To decide from observational data alone whether $X$ causes $Y$ or the reverse, one common approach fits a regression in each direction. It then asks which direction leaves a residual (the part of the output the regression cannot explain) that is statistically independent of the input. That check needs a kernel independence test: a measure of dependence that compares inputs and residuals through many non-linear features rather than through correlation alone. Analog in-memory chips store a fixed matrix as physical conductances in a crossbar of phase-change memory cells and multiply by it almost for free. They have already been shown to compute such random non-linear features with large energy savings. The obstacle is that the stored conductances slowly decay after programming. This "conductance drift" follows a power law, and each cell decays at its own rate.
This note argues that the obvious way to handle drift is aimed at the wrong target when the goal is causal direction. The obvious way is to restore the drifted weights to their programmed values, for example with a global rescaling. Per-cell drift applied to random projection weights does not mainly add error. It turns the feature map into a slightly different, but still valid, kernel. A kernel independence test stays valid under any fixed characteristic kernel, meaning one able to detect any form of dependence. Drift therefore harms causal orientation only when the two directions, or the test statistic and its null calibration, are computed through different drift states. The proposal is to build and stress-test a "matched-kernel" protocol, then check whether it keeps orientation accuracy and test calibration flat over months of simulated drift while naive analog execution degrades. Matched-kernel means three things: both directions are scored through the same physical feature map within one short time window, the null calibration by permutation runs on-chip in that same window, and directions are compared by calibrated p-values rather than by raw dependence magnitudes.
Core Claim
Let $\omega_{ij}$ be the Gaussian random projection weights programmed into the crossbar at time $t_0$, and let $\nu_{ij}$ be each cell's drift exponent. At read time $t$, the analog feature map computes
$$\phi_t(x) = \sqrt{2/D}\,\cos\!\big(\tilde{A}_t x + b\big), \qquad (\tilde{A}_t)_{ij} = \omega_{ij}\,(t/t_0)^{-\nu_{ij}},$$
where $D$ is the number of features and $b$ is a digital random phase. Assume the $\nu_{ij}$ are independent of the $\omega_{ij}$, which is a modelling assumption we will check against simulator traces. Then each row of $\tilde{A}_t$ is a draw from a symmetric scale mixture of Gaussians. The expected inner product $k_t(x,x') = \mathbb{E}[\phi_t(x)^\top \phi_t(x')]$ is then still a shift-invariant, positive-definite kernel whose spectral distribution has full support, and so it is still characteristic.
The mean drift widens this kernel's bandwidth, and the spread of $\nu$ makes its tails heavier. Neither change invalidates a Hilbert–Schmidt Independence Criterion (HSIC) test whose null distribution is estimated by permutation under the same $\phi_t$. What does break validity is comparing a statistic computed under $k_{t_1}$ with one computed under $k_{t_2}$. That mismatch arises across the forward and backward directions, and between the statistic and its permutation null.
The prediction follows directly. Under the matched protocol, false-positive rate and orientation accuracy should stay flat in drift time until read noise and analog-to-digital conversion (ADC) quantization, rather than drift, set the floor. Under unmatched execution, they should degrade along the power law.
Relation to Prior Work
The method comes from two sources. Kernel Approximation using Analog In-Memory Computing maps random Fourier feature projections onto phase-change crossbars of the 64-core HERMES chip and argues that random weights tolerate device error. It evaluates only kernel ridge classification and Performer attention, and it reports that analog approximation error saturates as $D$ grows. We take its hardware mapping and test whether its tolerance argument survives a task where the signal is a small difference between two dependence scores, not a wide class margin. The causal side is the additive-noise orientation rule summarized in Non-linear Causal Discovery via Functional Constraints. That rule picks the direction whose residual is independent of its input. We also use the squared-loss mutual information variant of LSIR as an alternative dependence score.
The drift-mitigation literature supplies baselines and constraints. Global drift compensation rescales outputs by a scalar estimated from reference devices, as in Toward Software-Equivalent Accuracy on Transformer-Based DNNs With Analog Memory Devices and the ALBERT 14nm chip demonstration. It corrects mean decay but not the per-cell spread. Bhattacharjee et al. show that digital normalization recalibration fails once weight distortion exceeds a second-principal-component threshold of about $10^{-2}$. That result motivates our view that the spread of $\nu$ cannot simply be absorbed, but it was measured only on vision classifiers. Hardware-aware training (Rasch et al.) does not apply here, because random projections are not trained.
The measurement-error line of causal discovery (Zhang et al. 2018, upper-bound analysis) studies noise added to the data and mostly assumes linear non-Gaussian models. Our corruption lives in the computing device, not in the data, and it acts on the kernel rather than on the variables. This distinction matters. Measurement error can flip the additive-noise asymmetry, whereas we argue that a consistent change of kernel cannot. CV-LR is the strongest digital non-linear scorer, but it assumes exact arithmetic. It serves as our digital reference, not as something we map to analog.
Research Program
We build an analog dependence engine in the IBM Analog Hardware Acceleration Kit (aihwkit). The engine uses that kit's phase-change memory model: power-law drift with per-device exponent spread, read noise, 1/f noise, programming error, and ADC and digital-to-analog conversion quantization. The projection matrix $\omega$ is programmed once and read at times from one second to one year. Regression residuals come from a digital Gaussian-process or random-feature ridge fit, so drift enters only through the independence test. A later ablation moves the ridge regression onto the crossbar as well.
The program has three parts. First, a measurement study: map the effective kernel $k_t$ empirically by comparing its empirical spectral distribution with the theoretical scale mixture, and record the HSIC test's false-positive rate and power as functions of $t$, the spread of $\nu$, and $D$. Second, the matched protocol itself: interleave forward, backward, and permutation evaluations within a window short relative to $t$, since drift slows logarithmically. Third, a stress regime that deliberately violates the independence assumption between $\nu$ and $\omega$, for example by making drift depend on conductance, as amorphous-state physics suggests. This tests whether the "still a valid kernel" argument fails, and where.
Baselines
The do-nothing reference is ideal digital floating-point execution of the same random-feature HSIC orientation, with identical $\omega$, $D$, and permutation count. It is the accuracy ceiling every analog variant is measured against. It is also the comparison point for energy, which we estimate from the HERMES efficiency figures rather than measure.
The naive analog reference runs the same pipeline on drifting hardware with no compensation. Forward and backward scores are computed at their natural times, and the null is calibrated once at programming time. This reference shows how much drift actually costs.
The strongest existing analog competitor is global drift compensation, implemented exactly as in the reference-device scalar scheme of Spoon et al.. It is the established production answer to drift and restores mean bandwidth. It gets the same reference-device budget and the same read schedule as our protocol. Beating it means a statistically significant gain in orientation accuracy, or a false-positive rate held within its nominal 5% level, at drift times where it fails. If no such times exist up to one year, we report that as the outcome.
A periodic refresh baseline reprograms $\omega$ at fixed intervals. We compare its accuracy at matched write energy, since writes are costly.
The strongest digital method on graphs is CV-LR. It reports an F1 of 0.94 on SACHS (CV-LR) and is run from its published setup as an upper reference.
An ablation isolates the key ingredient. It keeps global compensation but removes, one at a time, (a) same-window direction pairing and (b) on-chip permutation calibration. This shows which part of matching matters.
Why not just…?
Why not just run causal discovery on a GPU? For accuracy alone, a GPU is the correct choice, and our digital baseline is exactly that. The case for analog rests on energy: HERMES reports 6.2x to 12.4x better energy efficiency than an INT8 A100 for random-feature workloads. The case matters for many repeated independence tests, for example inside graph search or in always-on edge monitoring. For one bivariate test it is weak, and we say so.
Why not just apply global drift compensation and treat the chip as exact? This is the real threat. If the spread of $\nu$ is small, a scalar fix may leave $k_t$ close enough to the programmed kernel that unmatched comparisons barely matter. The ablation settles this directly. If compensation alone matches our protocol across the full drift range and the realistic exponent spread, the contribution shrinks to a characterization result: drift is harmless for this task. We would report that rather than claim a method.
Why not just periodically reprogram the weights? Refresh does remove drift, but each refresh costs write energy and time. The normalization study shows that distortions beyond mild levels cannot be absorbed without retraining or rewriting. Our protocol claims that refresh is unnecessary for test validity. The equal-write-energy comparison tests that claim.
Why not just use linear non-Gaussian methods with measurement-error corrections? Those methods model noise in the data and assume linear mechanisms. Linear continuous learners perform poorly on SACHS, at F1 0.19 for NOTEARS and 0.42 for DAGMA as reported in CV-LR. They also do not address corruption of the kernel inside the computing device.
Proposed Experiments
On synthetic additive-noise pairs, with polynomial, sigmoid, and sinusoidal $f$, Gaussian and uniform noise, and sample sizes from 200 to 5,000, we measure orientation accuracy and the area under the curve of the score difference against drift time. On pairs generated to be truly independent, we measure the test's false-positive rate. The decisive check is whether the matched protocol holds the false-positive rate within nominal level and keeps accuracy within two points of digital at one year, while naive execution does not.
On the Tübingen cause-effect pairs, with the standard weighted accuracy, the reference is entropic causal inference at roughly 64% accuracy (Entropic Causal Inference). That paper describes additive-noise methods as competitive on this benchmark, but no separate reported additive-noise accuracy was found on Lacuna. Our digital run therefore provides the number we compare against.
As a stretch experiment, we orient edges on a given SACHS skeleton using the analog test, against CV-LR's digital F1 of 0.94. The stress experiment with conductance-dependent drift is the falsification test for the central claim. If accuracy collapses there even under matching, the "drift is only a kernel change" view is wrong in that regime, and we will map exactly where.
References
- Kernel Approximation using Analog In-Memory Computing
- Identifying Total Causal Effects in Linear Models under Partial Homoscedasticity
- Least-squares independence regression for non-linear causal inference under non-Gaussian noise
- Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices
- Demonstration of transformer-based ALBERT model on a 14nm analog AI inference chip
- Examining the Role and Limits of Batchnorm Optimization to Mitigate Diverse Hardware-noise in In-memory Computing
- Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators
- Causal Discovery in the Presence of Measurement Error
- An Upper Bound for Random Measurement Error in Causal Discovery
- Fast Causal Discovery by Approximate Kernel-based Generalized Score Functions with Linear Computational Complexity
- Using the IBM analog in-memory hardware acceleration kit for neural network training and inference
- Entropic Causal Inference: Identifiability and Finite Sample Results