Independent scholarship · Open accessThe monthly digest

Editorially reviewed

A Higher Effective Topos for Certified Stochastic Optimization

We construct a higher exact completion of Hyland’s effective topos and recover it on zero-types. Action groupoids supply transport for stochastic optimization and for the latent coordinates of a 4→2→4 autoencoder. Lean 4 verifies arithmetic and training compilation, polynomial jets, and certificates for explicit 1- and 2-cell presentations.

Anonymous← All research

Higher exact completion

Fix a set universe for Hyland’s effective topos $\mathrm{Eff}$ and a larger universe in which it serves as a small site. On $\mathrm{Eff}$ take the topology generated by finite jointly epimorphic families and form the corresponding ∞-topos of space-valued sheaves. The bounded coherent objects form a higher exact completion. Its full subcategory of zero-types is equivalent to $\mathrm{Eff}$ via Yoneda. The proof uses representable covers of a bounded coherent zero-type and its kernel pair to construct an equivalence relation in $\mathrm{Eff}$, then identifies the sheaf quotient with the representable for the effective quotient. The argument uses the higher exact-completion framework of Stefanich. Hyland (1982)70129-6) · Stefanich

Stochastic optimization and paths

Finite rational probability weights specify a law on a sample space. An explicit finite sequence of sample indices determines a run. This allows exact evaluation of a trajectory and, for a finite horizon, of its law by summing the weights of sample sequences. The algorithms include first-order SGD, momentum with a velocity state, and coordinate Newton updates for diagonal quadratic sample losses. Polynomial jets give the Taylor coefficients needed for higher-order updates, including a scalar Halley step with an explicit singular case.

In the coordinate-exchange example the group of order two acts on parameters and sample labels. An equivariant update commutes with transport along the resulting action groupoid paths. The action groupoid has a nonidentity stabilizer loop at a diagonal parameter with equal coordinates. That loop acts on the fibers of sample or velocity data even though the underlying parameter is fixed. The manuscript proves equivariance for transitions, finite trajectories, and finite laws. It also exhibits a degree-two gluing obstruction for a Newton update with an extra certificate family over $K(C_2,2)$.

Autoencoder benchmark

The undercomplete autoencoder compresses four rational features into two leaky-ReLU activations and reconstructs the input with an affine decoder. An exact backpropagation rule updates both layers. The manuscript evaluates replayable stochastic training on signed bar images and reconstruction on a stratified held-out Iris split with training-only normalization. Swapping the two latent units swaps the learned code and leaves the output and loss invariant. Lean checks that this transport commutes with single steps, entire schedules and finite independent-draw laws, as well as correctness of the training bytecode relative to the numerical specification. Autoencoders · UCI Iris

Lean artifact

The accompanying Lean 4 project contains an intrinsically typed rational expression language, a compiler to typed stack instructions, and a kernel-checked theorem equating compiled evaluation with the source semantics. It includes a compiler for arbitrary finite-order polynomial jets, a guarded Halley update, an executable optimizer, and a proof language for the coordinate-exchange action groupoid. A separate strict two-cell presentation checks interchange and distinguishes the degree-two generator from the identity. Executable examples and proof statements are included with the source.

The Lean proofs cover the defined rational and path calculi. The identification of their denotations in the higher exact completion and the zero-type recovery theorem are mathematical arguments in the paper. The paper compares the construction with the elementary higher realizability program of Awodey and Emmenegger and with the Lean proof frontend SynthLean. Awodey and Emmenegger (2025) · Nawrocki et al. (2026)

Reproducibility

The attached PDF gives the definitions, proofs, and worked computations in full. The source archive contains the LaTeX manuscript, the pinned Lean project, and the build and verification instructions. The run results use exact rational arithmetic. Generative AI assisted in drafting the manuscript and implementing the artifact.

Reproducibility artifacts

higher-effective-topos.pdf · 557,078 bytes · Download
SHA-256 f4e159783b3e93f099117e5ed9815fb3032064c12ded08de41cb6c8516f51835

higher-effective-topos-source.zip · 59,852 bytes · Download
SHA-256 0cf299ed8993c49f599afa92c809ce77add55b45dfbe47ab6174c6e6c46798c7

Discussion

Open a careful, constructive conversation about this note.

Comments appear after editorial approval.