Solving Semidefinite Programs with Thermal Bosons
27 May 2026, Yanjiang
Thermal bosons minimize their free energy, providing a physical blueprint for solving semidefinite programs through a Bose-Einstein regularisation.
Imagine a crowd of quantum particles — photons, say, or the vibrational modes of a crystal — each with the freedom to hum at many different frequencies. Left alone, they will settle into the arrangement that minimizes their total energy while obeying whatever conservation laws they carry. That is thermodynamics at its most patient. And, if a new theoretical framework from Cornell and Shanghai Jiao Tong University is right, it is also a blueprint for solving some of the hardest optimization problems in mathematics.
In a preprint (arXiv:2605.27228), a team led by Mark M. Wilde at Cornell University, together with Michele Minervini at Cornell and Nana Liu at Shanghai Jiao Tong University, reveals that semidefinite programs — a broad and powerful class of optimization tasks — can be mapped directly onto systems of independent bosonic modes. The eigenvalues of the optimization variable become expected occupation numbers; the objective function becomes the expected energy; and the linear equality constraints become conserved charges. What emerges is not just a clever reformulation, but a thermodynamic identity: every semidefinite program over an unbounded positive semidefinite cone is mathematically equivalent to a bosonic gas that wants to minimize its free energy.
The Landscape Where Optimization Meets Physics
Semidefinite programs (SDPs) ask a deceptively simple question. Given a linear function to minimize and a collection of linear constraints, find the positive semidefinite matrix — one whose eigenvalues are all non‑negative — that does the job. SDPs crop up across quantum information theory, control engineering, machine learning, and combinatorial optimization. They are, in many ways, the workhorse of modern algorithmic science. But solving very large SDPs with standard interior‑point methods can feel like climbing a mountain whose height grows with the dimension of the problem. The worst‑case duality gap — the discrepancy between the best primal and dual solutions — scales linearly with the dimension, which can stymie progress on genuinely hard instances.
Wilde, Minervini, and Liu approach the same mountain from a different side. Their key insight is that the unknown matrix inside an SDP looks, under a thermodynamic lens, exactly like a collection of quantum harmonic oscillators. Each eigenvalue behaves like the expected number of bosons occupying a particular energy level. Suddenly, the cold algebra of optimization takes on body heat. The problem becomes a physical one: among all possible occupation patterns, which one minimises the total energy while respecting a set of conserved non‑commuting charges?
A Hint of Bose‑Einstein Fog
To make this connection rigorous, the team introduces a temperature. Not a real temperature measured in kelvin, but a mathematical one that blurs the precise occupation numbers with Bose‑Einstein statistics. They add a Bose‑Einstein entropy term to the SDP objective — a kind of quantum fog that smears the occupation numbers into a thermal distribution. At finite temperature, the problem is no longer the original SDP; it is a free‑energy minimisation over bosonic modes, regularised in a way that makes the solution uniquely expressible as a Bose‑Einstein thermal operator. Only in the zero‑temperature limit does the fog lift, and the original SDP snaps back into focus as the ground state of the thermodynamic system.
This regularisation is more than a mathematical trick. It endows the optimisation with a clean physical interpretation. The optimal primal variable takes the form
[
X^* = \frac{1}{\exp(\beta H_{\text{dual}}) – I},
]
Where (H_{\text{dual}}) is built from the dual variables and (\beta) is the inverse temperature. The expression is the bosonic analogue of a Fermi‑Dirac distribution — a fact that feels, in the hands of the authors, at once surprising and inevitable.
Where Spectral Gaps Open Doors
By working at positive temperature, the researchers derive an approximation‑error bound that depends on the spectral gap of a certain dual slack operator. In principle, this could be vastly better than the linear‑in‑dimension worst‑case bound of interior‑point methods — if the spectral gap is large. For many structured problems that physicists care about, such gaps naturally appear. Yet the paper stops short of demonstrating this advantage for any concrete SDP family.
An important question, sharpened by earlier work on quantum SDP solvers, is whether the spectral‑gap gain materialises for practical SDPs encountered in real applications. The BaCoN framework introduced by Carmon and colleagues (arXiv:2003.08078) highlighted that previous quantum solvers required an a priori upper bound on the trace of the primal variable — a burden that the thermodynamic approach ostensibly bypasses. But the precise circumstances under which the dual slack operator’s spectral structure translates into a genuine computational edge remain unexplored. The framework is conceptually elegant; its practical dividends are still a matter for future work.
A Divergence with Bosonic Manners
The paper’s second major contribution is a new measure of dissimilarity between positive operators: the Bose‑Einstein quantum relative entropy. When the operators are unnormalised — that is, their trace is not fixed to one — the standard Umegaki relative entropy can misbehave, even turning negative. The Bose‑Einstein version, generated by the negative Bose‑Einstein entropy, avoids that pathology. It acts as a proper Bregman divergence on the unbounded positive semidefinite cone, and it satisfies a restricted monotonicity property under bosonic Gaussian channels — operations that model, for example, the linear loss of photons or the amplification of signals. While this monotonicity falls short of a full data‑processing inequality, it carves out a meaningful territory where bosonic information can still be fruitfully compared.
Quantum Algorithms Without the Trace‑Bound Anchor
Perhaps the most tantalising practical prospect is a hybrid quantum‑classical algorithm that directly exploits the thermodynamic picture. The algorithm uses Hamiltonian simulation to prepare thermal states of the dual slack operator, Hadamard tests to estimate fidelities and overlaps, and classical sampling to update the dual variables. Crucially, unlike previous quantum SDP solvers, it never demands an a priori bound on the trace of the optimal primal matrix. Instead, the runtime depends on the spectral structure of the dual slack operator: its ground‑space degeneracy, the spectral gap, and the norms of its projected components.
This is a conceptual advance, but it comes with a caution. The run‑time also involves norms of the optimal dual variables that are not characterised a priori, a gap that echoes concerns raised by Patel and collaborators (arXiv:2410.12935) in their own quantum thermal approach to ground‑state estimation. Without tighter bounds on these quantities, the algorithm could still face exponential bottlenecks on hard instances. In its current form, the work offers a compelling blueprint rather than a ready‑to‑deploy quantum solver.
What It Means to Think Thermodynamically
The deepest contribution of this preprint may be less about any single algorithm and more about a way of thinking. For decades, optimisation and statistical mechanics have lived as parallel intellectual traditions, each borrowing metaphors from the other without a direct mathematical passport. By showing that the semi‑definite cone’s algebra is identical — term for term — to the algebra of bosonic thermal equilibrium, Wilde, Minervini, and Liu have built that passport.
Liu’s earlier thermodynamic framework for SDPs (arXiv:2505.04514) provided the initial spark; this paper extends the program into the unbounded bosonic regime. What emerges is a picture in which the universe’s preference for minimising free energy — its cosmic laziness, if you like — is not just a physical principle but a potential computational resource. Whether the spectral gaps will cooperate, and whether quantum devices of the future can prepare thermal states with the necessary fidelity, remain open questions. But the bridge between optimisation and thermodynamics now stands, and it is sturdier than anyone expected.
— Yanjiang
Yanjiang is an online editor of LoomSci.com.
References
- Minervini et al., Bose‑Einstein thermal operators for semidefinite optimization, arXiv:2605.27228
- Carmon et al., Acceleration with a Ball Optimization Oracle, arXiv:2003.08078
- Patel et al., Quantum Boltzmann machine learning of ground‑state energies, arXiv:2410.12935
- Liu et al., Quantum thermodynamics and semi‑definite optimization, arXiv:2505.04514