When Does the t-SNE Objective Have a Minimizer? Two Phases under Perplexity Calibration
Li and Mossel showed that global minimizers of the t-SNE objective can lose basic features of the data, mainly for a fixed Gaussian bandwidth. With bandwidths calibrated to a perplexity, as in practice, a prior question arises: does a global minimizer exist? We show that the answer depends on the sample, even as d→∞. For n points uniform on the sphere S^{d-1} with d→∞, both outcomes, a minimizer and no minimizer, have probability bounded away from zero: for every n≥6 at perplexities near an explicit polygon value p_n, and, by a computer-assisted proof in interval arithmetic, for n=1000 at every p∈[13.75,40] and near every integer from 4 to 100, and at p∈{10,30,50} for several hundred pairs (n,p) with 160≤n≤2000. The proof follows a critical point of a scale-free boundary functional away from a joint realized by a planar Cauchy kernel, and a realization lemma turns such joints into calibrated ones. Along the way we show that calibrated bandwidths grow like √d with a random constant and that the affinities converge to a non-uniform Gaussian-softmax law, give an exact existence criterion, classify existence completely for three points, and prove that the strict non-existence phase is open (non-existence alone is not). When no minimizer exists, every minimizing sequence spreads out without bound. The accompanying zip contains the interval-arithmetic certificate code and certified outputs.
Authors
- Xiang Li (ORCID: https://orcid.org/0000-0002-9059-4759)
Institutions
- Hanyang University (KR)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24349201
- Primary Topic
- Stochastic Gradient Optimization Techniques
- Type
- preprint