Statement

Let $f \in \{0,1\}^d$ be a binary feature vector with covariance $\Sigma = \mathbb{E}[ff^\top]$, encoded by a tied-weight autoencoder $W \in \mathbb{R}^{m\times d}$ ($m < d$) whose $i$-th reconstructed feature is $\hat f_i = \sigma\big(\|w_i\|^2 f_i + \sum_{j\neq i}\langle w_i,w_j\rangle f_j + b_i\big)$ — the second (interference) term $\mathcal I_i$. For a linear autoencoder ($\sigma = \mathrm{id}$) the loss-optimal $P=W^\top W$ is the orthogonal projector onto $\Sigma$‘s top-$m$ principal components (Baldi & Hornik, 1989); substituting this optimum back in gives $\mathcal I_i = (1-P_{ii})f_i - \varepsilon_i$, where $\varepsilon_i \to 0$ exactly when $\mathrm{rank}(\Sigma)\le m$ and shrinks with the residual spectral mass $\sum_{k>m}\lambda_k(\Sigma)$ (Eckart–Young) as $\Sigma$ becomes more nearly low-rank. In this regime interference is proportional to the signal itself, not opposed to it — “constructive” rather than adversarial — and the resulting weight geometry is exactly whatever shape $\Sigma$‘s leading eigenvectors have, rather than a shape chosen to minimize $|\langle w_i,w_j\rangle|$ pairwise. This constructive regime is favored over the classical interference-filtering regime (near-orthogonal $w_i$, $\|W\|_F^2\approx d$) whenever the bottleneck is tight ($m\ll d$) and/or weight decay penalizes $\|W\|_F^2$, since the constructive solution has $\|W\|_F^2=\mathrm{tr}(P)=m<d$ — strictly cheaper under an $L_2$ weight penalty.

Intuition

The classical “toy models of superposition” picture treats packing more features than dimensions as a purely adversarial problem: cram in extra near-orthogonal directions, accept a little crosstalk, and let a ReLU plus a negative bias filter the crosstalk back out — like assigning everyone a slightly-too-similar locker combination and hoping nobody’s actual combination gets triggered by someone else’s. This hypothesis says that when the features aren’t independent — when knowing one tells you something about another — a network doesn’t have to fight this correlation, it can spend it: reuse the same shared directions for features that tend to co-occur, the same way a lossy compressor doesn’t store every correlated pixel independently but exploits the redundancy directly. The “interference” from a correlated neighbor firing isn’t noise to be filtered anymore; it’s exactly the redundant information the compressed code is designed to carry.

Properties

• Not a claim about sparsity alone — about the covariance’s rank/spectrum. The classical near-orthogonal/regular-polytope regime (Elhage et al., 2022) was derived for sparse, weakly correlated features; this hypothesis’s distinguishing content is specifically about what changes when features are correlated — the relevant quantity is how concentrated $\Sigma$‘s eigenvalue spectrum is relative to the bottleneck width $m$, not how sparse individual features are.
• A concrete, checkable mechanism switch, not a vague “sometimes.” The weight-norm identity ($\|W\|_F^2\approx d$ for interference-filtering vs. $\|W\|_F^2 = m$ for the constructive/low-rank-projection regime) gives a single scalar diagnostic for which regime a trained network has settled into, and predicts the switch is driven jointly by compression ratio $m/d$ and the presence/strength of weight decay — both independently manipulable and testable.
• Explains cyclic and clustered geometry through a second, distinct mechanism from corpus-statistics-symmetry arguments. If $\Sigma$ itself happens to be (near-)circulant — as it is for calendar-month co-occurrence, where two months’ joint frequency depends mainly on how far apart they are in the year — this hypothesis predicts a trained, bottlenecked, weight-decayed autoencoder’s weight columns will recover a Circle shape as a consequence of compression, not because anything about the training objective explicitly asked for cyclic structure. See “Relative to translation-symmetry” below for how this relates to, rather than competes with, that shape’s other proposed origin.
• Distinct from, and must not be conflated with, value-coding. A real-valued quantity a network computes as part of its task (e.g. a running modular-addition sum, or a city’s coordinate) can also be linearly encoded as a circle or manifold with no correlated-feature packing, no bottleneck-driven compression, and no superposition involved at all — the originating paper is explicit that finding a circle empirically does not by itself distinguish “this is constructive-interference-driven superposition” from “this is a single value-coded scalar with an intrinsically periodic range.” Confirming this hypothesis specifically requires identifying a genuine packed, correlated feature set (not a single computed scalar) whose covariance predicts the observed shape.
• In tension with, and a refinement of, the naive polytope picture — not a resolution of the deeper LRH/superposition tension. Hypotheses already flags that a naive reading of superposition (packing $\gg d$ mutually orthogonal directions into $d$ dimensions) is in tension with Linear Representation Hypothesis; this hypothesis doesn’t resolve that tension in general, but it does show one concrete way many more-than-$d$ features can be packed with non-orthogonal, correlation-aligned weight columns while remaining linearly recoverable (via a linear or ReLU decoder) — a narrower, mechanistically specific instance of “superposition needn’t destroy linear decodability,” rather than a general proof that it never does.

Relative to translation-symmetry

Translation Symmetry Hypothesis explains why a co-occurrence/PMI matrix built directly from corpus statistics would be circulant or Toeplitz in the first place (translation-invariant relatedness $\Rightarrow$ Fourier-mode eigenvectors), operating at the level of the input statistics themselves. This hypothesis instead explains when a trained, bottlenecked autoencoder’s own weights will actually surface those top eigenmodes as its packing basis, as opposed to defaulting to the interference-avoiding near-orthogonal regime that ignores $\Sigma$‘s structure entirely — operating at the level of what an optimizer under a compression + weight-decay constraint does with that statistic. The two are complementary rather than competing: translation-symmetry supplies a candidate reason $\Sigma$ (or an input PMI matrix) might already have circular/Fourier structure; constructive-interference supplies the separate mechanism (tight bottleneck, $L_2$-penalized weights) by which a network’s learned representation would actually inherit that structure rather than override it.

Exercises

Base

1. In the linear-autoencoder identity $\mathcal I_i = (1-P_{ii})f_i - \varepsilon_i$, what does $\varepsilon_i \to 0$ require of $\mathrm{rank}(\Sigma)$ relative to the bottleneck width $m$, and why does this make the interference term exactly proportional to $f_i$ in that limit?

2. Why does the weight-norm identity $\|W\|_F^2 = \mathrm{tr}(P) = m$ for the constructive/low-rank-projection solution, versus $\|W\|_F^2 \approx d$ for interference-filtering, make weight decay a mechanism that biases training toward the constructive regime?

Middle

3. A researcher trains two autoencoders on the same correlated feature data — one with $m/d = 0.8$ (mild compression) and one with $m/d=0.02$ (severe compression) — with identical weight decay. Which is more likely to land in the constructive-interference regime, and why does the Eckart–Young bound make this a matter of degree rather than a sharp switch?

4. Suppose an empirical study finds a circle in a trained network’s residual stream for a cyclic real-world variable (e.g. months) but does not check the weight-norm diagnostic or verify a genuinely packed, correlated feature set is involved. Why is this alone insufficient to credit constructive-interference (rather than value-coding, or plain Translation Symmetry Hypothesis acting on some other representational pathway) as the mechanism?

Pro

5. The originating paper reports that a ReLU autoencoder and a linear autoencoder produce similar Gram-matrix (pairwise weight inner-product) patterns at small bottleneck width $m$, but diverge — with the ReLU version showing sparser interference patterns — at larger $m$. Using the interference-filtering-vs-constructive framing, explain why this divergence is expected specifically as $m$ grows, rather than at every compression level.