Certified Phase Transitions: Placing Electroweak Baryogenesis Inside a Decidable Logic

The universe contains far more matter than antimatter. One leading explanation, electroweak baryogenesis, needs the early universe to have cooled through a strongly first-order electroweak phase transition. In such a transition, bubbles of the new vacuum appear and expand, and the Higgs field jumps abruptly instead of changing smoothly. Whether a proposed extension of the Standard Model produces such a jump is currently decided by numerical scans. Physicists sample parameter points, look for local minima of a temperature-dependent energy function (the effective potential), and run shooting codes such as findbounce (Theoretical uncertainties in first-order electroweak phase transitions). These scans can miss a deeper vacuum hidden elsewhere in field space. They cannot prove that a parameter region has no viable transition. They also report results at one arbitrary choice of renormalization scale, even though that choice is the largest source of theoretical error.

This note proposes to take that numerical procedure and place it inside a logic where every question has a guaranteed yes-or-no algorithm. The logic is the first-order theory of the real numbers with addition, multiplication, and order. Tarski showed that this theory is decidable: an algorithm always terminates with a correct answer. The key observation is that the standard high-temperature, one-loop, daisy-resummed effective potential is semialgebraic. A semialgebraic set is one described by polynomial equalities and inequalities. Two substitutions make this work. The scale logarithm is treated as a bounded real variable, and each non-integer mass power is replaced by an auxiliary variable with a polynomial constraint. Under these substitutions, the question "does this model have a strongly first-order transition, with no deeper vacuum, for every renormalization scale in the accepted band?" becomes a decidable sentence. We would build the encoding, map exactly where the physics pipeline leaves the decidable fragment, and test whether certified answers change conclusions that numerical scans currently reach.

Core Claim

The claim has three parts.

First, there is a precise decidability frontier running through the standard phase-transition pipeline. On the decidable side sit three steps: finding the vacuum structure, the critical temperature at which two minima are degenerate, and the washout criterion $v_c/T_c > 1$. These lie in real-closed-field arithmetic once the potential is truncated at leading high-temperature order. On the other side sit two steps. The full thermal integrals are transcendental functions, and the nucleation condition depends on a differential equation (the bounce equation).

Second, the boundary can be pushed outward soundly. The transcendental pieces can be sandwiched between certified polynomial upper and lower bounds. The variational characterization of the bounce lets polynomial trial profiles give certified upper bounds on its action, and hence sufficient conditions for nucleation.

Third, inside the decidable fragment we can produce things scans cannot. These include proofs that a parameter box contains no strongly first-order point, exact boundary polynomials for the viable region, and verdicts that hold uniformly over the scale band rather than at one scale.

The encoding rests on a standard fact about the high-temperature expansion. For a boson of field-dependent mass squared $m^2(\phi)$, the one-loop thermal plus Coleman–Weinberg contribution is, to leading orders:

$$ \Delta V \simeq \frac{m^2 T^2}{24} - \frac{T\,(m^2)^{3/2}}{12\pi} + \frac{m^4}{64\pi^2}\Big(2L - \log c_B\Big). $$

Here $T$ is temperature, $c_B$ is a known numerical constant, and $L = \log(\mu/T)$, where $\mu$ is the renormalization scale. In this expansion the $m^4\log m^2$ pieces cancel between the zero-temperature and thermal parts. That cancellation leaves only $L$, which ranges over $[-\ln 2, \ln 2]$ when $\mu \in [T/2, 2T]$.

Two substitutions then make the whole expression polynomial. We introduce $w \ge 0$ with $w^2 = m^2$ to absorb the $3/2$ power, and we treat $L$ as an ordinary bounded real variable. Daisy resummation, which replaces $m^2$ by $m^2 + \Pi T^2$ for a thermal mass coefficient $\Pi$, keeps the expression polynomial. Fermion terms behave the same way, without the cubic term.

The robust washout criterion then reads, schematically: for all $L$ in the band, there exist $T$ and a broken minimum $\phi_c$ such that two conditions hold. The symmetric and broken minima are degenerate and $|\phi_c|/T > 1$. And for all field values $\phi'$, $V(\phi', T) \ge V(\phi_c, T)$. This is a sentence with alternating quantifiers over the reals, so Tarski's theorem applies.

One modeling choice must be stated openly. Field-dependent masses can turn negative (for example for Goldstone modes), which makes $(m^2)^{3/2}$ ill-defined. We adopt the standard practice of using daisy-improved masses restricted to their positive region, and we flag this as part of what "the fragment" means.

How This Relates to Prior Work

The decidability literature supplies the method and the way of thinking. Several papers carve decidable pieces out of undecidable formalisms by restricting syntax. On Decidable and Undecidable Extensions of Simply Typed Lambda Calculus shows that adding references or exceptions pushes program analysis past the frontier. Fragments of ML Decidable by Nested Data Class Memory Automata recovers decidability by limiting how state is used. Verifying higher-order concurrency with data automata imposes local boundedness and restricted semaphores. We apply the same move to a physics computation. Truncation order, number of scalar fields, and the treatment of thermal functions are the "syntactic" restrictions we vary. The frontier we chart concerns physical validity and computational cost, not program semantics.

The Counting Power of Transformers constrains our design. It shows that polynomial expressivity becomes undecidable once polynomial relations can encode Diophantine equations, via Hilbert's Tenth Problem. Our setting avoids that trap only because all variables range over the reals, not the integers. Keeping integer-valued quantities (such as particle multiplicities) as fixed constants rather than variables is a design rule that this paper forces on us.

Grounding Neural Inference with Satisfiability Modulo Theories embeds the Z3 solver in learning, but only over linear integer arithmetic, Booleans, and bit-vectors. We use nonlinear real arithmetic, and we use the solver as a certifier, not as a training signal.

On the physics side, Theoretical uncertainties in first-order electroweak phase transitions motivates the robustness quantifier. It finds that varying $\mu \in [T/2, 2T]$ dominates every other uncertainty in the transition parameters of the Standard Model effective field theory with a $(\Phi^\dagger\Phi)^3/\Lambda^2$ operator. That operator also makes it our first test model.

Electroweak baryogenesis and dark matter with an approximate R-symmetry and Probing the Electroweak Phase Transition with Exotic Higgs Decays build singlet extensions in which a tree-level barrier produces the strong transition. The second paper reports that viable light-scalar models imply $\text{Br}(h\to ss) \in [10^{-6}, 10^{-1}]$. Both were obtained by grid scans and forward simulation, so these are claims we can try to certify.

The same Carena et al. paper also limits our scope. Perturbation theory breaks down for $m_s < 5$ GeV. Our fragment certifies statements about the perturbative potential. It cannot make that potential physically trustworthy where it is not. We exclude that regime from any physical claim.

Research Program

The work has four parts.

The first part is the encoding. We write a translator from a model specification (fields, couplings, thermal mass coefficients) into first-order real-arithmetic sentences. The translator covers three models of increasing size. The first is the single-field effective theory with the $\phi^6$ operator. The second is the real-singlet extension with a $Z_2$ symmetry, where two-step transitions occur. The third is the general singlet extension with a tree-level barrier.

The second part is solving. We run two kinds of solver backends on identical sentences. One is a complete nonlinear-real-arithmetic decision procedure, such as Z3's nonlinear mode, for per-point and existential queries. The other uses cylindrical algebraic decomposition (a method that splits space into regions where polynomial signs are constant) for quantifier elimination over two- or three-parameter slices. Quantifier elimination yields the explicit polynomial boundaries of the viable region.

The third part pushes past the frontier. We compute certified polynomial enclosures of the full thermal function $J_B$ on bounded ranges of $m/T$, with rigorous remainder bounds. These give inner and outer approximations of the viable region under the exact one-loop potential. We also derive certified upper bounds on the bubble action $S_3$ from polynomial trial paths. This yields a sound sufficient condition for nucleation, $S_3/T < 140$, although it is not a necessary one.

The fourth part maps the frontier empirically. We measure solver time and success as the number of variables grows (fields, auxiliary mass variables, symbolic parameters). Cylindrical algebraic decomposition is doubly exponential in the number of variables, so we expect a sharp practical wall. Finding where that wall falls for realistic models is itself a result.

Baselines

The naive reference point is the closed-form criterion for single-field models. For a potential with cubic coefficient $E$ and quartic $\lambda$, $v_c/T_c = 2E/\lambda$. Our pipeline must reproduce this exactly on the single-field model. If it does not, the encoding is wrong. This baseline also marks where the proposal adds nothing: in the single-field case the method is a correctness check, not a contribution.

The strongest existing method is the standard numerical pipeline used in Theoretical uncertainties in first-order electroweak phase transitions. It combines dense parameter sampling, local minimization of the effective potential from multiple starting points, and findbounce for nucleation. It is the strongest competitor because it is what the field actually trusts, and it handles the exact one-loop potential, not a truncation.

We run it on the same models, the same parameter boxes, and the same truncated potential, with matched wall-clock budgets. We give it a generous number of random restarts and also run it at the three scales $\mu = T/2, T, 2T$. No reported result found on Lacuna for the misclassification rate of such scans on singlet-extension benchmarks. The file gives physics results but no accuracy figures for the scanning procedure itself, so we must measure that rate ourselves.

The method "beats" the scan in one of three ways. It may find points where the scan's label is provably wrong, for example a missed deeper vacuum or a spurious strong transition. It may certify empty boxes the scan can only call "nothing found." Or it may show that the scale-robust viable region differs materially from the single-scale region.

The second baseline is interval-arithmetic branch-and-bound. This method recursively subdivides parameter and field space and uses interval bounds to rule boxes in or out. It is also sound, so it is the real methodological competitor. We run it on identical sentences and budgets.

The ablation that isolates our key ingredient removes the scale variable $L$ and fixes $\mu = T$. Comparing the viable regions with and without the universal quantifier over $L$ measures what the robustness specification contributes.

Why not just…?

Why not just run denser numerical scans? Denser scans reduce, but never remove, two failures: a deeper vacuum missed by every starting point, and the inability to state that a region is empty. The Carena et al. analysis shows that viable regions can be thin slivers, with branching ratios down to $10^{-6}$, which is exactly where finite grids skip structure. That is a reason for concern, not proof. The decisive test is empirical. If, across the singlet models, we find zero certified mislabelings and the certified-empty boxes match what scans already believed, then exact solving buys only proofs, not new physics, and we will report that.

Why not just use interval branch-and-bound instead of exact decision procedures? This is a genuine threat. For pointwise verdicts and low-dimensional boxes, interval methods may be as fast and equally sound, and we treat them as a co-equal backend rather than a straw man. They have two weaknesses. Near degenerate minima they can subdivide indefinitely, and boundaries where $v_c/T_c = 1$ exactly, or where two minima are exactly degenerate, are where the physics lives. They also produce boxes, not the polynomial equations of the boundary. If interval methods match quantifier elimination on both coverage and time in the two-field models, the contribution narrows. It becomes the decidability classification and the encoding, and the solver choice becomes secondary. The experiments are designed to settle this directly.

Why not just train a neural surrogate or ask a language model to classify parameter points? A surrogate cannot certify absence. The file's evidence that learned reasoners are weak at exactly this kind of question is indirect but consistent. SATBench reports o4-mini accuracy dropping to 65.0% on hard unsatisfiable puzzles. EquiBench reports best category accuracy of 63.8% on dead-code elimination, against a 50% random baseline. A surrogate remains useful for proposing candidate regions that the solver then certifies, and we use one only in that role.

Proposed Experiments

The first experiment is the sanity check. On the single-field $\phi^6$ model, we recover $2E/\lambda$ symbolically by quantifier elimination. We then reproduce, at fixed $\mu$, the washout bound on $\Lambda$ that Theoretical uncertainties in first-order electroweak phase transitions obtains numerically, and report the certified $\Lambda$ interval valid across the full $\mu$ band.

The second experiment is the main comparison, on the $Z_2$ singlet model. Over fixed parameter boxes, we compare the numerical pipeline against certified labels. The metrics are the fraction of the box certified, the rate of certified scan mislabelings broken down by failure type, and time per certified point.

The third experiment targets the light-scalar model of Carena et al., restricted to $m_s \ge 5$ GeV. We compute the certified infimum of the Higgs–singlet coupling over the strongly first-order region and translate it into a certified lower bound on $\text{Br}(h\to ss)$ within the truncated potential. We compare this bound against the reported $10^{-6}$ floor. A materially different certified floor would be the most consequential outcome. Agreement would itself be a useful confirmation of a scan-derived bound.

The fourth experiment measures how much of the viable region survives the move to the exact one-loop potential. We report what fraction of the truncated-potential verdicts remain certified under the polynomial enclosures of $J_B$, and how many nucleation-sufficient points the trial-path action bounds certify. Together these quantify how far the frontier can be pushed soundly.

The fifth experiment is the scaling study. We plot solver success against the number of variables to locate the practical tractability wall for each backend.