WorksPhysics and Astronomy
Sharp bounds on the Eliashberg critical temperature: a comparison principle, convexity in the spectrum, and the Kiessling-Altshuler-Yuzbashyan conjectures
Abstract
In Eliashberg theory of phonon-mediated superconductivity with vanishing Coulomb pseudopotential, the critical temperature marks the linear instability of the normal state. Kiessling, Altshuler and Yuzbashyan (KAY) characterized it through the top eigenvalue of a compact operator on Matsubara-frequency sequences. They proved that is unique at strong coupling and conjectured for every coupling , with the second moment of the normalized phonon spectrum and sharp, a bound Leavens stated in 1977 on partly numerical grounds. A preprint by Kondratev, not yet peer reviewed, gives a proof of this bound and of all-coupling uniqueness. We give an independent proof, for every spectrum with finite second moment, based on a comparison principle that is also Kondratev's central step and was proved to first order by Bergmann and Rainer: strengthening the pairing interaction at every Matsubara transfer, strictly at the smallest one, strictly raises the top eigenvalue. At fixed and a single Einstein mode maximizes . Convexity of the top eigenvalue in the spectrum gives sharp bounds on by the moments , , including Leavens's zeroth-moment bound , numerically, optimal at moderate coupling. All bounds persist under a Coulomb pseudopotential with symmetric cutoff. A prefix decomposition of non-increasing vectors proves a third KAY conjecture, not treated in the preprint. The finite-dimensional core, including that conjecture, is formalized in Lean 4, and two computer-assisted proofs independent of the comparison principle confirm the Einstein bound.
MSC Classification: 82D55, 47A75, 15B48, 47B65, 68V15
Introduction
Background
Eliashberg theory [1] is the standard framework for computing the critical temperature of phonon-mediated superconductors [2, 3]. Its input is the Eliashberg spectral function , which measures how strongly electrons couple to lattice vibrations of frequency . Two numbers summarize it: the dimensionless coupling constant , and the second moment of the normalized spectral measure .
For weak coupling is exponentially small in . For strong coupling, Allen and Dynes [4] found that it grows without bound but only like a square root,
in units with . The right-hand side of (1) has long been regarded as an upper bound on at every coupling. Moussa and Cohen state it as one, for , citing Allen and Dynes and Leavens [4, 5, 6]; see also [7].
Allen and Dynes also argued that a single sharp phonon line (an Einstein spectrum) maximizes at fixed and : they proved that the Einstein spectrum maximizes each coefficient of their linearized gap equation at fixed and , and noted that the top eigenvalue, and with it , is maximized by making all as large as possible [4 Sect. II and App. A]. Bergmann and Rainer had shown that for added spectral weight at any frequency raises to first order, that is, that the functional derivative is positive [8].
Leavens stated the bound itself as a rigorous least upper bound for , in the form [6 eq. (20)]. The first moment equals , so his constant is in the units of (1). His derivation has two steps. A first-order variational argument based on the positivity of Bergmann and Rainer shows that a spectrum maximizing at fixed first moment can only be a single Einstein mode. This presupposes that a maximizer exists, whereas at fixed first moment the supremum is approached only by a mode of diverging weight at vanishing frequency, as he notes. The maximum over Einstein spectra is then read off from numerical solutions of the Eliashberg equations together with the strong-coupling asymptote [6 Sect. 2 and Fig. 1]; this second step is the content of Conjecture 1 of Part III below. Two years earlier he had stated, as rigorously shown, the zeroth-moment bound , with equality for the Einstein spectrum at frequency , that is at coupling [9 eqs. (13)–(14)]. Its derivation has the same two steps: the variational argument, which presupposes an optimal spectrum, followed by a numerical search over Einstein spectra [9 Sect. 2]; the 1977 paper restates the bound [6 eqs. (18)–(19)]. In both papers Leavens argued that the bounds persist for because the Coulomb repulsion lowers [6, 9 Sect. 3].
In a series of three papers [10, 11, 12], Kiessling, Altshuler and Yuzbashyan (KAY) made these questions precise. Building on the spin-chain formulation of Yuzbashyan and Altshuler [13], they showed that the normal state of the linearized, isotropic Eliashberg model with Coulomb pseudopotential is linearly unstable precisely when times the top eigenvalue of an explicit compact self-adjoint operator exceeds one. They identified the constant in (1) as , where is the top eigenvalue of the operator of their -model at [10], and they stated the following conjectures.
Conjecture 1 of Part III [12] (Einstein phonons of frequency ): for all and there is a critical temperature , and .
Conjecture 2 of Part II [11] (general spectra ): for all admissible and there is a critical temperature , and .
Conjecture 3 of Part II [11]: a sharp lower bound, attained exactly by constant sequences, for a finite-dimensional Rayleigh quotient on non-negative non-increasing sequences.
KAY proved that is unique when exceeds a spectrum-dependent threshold (for Einstein phonons, [12 Thm. 5]), that has the limit as , and an explicit upper bound on whose constant is about 2.03 times the conjectured one [11, 12]. They noted that a proof of uniqueness for all , which the thermodynamic picture of the transition presupposes, would be mathematically desirable [11, 12]. Their monotonicity argument needs the dimensionless phonon frequency to satisfy [11 Lemma 3], and they remarked that proving monotonicity for all “would require a different strategy” [12]. Elezaby and Abanov sharpened rigorous bounds for the -model [14].
In September 2026 Kondratev posted a preprint [15], not yet peer reviewed, that gives a proof of Conjecture 2 of Part II, and with it Conjecture 1 of Part III, for every non-zero phonon measure with finite coupling and second moment. It shows that the top eigenvalue decreases strictly with temperature on all of , so that the onset temperature is unique for every , and it also derives the optimality of Einstein spectra at fixed and [15 Sect. 12]. Its central step, a coefficient-by-coefficient comparison of two kernels on the ordered top eigenvector [15 Prop. 1], is in substance our Theorem 3. We found the argument presented here without knowledge of that preprint and learned of it while preparing this manuscript. § 1.4 compares the two treatments.
Results
We prove all three conjectures, together with uniqueness of at every coupling (Theorems 5, 10 and 4). Results 1–3 below are also obtained in [15]; results 4 and 5 are not treated there. Result 2 is the bound stated by Leavens, whose derivation is described above [6], and result 3 was argued by Allen and Dynes, with the monotonicity of the top eigenvalue in the coefficients read off from the equations rather than proved [4].
Uniqueness. For every phonon spectrum with finite second moment and every there is exactly one critical temperature: for and for , so for KAY-admissible the linearized problem has a superconducting instability below and none above. increases continuously and strictly with .
The bound. for every and every spectrum. The constant cannot be improved.
Einstein spectra are optimal. At fixed and , a single Einstein mode at frequency has a strictly higher than any other spectrum.
Conjecture 3 of Part II holds.
Moment bounds. For , , where is a constant defined by the Einstein problem, and the bound is the supremum of over Einstein spectra with the same moment (Theorem 8). For this is result 2; for it is Leavens’s bound [9], where in floating point.
The first three follow from a single statement, the comparison principle of Theorem 3: if the pairing interaction is increased at every Matsubara frequency transfer, and strictly at the smallest transfer, then the top eigenvalue strictly increases. Conjecture 3 is proved separately in § 6, by decomposing non-increasing vectors into prefix indicators. Result 5 rests on a second, elementary observation: the top eigenvalue is a supremum of functionals that are linear in the spectral measure, hence convex in it, so it is at most the average of the Einstein eigenvalues over the spectrum (Lemma 15), and at a fixed moment no spectrum does better than the supremum over Einstein spectra. This makes Leavens’s reduction to Einstein spectra rigorous without assuming that a maximizing spectrum exists. Combined with the computer-assisted proofs of the Einstein inequality, it also proves that the normal state is stable above the bound of result 2 without using the comparison principle. Uniqueness also yields these bounds for every instability temperature when a Coulomb pseudopotential with a symmetric frequency cutoff is added (Corollary 9), which makes Leavens’s remark on rigorous in that model. Apart from this corollary, these are theorems about the linearized, isotropic Eliashberg model with ; § 8 states what they do not cover.
The idea of the proof
The interaction between Matsubara frequencies and depends only on the transfer , so the KAY operator is a sum : one fixed operator per transfer , weighted by the interaction strength at that transfer. Heuristically, raising changes the top eigenvalue, to first order, by , where is the optimizer. Each has two competing parts (Figure 1): a pairing part, which couples frequencies a distance apart, directly or after reflecting one of them to , and a negative diagonal part, which comes from the mass renormalization and acts as pair breaking. On the general vector either part can win. On the optimizer the pairing part always wins: Lemma 11 writes as a sum of terms and , where is the gap divided by the frequency, and Lemma 12 shows that is non-negative and non-increasing. Every term is then non-negative. KAY instead differentiated with respect to , grouping the transfers with weights that decrease in only when [11 Lemma 3]. The transfer decomposition removes this restriction. For the -model, KAY had already expanded their Rayleigh quotient over transfers, with coefficients that are non-negative on non-increasing vectors, to prove that the quotient is positive [11 eqs. (72)–(73)]; here the same expansion is used to compare different kernels. The first-order version of the whole argument goes back to Bergmann and Rainer [8 App. A]. Working with the pair-breaking eigenvalue of the linearized equations at a finite frequency cutoff, they obtained from the Hellmann–Feynman theorem and showed that the coefficient of every transfer in it is non-negative, using that for the gap function is positive and decreasing in . Theorem 3 is the corresponding statement for finite changes of the kernel: it needs no cutoff, no differentiability of the top eigenvalue and no monotonicity of the larger kernel, and the ordering of the optimizer is proved in full (Lemma 12). It also supplies the step in the Allen–Dynes argument for Einstein optimality that they read off from their equations: the top eigenvalue increases when every does, although the pair-breaking diagonal terms grow with them. Likewise, Theorem 5 replaces both steps of Leavens’s derivation: it needs no maximizing spectrum, and the Einstein inequality that he read off from numerical solutions is proved.
Figure 1. Sign pattern of the transfer operator in (8) for on the first ten Matsubara frequencies. The panels show the pairing terms and , the diagonal term , and their sum. Positive entries are solid blue and negative entries are hatched orange. The operator is indefinite, but Lemma 11 shows that it is non-negative on the cone containing the optimizer.
Relation to Kondratev’s preprint
The proof of the bound in [15] and the one given here have the same structure. The expansion of the quadratic form over transfers [15 eqs. (9)–(10)] is our Lemma 11, the kernel comparison on the ordered top eigenvector [15 Prop. 1] is our Theorem 3, and both treatments obtain Einstein optimality from Jensen’s inequality, as Allen and Dynes did for the coefficients [4 App. A]. Both rest on the transfer-by-transfer positivity that Bergmann and Rainer used at first order [8]. The two treatments differ in four respects. First, the ordering of the top eigenvector is proved there by Perron–Frobenius theory and power iteration, and here by an extremal argument that avoids the Perron–Frobenius theorem (Lemma 12). Second, the passage to infinitely many frequencies uses norm convergence of the truncated eigenvectors there and weak compactness here. Third, the preprint encloses the constant, , in exact arithmetic, whereas our proofs use only through its definition as a top eigenvalue; our certificates instead prove the Einstein inequality itself without the comparison principle (Appendix § 10). Fourth, the preprint proves sharpness along a family of Einstein spectra, while Theorem 5(c) also gives the limit for each fixed spectrum with finite fourth moment. Conjecture 3 of Part II, the moment bounds of Theorem 8, Corollary 9 and the formalization of § 7 have no counterpart in [15].
Verification and organization
For every finite truncation, the cone lemma, transfer decomposition, comparison principle, Einstein bound, monotonicity and non-strict Einstein optimality for discrete spectra are formalized and machine-checked in Lean 4 with Mathlib [16, 17], as is Conjecture 3 in full (§ 7). The text supplies the compactness argument needed for infinitely many frequencies. Two further computer-assisted proofs of the Einstein bound, independent of the comparison principle, use interval and exact rational arithmetic (Appendix A).
Section § 2 derives the KAY operator from the linearized Eliashberg equations and collects its basic properties. Section § 3 states the results. Sections § 4–§ 6 contain the proofs. Section § 7 describes the formalization and Section § 8 the scope of the results and some open problems.
The linearized Eliashberg problem
From the Eliashberg equations to an eigenvalue problem
Let and , , be the fermionic Matsubara frequencies. The isotropic Eliashberg equations with , linearized in the gap function , read [2, 3, 18]
with the interaction
Since , the interaction depends only on the transfer . We write
so that and . Put and .
We look for even solutions, , and restrict to . Splitting the sum in (2) into and and telescoping gives , hence
In (3) the same splitting gives . Insert (5) and set . The term of the first sum contributes on the right, which cancels the term coming from the in on the left. What remains is
Dividing by and writing turns this into the symmetric eigenvalue problem
where, for any summable sequence , is the operator on real with matrix
The value does not enter. This is the operator of KAY [11 eqs. (39)–(42)], whose quadratic form agrees, up to a positive overall factor, with the second variation of their spin-chain free energy at the normal state. KAY proved [11 Thms. 1, 1+ and 3] that the normal state is linearly unstable against superconducting perturbations if and only if , where
is the largest eigenvalue. The critical temperature is therefore the temperature at which .
In KAY’s spin-chain formulation, is, to linear order, times the spin angle of [11, 13]. It is proportional to rather than to the gap itself. The condition means .
Kernels
Definition 1. A kernel is a sequence of positive numbers that is non-increasing in and summable.
We use three families of kernels.
General spectra. Let be the set of probability measures on with . For and , in (4) is a kernel, since . KAY’s admissible class [11], spectra supported in with density proportional to near , is contained in .
Einstein phonons. For the kernel depends only on , namely . We write ; this is the top eigenvalue of KAY’s Einstein operator [12].
The -model at . The kernel gives KAY’s -model operator at [10], and
This operator describes the strong-coupling regime: as one has , hence .
Basic properties of the operator
Write for the indicator of a statement. Every is a weighted sum of fixed operators, one per transfer:
Here (Toeplitz) couples frequencies at distance , (Hankel) couples to the reflected frequency at distance , and is the contribution of the transfer to the mass renormalization (5): enters for every . Figure 1 shows the pattern.
Lemma 2 Let be summable sequences of real numbers.
(a) , and ; hence , the series (8) converges in operator norm, and .
(b) is compact and self-adjoint, and for .
(c) If , then , so is the largest eigenvalue of .
(d) Let be the upper-left block of (the truncation to the first positive Matsubara frequencies) and its largest eigenvalue. Then , and as .
Proof. (a) , where maps and has norm . is a finite anti-diagonal matrix with entries . is diagonal with largest entry $2/w_j$. For the three bounds sum to , and all three decrease in . (b) is a weighted shift with weights tending to , has finite rank and is diagonal with entries tending to ; so each is compact, and so is the norm-convergent sum. (c) because $A_0=0. A compact self-adjoint operator whose spectrum contains points above 0 has its supremum as an eigenvalue. (d) is a compression of , so is non-decreasing in and . Every finitely supported unit vector lies in some truncation, and such vectors are dense in the unit sphere; since the quadratic form of the bounded operator is continuous, . □
Main results
Theorem 3 (Comparison principle) Let be a kernel and a sequence with for every , , and . Then
In Theorem 3 only needs to be non-increasing; need not be monotone. Monotonicity of is used only to locate the optimizer of (Lemma 12). We write and for the corresponding truncated eigenvalues.
Theorem 4 (Uniqueness of at every coupling)
(a) The Einstein eigenvalue is continuous and strictly increasing on , with and as .
(b) For every , is continuous and strictly decreasing on , with as and as .
(c) Consequently, for every and every there is exactly one with . One has for and for , so for KAY-admissible the normal state is linearly unstable exactly when . The map is a continuous, strictly increasing bijection of onto itself.
(d) For Einstein phonons of frequency , .
For every truncation , the maps and are likewise continuous and, respectively, strictly increasing and strictly decreasing.
Theorem 5 (The sharp bound; Conjecture 1 of Part III and Conjecture 2 of Part II)
(a) for every , and for every and .
(b) Consequently, for every ,
(c) The constant is sharp: as , hence as . The same limit holds for every with .
Theorem 6 (Einstein spectra are optimal). Let and . Then for every , with strict inequality unless . Hence for every unless .
Theorems 4 and 5 together prove Conjecture 1 of [12] and Conjecture 2 of [11], which KAY stated for their admissible class; we obtain them, as does [15], on the larger class . Theorem 6 says that at fixed and the Einstein spectrum at frequency has the largest . Figure 2 illustrates the Einstein case.
Figure 2. Floating-point illustration for Einstein phonons, with truncation . The curves are not used in the proofs. Left: , which Theorem 5 proves is below 1 and tends to 1 as (strong coupling). Right: against , with the ceiling .
Remark 7 (Adding spectral weight). Theorem 3 also gives the global form of the result of Bergmann and Rainer [8]. Write , so that and, since is linear in , . Let , where is a finite measure on with , and let and be the coupling and normalized spectrum of . At every each coefficient increases strictly, by at most , so Theorem 3 gives , and Theorem 4(b) gives . Adding phonon spectral weight at any frequencies strictly raises .
Theorem 8 (Sharp moment bounds). For let
Then and . For every , and ,
The constant is sharp: the right-hand side is the supremum of over Einstein spectra with the same , the coupling being free.
For , and Theorem 8 is the bound of Theorem 5, without strict inequality. For it is Leavens’s zeroth-moment bound [9 eq. (13)]. A floating-point evaluation of on truncations with and , which agree to , gives , attained at , so that
with equality, to the accuracy of these values, for the Einstein spectrum of weight at frequency , that is , close to Leavens’s numerical values 0.2309, 1.750 and 1.143 [9 Sect. 2]. These numerical values of and are not certified; the inequality holds with defined as the supremum. Unlike the first-moment bound, which is approached only as , the zeroth-moment bound is, numerically, attained at moderate coupling, where Migdal–Eliashberg theory is usually applied. For no such bound exists, since as forces , as Leavens observed [6 eq. (17)]; for finiteness of would need a growth bound on that we do not prove.
Corollary 9 (Coulomb pseudopotential with a symmetric cutoff) Let , , and . Replace in (3) by and leave (2) unchanged. The cutoff is placed symmetrically on both frequencies, so that the Coulomb correction is self-adjoint; as in the usual pseudopotential treatment, the Coulomb term does not enter the mass renormalization [2, 4]. If at some the resulting equations have a non-zero even solution with , then
and obeys the bounds of Theorem 8. The same holds for the truncation to the first positive frequencies, and for any repulsive term whose matrix in the variables is positive semidefinite.
Proof. For even solutions the Coulomb term adds to the right-hand side of the equation for in Section 2, so (6) becomes with , a vector with finitely many non-zero entries. The inner product with gives , hence . By Theorem 4(c) this excludes , and Theorem 5(b) bounds . For the truncation the same computation gives , and by Lemma 2(d). For a general positive semidefinite repulsion the inner product argument is unchanged. □
Corollary 9 makes rigorous, in this model, Leavens’s remark that the bounds for also apply for [6, 9 Sect. 3]. It says nothing about uniqueness of the instability temperature when : the Coulomb term is negative and frequency independent, and the optimizer need not lie in the cone .
The last result concerns the quadratic forms behind KAY’s finite-dimensional analysis. For , a sequence and put
With these are and , so is the Rayleigh quotient of written in the variables , which are proportional to KAY’s linearized spin angles (§ 2). For they are the numerator and denominator of [11 l.h.s. of (70)]. KAY proved the quotient positive on non-zero, non-negative, non-increasing , showed that constant sequences are local minimizers, and conjectured that they are global minimizers [11 Conjecture 3].
Theorem 10 (Conjecture 3 of Part II) Let and let be positive and strictly decreasing, for instance with . Put
Then for every non-zero, non-negative, non-increasing ,
with equality if and only if is constant.
Since the optimizer of lies in the same cone (Lemma 12), Theorem 10 says that on this cone the Rayleigh quotient ranges between , attained by constants, and , attained by the optimizer.
Proof of the comparison principle
For write , the gap divided by the frequency, up to normalization (§ 2). Let
and let be its finite-dimensional analogue. In physical terms, means with non-increasing.
Each transfer operator is non-negative on the cone
Lemma 11 If , then for every
and . The same holds for and the truncated forms , with the sums restricted to indices below .
Proof From (8), , and . All series converge absolutely because . Subtracting gives (9). On every summand is non-negative, and for the Hankel sum is . In the truncated case the Toeplitz and diagonal sums run over , and the same pairing applies. □
In (9), the pair-breaking term is dominated by the pairing term because . For and finite , the sum is KAY’s expansion [11 eqs. (72)–(73)]; the same identity is [15 eqs. (9)–(10)].
The optimizer lies in the cone
Lemma 12. Let be a kernel. Then has a unit eigenvector for , and . The same holds for every truncation , with an eigenvector for whose entries are all positive.
For finite this is [11 Prop. 8]. We give a self-contained proof that avoids the Perron–Frobenius theorem; it is the proof formalized in Lean (§ 7).
Proof. Finite . Write and ; the case is trivial, so let . The unit sphere is compact, so the quadratic form attains its maximum on it. Since all off-diagonal entries of are positive, , so there is a maximizer . A maximizer is an eigenvector: the function is non-positive and vanishes at , so its derivative there, , vanishes for every . If for some , the -th row of reads , which forces ; hence for all .
Put , and for . Multiplying the -th row of by and adding to both sides gives
and is increasing in . So the matrix has positive entries and satisfies .
maps the cone of non-negative non-increasing vectors into itself. The cone consists of non-negative combinations of the prefix indicators , , where is the indicator of , and
because and is non-increasing on (recall ). So , and multiplication by the positive non-increasing sequence preserves .
Now let and maximize over the compact set . The set is mapped into itself by , since . The maximum is positive, since a small constant vector lies in . Let attain it, so . If , then has all entries positive, so , a contradiction. Hence is non-increasing, that is, .
. By Lemma 2(d), . Let be the unit eigenvectors above, padded by zeros, and choose a subsequence with weakly. Since is compact, , and ; so . Thus , and as its Rayleigh quotient is at least , which forces and . Weak convergence implies convergence of each coordinate, which preserves and the monotonicity of . Finally , since a non-zero non-negative non-increasing sequence has a positive first entry.
Proof of Theorem 3
Proof. Take from Lemma 12 for the kernel . By (8) and Lemma 11,
The rearrangement is justified by . For a truncation the same computation applies, with and the eigenvector of ; there only occur.
Remark 13 Where is differentiable, the same computation gives . KAY grouped this derivative by the weights , which requires them to decrease in ; that holds if and only if [11 Lemma 3]. Grouping by the transfer needs no condition on the weights.
Proofs of uniqueness, the sharp bound and Einstein optimality
We use one elementary estimate twice.
Lemma 14 (Zero-temperature divergence) Let , , and let be a summable real sequence with for . Then , where .
Proof. Test with , , that is . The reference matrix is defined directly by the finite formula (7); its Rayleigh quotient at this vector is
Write with . The off-diagonal entries of are , of absolute value at most ; the diagonal entries are , of absolute value at most . At the test vector the quadratic form of the difference is therefore at most in absolute value, and dividing by gives the claim. The first inequality is Lemma 2(d).
Since , the right-hand side can be made as large as we like.
Proof of Theorem 4 Monotonicity. If , then for every , the differences are summable, and the Einstein kernels are kernels; Theorem 3 gives . If , then for every because gives no mass to , and ; so .
Continuity. By Lemma 2(a) and Weyl’s inequality, . For , each tends to as by dominated convergence, and is at most , which is summable in . Hence is continuous. The Einstein case is the special case .
Limits. As , Theorem 5(a), whose proof uses only Theorem 3, gives ; likewise as . As , for every by dominated convergence, so Lemma 14 gives for every , and . The same argument gives as .
Conclusion. Thus is a decreasing homeomorphism of onto itself, and is the unique solution of ; it increases continuously and strictly with . For KAY-admissible the stability statement is [11 Thm. 1+]. For Einstein phonons, gives Theorem 4(d).
Truncations. The same arguments, with the truncated version of Theorem 3 and the bound (Lemma 2(a) for the blocks), prove the monotonicity and continuity of and . Their limits are finite: as (compare [11 Thm. 5]). □
Proof of Theorem 5 (a) For the Einstein kernel satisfies for every , and Theorem 3 with and gives by Lemma 2(b). For general ,
for every , and the same argument gives .
(b) At we have , which rearranges to the claim; the Einstein case is .
(c) Let . Then
so by Lemma 2(a), . Dividing by shows that the ratio tends to as . As , by Theorem 4, and then gives . For Einstein phonons . □
Proof of Theorem 6 For fixed and the function is strictly concave on . By Jensen’s inequality,
which is the Einstein kernel at frequency , i.e. at . Equality holds for some (equivalently every) only if is -almost surely constant, i.e. . Otherwise the inequality is strict for every , and Theorem 3 gives . At this gives , so by Theorem 4 the Einstein system at frequency is still below its critical temperature: . □
Lemma 15 (Convexity in the spectrum) For every and , .
Proof. Write for the Einstein kernel, so that and . The function is continuous (Theorem 4; the continuity part of its proof uses only Lemma 2(a)) and satisfies by Lemma 2(a) with and ; since has a finite first moment, the integral is finite. Let have finitely many non-zero entries. Then is a finite linear combination of the , so
Such vectors are dense in and is bounded, so taking the supremum of the Rayleigh quotient over them gives the claim. □
Proof of Theorem 8. For , by Theorem 5(a), and for , by the estimate in the proof of Lemma 15. Theorem 5(a) and (c) give . By Lemma 15 and ,
At the left-hand side equals by Theorem 4, which gives the bound on . For sharpness, fix and , put and , and take the Einstein spectrum , whose -th moment is and whose coupling is . Then , so by Theorem 4(d). The supremum of over is . □
Proof of Theorem 10
The proof decomposes a non-increasing vector into the prefix indicators of § 4 and has three steps: the form on a single prefix, the cross terms between nested prefixes, and the strict decrease of in .
Proof. Write . For let be the indicator of , and let be the symmetric bilinear form of .
Prefixes. On a single prefix the diagonal subtraction cancels the Toeplitz part exactly, because . Hence , and . So .
Cross terms. For , the vector is supported on , disjoint from the support of . All matrix entries of that couple the two supports are off-diagonal and positive, and is diagonal. Therefore .
is strictly decreasing. Let be independent and uniform on . Then is uniform on , so . For , pointwise, with strict inequality when . Since is strictly decreasing, .
Conclusion. Every non-negative non-increasing is with (). Then
The diagonal terms with are , which are positive unless . So equality forces for all , i.e. constant. Conversely, a constant gives . □
Machine-checked and computer-assisted verification
Lean formalization
The finite-dimensional core of the argument is formalized in Lean 4 (v4.35.0) with Mathlib [16, 17]. For and an arbitrary real sequence , the development defines the matrix of (7) on Fin entry by entry, its quadratic form, and as the supremum of the quadratic form over the unit sphere. A kernel (IsKernel) is a sequence that is positive and non-increasing for ; summability plays no role at finite . The cone InCone is . Discrete phonon spectra are finite families of frequencies with non-negative weights summing to one. Table 1 lists the formalized theorems. Each depends only on Lean’s three standard axioms (propext, Classical.choice, Quot.sound), as reported by #print axioms; the development contains no sorry, no added axioms and no native decide.
| Lean name | Statement |
exists_max, qf_le_kappa | is attained on the unit sphere |
exists_cone_maximizer | Lemma 12 for finite : a maximizer lies in |
qf_sub, mu_nonneg, mu_one_ge | decomposition (8) and Lemma 11 |
comparison | Theorem 3 for |
kappa_pos | for a kernel , Lemma 2(c) |
kappa_smul | for |
einstein_lt | for , Theorem 5(a) |
einstein_strictMonoOn | is strictly increasing on |
disp_lt | Theorem 5(a) for discrete spectra |
disp_strictAntiOn | is strictly decreasing on , discrete |
disp_le_einstein | for , discrete |
conj3, conj3_eq_iff | Theorem 10, inequality and equality case |
Table 1. Lean theorems and the statements they formalize, all for the truncation to frequencies.
What is not formalized. (i) The passage : Lemma 2(d) and the case of Lemma 12 (its first assertion); (ii) continuity and the limits in Theorem 4, and the construction and properties of in its parts (c)–(d); (iii) the conversion of the spectral bound to a bound in Theorem 5(b), and the sharpness assertion in its part (c); (iv) spectra that are not finite sums of point masses, and the strict form of Theorem 6; (v) the derivation of Section 2, Lemma 15, Theorem 8 and Corollary 9; (vi) KAY’s identification of the instability with . The steps in (i)–(v) use standard facts about compact operators, dominated convergence and Jensen’s inequality and are proved in full in the text; step (vi) is the cited result of KAY [11 Thm. 1+] and is not reproved here. In particular, the formalized Einstein bound has the truncated constant ; the statement for the full operator needs the case of Lemma 12.
Computer-assisted proofs
Before the comparison principle was found, the Einstein inequality was proved using ball arithmetic. A second implementation, written from the list of lemmas alone, proves it in exact rational arithmetic. Neither uses Theorem 3; together they provide independent checks of Theorem 5 for Einstein phonons. Lemma 15 does not use Theorem 3 either, and with it gives for every . So the certificates also prove, independently of the comparison principle, that for every : the normal state is linearly stable above the bound for every spectrum. Appendix § 10 gives the details.
Use of large language models
This research used large language models (LLMs) extensively, and we describe their role as the journal’s policy requires. The author selected the problem, directed the research throughout, decided which approaches to pursue and which to abandon, reviewed the statements and proofs of all results in full, and is responsible for the content of this paper. LLM-based coding agents (Anthropic Claude and OpenAI GPT models, used through an agent harness) assisted, under the author’s direction, with the literature search, the exploration of proof strategies and the drafting of proofs, the certificate programs and the Lean formalization, adversarial audits of the arguments by models of different families, and drafting and editing the text. No output was accepted on the authority of a model. Every new result in the paper is proved in the text and checked by the author; its finite-dimensional core is in addition machine-checked in Lean, and the Einstein bound is confirmed by the certificates of Appendix § 10. No LLM is an author of this paper.
Scope and open problems
Apart from Corollary 9, all results are theorems about the linearized, isotropic Eliashberg model with , in the formulation of KAY [11, 12], where the critical temperature is the linear-stability boundary of the normal state. They do not address the nonlinear gap equations or the order of the transition, a Coulomb pseudopotential beyond the bound of Corollary 9, anisotropic or multiband pairing, vertex corrections, or the breakdown of Migdal–Eliashberg theory at large [19, 20]. That breakdown motivates bounds on of a different kind [21]. The bound is a statement about this model, not a prediction for any material.
Four questions seem natural. (i) With the kernel acquires a negative, frequency-independent part and the cone argument does not apply as it stands. Corollary 9 bounds every instability temperature, but whether that temperature is unique, and whether a comparison principle survives, is open. (ii) The comparison principle concerns the linear problem. Whether the full nonlinear gap equations have a non-trivial solution exactly for , so that the linear and thermodynamic transition temperatures coincide for all , is part of KAY’s broader programme [11]. (iii) The cone seems to be the natural home of physical gap functions in this model. Other monotonicity questions, for instance in the -model with general [10, 14], may yield to the same decomposition by transfer. (iv) Theorem 8 leaves the value of uncertified, which would need an upper bound on uniformly for large , and the moments open.
Supplementary information. The supplementary archive contains the Lean project (source and build instructions), the two certificate programs with their receipts, and the scripts that produce the figures, the values quoted in Appendix A and the floating-point value of after Theorem 8.
Declarations
Funding. This work was carried out at, and funded by, SuperMatics Inc. It received no external funding.
Competing interests. The author is a founder of SuperMatics Inc., which develops computational methods for materials discovery. The author declares no other competing interests.
Ethics approval and consent to participate. Not applicable.
Consent for publication. Not applicable.
Data availability. No experimental data were generated. The supplementary archive contains the programs that produce the paper’s own numerical results, together with the certificate receipts.
Materials availability. Not applicable.
Code availability. The Lean formalization, the certificate programs and the figure scripts are included in the supplementary archive, which is also distributed as the ancillary files of the arXiv version of this paper.
Author contribution. D.R. selected the problem, directed the research, reviewed and checked all statements and proofs, and wrote the paper, with the assistance of large language models described in § 7.3.
Computer-assisted proofs of the Einstein inequality
We prove for every , where and ; by Lemma 2(b), , so this is Theorem 5(a) for Einstein phonons. Note that . Decompose with entrywise and diagonal, and let , a decreasing function.
Lemma 16 (a) (Envelope.) If , then with .
(b) (Block.) If with , and , then
(c) (Cholesky certificate.) If is an exactly representable lower-triangular matrix and with , then .
(d) (Tails.) On the coordinates , , and
(e) (Large .) .
Proof. (a) , because each decreases in . (b) . (c) with . (d) is bounded above by the tail of ; the Schur test with weights bounds its norm by . The Frobenius bound uses and . (e) The same Schur test over all indices. □
Primary certificate (ball arithmetic).
A floating-point eigenvector of the truncation of , rounded to an exact dyadic vector and then normalized in Arb to a unit vector , gives the rigorous lower bound on and the residual . By min–max and Lemma 16(b)–(d) applied to , the second eigenvalue of is at most (its floating-point value is ), and the Davis–Kahan theorem [22] bounds the distance of to the unit top eigenvector of , with its sign chosen so that , by . The half-line is split into three regions.
Region I, . Here with and . For a unit vector with and , with , and the resulting positivity criterion , with , is at least .
Region II, , is divided into 48 cells with dyadic endpoints, obtained by rounding a geometric grid of ratio 1.06. In each cell, Lemma 16(a)–(d) with and a Cholesky certificate bound below . The smallest relative margin is 3.96%, on .
Region III, , is controlled by .
All bounds are evaluated in Arb ball arithmetic [23] at 200 bits, through python-flint 0.9.0 [24] on FLINT 3.6.0 [25]. Floating point only proposes trial objects (the unnormalized trial vector and the Cholesky factors ), which are then treated as exact dyadic data; the trial vector is normalized in Arb. The run takes about 2.1 s on a laptop. Table A1 lists the certified quantities.
| Quantity | Certified bound | Role |
| (Rayleigh quotient, ) | lower bound on | |
| (residual) | eigenvector error | |
| (second eigenvalue of ) | spectral gap | |
| () | Region I | |
| () | Region I | |
| Region I criterion | Region I | |
| Region II worst margin | (48 cells) | Region II |
| Region III |
Table A1. Bounds from the receipt of the primary certificate.
Independent certificate (exact rationals).
A second implementation works after the congruence , , in which all entries are rational, and checks every inequality in exact rational arithmetic (FLINT fmpq). It uses , the regions , (17 geometric cells of ratio ) and , the rational lower bound , and rational bounds on the integrals in place of coth. It certifies on the first region, a gap of at least below on the second and of at least on the third, in 0.25 s.
Provenance.
The plan of the primary certificate was fixed before it was run (SHA-256 197786ca6b2fb0676ffde142bf6c7bbec1b5c97f8e43d988bfcf53bd2decbaec). Primary certificate script cd41aff1b564efe08fa23eea1849b709fe1324661d20b26d55652bcd15df4d9d, receipt 9ce805079f1f3a1abce66b757210adba40ee647f656fddf4588387664d893bfe; independent certificate script 7242083ff94a7c1f939a39ea6df55f18b4398ae220fd7b87b9c28646356f8036, receipt 4a502d96d2cc311ba5948e67ebd861b36e913650778af7ab1fc46192c86c3e03.
References
- [1]Eliashberg, G.M.: Interactions between electrons and lattice vibrations in a superconductor. Soviet Physics JETP 11(3), 696–702 (1960)
- [2]Carbotte, J.P.: Properties of boson-exchange superconductors. Reviews of Modern Physics 62(4), 1027–1157 (1990) https://doi.org/10.1103/RevModPhys.62.1027
- [3]Marsiglio, F.: Eliashberg theory: A short review. Annals of Physics 417, 168102 (2020) https://doi.org/10.1016/j.aop.2020.168102
- [4]Allen, P.B., Dynes, R.C.: Transition temperature of strong-coupled superconductors reanalyzed. Physical Review B 12(3), 905–922 (1975) https://doi.org/10.1103/PhysRevB.12.905
- [5]Moussa, J.E., Cohen, M.L.: Two bounds on the maximum phonon-mediated superconducting transition temperature. Physical Review B 74(9), 094520 (2006) https://doi.org/10.1103/PhysRevB.74.094520
- [6]Leavens, C.R.: Important parameters in strong-coupling superconductivity. Journal of Physics F: Metal Physics 7, 1911–1922 (1977) https://doi.org/10.1088/0305-4608/7/9/034
- [7]Semenok, D.V., Altshuler, B.L., Yuzbashyan, E.A.: Fundamental limits on the electron–phonon coupling and superconducting T_c. Advanced Materials 37(40), 2507013 (2025) https://doi.org/10.1002/adma.202507013
- [8]Bergmann, G., Rainer, D.: The sensitivity of the transition temperature to changes in α^2F(ω). Zeitschrift für Physik 263, 59–68 (1973) https://doi.org/10.1007/BF02351862
- [9]Leavens, C.R.: A least upper bound on the superconducting transition temperature. Solid State Communications 17, 1499–1504 (1975) https://doi.org/10.1016/0038-1098(75)90982-5
- [10]Kiessling, M.K.-H., Altshuler, B.L., Yuzbashyan, E.A.: Bounds on T_c in the Eliashberg theory of superconductivity. I: The γ-model. Journal of Statistical Physics 192(5), 69 (2025) https://doi.org/10.1007/s10955-025-03446-5
- [11]Kiessling, M.K.-H., Altshuler, B.L., Yuzbashyan, E.A.: Bounds on T_c in the Eliashberg theory of superconductivity. II: Dispersive phonons. Journal of Statistical Physics 192(7), 94 (2025) https://doi.org/10.1007/s10955-025-03468-z
- [12]Kiessling, M.K.-H., Altshuler, B.L., Yuzbashyan, E.A.: Bounds on T_c in the Eliashberg theory of superconductivity. III: Einstein phonons. Journal of Statistical Physics 192(7), 93 (2025) https://doi.org/10.1007/s10955-025-03469-y
- [13]Yuzbashyan, E.A., Altshuler, B.L.: Migdal–Eliashberg theory as a classical spin chain. Physical Review B 106(1), 014512 (2022) https://doi.org/10.1103/PhysRevB.106.014512
- [14]Elezaby, A., Abanov, A.: Superconductivity near a quantum critical point: Bounds on the transition temperature in the γ-model. Physical Review B 114(2), 024512 (2026) https://doi.org/10.1103/wpn8-8j2t
- [15]Kondratev, O.: A proof of Conjecture 2 of Kiessling, Altshuler and Yuzbashyan: the sharp spectral-moment bound on the linear-stability onset temperature in Eliashberg theory. Zenodo preprint, not peer reviewed. Latest version, record 22848039, 19 September 2026 (2026). https://doi.org/10.5281/zenodo.22845696
- [16]de Moura, L., Ullrich, S.: The Lean 4 theorem prover and programming language. In: Automated Deduction – CADE 28. Lecture Notes in Computer Science, vol. 12699, pp. 625–635. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-79876-5_37
- [17]The mathlib Community: The Lean mathematical library. In: Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2020), pp. 367–381. ACM, New York (2020). https://doi.org/10.1145/3372885.3373824
- [18]Allen, P.B., Mitrović, B.: Theory of superconducting T_c. In: Ehrenreich, H., Seitz, F., Turnbull, D. (eds.) Solid State Physics vol. 37, pp. 1–92. Academic Press, New York (1983). https://doi.org/10.1016/S0081-1947(08)60665-7
- [19]Yuzbashyan, E.A., Altshuler, B.L.: Breakdown of the Migdal–Eliashberg theory and a theory of lattice–fermionic superfluidity. Physical Review B 106(5), 054518 (2022) https://doi.org/10.1103/PhysRevB.106.054518
- [20]Esterlis, I., Nosarzewski, B., Huang, E.W., Moritz, B., Devereaux, T.P., Scalapino, D.J., Kivelson, S.A.: Breakdown of the Migdal–Eliashberg theory: A determinant quantum Monte Carlo study. Physical Review B 97(14), 140501 (2018) https://doi.org/10.1103/PhysRevB.97.140501
- [21]Esterlis, I., Kivelson, S.A., Scalapino, D.J.: A bound on the superconducting transition temperature. npj Quantum Materials 3, 59 (2018) https://doi.org/10.1038/s41535-018-0133-0
- [22]Davis, C., Kahan, W.M.: The rotation of eigenvectors by a perturbation. III. SIAM Journal on Numerical Analysis 7(1), 1–46 (1970) https://doi.org/10.1137/0707001
- [23]Johansson, F.: Arb: Efficient arbitrary-precision midpoint-radius interval arithmetic. IEEE Transactions on Computers 66(8), 1281–1292 (2017) https://doi.org/10.1109/TC.2017.2690633
- [24]Johansson, F., Benjamin, O., et al.: python-flint: Python bindings for FLINT, version 0.9.0. https://github.com/flintlib/python-flint (2026)
- [25]The FLINT team: FLINT: Fast Library for Number Theory, version 3.6.0. https://flintlib.org (2026)

