We prove that the spherically symmetric SO(d) Yang–Mills equation admits a third smooth self-similar profile, in addition to the two known explicit profiles, in every even dimension d≥12. This complements the analogous existence theorem in odd dimensions d≥11 (arXiv:2610.36784). In contrast to the odd-dimensional case, where the classification of smooth self-similar profiles reduces to counting roots of an explicit resonance polynomial [Phys. Rev. D 114, 064021 (2026)], the even-dimensional proof requires solving a nonlinear connection problem. In d=10, where one explicit profile degenerates to the zero solution, we give a formal argument for the existence of a second nontrivial smooth profile in addition to the single nontrivial explicit one.
Introduction
Self-similar solutions of the Yang–Mills equation provide natural models of finite-time singularity formation. For such solutions a fixed spatial profile is compressed into a shrinking region around the blowup point. Constructing and classifying these profiles is a first step towards understanding the possible local blowup dynamics. Their role in the evolution depends on their stability.
Under the spherically symmetric magnetic ansatz with gauge group SO(d), the Yang–Mills equation in (d+1)-dimensional Minkowski spacetime reduces to the radial nonlinear wave equation
ϕtt=ϕrr+rd−3ϕr+r2d−2ϕ(1−ϕ2).(1)
The equation is invariant under ϕ(t,r)↦ϕ(t/λ,r/λ) for any λ>0. Solutions unchanged by this scaling are called self-similar. By time-translation invariance, for any T they take the form
ϕ(t,r)=u(y),y=T−tr,t<T.
Substituting this ansatz into (1) gives the profile equation
y2(1−y2)u′′+((d−3)y−2y3)u′+(d−2)u(1−u2)=0.(2)
The profiles of interest are nonconstant and smooth on 0≤y≤1, so the associated solutions are smooth in the causal past of (T,0), where they blow up as their spatial scale T−t shrinks to zero. Throughout the paper, profiles related by the reflection u↦−u are regarded as equivalent.
A standard way to solve this two-endpoint boundary-value problem is shooting. Near the origin, there is a one-parameter family of regular local solutions
u(y;a)=1−ay2+O(y4),a∈R.(3)
The shooting method consists in continuing these solutions towards y=1 and selecting the values of the parameter a for which the resulting shot is smooth at the light cone y=1.
As shown in [1], this shooting problem simplifies considerably when d≥10. More precisely, for each fixed d in this range and every a>0, the corresponding solution u(y;a) extends to the whole interval 0<y<1, where it is positive, strictly decreasing, and has a finite limit
c(a):=y→1−limu(y;a).
Moreover, the shooting map a↦c(a) is a strictly decreasing bijection from [0,∞) onto (0,1], with c(0)=1 and lima→∞c(a)=0. Thus, the oscillations and repeated endpoint values that complicate shooting in lower dimensions do not occur here: each endpoint value in (0,1] corresponds to exactly one shooting parameter. Below, we take these results as input. It remains to determine the values of the shooting parameter a for which the solution u(y;a) extends smoothly to y=1.
Two such solutions are known explicitly:
u±(y)=1−y2+β±α±y2,a±=β±α±,0<a+<a−,(4)
where, using r=3(d−4)(d−2),
α±=2±3(d−2)2r,β±=32(d−4)±r.
The solutions u+ and u− exist for d≥5 and d≥11, respectively.
The question is whether there are any other smooth profiles. The analysis depends crucially on the parity of d because the local behavior of solutions near y=1 is governed by the indicial exponents 0 and
ρ=2d−3.
The problem is therefore resonant when d is odd, in which case ρ is an integer, and nonresonant when d is even, in which case ρ is a half-integer. We discuss these two cases separately.
Odd dimensions For odd d=2ℓ+5≥11 the general local solution with finite endpoint value c has the expansion
u(y)=c+n=1∑∞αn(1−y)n+Aℓ(c)(1−y)ℓ+1log(1−y)+⋯,
where we omitted higher-order power–log terms (1−y)jlogk(1−y). Here c and αℓ+1 are free parameters, while α1,…,αℓ are determined recursively by c. The coefficient
Aℓ(c)=c(1−c2)Pℓ(c2),
measures the obstruction to continuing the Taylor expansion through the resonant order (1−y)ℓ+1. Here Pℓ is an explicitly computable polynomial of degree ℓ, defined up to a constant factor. If Pℓ(c2)=0, all logarithmic terms vanish and the expansion reduces to a convergent Taylor series. Since the origin-regular shots have 0<c<1, they are smooth at y=1 if and only if Pℓ(c2)=0. Combined with monotonicity of the shooting map, this reduces the classification of smooth self-similar profiles to counting distinct roots of Pℓ(c2) in 0<c2<1: each root determines one smooth profile, modulo reflection. Two roots correspond to the two explicit profiles. The existence of at least one additional root was recently proved in [3], establishing the existence of a third nonconstant smooth profile in every odd dimension d≥11 and confirming the conjecture formulated in [1].
Even dimensions. Write d=2m with an integer m≥5. Since ρ=m−23 is not an integer, each endpoint value c determines a unique local solution u0(y;c) analytic at y=1. The general local solution with the same endpoint value has an additional free parameter, the amplitude of its leading nonanalytic term. This contrasts with the logarithmic term in odd dimensions, whose coefficient is determined by c.
For an origin-regular shot, continuation from the origin to the light cone selects both the endpoint value c(a) and the nonanalytic amplitude Bm(a). The resulting endpoint expansion takes the form
u(y;a)=uan(y;a)+Bm(a)(1−y)ρg(y;a),(5)
where uan and g are analytic at y=1, with uan(1;a)=c(a) and g(1;a)=1. Here uan denotes the analytic component of the full solution, not in general the analytic solution u0(y;c(a)). The decomposition and its dependence on the shooting parameter are justified in Appendix § 4. By (5), the shot is smooth at the light cone if and only if Bm(a)=0. Unlike in odd dimensions, this criterion cannot be reduced to an algebraic condition on the endpoint value c. Instead, finding the zeros of Bm requires controlling the nonanalytic amplitude along the family of origin-regular shots. This is a genuinely nonlinear connection problem.
For m≥6, establishing a third smooth profile amounts to finding a zero of Bm distinct from the two explicit zeros,
Bm(a+)=Bm(a−)=0.
For m=5, only the explicit zero a+ remains, and we seek a second nontrivial smooth profile. By continuity of Bm, it is enough to find a sign change in an interval containing no explicit zero. The argument splits into three cases:
For m=6,7,8, Bm crosses both explicit zeros in the same direction, forcing an additional zero in (a+,a−).
For every integer m≥9, the sign of Bm immediately beyond a− differs from its sign for large a, forcing an additional zero in (a−,∞).
For m=5, the sign of B5 immediately beyond a+ differs from its sign for large a, forcing an additional zero in (a+,∞).
The first two cases are proved rigorously, while the third is established only at a formal level. Thus, we prove the existence of a third nontrivial smooth profile in every even dimension d=2m≥12. For m=5, where one explicit profile degenerates to the zero solution, we show by formal methods the existence of a second nontrivial smooth profile in addition to the remaining explicit one.
The rest of the paper is organized as follows. In § 2, we determine the large-a behavior of Bm(a), including the exceptional case m=5. In § 3, we determine the crossing directions at the explicit zeros and complete the existence proof for m≥6. The technical details are collected in the appendices.
Large-a behavior
The case m≥6
We use the shorthands
ν=m2−6m+6,k=m−2−ν,(6)
so that 1<k<2. As before, ρ=m−23 is the half-integer light-cone indicial exponent.
The origin expansion (1.3) identifies y∼a−1/2 as the scale on which the deviation from u=1 is no longer small. To resolve this shrinking region, let
The normalized solution Q is the limiting core profile: Ua→Q on compact x-intervals as a→∞. With τ=logx and q(τ)=Q(eτ), equation (7) becomes
q¨+(2m−4)q˙+(2m−2)q(1−q2)=0.
The origin expansion shows that the orbit enters the forward-invariant strip
0<q<1,−kq<q˙<0,
whose invariance follows by checking the vector field on its boundary. Thus Q is global, positive, and strictly decreasing to zero. Linearization at q=0 gives
Q(x)∼Cmx−kas x→∞,Cm>0.(8)
Indeed, the inequality q˙+kq>0 makes ekτq(τ) strictly increasing and hence rules out decay along the faster mode.
The outer field inherits its amplitude and small-y behavior from this tail in the overlap region a−1/2≪y≪1:
u(y;a)=Ua(ay)≃Q(ay)∼Cma−k/2y−k.
To extend this approximation beyond the overlap region, we substitute the ansatz
u(y;a)≃Cma−k/2Fm(y)(9)
into (1.2) and retain the leading terms as a→∞. This gives the linear equation
y2(1−y2)Fm′′+((2m−3)y−2y3)Fm′+(2m−2)Fm=0.(10)
Matching the behavior Fm(y)∼y−k for y→0 selects the solution
Fm(y)=y−k2F1(−2k,21−k;1+ν;y2).(11)
Thus the core fixes the amplitude, while the linear outer equation fixes the shape. Applying the hypergeometric connection formula [4 Eq. (15.8.4)] to (11) gives
Comparing this expansion with (5) and recalling the outer amplitude Cma−k/2, we obtain the asymptotic behaviour for a→∞
Bm(a)∼2ρCmKma−k/2.(14)
The matching and the passage to the light-cone coefficient are justified rigorously in Appendix § 5. It remains to determine the sign of Km. Since 1<k<2, both denominator Gamma factors in (12) are negative. Hence
sgnKm=sgnΓ(−ρ)=εm,εm:=(−1)m−1.(15)
Since Cm is positive, this is also the sign of Bm(a) for sufficiently large a: negative when m=6,8,10,…, and positive when m=7,9,11,….
The exceptional dimension d=10 (m=5)
For m=5, the limiting core profile is explicit:
Q(x)=1+x21.(16)
In the overlap region a−1/2≪y≪1,
Q(ay)=a−1y−2−a−2y−4+O(a−3y−6).
The leading outer term is therefore of order a−1. Since the cubic term in (2) first contributes at order a−3, the first two outer terms satisfy the linear equation
y2(1−y2)F′′+(7y−2y3)F′+8F=0.
One solution is
F5(y)=y−2+41.
Reduction of order gives a second linearly independent solution,
G(y)=−2F5(y)∫1yF5(s)2s−7(1−s2)5/2ds.(17)
The factor −2 normalizes the leading term at the origin to y−4, while the lower integration limit fixes the freedom to add a multiple of F5 by imposing G(1)=0. As y→0,
G(y)=y−4+6y−2logy+O(y−2).
Matching the leading core term selects a−1F5, while the term −a−2y−4 fixes the coefficient of G to be −a−2. The accompanying logarithmic term is matched by the first inner correction and introduces a logarithmic dependence on a in the coefficient of F5. Since F5 is analytic at y=1, this does not affect B5(a).
Using F5(y)=5/4+O(1−y) and 1−s2=2(1−s)(1+O(1−s)) in the integral representation of G, we obtain
G(y)=5322∫y1(1−s)5/2ds(1+O(1−y)).
Hence
G(y)∼35642(1−y)7/2as y→1−.(18)
and therefore
B5(a)∼−35642a−2as a→∞.(19)
Remark 2.1. A rigorous proof of (2.14) requires controlling the first inner correction and propagating the matching error to the outer region. These standard but lengthy estimates are omitted to avoid obscuring the main argument. We therefore state (2.14) as a formal asymptotic prediction.
Additional profiles from crossing directions
The direction in which Bm crosses an explicit zero is determined by the variation
w(y):=∂a∂u(y;a)a=a±,
so that
u(y;a±+δa)=u±(y)+δaw(y)+O(δa2).
Differentiating (1.2) with respect to a at a=a± gives
The origin expansion u(y;a)=1−ay2+O(y4) gives w(y)=−y2+O(y4). Similarly, differentiating the endpoint expansion (1.5) at a=a± and using Bm(a±)=0 gives
w(y)=wan(y)+Bm′(a±)(1−y)ρg(y;a±),(21)
where wan is analytic at y=1. Since g(1;a±)=1, the leading fractional coefficient of w is Bm′(a±). This coefficient determines the crossing direction whenever it is nonzero, because
Bm(a±+δa)=δaBm′(a±)+O(δa2).
We carry out this analysis for both explicit profiles when m≥6 and for the single explicit profile u+ when m=5.
For either branch, suppress the sign label and set
A=1+β,ξ=y2+βAy2,w=−Aβξf(ξ).(22)
This maps 0≤y≤1 to 0≤ξ≤1, and the origin condition w=−y2+O(y4) becomes f(0)=1. Inserting (22) into (20) and using the explicit profile parameters gives
Equation (23) is a Heun equation with four regular singular points at ξ=0,1,A,∞.
The origin-regular solution has the Taylor expansion
f(ξ)=n=0∑∞fnξn,f0=1.(24)
Since A>1, the nearest possible singularity is at ξ=1, so the series converges for 0≤ξ<1.
Near the light cone, the origin-regular solution defined by (24) has the representation
f(ξ)=fan(ξ)+C(1−ξ)ρh(ξ),h(1)=1,(25)
where fan and h are analytic at ξ=1. Comparing this with (21) using the transformation (22), we obtain, after restoring the branch labels,
Bm′(a±)=−A±β±(A±2β±)ρCm,±.(26)
Our task is to determine the sign of the connection coefficient C. To this end, we differentiate (25)m−1 times with respect to ξ. Recalling that ρ=m−23, we obtain
where the product is positive. Thus, if C=0, the derivative diverges with sign εmsgnC.
On the other hand, differentiating the Taylor series (24)m−1 times gives
f(m−1)(ξ)=n≥m−1∑(n−m+1)!n!fnξn−m+1.(28)
The analysis of the recurrence relation in Appendix C shows that all sufficiently large Taylor coefficients fn have the same sign σ∈{−1,1} and that f(m−1) diverges as ξ→1−. By (28), this divergence has sign σ, since the finitely many terms preceding the tail remain bounded. Comparing with (27), we conclude that C=0 and
sgnC=σεm.(29)
Since the prefactor in (26) is negative, we obtain
sgnBm′(a±)=−σεm.(30)
The appendix establishes σ=−1 on the plus branch for every m≥5 and on the minus branch for m=6,7,8, whereas σ=1 on the minus branch for m≥9. The resulting crossing signs are summarized below, together with the intervals containing an additional zero. For m≥9, the conclusion also uses (14) and (15); for m=5, it relies on the formal prediction (19).
m
sgnBm′(a+)
sgnBm′(a−)
interval containing a new zero
6,7,8
εm
εm
(a+,a−)
m≥9
εm
−εm
(a−,∞)
5
+1
—
(a+,∞)
Table 3.12.
For m≥6, these sign changes and continuity of Bm(a) yield the following existence result.
Theorem 3.1. For every integer m≥6 (spatial dimension d=2m), there are at least three nonconstant smooth self-similar profiles on [0,1], modulo reflection. Besides u±, there is a profile u⋆ with shooting parameter
a⋆∈(a+,a−)(m=6,7,8),a⋆∈(a−,∞)(m≥9).
Proof. For m=6,7,8, the derivatives of Bm at the two explicit zeros have the same nonzero sign, hence Bm(a++δ) and Bm(a−−δ) have opposite signs for sufficiently small δ>0. Continuity therefore gives an additional zero in (a+,a−).
For m≥9, we have sgnBm′(a−)=−εm, so Bm(a−+δ) has sign −εm for sufficiently small δ>0. On the other hand, (14) and (15) give sign εm for Bm(a) at a sufficiently large parameter. Continuity therefore gives an additional zero in (a−,∞).
In both cases, Theorem A.2 implies smoothness.
For m=5, the crossing sign B5′(a+)>0 gives B5(a++δ)>0 for sufficiently small δ>0, whereas (19) predicts that B5(a)<0 for large a. Taking this prediction as input, continuity forces an additional zero in (a+,∞). Theorem A.2 then gives a second nontrivial smooth profile. This conclusion remains formal because the error estimate needed to justify (19) has not been established.
Endpoint coordinates
We describe a shot near the light cone by two numbers: its limiting value c=u(1) and the amplitude B of the leading fractional-power term. Smooth profiles are precisely those for which B=0. We construct these endpoint coordinates and show that they depend analytically on the shooting parameter.
Local behaviour at the light cone
For a fixed integer m≥5, put
s=y1−1,v(s)=u(y),ρ=m−23.
The light cone corresponds to s=0, and the profile equation becomes
s(2+s)v′′−2(ρ−1)(1+s)v′+(2m−2)v(1−v2)=0.(31)
The indicial exponents are 0 and ρ.
We use the following theorem of Breitenlohner–Forgács–Maison (BFM) [2].
Theorem A.1 (BFM). The system
sX′=sf(s,X,Y),sY′=−ΛY+sg(s,X,Y),
where f and g are analytic and Λ is a positive constant, has a local solution near s=0 satisfying
X(0)=c,Y(0)=0.
The solution is analytic in s, c, and any additional parameters entering the equations analytically, and is unique among solutions continuous at s=0 with these limiting values.
We apply this theorem twice: first to construct the analytic solution with prescribed endpoint value c, and then to include the independent fractional-power correction.
The analytic solution. We first construct the solution analytic at s=0 with v(0)=c. To bring equation (A.1) into BFM form, we seek a polynomial
P(s,v)=j=0∑m−2sjpj(v)
which approximates the relation between v′ and v along analytic solutions. Substituting
This measures the failure of the relation v′=P(s,v) to satisfy the profile equation. We choose the coefficients pj(v) to cancel successive powers of s, so that R[P]=O(sm−1). Cancellation of the constant term gives
p0(v)=2m−52m−2v(1−v2),
while cancellation of the coefficient of s gives
2(2−ρ)p1+2p0p0′−2(ρ−1)p0=0,
which determines p1 from p0. More generally, suppose that p0,…,pj−1 have been chosen so that
The denominator is nonzero because ρ is a half-integer. Continuing through j=m−2, we obtain
R[P]=sm−1R(s,v),
with R analytic. Having constructed P, we introduce the remainder Y by
v′=P(s,v)+sm−2Y.
This brings (A.1) into BFM form:
sv′sY′=s(P(s,v)+sm−2Y),=−21Y+sG(s,v,Y),(32)
where
G(s,v,Y)=(2+sρ−1−Pv(s,v))Y−2+sR(s,v)
is analytic. Here the coefficient −21 follows from ρ−1−(m−2)=−21. Theorem A.1 gives a solution with v(0)=c and Y(0)=0, analytic in s and c. We denote the resulting analytic solution of Eq. (A.1) by v0(s;c). Since the Taylor recurrence for (A.1) has nonzero divisors 2n(n−ρ) for every integer n≥1, this is the unique analytic solution with endpoint value c.
The fractional-power family. To include the second free parameter, we write
v(s;c,B)=v0(s;c)+sρZ(σ;c,B),σ=s.(33)
In terms of σ, the factor sρ becomes the integer power σ2m−3. Substituting into (A.1) and writing W=dZ/dσ, we obtain
Since 2ρ=2m−3 is an integer, H is analytic. The BFM theorem gives a unique solution with
Z(0;c,B)=B,W(0;c,B)=0,
analytic in (σ,c,B). Thus c fixes the endpoint value, while B fixes the leading fractional-power correction. For B=0, uniqueness gives Z=W=0, and the solution reduces to v0(s;c).
To recover the decomposition used in the main text, split Z into its even and odd parts:
Z(σ;c,B)=Ze(σ2;c,B)+σZo(σ2;c,B).
Since ρ+21=m−1 is an integer, the odd part contributes only integer powers of s. Moreover, Ze vanishes at B=0, so it has an analytic factor B. Consequently, the endpoint family can be written as
where both functions van and G0 are analytic in s and the parameters. The analytic component need not be a solution, but equals v0 when B=0.
Returning to y, we obtain
u(y)=uan(y;c,B)+B(1−y)ρg(y;c,B),
where
uan(1;c,B)=c,g(1;c,B)=1.
Identification of origin-regular shots
We show that every origin-regular shot belongs to the family constructed above. Fix one such shot, suppress its parameter a, and let c be its endpoint value, which exists by [1]. Subtracting the equation for v0(s;c) from that for v shows that η=v−v0(s;c) satisfies the linear equation
We seek a solution η1=sρz with z(0)=1. Substitution gives
s(2+s)z′′+2(ρ+1+s)z′+[Q(s)−ρ(ρ−1)]z=0.
An elementary local existence argument gives such a solution with bounded z′, so η1(s)=sρ(1+O(s)). Reduction of order yields a second solution, linearly independent of η1,
These functions satisfy (A.4). The uniqueness assertion in Theorem A.1 therefore identifies the shot with the member of the family constructed above corresponding to (c,B).
Dependence on the shooting parameter
The endpoint value and fractional amplitude are genuine coordinates for the local solution, not merely formal expansion coefficients. To see this, fix a small s0>0 and consider
(c,B)⟼(v(s0;c,B),∂sv(s0;c,B)).
The endpoint expansion gives
∂cv=1+O(s),∂s∂cv=O(1),
and
∂Bv=sρ(1+O(s)),∂s∂Bv=ρsρ−1(1+O(s)).
Consequently,
det∂(c,B)∂(v,vs)=ρs0ρ−1(1+O(s0))=0.
The analytic inverse function theorem shows that (c,B) depend locally real analytically on the interior value and derivative. Since the origin-regular shot depends analytically on a at every interior point, we obtain the following result.
Theorem A.2. The endpoint parameters (c,B) depend locally real analytically on the value and derivative of the solution at a fixed interior point sufficiently close to y=1. In particular, for the origin-regular shooting family, c(a) and Bm(a) are locally real analytic functions of a>0. An origin-regular shot is smooth at y=1 if and only if Bm(a)=0. In that case, it is analytic there.
The smoothness criterion follows directly from the endpoint expansion. When B=0, the solution coincides with the analytic solution v0. When B=0, the term B(1−y)ρg(y), with g(1)=1, prevents smoothness because ρ is a half-integer.
Near the zero solution, linearizing the map from interior data to B gives the fractional-power coefficient of the corresponding linearized solution. This observation is used in Appendix B to transfer the outer approximation to the endpoint amplitude.
Matching the core to the outer solution
For m≥6, we justify the large-a approximation
u(y;a)∼Cma−k/2Fm(y)
and its consequence for the fractional-power coefficient,
Bm(a)∼2ρCmKma−k/2.
The main point is that the nonlinear core fixes the amplitude of the outer solution, while nonlinear effects outside the core produce only a smaller correction.
Matching in the overlap region. Choose a large number R and match at
yR=aR.
For large a, this point lies well inside the light cone but outside the core. Convergence to the core profile gives
u(yR;a)≃Q(R),yRuy(yR;a)≃RQ′(R).
Let
La(y)=αa,RFm(y)+βa,RGm(y)
be the solution of the linear outer equation with these same value and derivative at yR. Here
Fm(y)∼y−k,Gm(y)∼y−ℓ,ℓ=m−2+ν>k,
as y→0. Matching the two powers gives, for fixed R as a→∞,
ak/2αa,R⟶Cm,R:=Rkℓ−kℓQ(R)+RQ′(R),ak/2βa,R⟶0.
The core tail Q(R)∼CmR−k implies Cm,R→Cm as R→∞. Thus the core selects the outer mode Fm with amplitude Cma−k/2; the contribution of the faster mode Gm is negligible at fixed y>0.
Size of the nonlinear correction. At the matching point, the field is already small: u(yR;a)=O(R−k). The cubic term is therefore smaller than the linear term by a factor O(R−2k). To check that this small correction does not accumulate as we move outwards, write the equation in integral form:
u(y;a)−La(y)=∫yRyGm(y,s)u(s;a)3ds.
For yR≤s≤y≤y0, where y0>0 is fixed and small, the linear Green function satisfies
∣Gm(y,s)∣≤Cy−ksk−1.
Using the matched decay u(s;a)=O(a−k/2s−k) in this integral gives
∣u(y;a)−La(y)∣≤CR−2ka−k/2y−k.
For sufficiently large R, this estimate also makes the assumed decay self-consistent. The derivative is controlled in the same way at y=y0.
Hence the nonlinear correction can be made arbitrarily small relative to the leading outer field by choosing R large. Taking a→∞ first and then R→∞, we obtain
Away from the origin and the light cone, ordinary continuous dependence propagates this convergence. Thus
ak/2u(⋅;a)⟶CmFm
together with the first derivative on every closed subinterval of (0,1).
Transfer to the light cone. Convergence at interior points does not by itself determine the fractional-power coefficient at the singular endpoint. For this last step, we use the endpoint coordinates constructed in Appendix A.
Fix y1<1 sufficiently close to the light cone. Near the zero solution, the fractional-power coefficient is an analytic function B of the interior data (u(y1),uy(y1)). Its linear term gives the fractional-power coefficient of the corresponding solution of the linearized equation. Since
we need only the fractional-power coefficient of Fm. By (2.8), this is 2ρKm. Consequently,
Bm(a)=2ρCmKma−k/2+o(a−k/2),
which proves (2.9).
Taylor coefficient estimates
We establish the eventual coefficient signs and the divergence of f(m−1) used in Section 3. Throughout, ρ=m−23, and A denotes the parameter of the explicit branch under consideration.
Let gn denote either fn or −fn. It satisfies the same recurrence. To obtain lower bounds on gn/gn−1, choose a positive sequence rn and set
en=gn−rngn−1,En=Tn−rnCn−Dnrn+1.(38)
Then (C.1) gives the exact identity
Dnen+1=rnCnen+Engn.(39)
Suppose that for some N≥3,
gN−1>0,eN>0,En>0(n≥N).
Then gN=rNgN−1+eN>0, and (C.4) gives eN+1>0, since Cn,Dn>0 for n≥3. Induction therefore yields
gn>rngn−1>0(n≥N).(40)
Iterating this inequality gives
gn>gN−1j=N∏nrj.
If
rn=1−nα+O(n−2),
then taking logarithms and summing gives
logj=N∏nrj=−αlogn+O(1).
Consequently,
gn>cmn−α(n≥N)
for some cm>0. Thus it suffices to verify the starting inequalities gN−1>0 and eN>0, together with En>0 for n≥N. Here cm>0 is independent of n.
We treat the branches separately below. In each case, the comparison identity determines the eventual coefficient sign. For the negative tails, the resulting polynomial lower bound also proves endpoint divergence. For the positive tail, we first exclude analyticity at ξ=1 by comparing coefficient ratios, and then deduce divergence from the endpoint expansion.
Negative coefficient tails
The plus branch for m≥5. Set gn=−fn and choose
rn=n+m−25n−2,n≥3.
On this branch A>2m−3, so g2,g3>0. Moreover,
g2g3−r3=6A(m+3)(2m+1)18Am−21A+4m2+24m+11>0.
Thus the comparison starts at n=3.
A direct calculation gives
En=2(2n+2m−3)Smn+Qm,
where
Sm=A(10m−17)−16m+32,Qm=A(13−2m)−16m2+56m−48.
Using A>2m−3, we find
Sm>20m2−80m+83>0,3Sm+Qm>40m2−152m+162>0.
Hence
Smn+Qm=(n−3)Sm+(3Sm+Qm)>0(n≥3),
so En>0 for all n≥3. The comparison identity therefore gives
gn>rngn−1>0,n≥3.
Since
rn=1−nρ+1+O(n−2),
it follows that
−fn≥cmn−ρ−1(41)
for some cm>0 and all $n \ge 3.
The minus branch for m=6,7,8. Set gn=−fn and choose
rn=n−16+m−21n−16,n≥66.
The comparison starts later because f2,f3>0. For each m=6,7,8, we use exact arithmetic in Maple to verify the AI-suggested starting inequalities
g65>0,g65g66−r66>10001,(42)
so e66>0. The same calculation shows that
2(n−16)(2n+2m−31)En
is a polynomial in n−66 with strictly positive coefficients. Hence En>0 for all n≥66, and the comparison identity gives
gn>rngn−1>0,n≥66.
Since
rn=1−nρ+1+O(n−2),
we again obtain (C.6).
Nonvanishing of C and endpoint divergence. In all the negative-tail cases, the eventual coefficient sign is σ=−1. Moreover,
(n−m+1)!n!∼nm−1,m−2−ρ=−21.
It follows from (C.6) that
(n−m+1)!n!fn≤−cmn−1/2
for all sufficiently large n, with cm>0. Consequently, the tail of the differentiated Taylor series
f(m−1)(ξ)=n≥m−1∑(n−m+1)!n!fnξn−m+1
tends to mathord−∞ as ξ→1−. The finitely many preceding terms remain bounded, so
f(m−1)(ξ)⟶−∞as ξ→1−.
The endpoint expansion (3.8) therefore implies C=0.
The minus branch for m≥9
Set gn=fn, write
Rm=23(m−2)(m−1),A=34m−5−Rm,
and choose
rn=A(n+25)n+1(1+3n21),n≥3.(43)
We first prove positivity of the coefficients and then establish nonvanishing of C and endpoint divergence.
Starting the comparison. Here g2>0, and
g2g3−r3=594A(m+3)5(297A−50m−51).
Substituting the formula for A gives 297A−50m−51>0 for m≥9. Thus g3/g2>r3, initializing the comparison at n=3.
Positivity of the gap. We first establish the lower bound
A>L(m):=1322+234(m−9),m≥9.
Indeed, direct calculation gives A(9)=(31−442)/3>22/13 and
All coefficients of K are positive, as are J(3)=10036, J(4)=12453, J(5)=10848, and
J(6+j)=279j4+3162j3+9986j2+10404j+14149(j≥0).
Hence En>0 for every integer n≥3. The comparison identity therefore gives
gn>rngn−1>0,n≥3.
Since gn=fn, the eventual coefficient sign is σ=1. Moreover, the comparison product is asymptotic to a positive constant times A−nn−3/2, so
gn>cmA−nn−3/2n≥3,
where cm>0 is independent of n.
Nonvanishing of C and endpoint divergence. Returning to fn=gn, we have
fn>cmA−nn−3/2.
This bound does not by itself force divergence at ξ=1.
We prove that C=0 by contradiction. Suppose that C=0. Then f is analytic at ξ=1, so its Taylor series at 0 has radius at least A. The lower bound above shows that the radius is exactly A. Thus f is singular at ξ=A, where the local exponents are 0 and 21, and hence has a nonzero square-root component there. Standard coefficient asymptotics give a ratio expansion whose n−2 coefficient is determined by the recurrence:
fn−1fn=A1(1−2n3+4(A−1)n22m−4−3A+O(n−3)).
On the other hand,
rn=A1(1−2n3+12n249+O(n−3)).
The n−2 coefficient in the ratio expansion is strictly smaller than that in rn, since
1249−4(A−1)2m−4−3A=12(A−1)58A−6m−37>0,
where the last inequality follows from A>L(m). Thus fn/fn−1<rn for all sufficiently large n, contrary to the comparison inequality. Hence C=0.
By (27) and eventual positivity of the Taylor coefficients,
f(m−1)(ξ)⟶+∞asξ→1−.
This completes the coefficient and divergence estimates used in § 3.
Acknowledgements
AI assistance was used in Appendix C to explore recurrence relations, choose comparison ratios, initialize the induction arguments, and perform exact-arithmetic calculations. The author takes full responsibility for the mathematical content and final text.
References
[1]P. Bizoń, I. Głogić, and A. Wasserman, Three self-similar solutions of Yang–Mills equations in high odd dimensions, Phys. Rev. D 114, 064021 (2026).
[2]P. Breitenlohner, P. Forgács, and D. Maison, Static Spherically Symmetric Solutions of the Einstein–Yang–Mills Equations, Commun. Math. Phys. 163 (1994), 141–172.DOI
[3]P. Bizoń and R. Kluszczyński, Existence of a third self-similar solution of the Yang–Mills equation in high odd dimensions, arXiv:2610.36784 (2026).arxiv.org/abs/2610.36784
[4]NIST Digital Library of Mathematical Functions, https://dlmf.nist.gov/, Eq. (15.8.4).