Definition

A paraboloid of revolution is generated by rotating a profile curve $r = f(h)$ around an axis, giving a surface parametrised by an angular coordinate $\theta \in S^1$ and an axial coordinate $h$:

$\big(f(h)\cos\theta,\ f(h)\sin\theta,\ h\big), \qquad \theta \in [0, 2\pi)$

The defining feature relative to a cylinder ($f(h) \equiv \text{const}$, i.e. $S^1 \times \mathbb{R}$, where the circle’s radius never changes) is that $f$ is genuinely a function of $h$ — the circle’s radius grows or shrinks as $h$ varies, and can shrink to a single point ($f(h)=0$) at one or both ends of $h$‘s range. The classical paraboloid $z = x^2+y^2$ is the special case $f(h) = \sqrt{h}$, one-sided and unbounded; the map’s use of the term (per Bhalla et al., 2026, below) is the qualitative shape — radius as a non-constant function of the axial coordinate — rather than this specific quadratic profile, which the source paper does not state in closed form.

Why this structure / Why it matters

A cylinder ($S^1 \times \mathbb{R}$) is the “independent” combination of a circular and a linear factor — Circle and 1D continuum manifold/Affine Subspace simply stacked with no interaction between them. A paraboloid is the same two coordinates coupled: how much the circular factor can vary depends on where you are along the axial one. This matters representationally because it means the two “factors” of a concept (e.g. hue and lightness) are not cleanly separable/orthogonal in the way a Cartesian product would be — probing or steering along the circular coordinate has different effective range depending on the axial coordinate’s value.

Relative to manifolds-hyperboloid

Both are quadric-family local shapes under Manifolds, but arise from different detection methods and different domains: this node’s paraboloid shape is read directly off PCA loop-radius behavior (a parametrization-mode claim, circular hue × continuous lightness in color representations), whereas Local hyperboloid-type quadric patch‘s shape is established via implicit polynomial surface fitting and eigenvalue- based quadric-type classification on controlled sentence-paraphrase neighbourhoods. Neither generalizes the other; they are sibling instances of the same quadric-surface family found by different means in different representational domains.

Key evidence

Bhalla et al. (2026) report that color representations in Llama-3.1-8B (residual stream, layer 19) organize along “a paraboloid with circular hue and lightness dimensions” (their Sec. 3, Fig. 2 left) rather than a flat hue-circle × lightness-line cylinder — i.e. the spread of hue-distinct representations is not constant across the lightness coordinate. They validate this is not a merely visual/PCA artifact: activation-patching steering interventions along the manifold’s principal components (moving between hex-code prompts’ PCA centroids) produce smooth, monotonic changes in the model’s predicted color-name tokens (their App. B.1, Table 2), indicating the structure is causally load-bearing, not incidental. See color-paraboloid for the full Observation.

How to detect it

Fit PCA (or another dimensionality-reduction method) to activations varying a single concept with two known sub-factors (one cyclic, e.g. hue; one continuous, e.g. lightness). Check the resulting 2-3-component projection for (a) a closed loop at each fixed value of the continuous factor (confirming the circular part), and (b) whether that loop’s radius changes across the continuous factor’s range — constant radius indicates a cylinder, non-constant (especially radius $\to 0$ at an extreme) indicates a paraboloid-like coupling instead.

Key papers

• Bhalla, Fel, Rager, Feucht, Haklay, Wurgaft, Boppana, Kowal, Shyam, Lewis, McGrath, Merullo, Geiger & Lubana (2026). Do Sparse Autoencoders Capture Concept Manifolds? arXiv:2604.28119 — origin of this characterization in this map’s corpus, for color representations.

Exercises

Base

1. Is a cylinder $S^1 \times \mathbb{R}$ a special case of the paraboloid parametrisation above? If so, what is $f(h)$?

2. If $f(h_0) = 0$ for some $h_0$, what happens to the surface at that value of $h$, and is the result still a manifold everywhere?

Middle

3. A researcher measures the PCA radius of the hue loop at three lightness values and finds radii $(0.9, 0.9, 0.9)$ (within noise). Does this support a paraboloid or a cylinder account of the underlying geometry? What additional lightness values would make the test more conclusive?