Butterfly Certificates for Matrix Convexity on Free Sets

Butterfly Certificates for Matrix Convexity on Free Sets

A weird-but-beautiful idea sits at the heart of new research from the original paper: you can get global convexity extensions from only local data—but only after you “lift” everything to the right operator level. The paper blends two themes: (1) Cartan convexity for noncommutative (free) functions and (2) a “butterfly” style realization formula that turns local positive constraints into an extension on a whole neighborhood of a noncommutative hull.

And yes, there’s also a meta-layer: the author uses the concept of a mathematical koan—a compact, checkable specification of definitions, proof mechanisms, and examples—explicitly in response to how AI can help reconstruct papers from smaller cores.

Why This Matters: Convex Extensions Without Patching, Thanks to “Universal Points”

This is significant right now because AI systems increasingly propose solutions by assembling local fragments—then hope everything stitches together. But in noncommutative geometry, “stitching” is not free: compatibility across matrix levels is the hard part. This paper gives a structural reason why compatible global behavior can be forced: real free sets admit a universal point constructed in a specific Hilbert-space dimension, and that universality removes the need for traditional compactness/patching arguments.

A concrete “today” scenario (even if you never touch operator algebras directly): suppose you’re doing model reduction or robust control design where the constraints naturally live at many system sizes. You might have a library of local inequalities—valid near certain operating points or on certain compressions—and you want to know whether there exists a single certificate (a single extension) that keeps the inequalities true on the entire admissible hull. The paper’s mechanism says: if your data come from the right kind of free set and your function satisfies a Cartan convexity condition (local matrix-convex liftability), then you can extend it to a neighborhood—and you can do it via a stable “butterfly” representation where a denominator stays uniformly positive across the hull.

How does this compare to earlier AI research? A lot of AI work in math focuses on producing long explanations or expanding a sketch into a full proof. This paper takes a different stance: it proposes that the right unit of publication for AI-friendly mathematics is not prose, but a koan—a compact, operational core that uniquely pins down the proof structure. That’s a direct answer to the “AI gives you plausible continuity but not guaranteed correctness” problem the author emphasizes in the methodology section.

Here’s the core comparison idea, expressed plainly: the paper’s proof strategy avoids “gluing” local lifts across infinitely many levels, by using a single universal direct sum object so the compatibility is automatic.

Universal Points: The Trick That Makes Compatibility Automatic

The paper’s central observation is easy to miss if you’re skimming: it’s not enough that a function behaves nicely locally. The local hypotheses must be represented in a way that ensures they remain compatible across all amplifications and compressions.

Real free sets let you “sum the world” at one Hilbert space level

A real free set is, roughly, a collection of self-adjoint operator tuples that’s stable under three operations:

1. Unitary conjugation
2. Reducing subspaces
3. Direct sums (with a controlled bound on how many summands you allow)

Given a bounded real free set (K), the paper constructs a universal point ( \mathbf{X} \in K ). This is the key object: every point (X \in K) becomes (up to unitary equivalence) a reducing summand of an amplification of (\mathbf{X}).

The punchline is operational:

• If you can build a local matrix-convex lift near (\mathbf{X}),
• then positivity properties in that lift can be propagated to every compression of every amplification,
• which means you automatically get a consistent extension near the hull.

Why the Hilbert-space dimension matters (and isn’t negotiable)

The universal point doesn’t just exist abstractly; the paper gives a cardinal bound for the Hilbert-space dimension needed for the universal sum. In the finite-coordinate part, it defines
[
\lambda = \max{|\mathcal{I}|,\aleph0}.
]
Then it shows that on standard Hilbert spaces (\ell2(\kappa)) with (\kappa \le \lambda), there are at most (2^\lambda) kinds of tuples to worry about at each dimension level, and only (\lambda) many levels—leading to a bound of the form (2^{\lambda}) times another (\lambda)-dependent factor. In the final statement, it’s enough to take the universal point on a Hilbert space of dimension at most
[
2^\lambda \cdot \lambda
]
(up to the paper’s more careful expression (2\lambda^2^\lambda) in the koan/core statement).

Why do we care? Because universality depends on capturing all small reducing components that might appear in any point of the free set. If you only did this at finitely many levels, the paper explains a scalar counterexample shows you can lose extension completely.

Butterfly Realizations: Turning Positivity at a Point into Positivity on a Hull

Now for the “butterfly” part, which is both a proof mechanism and a style of certificate.

Matrix convexity is the natural language for this propagation

Matrix convexity is perfect here because its closure properties match the operations you’re allowed to perform on free sets:

• direct sums
• isometric compressions
• compatibility across matrix levels

Also, locally bounded matrix-convex free functions behave well analytically (real-analytic behavior shows up in the realization theory story).

The noncommutative Kraus-butterfly theorem supplies the denominator positivity

The extension uses a noncommutative Kraus/butterfly realization theorem (credited to Pascoe and Tully-Doyle in the paper). The format is schematic but powerful: a function can be realized as something like
[
F(X)=\text{(affine numerator)} \cdot P(X)^{-1} \cdot \text{(affine denominator)},
]
with the crucial property that
[
P(X) \succeq \varepsilon I
]
uniformly on the set of interest.

The paper’s main engineering step is this:

• Once positivity (P(\mathbf{X}) \succeq \varepsilon I) holds at the universal point,
• affineness and matrix-convexity force the inequality to persist for every point in the noncommutative convex hull,
• so the “resolvent denominator” never collapses.

That’s exactly why the extension domain becomes a genuine open neighborhood of the hull:
[
{P \succ 0}
]
(rather than just an existence statement that might fail for denominator reasons).

What “Cartan convexity” means here: local agreement with a matrix-convex lift

The paper defines a function (f) on a bounded real free set (K) to be Cartan convex if, near every point (X \in K), it matches a locally bounded matrix-convex free function on a neighborhood. “Agree near the point” is important: it’s a germ condition, not just equality of values at one spot.

Then the main Cartan extension theorem says (informally):
- every Cartan convex free function extends to a locally bounded matrix-convex function
- on an open neighborhood of the noncommutative convex hull of (K)
- and that extension admits a butterfly realization.

And because the universal point contains every relevant reducing component directly, the realized extension recovers the original function exactly on all of (K).

Graphs and (\Gamma)-Convexity: What Compressions Are Allowed to Preserve

So far, we’ve talked about convexity “in the ambient coordinates.” The paper then introduces a graph map (\Gamma) to control what information is preserved by admissible compressions.

(\Gamma)-pairs encode which moments survive a compression

A graph map (\Gamma = (\gamma1,\dots,\gammar)) is a tuple of self-adjoint free functions (often polynomials, but not only). A (\Gamma)-pair ((X,V)) means, roughly:

• (V:\mathcal{H}\to\mathcal{K}) is an isometry,
• and the compression preserves the (\Gamma)-coordinates:
  [
  \Gamma(V^XV)=V^\Gamma(X)V.
  ]

This lets you define a (\Gamma)-convex hull: the smallest (\Gamma)-convex set containing your original (K).

Intrinsic (\Gamma)-convexity vs lift-Cartan (\Gamma)-convexity: not the same

The paper is careful to distinguish two notions:

• Intrinsic (\Gamma)-convexity: Jensen-type inequalities are tested only on (\Gamma)-pairs.
• Lift-Cartan (\Gamma)-convexity: the function admits a local matrix-convex lift in the lifted coordinates given by (\Gamma).

A striking warning appears: it’s possible to satisfy the intrinsic Jensen inequalities but still fail to have any matrix-convex lift in the lifted independent coordinates.

The paper illustrates this with a partial-convexity graph (\Gamma_{\text{par}}(x,y)=(x,y,y^2)) and a function whose obstruction shows up through a mixed Hessian behavior. The upshot is: intrinsic tests alone do not guarantee liftability away from the graph.

Four “Standard” Geometries: Quadratic, Holomorphic, Partial, and Jordan-Product Graphs

This section is where the paper gets especially geometric and intuitive—almost like watching the same mechanism in different costumes.

Quadratic graph (\Gamma{\square}(X)=(X,\sumj X_j^2)): real analyticity becomes the criterion

When (\Gamma) is the quadratic coordinate, lift-Cartan (\Gamma)-convex functions turn out to be exactly the real analytic self-adjoint free functions in the right category. And the paper claims you only need the sum of squares—adjoining all individual squares isn’t necessary for that analytic category.

Partial convexity: (\Gamma) that encodes (x,y,y^2) doesn’t force global liftability

Partial convexity is where the story becomes subtle: you may have a noncommutative Jensen inequality in the intrinsic sense, but the butterfly realization can involve parameter-dependent affine tails, and a mixed-Hessian freedom can prevent a clean convex lift.

So: partial convexity is weaker than lift-Cartan convexity in general.

Full holomorphic graph: the graph retains “triangular information” needed for plushness

At the other extreme, if (\Gamma) includes all analytic and coanalytic monomials (the “full holomorphic graph”), the intrinsic (\Gamma)-convexity condition becomes strong enough to imply free plurisubharmonicity (free plushness). The paper explains this via triangular perturbations: upper vs lower triangular choices separate the complex Hessian into positive orderings.

So in the holomorphic setting, intrinsic Jensen convexity lines up with plushness and convex liftability.

Jordan-product graph (\Gamma_J(x,y)=(x,y,xy+yx)): a probability/covariance geometry pops out

One of the most approachable examples in the paper is the Jordan-product graph. Here the (\Gamma_J)-pair condition forces a zero Jordan covariance condition, and the scalar geometry becomes a hyperbolic paraboloid and rectangular-hyperbola boundaries.

Even more concrete: for scalar points ((x,y)), membership in the (\Gamma_J)-convex hull corresponds to probability mixtures where the random variables (x) and (y) have covariance zero. The paper even identifies a sharp scalar Carathéodory number of four: any point in the hull can be represented using at most four generating atoms, and four is best possible.

That’s a real “geometry-to-probability” bridge: the graph choice determines what statistical constraint you’re encoding.

Key Takeaways

Key Takeaways

• Universal points are the compatibility engine. For bounded real free sets, the paper constructs a universal Hilbert-space point (\mathbf{X}) so every (X\in K) appears as a reducing summand of an amplification of (\mathbf{X}).
• Cartan convexity is the right local hypothesis. Cartan convex functions are locally equal to locally bounded matrix-convex lifts; this local agreement (germ-level) is what makes extension work.
• Butterfly realizations turn local positivity into global extension. Using a Kraus-butterfly mechanism, the extension keeps a denominator (P(X)) uniformly positive across the nc convex hull neighborhood.
• Intrinsic Jensen tests may not imply liftability. For nonlinear graphs (\Gamma), passing intrinsic (\Gamma)-Jensen inequalities can fail to produce any matrix-convex germ in lifted coordinates—mixed Hessian obstructions can remain hidden.
• The choice of (\Gamma) determines the geometry. Quadratic graphs align with real analyticity; full holomorphic graphs align with plushness; Jordan-product graphs translate into covariance-zero probability geometry with a sharp Carathéodory number of four.
• For applied readers: if your constraints live naturally across matrix levels (systems of different sizes), this framework suggests how to demand “right kind of liftable convexity” so local certificates can extend without fragile patching.

If you want, I can also turn one specific example (like the Jordan-product / covariance geometry or the holomorphic plush case) into a more narrative, almost “proof-sketch” style walkthrough with less operator-algebra vocabulary.