm simultaneously imputes missing values and can flag rows by aggregating cellwise residual evidence (Rousseeuw et al., 2016).

These two data-analytic uses are complementary. The former adds noise in the residual space to preserve one specified linear analysis exactly; the latter identifies cells or columns whose observed values depart from what the rest of the data predict.

6. Sedimentation columns and turbidity control

In process systems, ColumnDisturb refers to disturbance handling in a sedimentation column used for water recovery from slurries. The measured output is turbidity at the top of the column, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$00, and the manipulated variable is the net flow command $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$01. Because the phenomenological model is considered too complex for control design and difficult to identify from plant data, the paper proposes an empirical direction-dependent piecewise time-delay model,

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$02

where $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$03 for $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$04, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$05 for $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$06, and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$07 for $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$08. With sampling time $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$09, the identified parameters are $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$10, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$11, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$12, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$13 for increasing input and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$14, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$15, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$16, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$17 for decreasing input (Toledo-Zucco et al., 2023).

The controller is a discrete PI law,

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$18

with gains chosen by a Common Lyapunov–Krasovskii Functional construction. Proposition 1 provides LMIs that certify asymptotic stability of the switched delayed closed loop, and Proposition 2 augments the LMIs with a guaranteed-cost bound

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$19

The approach is validated on a pilot plant with a sedimentation column of height $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$20 and cross-section $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$21, an Allen-Bradley ControlLogix PLC, and a turbidity probe at the top. Reported behavior includes stable regulation around the setpoint, first-order-like response when turbidity increases with effective closed-loop constant around $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$22, somewhat slower response when turbidity decreases, and bounded control effort under inflow disturbances (Toledo-Zucco et al., 2023).

Relative to the other meanings of ColumnDisturb, this one is operational rather than geometric. The “column” is the process vessel itself, and the disturbance problem is cast as switched-delay control under asymmetric dynamics.

7. DRAM read disturbance and mitigation

In computer architecture, ColumnDisturb denotes a newly demonstrated DRAM read-disturbance phenomenon that operates through columns, that is, bitlines, rather than rows. Repeatedly activating a single aggressor row or keeping it open for extended time disturbs cells sharing the same columns across multiple DRAM subarrays because of the open-bitline architecture and shared sense amplifiers. The characterization study evaluates $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$23 DDR4 chips and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$24 HBM2 chips from three major manufacturers. A single aggressor can disturb cells across three consecutive subarrays, as many as $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$25 rows in the tested chips, and no ColumnDisturb bitflips are observed in subarrays that do not share columns with the aggressor. The disturbance worsens with technology scaling; minimum time to first bitflip reduces by up to $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$26, and in one Micron 16 Gb F-die module multiple cells fail within $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$27 at $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$28, that is, within the nominal DDR4 refresh window. The phenomenon induces only $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$29 bitflips and affects up to $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$30 more rows than retention across the tested temperature levels and intervals (Yüksel et al., 16 Oct 2025).

The paper models the average column voltage during disturbance as

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$31

and reports strong sensitivity to temperature, aggressor row on-time, and data pattern. Lower $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$32 substantially increases vulnerability; at $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$33, reducing $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$34 increases failures by $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$35 for SK Hynix, $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$36 for Micron, and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$37 for Samsung. These observations differ qualitatively from RowHammer and RowPress because the blast radius is column-centric and extends across multiple subarrays rather than to a few neighboring rows (Yüksel et al., 16 Oct 2025).

The first dedicated mitigations are ColumnKeeper-D and ColumnKeeper-P. CK-D is deterministic: it uses two counters per subarray to track activations affecting the odd and even columns, triggers a preventive refresh of one row in a subarray when either counter reaches

$$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$38

and advances a round-robin row pointer table. CK-P is probabilistic: on each activation to the middle subarray, it refreshes one row in three consecutive subarrays with probability $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$39, where $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$40 is selected from the Binomial-CDF analysis to meet a target yearly success probability. At the current experimentally demonstrated threshold $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$41, CK-D and CK-P incur average single-core performance overheads of $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$42 and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$43, respectively; at $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$44 these rise to $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$45 and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$46. The reported area overheads are $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$47 for CK-D and $$\partial_t u + (u\cdot\nabla)u + f \hat z \times u = -\nabla \pi + b \hat z + \nu \nabla^2 u + F(x,t),$$48 for CK-P (Kakolyris et al., 21 Jun 2026).

This hardware usage is the most literal reading of “column disturbance.” It is also the one with the clearest systems implication: existing row-oriented disturbance models and mitigations do not directly generalize to a column path that spans three subarrays. The cited work therefore reframes disturbance management in DRAM from row locality to column sharing.