editors: nobody
% =====================================================================
% Equality cases in the sharp one-dimensional matrix Lieb--Thirring
% ineuality: a complete, self-contained proof.
%
% This file is written to be read straight through. Each technical
% object is introduced with a plain sentence saying what it is for,
% then used. The two analytic machines (the "trace engine" in
% Section 4 and the "field/path engine" in Section 5) are proved in
% full, so the main theorem never relies on an unproved black box.
%
% Road map:
% Section 1 the problem, the answer, and the idea of the proof
% Section 2 notation and the four constants
% Section 3 two elementary lemmas (spectrum, Young ineuality)
% Section 4 the trace engine: an exact integrated deficit identity
% Section 5 the field and the connecting path (existence)
% Section 6 the finite-interval action estimate
% Section 7 the sharp ineuality itself
% Section 8 from equality to pointwise matrix relations
% Section 9 fixed channels, the profile, complex W, attainability
% =====================================================================
allowdisplaybreaks[2]
# Equality cases in the sharp one-dimensional[2pt]
matrix Lieb--Thirring ineuality[6pt]large a complete proof
maketitle
Abstract:
We classify all equality cases in the sharp one-dimensional Lieb--Thirring
ineuality, for every finite matrix size and every exponent
1/2<gamma<3/2. For a measurable Hermitian positive semidefinite
potential W with int_Rtr(W^{gamma+1/2})<∞, equality holds
exactly when, in one constant unitary basis, W is a direct sum of scalar
one-bound-state solitons and zero channels, with independent scales and
centres. Every step is proved here.
tableofcontents
% =====================================================================
# The problem, the answer, and the idea
% =====================================================================
## What is being estimated
Fix an integer mge1 (the number of internal ``channels''). A
potential is a measurable function W:RtoC^{mtimes m} whose
values are Hermitian and positive semidefinite. Think of -Delta acting
on vector-valued functions psi:RtoC^m, and of W as a position
dependent multiplication operator. The Schr"odinger operator
H_W=-(d^2)/(dx^2)-W
is unbounded below when W is large, but it has only finitely many
negative eigenvalues below any fixed level (Lemma~lem:spectral).
Write lambda_j(H_W) for the eigenvalues, counted with multiplicity, and
for 1/2<gamma<3/2 set
Tr(H_W)_-^gamma
=sum_{lambda_j(H_W)<0}|lambda_j(H_W)|^gamma .
Here tr is the unnormalised trace on C^{mtimes m}, and W^p is
defined by spectral calculus.
## The ineuality and its constant
The sharp Lieb--Thirring ineuality in one dimension is
Tr(H_W)_-^gamma
le C_gammaint_Rtr(W(x)^p),dx,
p=gamma+1/2,
with the explicit constant
C_gamma
=((gamma-1/2)/(gamma+1/2))^{gamma-1/2}
(Gamma(gamma+1))/(sqrtpi,Gamma(gamma+3/2)) .
The ineuality itself is due to Lieb and Thirring and, in this matrix
form, to the companion paper [cite]. The new content here is the
equality statement, proved together with the ineuality so that
both rest on the same estimates.
## The answer
Set
r=(1)/(gamma-1/2)>1 .
The extremal potentials are built from the single scalar profile
w(x)=(r+1)a^2,sech^2!(r,a,(x-x_0)),
a>0, x_0inR ,
which has exactly one negative eigenvalue, namely -a^2, and saturates
(eq:sharp) by itself.
[Equality classification]
Let mge1, 1/2<gamma<3/2, p=gamma+1/2, and let W be a
measurable Hermitian positive semidefinite potential with
int_Rtr(W^p)<∞. Then (eq:sharp) holds. Moreover,
equality holds if and only if there are an integer kin{0,...,m},
a constant unitary matrix U, positive numbers a_1,...,a_k, and real
centres x_1,...,x_k such that, for almost every x,
W(x)=U,diag(w_1(x),...,w_k(x),0,...,0),U^* ,
w_j(x)=(r+1)a_j^2,sech^2!(r a_j (x-x_j)).
The case k=0 means W=0. Each nonzero channel carries exactly one
negative eigenvalue -a_j^2, so every extremal potential has at most m
negative eigenvalues counted with multiplicity. The scales a_j and
centres x_j are independent, and coincident profiles are allowed.
## The idea of the proof
The proof has six moves. We state them now so the reader knows where
each later section is going.
[(1)] Linearise the spectrum. Approximate finitely many
negative eigenfunctions by smooth compactly supported trial functions and
arrange them as the rows of a matrix U(x). A lower triangular matrix
C keeps the rows in successive spans; this is what makes the energies
ell_i^2 appear in order. This is the standard trial-function step of
the ineuality.
[(2)] Build an action. For a diagonal matrix K of trial
root-energies, define the energy
E_K(A)=|A'|_{HS}^2+tr(K^2AA^T)-tr((AA^T)^q)). It is
the trial-function energy written in matrix language.
[(3)] Insert a matrix path. A symmetric path B(x) running
from K at the left end of the interval to -K at the right end turns
the action into a square plus a boundary term (complete the square):
[
E_K(A)=(boundary)+E_K(A)=(boundary)+int(|A'-BA|_{HS}^2-F(A,B)).
]
The boundary term produces exactly b_sigmasumell_i^{2gamma}.
[(4)] Keep the defect. Equality forces the square
|A'-BA|_{HS}^2 and an algebraic defect
Phi_K(A,B)=|[B,K^2]|_{op}^2+|K^2-B^2-(AA^T)^r|_{op}^2 to vanish.
The trace engine of Section~4 is what makes the coefficient of Phi_K
uniform in the number of rows. This uniformity is essential because the
number of negative eigenvalues is not known in advance to be finite.
[(5)] Pass to a limit. The number of rows may grow without
bound while the number m of columns stays fixed. The fixed column
count bounds the tail rows by the small energies, giving strong pointwise
convergence and exact algebraic relations in the limit.
[(6)] Solve the leftover ODE. The limit relations force the
channel subspaces to be constant in x. On each channel a matrix
Riccati equation reduces to a scalar equation whose only bounded
solutions are the profiles (eq:profile). This yields
Theorem~thm:main, and a direct factorization check proves that every
listed potential attains equality.
## Brief history
Lieb and Thirring introduced spectral moment ineualities for the kinetic
energy of fermions [cite]. The sharp
scalar endpoint gamma=1/2 is due to Hundertmark--Lieb--Thomas
[cite], extended to operator-valued potentials by
Hundertmark--Laptev--Weidl [cite]. The semiclassical constant is
sharp for gammage3/2; the relevant tools are the moment lifting of
Aizenman--Lieb [cite] and the operator-valued ineualities
of Laptev--Weidl [cite]. Benguria--Loss [cite]
gave a commutation proof at gamma=3/2. The one-state scalar profile
(eq:profile) goes back to Keller [cite]; its role in the
Lieb--Thirring problem is explained in cite[Section~4]{LiebThirring1976}.
At the endpoint gamma=3/2 the equality structure changes: Frank, Gontier
and Lewin [cite] show that several bound states can
share a single scalar channel (KdV multisolitons). The strict ineuality
r>1 used below is exactly what rules that out in the open interval.
% =====================================================================
# Notation and the four constants
% =====================================================================
Throughout,
sigma=gamma-1/2in(0,1),
p=1+sigma=gamma+1/2,
r=frac1sigma>1,
q=r+1 .
Note the identities
2sigma+1=2gamma, rp=q=1+r .
Two constants recur:
D_sigma=frac{sigma^sigma}{(1+sigma)^{1+sigma}},
b_sigma=int_{-1}^{1}(1-s^2)^sigma,ds .
They are linked to (eq:Cgamma) by the beta integral:
C_gamma=(D_sigma)/(b_sigma).
We use the operator norm |·|_{op} and the Hilbert--Schmidt norm
|A|_{HS}^2=tr(A^*A). For real matrices T also denotes transpose.
Sym_n is the space of real symmetric ntimes n matrices with inner
product tr(BH); the commutator is [B,H]=BH-HB. The identity matrix
or operator has the size forced by context.
% =====================================================================
# Two elementary lemmas
% =====================================================================
## The spectrum is discrete and the form is closed
The first lemma says we may freely talk about ``the'' negative eigenvalues,
even when W is only measurable.
[Spectral facts]
Let p>1 and let W:RtoC^{mtimes m} be measurable, Hermitian and
positive semidefinite, with int_Rtr(W^p)<∞. The quadratic form
h_W[psi]=int_R|psi'|^2-int_R<psi,Wpsi>,
psiin H^1(R;C^m),
is closed and bounded below. Multiplication by W^{1/2} is compact from
H^1 to L^2. The negative spectrum of H_W consists of isolated
eigenvalues of finite multiplicity, accumulating at most at 0. Every
eigenfunction lies in H^1 and solves -psi''-Wpsi=lambdapsi in
distributions.
Proof.
Put w=|W|_{op}. Then 0le w^pletr(W^p), so win L^p(R). The
one-dimensional Sobolev bound, applied to the vector x|->|psi(x)|,
reads |psi|_∞^2le 2|psi|_2|psi'|_2. Hence, by
H"older and Young,
int_R<psi,Wpsi>
le 2^{1/p}|w|_p|psi'|_2^{1/p}|psi|_2^{2-1/p}
le ε|psi'|_2^2
+C_ε|psi|_2^2 .
So a suitably shifted form norm is equivalent to the H^1 norm: the form
is closed and bounded below, and by polarization the potential term is
continuous on H^1.
Let Tpsi=W^{1/2}psi. For R,L>0 define the truncated multiplier
T_{R,L} by keeping only {|x|le R, wle L}. By the compact
embedding H^1([-R,R])↪ L^2([-R,R]), each T_{R,L} is
compact, and
[
|(T-T_{R,L})psi|_2^2
le C,bigl|w,1_bf1_{{|x|>R}cup{w>L}}bigr|_p
|psi|_{H^1}^2 .
]
The prefactor tends to zero as R,Lto∞, so T is a norm limit of
compact operators, hence compact.
Now fix a>0. If the spectral subspace of H_W in (-∞,-a] were
infinite dimensional, it would contain an orthonormal sequence bounded in
H^1 by (eq:formbound). A subsequence converges weakly to zero in
H^1, so its T-images converge strongly to zero; but for unit vectors
in that subspace h_W[psi]ge-age-|Tpsi|_2^2, forcing
|Tpsi|_2^2ge a, a contradiction. Thus every such subspace is finite
dimensional.
Finally test the weak form equation against smooth compactly supported
functions: this gives the stated distributional equation. Its potential
term is locally integrable because Win L^p_{loc} and psi is
locally bounded.
## Young's ineuality for matrices
This is the scalar ineuality yz-z^qle D_sigma y^p promoted to
matrices, together with its equality condition.
[Trace Young ineuality]
For positive semidefinite matrices W,Z of the same finite size, put
Y(W,Z)=D_sigmatr(W^p)-tr(WZ)+tr(Z^q).
Then Y(W,Z)ge0, with equality if and only if W=qZ^r. Moreover
Y(W,Z)ge1/2tr(Z^q)
-(2^{p-1}-1)D_sigmatr(W^p).
Proof.
Maximise z|-> yz-z^q over zge0 for fixed yge0. The maximum
D_sigma y^p is attained exactly at y=qz^r; this is the scalar Young
ineuality behind D_sigma in (eq:Dconst). Choose an orthonormal
eigenbasis (v_j) of Z with eigenvalues z_j, and set
y_j=< v_j,Wv_j>. By (eq:youngdefect) and the scalar
ineuality,
[
tr(WZ)-tr(Z^q)
=sum_j (y_j z_j - z_j^q)
le D_sigmasum_j y_j^p .
]
Because t^p is strictly convex on [0,∞),
sum_j y_j^plesum_j< v_j,W^p v_j>=tr(W^p).
Thus Y(W,Z)ge0. Equality in the convexity step forces every v_j to
lie in an eigenspace of W; equality in the scalar step then forces
Wv_j=qz_j^r v_j for every j, i.e. W=qZ^r. This covers zero
eigenvalues as well.
For (eq:youngcoercive), apply the already proven Yge0 to the pair
(2W,Z) and expand:
tr(2WZ)le1/2tr(Z^q)+2^{p-1}D_sigmatr(W^p), which rearranges
to (eq:youngcoercive).
% =====================================================================
# The trace engine: an exact integrated deficit
% =====================================================================
The action estimate compares a matrix N against its power N^q. The
classical ineuality is tr(NV)getr(N^q) under a symbol condition, but
equality analysis needs the exact size of the gap. The engine below gives
that gap as an integral of two manifestly nonnegative squares. The point
is that it is an identity, not just an ineuality, so a vanishing total
gap forces each square to vanish.
Throughout this section 0<sigma<1, beta=sigma-1/2, and
p=1+sigma, r=1/sigma, q=1+r as in (eq:params).
## The normalisation constant
Define c_sigma>0 by
c_sigma^{-1}
=int_0^∞ t^beta
(t^{-1/2}-(t+1)^{-1/2}),dt .
The integrand is O(t^{sigma-1}) near 0 and O(t^{sigma-2}) near
∞, so the integral converges (this uses 0<sigma<1). A change of
variable gives the scaling identity we will use constantly:
c_sigmaint_0^∞ t^beta
(t^{-1/2}-(t+y)^{-1/2}),dt=y^sigma,
yge0 .
Indeed, substitute t=yu; the left side becomes
c_sigma y^{beta+1/2}int_0^∞ u^beta(u^{-1/2}-(u+1)^{-1/2}),du
=y^sigma by (eq:trace:normalization) and beta+1/2=sigma.
## Statement of the engine
[Exact integrated trace deficit]
Let nge1 and let B,VinC^{ntimes n} be Hermitian. Suppose
Y_s=(sI-B)^2, X_s=Y_s+Vge bIquad(sinR)
for some b>0. Define
R(t)=frac1piint_R(X_s+tI)^{-1},ds,
P(t)=t^{-1/2}I-R(t), t>0,
and
N=c_sigmaint_0^∞ t^beta P(t),dt .
Then these integrals converge absolutely in matrix norm, R(t)>0, and
there are continuous Hermitian matrices L(t)=R(t)^{-1}>0 and Z(t)
with
V+tI=L^2+Z^2+i[B,Z],
LZ+ZL+i[B,L]=0 .
If Nge0, then
tr(NV)-tr(N^q)
=c_sigmaint_0^∞ t^beta(
&tr(RZ^2)
&+qbigl|R^{-1/2}-R^{1/2}(tI+N^r)^{1/2}bigr|_{HS}^2
),dt .
In particular tr(NV)getr(N^q), and equality holds if and only if both
integrands vanish for every t>0.
The hypothesis concerns the full symbol X_s; the perturbation V itself
need not be positive. The proof has three parts: recover L,Z from
half-line boundary values, evaluate a logarithmic determinant integral,
and complete a matrix square.
## Step 1: half-line boundary maps
Quadratic growth of X_s together with (eq:trace:symbol) gives an
integrable bound for X_s^{-1}, hence R(t)>0 and R(t)=O(1) as
t↓0. Diagonalising B gives the free identity
frac1piint_R(Y_s+tI)^{-1},ds=t^{-1/2}I. For tge1 both
X_s+tI and Y_s+tI dominate c(s^2+t)I with c>0 independent of
s,t; the resolvent identity bounds the difference of inverses by
C(s^2+t)^{-2}. Consequently
|P(t)|_{op}=
O(t^{-1/2}), & t↓0,
O(t^{-3/2}), & tto∞,
which proves convergence of N in (eq:trace:N).
Fix t>0 and write nabla=∂_x-iB. Consider
(-nabla^2+V+tI)psi=0 .
For each vinC^n there is a unique H^1 solution on either half-line
with psi(0)=v. The sesquilinear form
a_t(phi,psi)=int(<nablaphi,nablapsi>
+<phi,(V+tI)psi>),dx is coercive on H^1(R;C^n), since
its Fourier symbol X_s+tI dominates c_t(1+s^2)I; zero extension gives
coercivity on the trace-zero subspaces of each half-line, and the Riesz
representation theorem supplies the unique weak solution. The equation
gives psi''in L^2, so psiin H^2, and both psi and psi'
vanish at the infinite end.
Define the right and left boundary matrices S,T by
nablapsi(0+)=-Sv, nablapsi(0-)=Tv .
Integrating by parts gives
a_t(psi_v,psi_w)=< v,Sw> on the right and
=< v,Tw> on the left, so S,T are Hermitian. Translation
and uniqueness show nablapsi=-Spsi throughout the right half-line
and nablapsi=Tpsi throughout the left half-line. Substituting into
(eq:trace:ode) with arbitrary boundary data,
V+tI=S^2+i[B,S]=T^2-i[B,T] .
Glue the two half-line solutions with the same boundary value v. The
result is continuous with derivative jump -(S+T)v, so
(-nabla^2+V+tI)psi=delta_0(S+T)v. Fourier inversion with kernel
e^{-isx} gives
[
v=frac1{2pi}int_R(X_s+tI)^{-1},ds,(S+T)v
=1/2R(t)(S+T)v .
]
The inverse symbol is O(s^{-2}), so its Fourier inverse is continuous
and evaluation at zero is legitimate. Therefore
L:=1/2(S+T)=R^{-1}>0,
Z:=1/2(S-T)
satisfy (eq:trace:LZ): adding and subtracting the two halves of
(eq:trace:riccati) gives the first equation, and the second holds
because LZ+ZL=1/2(S^2-T^2) and i[B,L]=tfrac i2[B,S+T]. The
equation LZ+ZL+i[B,L]=0 determines Z uniquely from L>0 (in an
eigenbasis of L the map Z|-> LZ+ZL multiplies the (j,k) entry
by L_{jj}+L_{kk}>0), which also proves continuity of Z(t).
Two consequences will be used. First, the diagonal of the second
relation gives tr Z=0. Second, multiplying i[R,B]=-(ZR+RZ) (the
R-conjugate of the second relation) into the trace of R(V+tI) gives
the signed identity
tr(R(V+tI))=tr L-tr(RZ^2),
where the sign comes from i,tr(R[B,Z])=-2tr(RZ^2).
## Step 2: a logarithmic determinant integral
Define
F(t)=(tr V)/(sqrt t)
-frac1piint_R
log(det(X_s+tI))/(det(Y_s+tI)),ds .
Both determinants are positive and the logarithmic ratio is
O(|s|^{-2}) at infinity; its t-derivative is the trace of a resolvent
difference with an integrable bound, locally uniform in t>0. Hence F
is continuously differentiable. The first relation in
(eq:trace:riccati) factors the symbol,
[
X_s+tI=(sI-B+iS)(sI-B-iS).
]
If u_j+iv_j are the eigenvalues of B+iS, counted with algebraic
multiplicity, then the square-integrability argument behind
(eq:trace:ode) forces v_j>0, and
det(X_s+tI)=prod_j((s-u_j)^2+v_j^2). The denominator has the same
form with centres the eigenvalues of B and widths sqrt t.
Translations of individual logarithms only change symmetric-cutoff
integrals by o(1) (the derivative
(d)/(da)int_{-M}^M h(s-a),ds=h(M+a)-h(M-a)to0 uniformly for
bounded a). Using the one-factor integral
int_Rlog(s^2+v^2)/(s^2+w^2),ds=2pi(v-w) for v,w>0, and
sum_jv_j=tr S=tr L, we obtain
F(t)=(tr V)/(sqrt t)-2tr(L-sqrt t I).
Set
G(t)=tr(L-2sqrt t I+tR).
Both F,Gge0: for G apply lambda-2sqrt t+t/lambdage0 to the
positive eigenvalues lambda of L; for F use
(eq:trace:Feval) with tr V=tr(L^2+Z^2)-nt (the trace of the
first relation in (eq:trace:LZ)) to get
F(t)=t^{-1/2}tr((L-sqrt t I)^2+Z^2)ge0. Combining
(eq:trace:RV) and (eq:trace:Feval) yields the identity
tr(PV)=F+G+tr(RZ^2),
with every term on the right nonnegative. By (eq:trace:Pbounds),
F, G, and tr(RZ^2) are integrable against t^beta dt, and
t^{beta+1}F(t) vanishes at both ends. Differentiating
(eq:trace:F) and using (eq:trace:Feval)--(eq:trace:G),
[
F'(t)=-frac{tr V}{2t^{3/2}}+tr P(t),
tF'(t)=-1/2F(t)-G(t).
]
Integration by parts with the vanishing endpoint limits gives
int_0^∞ t^beta G,dt=sigmaint_0^∞ t^beta F,dt; since
sigma=1/r and int F=rint G,
tr(NV)=q,c_sigmaint_0^∞ t^beta G(t),dt
+c_sigmaint_0^∞ t^betatr(RZ^2),dt .
## Step 3: the noncommutative square
For a positive semidefinite matrix D, set
[
F_0(t,D)=(tr D)/(sqrt t)
-2tr((tI+D)^{1/2}-sqrt t I).
]
For a scalar eigenvalue dge0 the summand is
int_0^d(t^{-1/2}-(t+y)^{-1/2}),dy; Tonelli's theorem and
(eq:trace:scaling) give
c_sigmaint_0^∞ t^beta F_0(t,D),dt=frac1ptr(D^p).
Now expand the Hilbert--Schmidt square without any commutation (the cross
factors R^{-1/2}R^{1/2} cancel, and cyclicity handles the rest):
D_D(t)
&=bigl|R^{-1/2}-R^{1/2}(tI+D)^{1/2}bigr|_{HS}^2 notag
&=tr(R^{-1}+R(tI+D)-2(tI+D)^{1/2}) notag
&=G(t)-tr(P(t)D)+F_0(t,D).
All three terms are integrable by (eq:trace:Pbounds),
(eq:trace:FG), and (eq:trace:F0). Integrating
(eq:trace:square) with weight t^beta and using
c_sigmaint t^beta P= N gives
c_sigmaint_0^∞ t^beta G,dt
=tr(ND)-frac1ptr(D^p)
+c_sigmaint_0^∞ t^betaD_D,dt .
When Nge0 choose D=N^r. Since rp=q, the first two terms equal
(1-1/p)tr(N^q)=tr(N^q)/q. Substituting into
(eq:trace:weighted) proves (eq:trace:deficit). Both integrands
in (eq:trace:deficit) are nonnegative, so equality forces them to
vanish for almost every t, hence by continuity for every t>0.
Finally, if B and V commute the symbol condition implies Vge bI;
simultaneous diagonalisation gives R(t)=(tI+V)^{-1/2}, N=V^sigma,
Z=0, and both remainders vanish, as scalar equality requires.
% =====================================================================
# The field and the connecting path
% =====================================================================
To run the complete-the-square step of the action estimate we need two
objects: a symmetric matrix field M_delta that mimics the power
(K^2-B^2)^sigma near values that matter, and a symmetric path B(x)
running from K to -K that solves an ODE of the form
ε B'=M_delta(B)-AA^T. This section constructs both. Set
delta>0 and recall beta=sigma-1/2, and let c_sigma be the
constant of (eq:trace:normalization).
## A scalar building block
On yge0, define
f_delta(y)=(c_sigma)/(pi)int_0^∞ t^beta
(frac1{t+delta}-frac1{t+delta+y}),dt,
and extend to y<0 by the tangent line f_delta'(0)y. For yge0,
f_delta'(y)=a_sigma(y+delta)^{beta-1},
a_sigma=(c_sigma)/(pi)int_0^∞
(u^beta)/((1+u)^2),du>0 .
Both integrals converge because -1/2<beta<1/2. Thus f_delta
is C^1, strictly increasing, and concave on yge0; the tangent
extension makes it C^1 on all of R and lets the functional calculus
below accept arguments with negative eigenvalues.
Set
Q_s(B)=s^2I-2sB+K^2,
and
M_delta(B)=lim_{Rto∞}int_{-R}^{R}
(f_delta(Q_s(B))-f_delta(s^2)I),ds,
BinSym_n,
where the symmetric cutoff cancels the leading odd term
-2sf_delta'(s^2)B. The scalar counterpart is
mu_delta(z)=int_R(f_delta(s^2+z)-f_delta(s^2)),ds,
zinR .
[Regularised field]
The limit (eq:field:M) exists locally uniformly, and
M_delta:Sym_ntoSym_n is locally Lipschitz with symmetric values.
There is a C^1 function J_delta:Sym_ntoR, J_delta(0)=0, with
dJ_delta(B)[H]=-tr(M_delta(B)H).
The scalar function mu_delta is locally Lipschitz and strictly
increasing with mu_delta(0)=0, and for zge0
mu_delta(z)=(z+delta)^sigma-delta^sigma .
For every real unit vector e and every standard basis vector e_i,
e^{T}
&lemu_delta(e^{T}-(e^{T},
Be_i=ye_iquad&⇒
M_delta(B)e_i=mu_delta(k_i^2-y^2),e_i .
For every diagonal sign matrix S,
M_delta(SBS)=S,M_delta(B),S .
Finally,
J_delta(-K)-J_delta(K)
=sum_{i=1}^nint_{-k_i}^{k_i}mu_delta(k_i^2-s^2),ds
;→;
b_sigmasum_{i=1}^n k_i^{2sigma+1}
(delta↓0).
Proof.
Matrix calculus bound. If h is Lipschitz on an interval
containing the spectra of Hermitian X,Y, then
|h(X)-h(Y)|_{HS}leLip(h),|X-Y|_{HS}.
Indeed, take eigenbases e_i of X and v_j of Y; the mixed entries
of h(X)-h(Y) and of X-Y are (h(lambda_i)-h(nu_j))< e_i,v_j>
and (lambda_i-nu_j)< e_i,v_j>; square the scalar Lipschitz
ineuality and sum.
Convergence of the field. On a bounded set of B's the spectrum
of E_s(B)=-2sB+K^2 lies in [-a|s|,a|s|] for all large |s|. For
h_s(y)=f_delta(s^2+y)-f_delta(s^2)-f_delta'(s^2)y,
(eq:field:fderivative) gives
sup_{|y|le a|s|}|h_s'(y)|=O(|s|^{2beta-3}). By (eq:field:HS),
[
f_delta(Q_s(B))-f_delta(s^2)I=-2s f_delta'(s^2)B+T_s(B),
]
where |T_s(B)|_{HS} and the local Lipschitz constant of T_s are
O(|s|^{2beta-2}) (include f_delta'(s^2)K^2 in T_s for the size
bound, and apply (eq:field:HS) to h_s(E_s(B)) using
E_s(B_1)-E_s(B_2)=-2s(B_1-B_2)). Since 2beta-2<-1, these remainders
are integrable at infinity; the odd term cancels at each symmetric cutoff,
and on bounded s-intervals (eq:field:HS) applies directly to
f_delta. This gives the stated convergence and regularity.
Primitive. Let g_delta'=f_delta. Trace differentiation gives
d,tr(g_delta(X))[H]=tr(f_delta(X)H) (true for polynomials by
cyclicity, then for g_delta by C^1 approximation on compact spectral
intervals). For sne0 the function
[
B|->(tr(g_delta(Q_s(B))-g_delta(K^2)))/(2s)
+f_delta(s^2)tr B
]
has differential -tr((f_delta(Q_s(B))-f_delta(s^2)I)H); at
s=0 the differential is constant in B and also has a primitive. Hence
the negative of each finite-cutoff field has vanishing integral around
every closed piecewise smooth curve in Sym_n, i.e. it is a gradient.
Local uniform convergence passes this to -M_delta, and path integration
from 0 defines J_delta, proving (eq:field:primitive).
Scalar bounds and profile. For bounded z, the integrand in
(eq:field:mu) and its Lipschitz constant in z are O(|s|^{2beta-2})
at infinity, so mu_delta is locally Lipschitz and absolutely
convergent; strict increase and mu_delta(0)=0 follow from f_delta.
For zge0, Tonelli applied to (eq:field:f) gives
[
mu_delta(z)=c_sigmaint_0^∞ t^beta
((t+delta)^{-1/2}-(t+delta+z)^{-1/2}),dt,
]
which the scaling identity (eq:trace:scaling) evaluates as
(eq:field:mupositive).
Direction and coordinate. For fixed a,z, translations under a
symmetric cutoff are legitimate:
lim_{Rto∞}int_{-R}^{R}
(f_delta((s-a)^2+z)-f_delta(s^2)),ds=mu_delta(z),
because F_z(s)=f_delta(s^2+z) is even and F_z'(s)=O(|s|^{2beta-1})to0,
so the difference of shifted and unshifted cutoffs is bounded by
|a|^2sup_{R-|a|le ule R+|a|}|F_z'(u)|to0. Concavity of f_delta
in an eigenbasis of Q_s(B) gives
e^{T} f_delta(e^{T}, whose
scalar argument is (s-e^{T}+e^{T}-(e^{T}
Integrating with (eq:field:shift) gives
(eq:field:direction); if Be_i=ye_i, the functional calculus is
exact on e_i, giving (eq:field:coordinate). Since SK^2S=K^2,
conjugation inside the integral gives (eq:field:equivariance).
Endpoints. On diagonal arguments, (eq:field:primitive) and
(eq:field:coordinate) read
dJ_delta(diag(b_1,...,b_n))=-sum_imu_delta(k_i^2-b_i^2),db_i.
Integrating from K to -K gives the first equality in
(eq:field:endpoints). For zge0, 0lemu_delta(z)le z^sigma
and mu_delta(z)to z^sigma; dominated convergence and the substitution
s=k_iu give int_{-k_i}^{k_i}(k_i^2-s^2)^sigma ds=b_sigma k_i^{2sigma+1},
hence the limit.
## Positive values of the field
The next proposition is the only place where the sign of an argument of
M_delta matters. It says that if M_delta(B) is itself a positive
matrix (as it will be once it equals AA^{T}, the whole symbol
Q_s(B) is positive definite and the trace engine applies.
[Positive values of the field]
If M=M_delta(B)ge0, then
Q_s(B)ge0 (sinR),
0le Mlekappa^{2sigma}I .
Set V=K^2-B^2+delta I. Then (sI-B)^2+V=Q_s(B)+delta Igedelta I,
and with R(t) and N as in (eq:trace:resolvents)--(eq:trace:N)
one has N=M+delta^sigma I. In particular the hypotheses of
Theorem~thm:trace hold with this B and V.
Proof.
By (eq:field:direction) and strict increase of mu_delta,
positivity of M forces e^{T}{T} for every
unit e. Completing the scalar square,
[
e^{T}-e^{T}
+e^{T}-(e^{T},
]
so Q_s(B)ge0 as a Hermitian matrix on C^n. Also
Mlemu_delta(kappa^2)Ilekappa^{2sigma}I by (eq:field:direction)
and (eq:field:mupositive).
Write Y_s=(sI-B)^2. By diagonalisation of B and (eq:field:shift),
the scalar subtraction f_delta(s^2)I in (eq:field:M) may be
replaced by f_delta(Y_s). Since Q_s(B) and Y_s are both positive,
(eq:field:f) gives their difference under f_delta as a Stieltjes
integral
[
(c_sigma)/(pi)int_0^∞ t^beta
((Y_s+(t+delta)I)^{-1}-(Q_s(B)+(t+delta)I)^{-1}),dt .
]
For fixed B,delta both inverted matrices are bounded below by
c(1+t+s^2)I, and their difference is K^2-B^2, so the resolvent
identity bounds the integrand by Ct^beta(1+t+s^2)^{-2}; integrating in
s gives C't^beta(1+t)^{-3/2}, integrable on (0,∞). Fubini
therefore applies and yields
[
M=c_sigmaint_0^∞ t^beta
((t+delta)^{-1/2}I-R(t)),dt .
]
Subtracting this from (eq:trace:N) and using (eq:trace:scaling)
gives N=M+delta^sigma I.
## The connecting path
The final object of this section walks B from K to -K while tracking a
prescribed density AA^{T} This is where topology enters: a
nonlinear equation for C must be solved, and the only obstruction would
be a parity count of zeros.
[Connecting path]
Let Ω=(x_-,x_+) be bounded and let Uin H^1_0(Ω;R^{ntimes d}),
n,dge1, satisfy int_Ω UU^{T},dx=I_n. For every
delta,ε>0 there are a real lower triangular matrix
C_εinR^{ntimes n} and a path
B_εin C^1(overlineoverlineΩ;Ω;Sym_n) such that, with
A_ε=C_ε U,
ε B_ε'
=M_delta(B_ε)-A_ε A_ε^{T},
B_ε(x_-)=K, B_ε(x_+)=-K,
and
sup_{xinoverlineoverlineΩ}Ω|B_ε(x)|_{op}lekappa,
sup_{0<εle1}|C_ε|_{HS}<∞,
the second bound for fixed delta,K,Ω.
Proof.
A cutoff and a homotopy. Use the continuous representative of U.
Choose a smooth nonnegative compactly supported cutoff chi on Sym_n
equal to 1 on {B:|B|_{op}lekappa} and invariant under conjugation
by diagonal sign matrices (averaging over the finite sign group produces
one). The truncated field Mhat=M=chi M_delta is bounded and
globally Lipschitz. Fix ε>0. For C lower triangular and
thetain[0,1], the equation
ε B'=thetaMhat(B)-(CU)(CU)^M(B)-(CU)(CU)^{T},
B(x_-)=K,
has a unique C^1 solution on overlineoverlineΩ;Ω; define the continuous
endpoint map Phi(C,theta)=B(x_+)+KinSym_n.
Sign symmetry. Let G be the group of diagonal sign
matrices, acting by C|-> SC on lower triangular matrices and by
conjugation on Sym_n. Equivariance of the field and uniqueness of the
ODE give
Phi(SC,theta)=S,Phi(C,theta),S .
Three facts about the zero set.
(i) Compactness. At a zero, integrating (eq:path:homotopy)
gives CC^{T}ε K+thetaint_ΩMhat(B(x))M(B(x)),dx,
so the zeros lie in a bounded set and their joint set in
R^{ntimes n}_{low}times[0,1] is compact. For
0<εle1 this also gives the uniform bound
|C|_{HS}^2le2tr K
+n|Ω|sup_{BinSym_n}|Mhat(B)M(B)|_{op}.
(ii) Freeness. No row of C can vanish at a zero. Otherwise the
reflection changing only row i would fix C; by (eq:path:equivariance)
and ODE uniqueness Be_i=ye_i throughout, and (eq:field:coordinate)
gives ε y'=thetachi(B(x))mu_delta(k_i^2-y^2) with y(x_-)=k_i.
Uniqueness forces yequiv k_i, contradicting y(x_+)=-k_i. Thus
G acts freely on the zero set.
(iii) One regular orbit at theta=0. There
Phi(C,0)=2K-ε^{-1}CC^{T} Its lower triangular zeros
are exactly C=diag(d_1,...,d_n) with d_i=pm√(2ε k_i):
the first row of CC^{T}ε K fixes |d_1| and forces the
rest of the first column to vanish, and induction proceeds. These 2^n
zeros form one sign orbit. At each, the derivative in a lower triangular
direction H has diagonal entries -2ε^{-1}d_iH_{ii} and lower
off-diagonal entries -ε^{-1}d_jH_{ij} (i>j), an isomorphism
R^{ntimes n}_{low}toSym_n.
Continuation by parity. Suppose there were no zero at theta=1.
Compactness and freeness give a bounded invariant open set
OsubsetR^{ntimes n}_{low} containing the projection of every
zero, with closure where all rows are nonzero, and a positive lower bound
for |Phi|_{HS} on (∂ Otimes[0,1])cup(overlineoverline OOtimes{1}).
Reparametrise theta to agree with the initial map on an initial collar
and retain this gap; extend F constantly past both ends, smooth by
convolution, splice to the initial map on a smaller collar, and average in
the form Psi(C,theta)=frac1{|G|}G|sum_S SPsi_0(SC,theta)S to
get a smooth equivariant approximation Psi still equal to the initial map
on the collar and close enough to preserve the gap.
Freeness allows equivariant perturbations with arbitrary target directions:
at C_0inoverlineoverline OO choose a ball whose sign translates are disjoint and
a bump b supported there with b(C_0)=1; then
V_H(C)=sum_S b(SC)SHS is equivariant with V_H(C_0)=H. Finitely many
such maps V_1,...,V_N span Sym_n at every point. Let hge0 be
smooth, vanish near 0, and positive off the collar, and set
Psi_alpha=Psi+h(theta)sum_jalpha_jV_j. For small alpha the gap
persists; at a zero with 0<theta<1 the full differential in
(C,theta,alpha) is onto (parameter directions span the target where
h>0; the C-differential is the above isomorphism where h=0). Hence
the interior zero set is a smooth manifold, and Sard's theorem
[cite] applied to projection onto the parameter ball gives
arbitrarily small alpha for which 0 is a regular value.
For such alpha the zero set in Otimes[0,1] is a compact smooth
one-dimensional manifold with boundary. The gap excludes boundary zeros
except at theta=0, where the initial collar gives exactly the 2^n
regular zeros. The free action has a quotient that is again a compact
one-dimensional manifold with boundary (a neighbourhood disjoint from its
nontrivial translates gives a local interval or half-interval chart); all
2^n boundary points form one orbit, so the quotient has exactly one
boundary point. A compact one-dimensional manifold is a finite union of
circles and closed intervals, which cannot have exactly one boundary point.
Contradiction. Hence Phi(·,1) has a zero.
Removing the cutoff. For the resulting path and any real unit e,
put y=e^{T} Whenever |y|>kappa, (eq:field:direction)
gives e^{T}{T}-y^2)<0.
Since chige0 and (CU)(CU)^{T}, (eq:path:ode) implies
y'le0 there. Both endpoints lie in [-kappa,kappa]; an excursion
above kappa cannot start from kappa while y is nonincreasing, and an
excursion below -kappa cannot return to -kappa while y is
nonincreasing. So |B(x)|_{op}lekappa for all x, which forces
chi(B)=1 and hence (eq:path:ode) with the untruncated field.
Equation~(eq:path:Cbound) supplies the remaining bound in
(eq:path:bounds).
% =====================================================================
# The finite-interval action estimate
% =====================================================================
This is the central estimate. It turns the trial functions into a
quantitative lower bound that measures, in operator norm, how far the
data are from the two algebraic equality relations. The uniformity of the
coefficient c_* in the row count is what will later allow an
a-priori-infinite eigenvalue list.
Fix positive integers n,d, a bounded interval Ω=(x_-,x_+), and a
diagonal K=diag(ell_1,...,ell_n) with 0<ell_ilekappa. For
Ain H^1_0(Ω;R^{ntimes d}) set rho=AA^{T} and
E_K(A)=Ω
(|A'|_{HS}^2+tr(K^2rho)-tr(rho^q)),dx .
For pointwise values AinR^{ntimes d} and BinSym_n, define the
algebraic defect
Phi_K(A,B)=minBigl{1,;
|[B,K^2]|_{op}^2+bigl|K^2-B^2-(AA^{T}|_{op}^2
Bigr}.
Thus Phi_K=0 encodes the two relations [B,K^2]=0 and
rho=(K^2-B^2)^sigma, while A'-BA=0 is the first-order relation.
[Quantitative action estimate]
Let 0<sigma<1 and kappa>0. There is c_*=c_*(sigma,kappa)>0 with
the following property. For n,d,Ω,K as above, suppose
Uin H^1_0(Ω;R^{ntimes d}) satisfies
int_Ω UU^{T},dx=I_n. For every eta>0 there exist a
real lower triangular CinR^{ntimes n} and a path
Bin C^1(overlineoverlineΩ;Ω;Sym_n) with
B(x_-)=K, B(x_+)=-K,
sup_{xinoverlineoverlineΩ}Ω|B(x)|_{op}lekappa,
such that, for A=CU,
E_K(A) b_sigmasum_{i=1}^nell_i^{2gamma}
+int_Ω(|A'-BA|_{HS}^2+c_*Phi_K(A,B)),dx-eta .
The constant c_* is independent of n,d,Ω,U and of
min_iell_i.
The proof has two ingredients. The first is a constraint lemma, obtained
by feeding the trace engine (Proposition~prop:field-pos plus
Theorem~thm:trace) into a penalty. The second completes the square
along the path of Proposition~prop:path; the boundary term there is
exactly J_delta(-K)-J_delta(K), which tends to
b_sigmasumell_i^{2gamma}. We prove the constraint lemma first.
## The nonlinear term is controlled by the defect
For 0<deltale1, recall the field M_delta of Section~sec:field
and the function F(A,B)=tr(rho^q)-tr((K^2-B^2)rho),
rho=AA^{T}
[Constraint estimate]
The constant c_*(sigma,kappa)>0 can be chosen so that whenever
0<deltale1, |B|_{op}lekappa, and M_delta(B)=AA^{T},
F(A,B)+c_*Phi_K(A,B)le E_{n,delta},
E_{n,delta}
=n(deltakappa^{2sigma}+delta^sigma(kappa^2+delta))
+delta^{2sigma}.
Proof.
Write rho=M_delta(B)ge0. By Proposition~prop:field-pos,
Q_s:=s^2I-2sB+K^2ge0 (sinR),
0lerholekappa^{2sigma}I .
Put V=K^2-B^2+delta I. With R(t), N, c_sigma from
(eq:trace:resolvents)--(eq:trace:N), Proposition~prop:field-pos
also gives N=rho+delta^sigma I. Since
(sI-B)^2+V=Q_s+delta Igedelta I and Nge0,
Theorem~thm:trace applies and supplies L(t)=R(t)^{-1}>0 and
Hermitian Z(t) with
V+tI=L^2+Z^2+i[B,Z], LZ+ZL+i[B,L]=0,
and the exact identity
Delta:=tr(NV)-tr(N^q)
=c_sigmaint_0^∞ t^beta(tr(RZ^2)
+qbigl|R^{-1/2}-R^{1/2}(tI+N^r)^{1/2}bigr|_{HS}^2),dtge0 .
Here the matrices may be complex Hermitian although A,B are real; i is
the imaginary unit.
A dimension-free bound from [1,2]. On 1le tle2 one has
a_kappa Ile R(t)le A_kappa I:
for the upper bound use Q_sge0 on a bounded interval and
Q_sge1/2s^2I for |s|gemax{1,4kappa}; for the lower bound
restrict to |s|le1 where Q_s+(delta+t)Ile(4+2kappa+kappa^2)I.
These constants depend on kappa alone. Also
|N|_{op}lekappa^{2sigma}+1, so H_t=(tI+N^r)^{1/2} and L(t) are
bounded in operator norm by constants depending only on sigma,kappa.
The two integrands control |Z(t)|_{HS}^2 and |L(t)-H_t|_{HS}^2
(use tr(RZ^2)=tr(ZRZ) and
R^{-1/2}-R^{1/2}H_t=R^{1/2}(L-H_t)). Averaging over [1,2] gives some
t in that interval with
|Z(t)|_{op}+|L(t)-H_t|_{op}le C_1sqrtDelta,
C_1=C_1(sigma,kappa); when Delta=0 take a point where both
nonnegative integrands vanish.
Reading off the defect. Suppose Deltale1 and use this t. The
first relation of (eq:act:LZ) gives
V-N^r=(L^2-H_t^2)+Z^2+i[B,Z], whose norm is le C_2sqrtDelta by the
bounds above and (eq:act:smallsquares). The second relation gives
|[B,L]|_{op}le2|L|_{op}|Z|_{op}, and
[B,V]=[B,L^2]+[B,Z^2]+i[B,[B,Z]] has the same bound. Thus
|V-N^r|_{op}+|[B,V]|_{op}le C_3sqrtDelta .
Now N=rho+delta^sigma I commutes with rho, and y|-> y^r is
Lipschitz on [0,kappa^{2sigma}+1] because r>1, so
|N^r-rho^r|_{op}le C_4delta^sigma. Since [B,V]=[B,K^2] and
deltaledelta^sigma, (eq:act:Vcoercivity) yields
[
|[B,K^2]|_{op}^2+|K^2-B^2-rho^r|_{op}^2le C_5(Delta+delta^{2sigma}).
]
Choosing c_*>0 small (also c_*le1), for Deltale1
c_*Phi_K(A,B)leDelta+delta^{2sigma},
and for Delta>1 this holds because Phi_Kle1.
Relating Delta to F. Expand N and V:
[
F(A,B)=-Delta+tr(rho^q)-tr(N^q)+deltatrrho+delta^sigmatr V .
]
Since N=rho+delta^sigma I and r>1,
tr(rho^q)letr(N^q); also
trrhole nkappa^{2sigma} and Vle(kappa^2+delta)I. Adding
(eq:act:coercivity) proves (eq:act:constraint-bound).
## The penalty limit and the proof of the action estimate
Proof. [Proof of Theorem~thm:action]
Fix 0<deltale1. Proposition~prop:path gives, for every
0<εle1, a lower triangular C_ε and a path
B_ε with the endpoints and norm bound of
(eq:act:endpointB) such that, for A_ε=C_ε U,
ε B_ε'
=M_delta(B_ε)-A_ε A_ε^{T}
For fixed delta,K,Ω,U the matrices C_ε are bounded
uniformly in 0<εle1. Suppress the subscript and write
u_delta(A,B)=|M_delta(B)-AA^{T}|_{HS}^2. Since
Ain H^1_0, Bin C^1, and M_delta has the primitive J_delta of
(eq:field:primitive), the product and chain rules for absolutely
continuous functions and (eq:act:path) give, almost everywhere,
(J_delta(B)+tr(A^{T}'
&=2tr((A')^{T}
-tr((M_delta(B)-AA^{T}')
&=2tr((A')^{T}-ε^{-1}u_delta(A,B).
Completing the square in the action and integrating, using A=0 at both
endpoints, gives the exact formula
E_K(A)
=J_delta(-K)-J_delta(K)
+int_Ω(|A'-BA|_{HS}^2-F(A,B)
+ε^{-1}u_delta(A,B)),dx .
Penalty limit. The continuous representative of U is bounded on
overlineoverlineΩ;Ω; the bounds on C_ε and |B_ε|_{op}
place all pairs (A_ε(x),B_ε(x)) in one compact set
K_{ntimes d}timesSym_n with
|B|_{op}lekappa, on which F,Phi_K,u_delta are continuous.
Lemma~lem:constraint then implies
limsup_{ε↓0}sup_{(A,B)inK_}
(F(A,B)+c_*Phi_K(A,B)-ε^{-1}u_delta(A,B))
le E_{n,delta}.
Otherwise some sequence ε_j↓0 and pairs in the
compact set make the expression at least E_{n,delta}+tau, tau>0;
boundedness of F+c_*Phi_K would force u_delta=O(ε_j), and a
convergent subsequence would give a limit pair with u_delta=0 and
F+c_*Phi_Kge E_{n,delta}+tau, contradicting Lemma~lem:constraint.
For any tau>0 choose ε so the supremum in
(eq:act:penalty) is at most E_{n,delta}+tau. Substituting into
(eq:act:identity),
E_K(A_ε)E_K(A_ε)
ge{}&J_delta(-K)-J_delta(K)-|Ω|(E_{n,delta}+tau)
&+int_Ω(|A_ε'-B_ε A_ε|_{HS}^2
+c_*Phi_K(A_ε,B_ε)),dx .
By (eq:field:endpoints) the endpoint term exceeds
b_sigmasumell_i^{2gamma} up to an error going to 0 with delta,
and |Ω|E_{n,delta}to0 as delta↓0. First choose
delta so that both are below eta/2, then tau with
|Ω|tau<eta/2, then the required ε. This proves
(eq:act:bound). All compactness arguments are made after fixing
delta; no bound uniform in delta is needed.
% =====================================================================
# The sharp ineuality itself
% =====================================================================
Before turning to equality we record that the action estimate already
gives the sharp bound (eq:sharp); this is the ineuality half of
Theorem~thm:main.
[Sharp spectral bound]
Every potential in Theorem~thm:main satisfies
Tr(H_W)_-^gammale(D_sigma)/(b_sigma)int_Rtr(W^p).
Proof.
First suppose W is real symmetric. Choose any finite list of negative
eigenvalues -k_1^2,...,-k_n^2 with k_1ge...ge k_n>0 and real
orthonormal eigenfunctions phi_i. For 0<xi<k_n^2 put
ell_i=(k_i^2-xi)^{1/2} and K=diag(ell_1,...,ell_n). There are
real smooth compactly supported orthonormal u_i with form matrix
(h_W[u_i,u_j])_{i,j=1}^nle-K^2 .
Indeed, approximate the phi_i in H^1 and multiply the family by the
inverse square root of its Gram matrix: the Gram matrix tends to I_n and
the form matrix tends to -diag(k_i^2) by Lemma~lem:spectral, so an
operator-norm error below xi gives (eq:bound:trial).
Let U have rows u_i and let Ω contain their supports. For every
lower triangular C, the ith row a_i of A=CU satisfies
[
h_W[a_i]+ell_i^2|a_i|_2^2
lesum_{lle i}C_{il}^2(ell_i^2-ell_l^2)le0 .
]
Set Z=A^TA. Since AA^T and Z have the same nonzero eigenvalues with
multiplicity, summing the preceding bound and applying Lemma~lem:young
gives
[
E_K(A){tr(WZ)-tr(Z^q)}le D_sigmaint_Rtr(W^p).
]
Theorem~thm:action, with its nonnegative defects discarded, gives
b_sigmasum_iell_i^{2gamma}le D_sigmainttr(W^p)+eta for every
eta>0. Let eta↓0, then xi↓0, and finally exhaust
the negative eigenvalues. The case of empty negative spectrum is immediate.
For complex Hermitian W=E+iF with E,F real, the realification
widetilde W=E&-FF&E
is real symmetric and positive semidefinite; the constant unitary matrix
2^{-1/2}(I&iII&-iI)
conjugates widetilde W to W ⊕ overlineoverline W.W. Hence
tr(widetilde W^p)=2tr(W^p) and
Tr(H_{widetilde W})_-^gamma=2Tr(H_W)_-^gamma (conjugation identifies
the spectra of H_W and H_{overlineoverline W}W). Applying the real ineuality
to widetilde W and dividing by two proves (eq:bound:sharp).
% =====================================================================
# From equality to pointwise matrix relations
% =====================================================================
We now assume equality holds and extract information. The difficulty is
that the number of negative eigenfunctions may be infinite while the
potential has only m channels. The fixed column count m is what keeps
the tails under control.
Throughout this section W is real symmetric, Wge0 almost everywhere,
and 0<int_Rtr(W^p)<∞. Suppose equality holds in the form
b_sigmasum_{iinI}{2gamma}
=D_sigmaint_Rtr(W^p),dx,
where -k_i^2 are the negative eigenvalues of H_W, repeated according to
multiplicity, and k_1ge k_2ge...>0. By Lemma~lem:spectral
the index set I is a nonempty finite initial segment of N
or all of N, each eigenvalue has finite multiplicity, and k_ito0
in the infinite case. Fix real orthonormal eigenfunctions
phi_iin H^1(R;R^m), viewed as row vectors. Write
H=;R),
Λ=diag(k_i:iinI),
kappa=k_1 .
For operators between real Hilbert spaces the superscript T also denotes
the adjoint.
[Pointwise relations at equality]
Under these assumptions there are rows a_iin H^1(R;R^m),
iinI, each a finite linear combination of eigenfunctions with
eigenvalue -k_i^2, such that
-a_i''-a_iW=-k_i^2a_i quadin distributions.
The rows a_i may vanish. On a common set of full measure they define a
Hilbert--Schmidt operator A(x):R^mtoH, and there is a bounded
self-adjoint operator B on H, |B|_{op}lekappa, with
[B,Λ^2]=0,
Λ^2-B^2=(AA^T)^r,
a_i'(x)=(BA)_{i,·}(iinI).
Moreover
W(x)=q(A(x)^TA(x))^r quadalmost everywhere.
The operator B exists separately at each such point; no measurable choice
is required.
## Finite spectral approximations and vanishing errors
Let n_j=j when I= NbbN and n_j=|I|I| when
I is finite. Choose 0<xi_j<1/j with
ell_{j,i}:=(k_i^2-xi_j)^{1/2}>0 (1le ile n_j),
sum_{ile n_j}(k_i^{2gamma}-ell_{j,i}^{2gamma})<1/j .
The same shift xi_j is used for every eigenvalue, so the differences of
their squares are unchanged. Set
K_j=diag(ell_{j,1},...,ell_{j,n_j}).
By (eq:bound:trial) we can choose real smooth compactly supported
orthonormal u_{j,1},...,u_{j,n_j} with H^1 errors below 1/j and
form-matrix error below xi_j:
|u_{j,i}-phi_i|_{H^1}<1/j,
(h_W[u_{j,i},u_{j,l}])_{i,l=1}^{n_j}le-K_j^2 .
Let U_j have rows u_{j,i} and choose a bounded interval Ω_j
containing their supports. Apply Theorem~thm:action with
K=K_j, d=m, kappa=k_1, and eta=1/j: it gives a lower triangular
C_j, a symmetric path B_j with |B_j|_{op}lekappa, and
A_j=C_jU_j. Extend A_j by zero outside Ω_j and B_j by K_j
to the left and -K_j to the right; all errors vanish outside Ω_j.
The lower triangular form of C_j gives an extra nonnegative error. Put
Z_j=A_j^TA_j, an mtimes m positive semidefinite matrix. By
(eq:comp:form),
E_{K_j}(A_j)
leint_R(tr(WZ_j)-tr(Z_j^q)),dx-T_j,
T_j=sum_{lle ile n_j}(C_j)_{il}^{,2}(k_l^2-k_i^2)ge0 .
Indeed, the ith row of A_j is sum_{lle i}(C_j)_{il}u_{j,l}, whose
quadratic form plus ell_{j,i}^2 times its squared L^2 norm is at most
sum_{lle i}(C_j)_{il}^{,2}(ell_{j,i}^2-ell_{j,l}^2)
=-sum_{lle i}(C_j)_{il}^{,2}(k_l^2-k_i^2); summing over i and using
that A_jA_j^T and Z_j share nonzero eigenvalues gives
(eq:comp:triangular). The ordering of the k_i makes each summand
of T_j nonnegative.
Recall the Young deficit (eq:youngdefect); by
Lemma~lem:young, Y(W,Z)=0 iff W=qZ^r. Combining
Theorem~thm:action, (eq:comp:equality), and
(eq:comp:triangular),
0le T_j+int_R(
Y(W,Z_j)+|A_j'-B_jA_j|_{HS}^2+c_*Phi_{K_j}(A_j,B_j)),dx
&le d_j,
d_j:=b_sigma(sum_{iinI}{2gamma}
-sum_{ile n_j}ell_{j,i}^{2gamma})+1/j
&→0 .notag
Here c_*>0 is fixed; its independence of n_j is essential. Hence all
integral errors and T_j tend to zero.
A uniform bound on A_j will also be needed. The coercive estimate
(eq:youngcoercive) gives
Y(W,Z_j)ge1/2tr(Z_j^q)-(2^{p-1}-1)D_sigmatr(W^p),
and since Z_j is mtimes m, scalar H"older gives
|A_j|_{HS}^{2q}=(tr Z_j)^qle m^{q-1}tr(Z_j^q). With
(eq:comp:vanish) this bounds A_j uniformly in L^{2q}(R) in the
Hilbert--Schmidt norm, independently of its number of rows.
## Convergence of each fixed row
For fixed iinI, the coefficients (C_j)_{il}, lle i, are
bounded as jto∞. Choose a bounded interval J_i so large that
[
(int_{J_i}phi_l(x)phi_s(x)^T,dx)_{l,s=1}^{i}
ge3/4 I_i,
]
possible because the full Gram matrix is the identity. By
(eq:comp:form) the corresponding Gram matrix of the u_{j,l} is at
least 1/2 I_i for large j, so
[
frac12sum_{lle i}(C_j)_{il}^{,2}
leint_{J_i}|(A_j)_{i,·}|^2,dx
le|J_i|^{1-1/q}(int_{J_i}|A_j|_{HS}^{2q},dx)^{1/q},
]
uniformly bounded. Lower triangularity keeps the number of coefficients
in a fixed row finite. A diagonal subsequence makes each (C_j)_{il}
converge to a limit c_{il}, lle i. From T_jto0 and
(eq:comp:triangular),
c_{il}=0quadwhenever k_l>k_i .
Define a_i=sum_{lle i}c_{il}phi_l, a finite sum; then
(eq:comp:form) gives
(A_j)_{i,·}→ a_iquadin H^1(R;R^m)
quadfor every fixed i.
The ordering of the eigenvalues together with
(eq:comp:gapsuppress) shows only eigenfunctions with eigenvalue
-k_i^2 occur in a_i, proving (eq:comp:roweig) (the distributional
form equation follows by testing against smooth compactly supported
functions; the potential is locally integrable and the rows locally
bounded).
Using the continuous, locally absolutely continuous representatives and
H^1(R)subset C_b(R), (eq:comp:rowconv) also gives pointwise
convergence at every x. Passing to a further subsequence we may arrange
that the integrals of the nonnegative integrand in (eq:comp:vanish)
are summable and that, for every fixed i, the squared L^2 errors of
the derivatives in (eq:comp:rowconv) are summable (a diagonal choice
successively takes the first j available derivative errors and the
integral error below 2^{-j}). Consequently, on a common set of full
measure,
Y(W,Z_j)to0,
|A_j'-B_jA_j|_{HS}to0,
Phi_{K_j}(A_j,B_j)to0,
(A_j')_{i,·}(x)to a_i'(x) ∀ i .
We also exclude the null set where W is not a finite positive semidefinite
matrix.
## Tail control in the row-index space
Pad A_j,B_j,K_j with zeros in the unused coordinates of H, so
A_j:R^mtoH and B_j,K_j are self-adjoint finite-rank operators.
Since the omitted k_i tend to zero,
|K_j^2-Λ^2|_{op}
lemax{xi_j,sup_{i>n_j}k_i^2}→0 .
Fix x in the full-measure set of (eq:comp:pointwise) and suppress
it. Write rho_j=A_jA_j^T. Convergence of Phi to zero gives
|[B_j,K_j^2]|_{op}to0,
e_j:=|K_j^2-B_j^2-rho_j^r|_{op}to0 .
For hge1 let P_h project onto the first h coordinates, Q_h=I-P_h,
and s_h=sup_{i>h}k_i^2to0. For unit v in the range of Q_h,
|B_jv|^2+< v,rho_j^rv>le s_h+e_j .
By Jensen for the spectral measure of rho_j (r>1),
< v,rho_jv>^rle< v,rho_j^rv>, so
|Q_hrho_jQ_h|_{op}le(s_h+e_j)^{1/r}. Although H may be
infinite dimensional, Q_hrho_jQ_h has rank at most m because A_j has
m columns; its trace is at most m times its operator norm. Hence
|Q_hA_j|_{HS}^2le m(s_h+e_j)^{1/r},
|B_jQ_h|_{op}le(s_h+e_j)^{1/2}.
The first bound converts decay of the eigenvalue parameters into decay of
the total mass of all omitted rows, with factor the fixed physical
dimension m rather than the number of retained eigenfunctions.
For fixed h the first h rows of A_j converge by
(eq:comp:rowconv); (eq:comp:tailbounds) makes the remaining rows
uniformly small by taking h large then j large. Thus A_j is Cauchy in
Hilbert--Schmidt norm and converges to an operator A:R^mtoH
whose rows are the a_i(x):
A_j→ Aquadin Hilbert--Schmidt norm.
This holds for the whole already-chosen subsequence and needs no further
choice depending on x.
For B_j only pointwise compactness is needed. Finite compressions
P_hB_jP_h are bounded in finite-dimensional spaces, so a diagonal
subsequence makes them converge for every h; self-adjointness and the
second bound in (eq:comp:tailbounds) give
|B_j-P_hB_jP_h|_{op}le2|B_jQ_h|_{op}le2(s_h+e_j)^{1/2}. The
subsequence converges in operator norm to a self-adjoint B,
|B|_{op}lekappa (possibly depending on x).
Now pass to all products. Hilbert--Schmidt convergence of A_j implies
operator-norm convergence of A_jA_j^T and A_j^TA_j; continuous
functional calculus for nonnegative operators gives convergence of their
rth powers. With (eq:comp:diagonal) and
(eq:comp:algebraerr) this proves
[B,Λ^2]=0 and Λ^2-B^2=(AA^T)^r. Also
B_jA_jto BA in Hilbert--Schmidt norm; taking the ith row and using the
derivative convergence and the first-order error in
(eq:comp:pointwise) gives a_i'(x)=(BA)_{i,·} for every i.
Finally continuity of the finite-dimensional Young deficit and
Y(W,A_j^TA_j)to0 give Y(W,A^TA)=0, and Lemma~lem:young gives
W=q(A^TA)^r.
All statements hold on one common set of full measure. No function
x|-> B(x) is selected; only the existence of a suitable operator at
each such point enters later. This proves Proposition~prop:relations.
% =====================================================================
# Fixed channels and the equality profiles
% =====================================================================
The pointwise relations of Proposition~prop:relations involve an
auxiliary operator B that may depend on x. We now show the channel
subspaces are constant, so that on each channel B becomes a fixed matrix
obeying an ODE whose solutions are the profiles of (eq:profile).
## The channel subspaces are constant
Assume first that W is real and is a nonzero equality potential. Use the
rows a_i, the numbers k_i, and Λ=diag(k_i) from
Proposition~prop:relations. For each distinct value alpha>0 among
the k_i, let n_alpha be its finite multiplicity and let G_alpha be
the n_alphatimes m matrix whose rows are the a_i with k_i=alpha;
each entry is in H^1(R) and we use its continuous representative.
At every point of the common full-measure set of Proposition~prop:relations
choose one of the operators B it supplies. Since B commutes with
Λ^2 it preserves each eigenspace, and
AA^T=(Λ^2-B^2)^{1/r} commutes with B and Λ^2. Writing
B_alpha for the block on the alpha^2-eigenspace,
G_alpha G_beta^T&=0quad(alphanebeta),
G_alpha'&=B_alpha G_alpha,
& [B_alpha,G_alpha G_alpha^T]&=0nonumber
alpha^2I-B_alpha^2&=(G_alpha G_alpha^T)^r,
& |B_alpha|_{op}&lekappa .
These hold almost everywhere with no compatibility between the choices of
B at different points.
For every alpha the subspace R_ G_alpha(x))^perp
subsetR^m is independent of x. The nonzero R_ are
mutually orthogonal and
sum_alphadimR_ m .
In particular only finitely many G_alpha are nonzero. The potential W
has a continuous representative, equal to zero on the orthogonal complement
of the R_ and to q(G_alpha^TG_alpha)^r on
R_ With this representative every G_alpha is C^2.
Proof.
Fix vinR^m. From (eq:rig:blocksquare),
|(G_alpha v)'|lekappa|G_alpha v| almost everywhere, and G_alpha v
is locally absolutely continuous. If it vanishes at x_0, integration and
Gronwall's ineuality, first right and then left of x_0, show it vanishes
everywhere; hence ker G_alpha(x) is constant, including off the original
full-measure set.
The products in (eq:rig:orthogonal) are continuous, so their
almost-everywhere vanishing is everywhere; thus the constant row spaces
R_ are mutually orthogonal, giving
(eq:rig:dimbound) and finiteness of the nonzero blocks. On the
full-measure set, A^TA=sum_alpha G_alpha^TG_alpha with only finitely
many nonzero summands, each acting on its own R_, so
(eq:comp:potential) gives
W=qsum_alpha(G_alpha^TG_alpha)^r quadalmost everywhere.
The right side is continuous; we use it as the representative of W. The
row equation (eq:comp:roweig) then reads
G_alpha''=alpha^2G_alpha-G_alpha W in distributions, whose right side
is continuous; integrating twice, G_alpha is C^2.
The row spaces split R^m into finitely many constant channels. To find
the potential on one channel we also need a constant range for G_alpha in
the eigenfunction index space.
For alpha with ell:=dimR_>0, the range of
G_alpha(x) in R^{n_alpha} is independent of x. In constant
orthonormal bases of this range and of R_ the restriction
of G_alpha is an invertible C^2 matrix D:RtoR^{elltimesell}, and
B_0=D'D^{-1} is C^1 with, everywhere,
B_0=B_0^T,
alpha^2I-B_0^2=(DD^T)^r>0,
B_0'=r(B_0^2-alpha^2I).
Proof.
Choose a constant orthonormal basis of R_ and let E(x) be
the matrix of G_alpha(x) restricted to it; it has full column rank ell
everywhere. Its range projection is P=E(E^TE)^{-1}E^T. The commutator in
(eq:rig:blocksquare) shows B_alpha preserves the range of
G_alpha G_alpha^T, which is the range of E; thus E'=B_alpha E has
columns in this range almost everywhere. Differentiating P gives
P'=(I-P)E'(E^TE)^{-1}E^T+E(E^TE)^{-1}(E')^T(I-P)=0 almost everywhere,
and since P is C^1 it is constant.
With constant orthonormal bases of both spaces, D is invertible C^2 and
D'D^{-1} is, almost everywhere, the restriction of B_alpha to its
invariant range. Then the symmetry and the square identity in
(eq:rig:active) hold almost everywhere and extend by continuity;
positivity is strict because D is invertible. On R_,
(eq:rig:contpot) makes the potential q(D^TD)^r, so
(eq:comp:roweig) becomes D''=alpha^2D-q(DD^T)^rD (using a
singular-value decomposition for the second form). Differentiating
B_0=D'D^{-1} with q=r+1,
[
B_0'=D''D^{-1}-B_0^2
=alpha^2I-q(alpha^2I-B_0^2)-B_0^2
=r(B_0^2-alpha^2I).
]
## The matrix Riccati equation separates the profiles
Fix a nonzero block and the matrix B_0. Orthogonally diagonalise the
symmetric B_0(x_*) at one point. The right side of the Riccati equation is
a polynomial in B_0, so uniqueness of the matrix ODE preserves the
diagonal form. The strict ineuality in (eq:rig:active) places each
initial diagonal entry b_j(x_*) in (-alpha,alpha), and the scalar
equation b_j'=r(b_j^2-alpha^2) has the unique solution with that value
b_j(x)=-alphatanh!(ralpha(x-y_j)), 1le jleell,
for a unique centre y_j. The diagonal matrix with these entries solves the
same initial-value problem as B_0 on all of R, so uniqueness gives
equality everywhere.
Now (eq:rig:active) shows DD^T is diagonal in these coordinates.
Since D'=diag(b_j)D, each row of D is a positive scalar multiple of its
value at x_*; the rows are nonzero and mutually orthogonal (as DD^T is
diagonal). Taking their normalised directions as a constant orthonormal
basis of R_ makes D diagonal with positive entries. The
potential in this fixed basis is diagonal, with entries
q(alpha^2-b_j(x)^2)
=(r+1)alpha^2,sech^2!(ralpha(x-y_j)).
Combining these bases over the finitely many nonzero R_ and
adding a basis of their orthogonal complement gives, for real W, the form
(eq:classification) with at most m nonzero profiles by
(eq:rig:dimbound). The zero potential corresponds to k=0.
## Complex Hermitian potentials
For a complex Hermitian W, take the realification widetilde W from
(eq:bound:realification), which is equivalent to W ⊕ overlineoverline WW
by a constant unitary matrix, with both inttr(W^p) and the negative
moment doubled. Equality for W implies equality for the real
widetilde W, so the real result gives a constant unitary V with
[
W(x) ⊕ overlineoverline W(x)W(x)
=Vdiag(f_1(x),...,f_{2m}(x))V^*
quadalmost everywhere,
]
each f_j a profile (eq:rig:profiles) or zero. Let P be the
projection onto the first summand and Phat=V^*PV.P=V^*PV. Since P
commutes with W ⊕ overlineoverline W,W, we have
(f_i-f_j)Phat_P_{ij}=0 almost everywhere, so Phat_P_{ij}=0
whenever f_i,f_j are not equal almost everywhere. Partition the indices
into groups of identical functions (including the zero function). On each
group the diagonal potential is a scalar multiple of the identity, so any
constant orthonormal basis of the intersection of that coordinate group with
ranPPhat diagonalises it. Transporting these bases by
V gives a constant unitary basis of C^m in which W has the asserted
profiles. This proves necessity for complex Hermitian W.
## Every listed potential attains equality
It remains to verify attainment directly. Fix a>0 and x_0inR and set
[
V(x)=(r+1)a^2sech^2!(ra(x-x_0)),
H=-∂_x^2-V .
]
Put w(x)=atanh(ra(x-x_0)) and Q=∂_x+w on
H^1(R). Since w,w' are bounded, direct differentiation gives
Q^*Q&=H+a^2,
QQ^*&=-∂_x^2+a^2+(r-1)a^2sech^2!(ra(x-x_0))ge a^2,
the last ineuality because r>1. The equation Qpsi=0 has the
one-dimensional L^2 solution space
[
ker Q=spanbigl{sech^{1/r}(ra(x-x_0))bigr},
]
so -a^2 is a simple eigenvalue of H, and (eq:rig:factor) excludes
eigenvalues below it. If Hphi=lambdaphi with -a^2<lambda<0, put
mu=lambda+a^2>0, g=Qphi; then |g|_2^2=mu|phi|_2^2>0,
Q^*Qphi=muphiinDomQ, so ginDom(QQ^*)
and QQ^*g=mu g, contradicting (eq:rig:partner) because mu<a^2.
Hence -a^2 is the only negative eigenvalue.
Finally 2p-1=2gamma and the substitution s=tanh y give
int_R V(x)^p,dx
&=((r+1)^p)/(r)a^{2gamma}int_Rsech^{2p}y,dy
&=((r+1)^p)/(r)a^{2gamma}int_{-1}^{1}(1-s^2)^{p-1},ds
=(b_sigma)/(D_sigma)a^{2gamma},
using (r+1)^p/r=D_sigma^{-1} and p-1=sigma. Therefore
Tr(H)_-^gamma=a^{2gamma}=C_gammaint_R V^p. The zero potential
contributes zero to both sides; finite orthogonal direct sums add both
spectral moments and trace integrals, and constant unitary conjugation
changes neither. So every potential listed in Theorem~thm:main
attains equality, completing the proof.
% =====================================================================
{99}
raggedright
Michael Aizenman and Elliott H. Lieb,
On semi-classical bounds for eigenvalues of Schr"odinger operators,
Physics Letters A 66 (1978), no.~6, 427--429.
Rafael Benguria and Michael Loss,
A simple proof of a theorem of Laptev and Weidl,
Mathematical Research Letters 7 (2000), no.~2--3, 195--203.
Rupert L. Frank, David Gontier and Mathieu Lewin,
Optimizers for the finite-rank Lieb--Thirring ineuality,
American Journal of Mathematics 147 (2025), no.~2, 503--560.
Dirk Hundertmark, Ari Laptev and Timo Weidl,
New bounds on the Lieb--Thirring constants,
Inventiones Mathematicae 140 (2000), no.~3, 693--704.
Dirk Hundertmark, Elliott H. Lieb and Lawrence E. Thomas,
A sharp bound for an eigenvalue moment of the one-dimensional
Schr"odinger operator,
Advances in Theoretical and Mathematical Physics 2 (1998),
no.~4, 719--731.
Joseph B. Keller,
Lower bounds and isoperimetric ineualities for eigenvalues of the
Schr"odinger equation,
Journal of Mathematical Physics 2 (1961), no.~2, 262--266.
Ari Laptev and Timo Weidl,
Sharp Lieb--Thirring ineualities in high dimensions,
Acta Mathematica 184 (2000), no.~1, 87--111.
Elliott H. Lieb and Walter E. Thirring,
Bound for the kinetic energy of fermions which proves the stability
of matter,
Physical Review Letters 35 (1975), no.~11, 687--689.
Elliott H. Lieb and Walter E. Thirring,
Ineualities for the moments of the eigenvalues of the Schr"odinger
Hamiltonian and their relation to Sobolev ineualities,
in Studies in Mathematical Physics: Essays in Honor of Valentine
Bargmann, Princeton University Press, 1976, 269--303.
Sharp one-dimensional Lieb--Thirring ineualities for matrix
potentials, companion preprint, 2026.
Arthur Sard,
The measure of the critical values of differentiable maps,
Bulletin of the American Mathematical Society 48 (1942),
883--890.