How Optimality Structures Sparse Dictionaries: A Theory for Understanding SAE Representations

Abstract: Sparse Autoencoders (SAEs) have found success parsing neural representations into interpretable concepts, providing a basis for understanding and control. However, what exactly SAEs extract, and, correspondingly, the scientific conclusions we can draw from them, are not obvious. Empirically, the proof is in the pudding: SAEs learn interpretable features. Theoretically, we lack a clear account of what properties a 'concept' must satisfy for an SAE to extract it. There has been extensive identifiability work studying the conditions under which sparse coding recovers ground-truth features; however, these approaches tends to focus on simple data-generating models (e.g. sparse independent features) which poorly approximate the internet-swallowing language-model representations on which SAEs are trained. Here, avoiding data-generating models, we ask simply what properties any dictionary learning optimum must satisfy. Concretely, we extend local optimality analyses (Gribonval & Schnass, 2010) to the nonnegative joint-optimisation problem that vanilla SAEs approximate, and derive constraints relating optimal SAE features to their distributions. We use these constraints to explain a range of observed SAE behaviours - hierarchical splitting & absorption, the structure of residuals, and dense antipodal features - each reflecting how L1+nonnegativity interact with data to structure optimal dictionaries. Finally, we construct a novel large-dictionary convex problem and explore the wide atom-per-datapoint limit. In sum, we hope to tease model assumptions from unexpected observations, letting us learn more from SAEs' successes and provide principles for designing their successors.

• The paper establishes mathematical optimality conditions that govern the emergent structure of sparse autoencoder dictionaries, explaining phenomena like feature splitting, absorption, and dense antipodal pairs.
• It reformulates the sparse autoencoder objective into equivalent convex and regularized problems, linking data geometry with representation sparsity.
• The study shows how nonnegativity and L1 sparsification induce dense antipodal pairs and influence decoding dynamics, offering actionable insights for network design.

Optimality-Driven Structure in Sparse Autoencoder Dictionaries

Overview

This work formulates a principled mathematical theory describing how the optimality conditions inherent in sparse autoencoders (SAEs) shape the structure of learnt dictionaries. The central motivation is the empirical success of SAEs in parsing neural network representations, particularly for interpretability and control tasks, coupled with empirical phenomena—such as feature splitting and absorption—whose explanations have been ad hoc in prior literature. Rather than presupposing a data-generating process, the analysis proceeds directly from the optimality properties of the underlying dictionary learning problem that SAEs approximate. This synthesis enables a precise account of which properties of "concepts" make them recoverable by SAEs, the generic emergence of feature splitting and dense antipodal pairs, and the necessary instability of hierarchically overlapping features under sparse, nonnegative coding constraints.

Problem Formulation and Reformulations

The paper formalizes the SAE objective as a matrix factorization problem:

$$\min_{W, \{z^{[i]}\}, b} \mathbb{E}_i \left[\lVert W z^{[i]} + b - x^{[i]} \rVert_2^2 + 2 \lambda \lVert z^{[i]} \rVert_1\right]$$

subject to $z^{[i]} \geq 0$ and $\lVert W_d \rVert_2 = 1$ for each dictionary column $d$.

Crucially, the analysis is not reliant on assumptions about data independence or the existence of a "true" underlying factorization. Instead, it pivots on what must be true for any solution to be a local or global optimum.

The authors derive several mathematically equivalent reformulations, including:

• Regularization-based versions with unconstrained dictionary scaling,
• A convex relaxation (in the regime where the number of features exceeds data points) expressed as an optimization over the completely positive cone of mean-zero representational similarity matrices,
• Formulations that fully optimize out the decoder, yielding an objective purely in latent codes or their correlations.

These formulations establish the direct link between the constraints imposed by the optimization, the resulting geometry of the data representations, and the emergent feature structure.

Optimality Conditions and Feature Interactions

The centerpiece of the analysis is a set of necessary optimality (stationarity) conditions. These extend the classical results for local optimality in nonnegative sparse coding (e.g. Gribonval & Schnass 2010) to the nonnegative, joint-optimized autoencoder objective.

Feature-Feature Constraints

The conditions connect the cosine similarities between dictionary atoms to the convex hull of input vectors on subsets of the data where particular features are inactive. Explicitly, for each pair of dictionary atoms, the vector of cosines must lie within the convex hull formed by suitably normalized data points with the corresponding feature off. This generalizes the notion of "separability" prevalent in NMF theory and allows non-orthogonal, highly correlated features to coexist stably in the dictionary—provided their zero sets sufficiently cover the input's range.

Several implications are demonstrated:

• Mutually exclusive ("categorical") features are always permitted and naturally tend to be selected.
• In the presence of hierarchical structure (features such as "dog" and "labrador", where one is always active with the other), the required convex hull degenerates, making such solutions always unstable except for the trivial case of identical features—hence "hierarchical splitting" or "feature absorption" is generically forced.
• Feature splitting (broad concepts splitting into finer ones as dictionary width increases) and absorption (removal of hierarchical overlap by absorbing broad concepts into specific ones) are thus not arbitrary or artifacts, but direct results of the interplay between optimality and nonnegativity/sparsity constraints, independent of encoder architecture details.

Feature-Residual Relationships

Additional first-order condition restrict the correlation between learned features and data residuals (what is left unexplained by the dictionary). If a missing feature would correlate too strongly with a dictionary entry in the set where that entry is inactive, including it as a separately coded feature would yield a lower cost solution. The constraint scales with the regularization parameter $\lambda$, implying that higher sparsity pressures allow larger residuals.

This insight means mutually exclusive features can always be left in the residuals. Yet, for hierarchically overlapping or range-limited features, the optimal dictionary must restructure itself such that these directions are directly encoded.

Dense Antipodal Pairs and the Role of Nonnegativity

A particular empirical observation is the emergence of dense antipodal feature pairs when training SAEs (or similar methods with strictly nonnegative activations such as TopK or JumpReLU) on data with dense latent structure. The mathematical analysis shows that for a single dense direction in the data, nonnegative codes necessarily yield at most two active neurons per data point, with their decoders pointing in exactly opposite directions. The threshold for splitting into two antipodal atoms is determined by the sparsity/mass of the activation distribution—a feature need only be split if it is active more than half the time.

This accounts for why alternative encoders (e.g., AbsoluteTopK allowing both positive and negative activations) eliminate the antipodal phenomenon.

The Wide-Dictionary Limit and Ray Clustering

In the regime where the number of dictionary atoms exceeds the number of data points (i.e., the "wide" or "overcomplete" limit), the convex relaxation becomes tight, and the structure of global optima is tightly constrained. The authors show that a global minimum always exists where each data point is represented by at most one active feature—i.e., the code is maximally sparse, and each ray from the optimal bias to a data point receives its own dictionary atom unless it lies within a certain $\lambda$-ball of the bias. This links directly to the notion of geometric medians and $1$-sparse representations in classical dictionary learning.

Thus, SAEs generically do not "recover" true overlapping or hierarchical latent structure unless this is also optimal under the imposed coding constraints, providing theoretical support for why feature splitting and absorption are unavoidable as dictionary width increases and reconstruction error is driven to zero.

Theoretical and Practical Implications

The central implication is that the structure of learnt features in SAEs is not a transparent window onto the latent structure present in the data or underlying network, but reflects a nontrivial interaction between data geometry, the nonnegativity constraint, and L1 sparsification. Interpretability claims or attempts at network control need to account for these effects. Unusual empirical behaviors such as, e.g., missing hierarchical features or the clustering of features into fine-grained, mutually exclusive subtypes, are predicted consequences of the optimization structure, not aberrations.

In terms of methodology, the convex reformulations and first-principles stationary conditions developed here offer a new analytical toolbox for guiding future SAE design. Specifically, modifying the encoder/decoder constraints, loss, or regularization (e.g., relaxing nonnegativity, introducing structured sparsity, or explicit hierarchical priors) could be analyzed at the same level of theoretical detail, enabling the principled development of successor architectures that better align with the act of concept extraction and mechanistic interpretability.

Conclusion

This work offers a rigorous, optimization-theoretic foundation for understanding feature structure in sparse autoencoders used for neural interpretability. By eschewing data model assumptions and instead deriving the necessary and sufficient optimality conditions for the full dictionary learning problem (with nonnegativity and sparsity), the analysis provides closed-form explanations for empirical phenomena and clarifies the limits and capabilities of current SAE approaches. This establishes a firm basis for both practical application and theoretical investigation into more expressive and robust unsupervised interpretable representation learning.

For further mathematical details, appendices and code, the author provides a resource at (2606.02385).