We show that the norm obtained by summing the L1 norms of all weak derivatives through order k, including the function itself, determines both n and k on the real Sobolev space Wk,1(Rn). Every surjective linear isometry for this norm is given by a translation, a signed permutation of the coordinates and multiplication by a fixed sign c∈{−1,1}.
Introduction
Throughout, functions are real-valued and operators are real-linear. For integers n,k≥1, we equip Wk,1(Rn) with the norm
Here ∣α∣=α1+⋯+αn, and D0f=f. In particular, for k=1 the derivative part is ∑j∥∂jf∥1. A signed permutation of the axes permutes these terms. We shall see that the geometry of the norm recovers the axes even when neither the dimension nor the order of the target space is fixed in advance.
Theorem 1.1.Suppose that n,m,k,s are positive integers and
T:(Wk,1(Rn),Nn,k)⟶(Ws,1(Rm),Nm,s)
is a surjective linear isometry. Then m=n and s=k. There exist a signed permutation matrix P, a vector b∈Rn and a sign c∈{−1,1} such that, for every f∈Wk,1(Rn),
(Tf)(y)=cf(Py+b)for almost every y∈Rn.(2)
A related question is whether, for n,m≥2 and k,s≥1, a Banach-space isomorphism
Wk,1(Rn)≅Ws,1(Rm)
necessarily implies n=m and k=s. For background on the isomorphic classification of Sobolev spaces, see Pełczyński and Wojciechowski [1]. Diestel [4 Theorem 1] used an Lp extension theorem to classify surjective linear isometries of Wk,p(U) on bounded connected Lipschitz domains U⊂Rn in the sense of (2). The result applies for 1≤p<∞ with p=2, provided that the isometry maps Wk,p(U)∩C(U) onto itself. When k>n/p, every function in Wk,p(U) has a continuous representative on U by Sobolev embedding, so this condition is automatic [4 Corollary 1]. M. Biegert and R. Nittka [3] proved that, for 1≤p<∞ with p=2, every linear isometry T:W1,p(Ω1)→W1,p(Ω2) preserves disjointness. Here Ω1,Ω2⊂Rn are open and the Sobolev norms use the Euclidean length of the gradient. If both sets are bounded and T is order bounded as a map from W1,p(Ω1) into Lp(Ω2), they also proved that there exist a measurable map ξ:Ω2→Ω1 and a real-valued weight g on Ω2 such that, for each u∈W1,p(Ω1),
We use the Lebesgue measure on Euclidean spaces and identify functions modulo null sets. The support of a measurable function will always mean the class of its nonzero set. Two functions are disjoint if their product vanishes almost everywhere.
Fact 2.1.[11 Theorem 6.19] If u∈Wloc1,1(Rn;R), then ∇u=0 almost everywhere on {u=0}. Consequently, for f∈Wk,1(Rn) every Dαf with ∣α∣≤k vanishes almost everywhere on {f=0}.
To see this, suppose that fg=0 almost everywhere. By Fact 2.1, Dαf vanishes almost everywhere on {f=0}, and Dαg vanishes almost everywhere on {g=0}. Since these two sets cover Rn up to a null set, DαfDαg=0 almost everywhere for every ∣α∣≤k. Lamperti’s equality criterion [10 Corollary 2.1] therefore gives
∥Dαf+Dαg∥1+∥Dαf−Dαg∥1=2∥Dαf∥1+2∥Dαg∥1.
Summing over α gives the norm identity in (3). Conversely, the triangle inequality bounds each term on the left by the corresponding term on the right. Equality in the sum thus implies equality for α=0. Since D0f=f and D0g=g, Lamperti’s criterion gives fg=0 almost everywhere. Both T and T−1 preserve the norm identity, so both preserve disjointness.
The Sobolev jet
We use the product of Lebesgue and counting measure on Rn×An,k. A sheet is a set Rn×{α} with its index fixed. We use γ≤α for coordinatewise comparison, α!=∏iαi! and (γα)=∏i(γiαi). Our approach of embedding Sobolev spaces into scalar Lp spaces, extending the induced isometry, and applying Lamperti’s theorem follows the strategy of Diestel [4 Sections 2–3].
Lemma 3.1. Give An,k counting measure. The map
Jn,kf(x,α)=Dαf(x),Jn,k:Wk,1(Rn)⟶L1(Rn×An,k),
is an isometry onto a closed subspace. For g(x)=e−∣x∣2, the jet Jn,kg is nonzero almost everywhere. The ratios Jn,kf/Jn,kg, as f ranges over Wk,1(Rn), generate the full product σ-algebra of Rn×An,k modulo null sets.
Proof. Integrating ∣Jn,kf∣ over the sheet Rn×{α} gives ∥Dαf∥1. Summing over the sheets gives precisely Nn,k(f), i.e., Jn,k is an isometry. Closedness now follows from completeness. We must find one jet which is nonzero almost everywhere on every sheet, and show that ratios by this jet distinguish the measurable sets of the whole product.
Let g(x)=e−∣x∣2. We show by induction on ∣α∣ that
Dαg(x)=Hα(x)g(x),
where Hα is a polynomial whose unique term of highest total degree is (−2)∣α∣xα. For α=0 this holds with H0=1, since D0g=g.
Suppose the assertion holds for α and write d=∣α∣. Since ∂jg=−2xjg, the product rule gives
Thus Hα+ej=∂jHα−2xjHα is again a polynomial. To identify its highest-degree term, note that ∂jHα contains no term of degree d+1. In −2xjHα, the only term of degree d+1 comes from multiplying the highest-degree term of Hα by −2xj. This gives
(−2xj)(−2)dxα=(−2)d+1xα+ej.(4)
All other terms have degree at most d, so this term cannot cancel. This proves the induction step. In particular, Hα is a nonzero polynomial of degree ∣α∣ for every α. Remove its zero set for each of the finitely many α, and let G be the common remainder. Since each Hα is a nonzero polynomial, G has full Lebesgue measure. On G the jet of g supplies a nonzero denominator on every sheet.
For a∈Qn put ga(x)=ea⋅xg(x). It is easy to see that ga∈Wk,1. On G the product rule gives
Suppose two pairs (x,α) and (x′,α′) give the same quotients for all a∈Qn. The expressions are continuous in a, so they agree for all a∈Rn. If x=x′, set v=x−x′ and substitute a=tv in the equality of the two quotients. By (3.1), this gives
etv⋅xPx,α(tv)=etv⋅x′Px′,α′(tv)(t∈R).
Dividing both sides by etv⋅x′ and using v⋅x−v⋅x′=v⋅(x−x′)=v⋅v=∣v∣2, we obtain
et∣v∣2Px,α(tv)=Px′,α′(tv)(t∈R).
Both sides have a polynomial factor taking value 1 at t=0. As t→+∞, the left side grows in absolute value faster than any polynomial, a contradiction. Thus the quotients determine x. With x=x′ the polynomial factors themselves agree. The unique term of highest total degree in Px,α(a) is aαg(x)/Dαg(x), with nonzero coefficient. Comparing its multiindex with that on the other sheet yields α=α′.
The quotients in (3.1) therefore define an injective Borel map
Φ:X:=G×An,k⟶RQn,Φ(x,α)=(Dαg(x)Dαga(x))a∈Qn.
By the Lusin–Souslin theorem in its standard Borel form [9 Corollary 15.2], the image Φ(X) is Borel in RQn, and Φ is a Borel isomorphism from X onto Φ(X).
Let
S:=σ((x,α)⟼Dαg(x)Dαga(x):a∈Qn).
We claim that S=B(X), where B(X) denotes the Borel σ-algebra of X. Since each quotient is Borel measurable, we immediately have S⊆B(X).
For the reverse inclusion, let E∈B(X). The Borel isomorphism property, together with the fact that Φ(X) is Borel, implies that Φ(E) is Borel in RQn. Moreover, injectivity gives
E=Φ−1(Φ(E)).
Since Qn is countable, the Borel σ-algebra of RQn is generated by the coordinate projections πa, a∈Qn. For each such projection,
(πa∘Φ)(x,α)=Dαg(x)Dαga(x).
Consequently, the inverse image under Φ of every Borel subset of RQn belongs to S. In particular, E=Φ−1(Φ(E))∈S. Thus B(X)⊆S, proving that the coordinate quotients generate the Borel σ-algebra on G×An,k. The omitted part of the product is null. This proves the claimed generation modulo null sets. □
By Lemma 3.1 and Hardin’s extension theorem [7 Theorem 4.2], the isometry Jm,sTJn,k−1 between the jet ranges extends to a surjective L1 isometry U of the ambient spaces satisfying
UJn,k=Jm,sT.(6)
Put X=Rn×An,k and Y=Rm×Am,s, and denote by μ and ν their respective products of Lebesgue measure and counting measure. Write MX for the algebra of measurable subsets of $X modulo μ-null sets. Thus [E]=[F] in MX precisely when μ(E△F)=0. Define MY in the same way using ν.
By Lamperti’s representation theorem [10 Theorem 3.1], the surjective linear isometry U:L1(X,μ)→L1(Y,ν) admits a representation
Uu=hτu
where
τ:MX⟶MY
is a Boolean σ-isomorphism and τ is the induced transformation on measurable functions, characterized by
τ1E=1Fwhenever [F]=τ([E]).
Surjectivity implies that h=0ν-almost everywhere. Indeed, every function in the range of U vanishes almost everywhere on Z:={h=0}. If ν(Z)>0, σ-finiteness provides a measurable set B⊆Z with 0<ν(B)<∞. Then 1B∈L1(Y,ν) cannot belong to the range of U, contradicting surjectivity.
By construction, the induced transformation transports measurable supports according to τ:
esssupp(τu)=τ(esssuppu),esssuppu:=[{u=0}].
Since multiplication by h does not change the measurable support, we conclude that
esssupp(Uu)=τ(esssuppu)(u∈L1(X,μ)).
Recovering the spatial algebra and the finite fibers
Let (Ei)i∈I be a family of measurable subsets of (X,μ). Its essential union is the least measurable upper bound modulo null sets: it is represented by a measurable set E such that
μ(Ei∖E)=0(i∈I),
and whenever a measurable set F satisfies μ(Ei∖F)=0 for every i∈I, one has μ(E∖F)=0. Such an E is unique modulo null sets. On a σ-finite space, the essential union exists and is represented by the union of a countable subfamily [5 Proposition 215B(iv)].
Lemma 4.1. Let F⊂Ws,1(Rm) be stable under multiplication by Cc∞(Rm) functions. If E is the essential union of the supports of the functions in F, then E×Am,s is the essential union of the supports of their jets.
Proof. For f∈F, Fact 2.1 gives Dβf=0 almost everywhere off E for every β. This proves containment in E×Am,s. For the reverse containment, the difficulty is that f(y)=0 need not imply Dβf(y)=0. We may, however, replace f by an appropriate smooth multiple without leaving F.
Choose fℓ∈F whose supports realize E, the case of empty E being immediate. For r∈N take χr∈Cc∞(Rm) equal to one on B(0,r). For every t∈Qm the multiplier y↦et⋅yχr(y) is smooth with compact support. Fix r,ℓ∈N, t∈Qm, and β∈Am,s. Since χr=1 on B(0,r), the weak product rule gives
Dβ(et⋅(⋅)χrfℓ)(y)=et⋅yγ≤β∑(γβ)tβ−γDγfℓ(y).
for almost every y∈B(0,r). The exceptional set may depend on r,ℓ,t,β. Since there are only countably many such choices, the union of all exceptional sets is a null set N⊂Rm. Thus, for each y∈/N, the identity holds for every ℓ∈N, t∈Qm, and β∈Am,s whenever r>∣y∣. If fℓ(y)=0, the coefficient of tβ in the sum is fℓ(y), so the polynomial in t is nonzero for some rational t. Choosing r>∣y∣, we have found a member of F whose jet is nonzero at (y,β). Since r∈N was arbitrary and almost every y∈E∖N satisfies fp(y)=0 for some p∈N, we obtain the conclusion.
For a family F⊂Wk,1(Rn), define its disjoint complement in Wk,1(Rn) by
Fd={u∈Wk,1(Rn):uf=0a.e. for every f∈F}.
For families in Ws,1(Rm), the disjoint complement is defined analogously within Ws,1(Rm). For open O⊂Rn let XO consist of the Sobolev functions vanishing almost everywhere outside O. For an open ball B one has
XB=(XRn∖B)d.(7)
Indeed, membership in XB implies disjointness from the exterior family. Conversely, choose smooth bumps supported in the exterior whose positive sets cover it, using a countable family of balls with closure in the exterior. A function disjoint from these bumps vanishes almost everywhere there. The only remaining set outside B is ∂B, which is null.
By surjectivity and Lemma 2.2, T(Fd)=(TF)d. In particular TXB is a disjoint complement by (7). Such a complement is stable under multiplication by smooth compactly supported functions, since multiplying u by a smooth factor preserves every relation uf=0. On the source side, XB already has this stability and its essential support union is B, as is seen using bumps inside B. Write EB for the essential support union of TXB. Fix f∈XB. Applying Lamperti’s support formula τ(esssuppu)=esssupp(Uu) to the function u=Jn,kf gives
τ(esssupp(Jn,kf))=esssupp(U(Jn,kf)).
By (3.2), U(Jn,kf) and Jm,s(Tf) agree almost everywhere on Rm×Am,s. Thus, for every f∈XB,
τ(esssupp(Jn,kf))=esssupp(Jm,s(Tf)).
We now take the essential union over f∈XB. Since τ is an order isomorphism of the measure algebras, it carries the least upper bound of the supports on the left to the least upper bound of their images. By Lemma 4.1, the essential union of the supports of Jn,kf for f∈XB is B×An,k. As f ranges over XB, Tf ranges over TXB, so the same lemma on the target gives EB×Am,s as the essential union of the supports of Jm,s(Tf). Consequently,
τ(B×An,k)=EB×Am,smodulo null sets.
We claim that the same formula holds for every Lebesgue measurable set C⊂Rn. Indeed, the sets C for which there exists a measurable D⊂Rm satisfying
τ(C×An,k)=D×Am,smodulo null sets
form a σ-algebra which contains the balls with centers in Qn and positive rational radii, so it contains every Borel subset of Rn. Every Lebesgue measurable set agrees with a Borel set modulo a null set, and taking the product with the finite set An,k preserves this agreement. This proves the claim.
Conversely, let D⊂Rm be Lebesgue measurable. Applying the same argument to T−1 and its extension on jets U−1 gives a measurable set C⊂Rn such that
τ−1(D×Am,s)=C×An,kmodulo null sets.
Thus every set of the form D×Am,s is the image under τ of a set of the form C×An,k, modulo null sets.
Lemma 4.2 (The permutation of the derivative indices). The isomorphism τ on the jet spaces acts by inverse images under a map of the form
(y,β)⟼(φ(y),πy(β)).
Here φ:Rm→Rn is a nonsingular isomorphism modulo null sets and πy:Am,s→An,k is a bijection for almost every y.
Proof. The preceding argument gives an isomorphism between the Lebesgue measure algebras of Rn and Rm, determined by
τ(C×An,k)=D×Am,s.
By [6 Theorem 343B(v) and Proposition 343F], the isomorphism between the Lebesgue measure algebras of Rn and Rm induced by τ is represented by mutually inverse nonsingular measurable maps
φ:Rm→Rn,ψ:Rn→Rm
modulo null sets. In particular, for all Lebesgue measurable sets C⊂Rn and D⊂Rm,
τ(C×An,k)=φ−1(C)×Am,s,
τ−1(D×Am,s)=ψ−1(D)×An,k,
modulo null sets, and
ψ(φ(y))=yfor almost every y∈Rm,φ(ψ(x))=xfor almost every x∈Rn.
For each α∈An,k, choose a measurable representative Eα of τ(Rn×{α}). Since τ preserves unions and intersections, the sets Eα are pairwise disjoint modulo null sets and cover Rm×Am,s modulo a null set. There are only finitely many α, so we may modify these sets on a single null set to obtain an actual partition. Define πy:Am,s→An,k by
πy(β)=α⟺(y,β)∈Eα.
For every measurable C⊂Rn, we have
C×{α}=(C×An,k)∩(Rn×{α}),
and hence
τ(C×{α})=(φ−1(C)×Am,s)∩Eαmodulo null sets.
The expression on the right is precisely the inverse image of C×{α} under the measurable map
F(y,β)=(φ(y),πy(β)).
Every measurable subset of Rn×An,k agrees modulo a null set with a finite union of sets Cα×{α}. Thus F represents τ by inverse images.
Applying the same construction to τ−1 gives a measurable map
G(x,α)=(ψ(x),ρx(α)),ρx:An,k→Am,s,
which represents τ−1 by inverse images. Consequently, G∘F and F∘G induce the identity on the respective measure algebras. Each composition therefore equals the identity almost everywhere. Indeed, since G∘F induces the identity on the measure algebra of Rm×Am,s, for every measurable set D⊂Rm×Am,s the symmetric difference
ND=(G∘F)−1(D)△D
is a null set with respect to Lebesgue measure times counting measure. Thus, for every z∈/ND,
z∈D⟺G(F(z))∈D.
The exceptional set ND may depend on D. To obtain this equivalence simultaneously for a family of sets that distinguishes every pair of points, consider
D={B(q,r)×{β}:q∈Qm,r∈Q,r>0,β∈Am,s}.
Here B(q,r) denotes the open ball with center q and radius r. This family is countable, since the rational centers and radii form countable sets and Am,s is finite. It also separates points. The countable union
N=D∈D⋃ND
is a null set. Consequently, for every z∈/N and every D∈D,
z∈D⟺G(F(z))∈D.
If G(F(z))=z, the separating property of D gives a set D∈D containing exactly one of these two points, contradicting this equivalence. Hence G(F(z))=z for every z∈/N, which proves
G∘F=idRm×Am,s
almost everywhere.
Applying the same argument to F∘G, using rational balls in Rn and the finite index set An,k, gives
F∘G=idRn×An,k
almost everywhere.
Since Am,s is finite, for almost every y∈Rm we have
G(F(y,β))=(y,β)
for every β∈Am,s.
Comparing the second coordinates gives
ρφ(y)∘πy=idAm,s.
Similarly, for almost every x∈Rn,
F(G(x,α))=(x,α)
for every α∈An,k, so
πψ(x)∘ρx=idAn,k.
By nonsingularity of φ, the latter identity remains valid for almost every y after substituting x=φ(y). Using ψ(φ(y))=y almost everywhere, we obtain
πy∘ρφ(y)=idAn,k.
Thus πy and ρφ(y) are mutually inverse for almost every y, which proves the asserted bijectivity. □
The weighted-composition representation of the ambient isometry can now be written, for u∈L1(Rn×An,k), as
(Uu)(y,β)=hβ(y)u(φ(y),πy(β)).
For a=Jn,kg from Lemma 3.1, the weight on sheet β is
hβ(y)=a(φ(y),πy(β))(Ua)(y,β).
Both a and Ua have full support, so every hβ is finite and nonzero almost everywhere. Taking u=Jn,kf and using (3.2) gives for almost all y∈Rm:
Dβ(Tf)(y)=hβ(y)Dπy(β)f(φ(y)),β∈Am,s.(8)
Also ∣An,k∣=∣Am,s∣. Appending the coordinate k−∣α∣ to α∈An,k gives a bijection with the (n+1)-tuples of nonnegative integers of sum k. The fiber count is therefore
(kn+k)=(sm+s).(9)
In (4.2) the exceptional null set may depend on f. We shall compare this identity only for countable families of tests.
The zeroth and first derivative levels
So far the bijection πy may interchange derivatives of different orders. Even the cardinality in (4.3) does not determine n and k separately. We first identify the index 0, and then identify the indices of length one.
Lemma 5.1. One has πy(0)=0 almost everywhere. In particular, if h=h0, then for every f∈Wk,1(Rn),
(Tf)(y)=h(y)f(φ(y))(10)
for almost every y.
Proof. Take a ball B with rational center and radius. Choose a cutoff θ∈Cc∞(Rn) equal to one on B. Consider
E=φ−1(B)∩{y:πy(0)=0}.
On E the target value Tθ reads a positive-order derivative of θ by (8). Hence Tθ=0 almost everywhere on E. By locality all its derivatives vanish there as well.
On the other hand, partition E into the finitely many sets Eβ=E∩{y:πy(β)=0}. They cover E up to a null set because πy is onto. On Eβ, the same jet formula gives DβTθ=hβθ∘φ=hβ. Since hβ is nonzero almost everywhere, Eβ is null. We conclude that E is null. Choosing a cover of Rn by balls with rational centers and positive rational radii gives the conclusion. □
Lemma 5.2. For 1≤j≤m, one has ∣πy(ej)∣=1 almost everywhere. Consequently m=n and s=k.
Proof. Fix j, a rational ball B⊂Rn, and a cutoff θ as in the preceding proof. Applying (10) to n+1 functions gives
v0=Tθ=h,vi=T(xiθ)=hφi(1≤i≤n)a.e. on φ−1(B).
All the vi belong to W1,1. These identities allow us to compute ∂jTf locally without assuming weak differentiability of h or of φ.
Define the vector-valued function v=(v1,…,vn) and fix ε>0. We work on Eε=φ−1(B)∩{∣h∣≥ε}. Fix f∈C∞(Rn). Choose a smooth cutoff ηε on [0,∞) equal to zero on [0,ε/2] and to one on [ε,∞) with 0≤ηε≤1. For (a,b)∈R×Rn put
Fε(a,b)={ηε(∣a∣)af(b/a),0,a=0,a=0.
To apply the Sobolev chain rule [2 Theorem 4.2], we show that Fε is globally Lipschitz by establishing the estimate
∣DFε(a,b)[u,w]∣≤Cε(∣u∣+∣w∣)(11)
for all (a,b)∈R×Rn and all directions (u,w)∈R×Rn. The constant may depend on f, ε and the chosen cutoff, but is independent of a,b,u,w. First fix a=0. To compute the derivative in the direction (u,w), consider the curve
t⟼(a+tu,b+tw)
and set
z(t)=a+tub+tw,z=z(0)=ab.
The denominator is nonzero for all sufficiently small t. Applying the quotient rule to each coordinate gives
z′(0)=a2aw−ub=aw−uz.
The product rule and the chain rule therefore yield
This proves the desired estimate (∗) for a=0. For ∣a∣<ε/2, the function Fε(a,b) vanishes for every b∈Rn. Hence Fε is also smooth in a neighborhood of every point (0,b), with derivative zero there. Thus (∗) holds on all of Rn+1.
Integrating the bounded derivative along line segments shows that Fε is globally Lipschitz. Since Fε(0,0)=0, the Sobolev chain rule [2 Theorem 4.2] gives
Here we used v0=h, v/v0=φ, and ηε(∣h∣)=1 on Eε, since ∣h∣≥ε there. Both Tf and Fε∘(v0,v) belong to W1,1(Rm), and their difference vanishes almost everywhere on Eε. By locality, the first weak derivatives of this difference also vanish almost everywhere on Eε. Hence
∂j(Tf)(y)=∂j(Fε∘(v0,v))(y)
for almost every y∈Eε.
Combining this identity with the Sobolev chain rule gives
∂j(Tf)(y)=DFε(v0(y),v(y))[∂jv0(y),∂jv(y)]
for almost every y∈Eε.
On Eε, we have
ηε(∣v0(y)∣)=1,ηε′(∣v0(y)∣)=0.
Consequently, the formula for DFε(a,b)[u,w] delivers
∂j(Tf)(y)=a0(y)f(φ(y))+i=1∑nai(y)(∂if)(φ(y))a.e. on Eε.(12)
For the fixed ball B, cutoff θ and index j, the functions v0=Tθ and vi=T(xiθ) are fixed independently of f. The formulas defining a0,…,an therefore give the same coefficient functions for every f∈Cc∞(Rn). For each such f, identity (12) holds almost everywhere on Eε.
Let α∈An,k have length at least two and consider E=Eε∩{y:πy(ej)=α}. On this set (8) identifies the left side of (12) with hej(Dαf)∘φ. Take successively the following test functions.
(i) For f=θ, Dαf and all first derivatives vanish on B, while f=1 there. Thus a0=0 on E almost everywhere.
(ii) For f=xiθ, the derivative Dαf still vanishes on B, whereas ∂ℓf=δiℓ there. With a0=0, the identity gives ai=0 for every i.
(iii) For f=xαθ, the right side of (12) is now zero on E, while Dαf=α! on B. Hence hejα!=0 almost everywhere on E.
Since hej is nonzero almost everywhere, E has measure zero. Only finitely many tests were needed for each fixed B, j, α and ε. Take ε=1/l for l∈N and let B range over the rational balls. The sets Eε obtained in this way cover the target up to a null set because h=0 almost everywhere. This excludes every α of length at least two. The value 0 is already used by πy(0), so injectivity gives ∣πy(ej)∣=1.
The m distinct first-order target indices consequently have distinct images among the n first-order source indices. Hence m≤n. Applying the same reasoning to T−1 gives n≤m. For fixed n≥1 the number ∣An,k∣ is strictly increasing in k, since each higher order adds at least one multiindex. Thus (9) also gives s=k. □
Henceforth m=n and s=k. There is a measurable permutation σy of {1,…,n} at almost every y such that
For α=0 this is (6.1), since β0(y)=0. Monotone convergence extends the formula from integrable v to all nonnegative v.
For any R>0, use as v the indicator of a measurable subset of B(0,R). Comparing the formulas for α and 0 shows that the densities ∣hβα(y)(y)∣ and ∣h(y)∣ have equal integrals on each such preimage. Since φ is invertible modulo null sets, these preimages include, up to null sets, every measurable subset of φ−1(B(0,R)). Both densities have finite integral on this set, so they agree almost everywhere there. Letting R grow yields the equality for a fixed α, i.e., for each α∈An,k, we have a null set Nα⊂Rm such that
∣hβα(y)(y)∣=∣h(y)∣(y∈/Nα).
Since An,k is finite, the union
N:=α∈An,k⋃Nα
is again a null set. Enlarge N by the null set on which πy fails to be a bijection. Thus, for every y∈/N, the preceding equality holds simultaneously for all α∈An,k.
Now fix β∈Am,s and y∈/N, and set αy:=πy(β). By the definition of the inverse indices,
βαy(y)=πy−1(πy(β))=β.
Since the equality is valid for every source index at this point y, we may apply it with α=αy to obtain
∣hβ(y)∣=∣hβαy(y)(y)∣=∣h(y)∣.
Consequently, ∣hβ∣=∣h∣ almost everywhere for every β∈Am,s, with a common exceptional null set.
Lemma 6.2.The isometry T extends to a surjective linear isometry of (W1,1(Rn),Nn,1). The extension satisfies (5.1) and (5.3) with the same base map and weights.
Proof. For f∈Wk,1, (6.1) gives ∥Tf∥1=∥f∥1. Moreover, summing (5.3) over j and using ∣hj∣=∣h∣ and the fact that σy is a permutation for almost all y, it follows that
where the last equality follows from (14). Thus T also preserves Nn,1 on Wk,1. This subspace is dense in W1,1 because it contains Cc∞. Extending T and T−1 by continuity for the first-order norm gives inverse isometries, as the compositions agree with the identity on Wk,1. Write T for the continuous extension constructed above. Fix f∈W1,1(Rn) and choose fℓ∈Cc∞(Rn) such that fℓ→f in W1,1(Rn).
Thus q↦h(q∘φ) is a continuous linear map on L1(Rn). Moreover, for any tuple (qi)i=1n with qi∈L1(Rn), the identities ∣hj∣=∣h∣ and the fact that σy is a permutation for almost every y give
The preceding identities show that g,G1,…,Gn belong to L1(Rn). The value and derivative formulas already established for T apply to every fℓ, since fℓ∈Wk,1(Rn). Applying the first norm identity to q=fℓ−f, we therefore obtain
∥Tfℓ−g∥1=∥fℓ−f∥1⟶0.
Applying the identity for tuples to qi=∂ifℓ−∂if gives
j=1∑n∥∂j(Tfℓ)−Gj∥1=i=1∑n∥∂ifℓ−∂if∥1⟶0.
In particular, ∂j(Tfℓ)→Gj in L1(Rn) for each j.
On the other hand, the continuity of T on W1,1(Rn) and its agreement with T on Wk,1(Rn) imply
Tfℓ=Tfℓ⟶Tfin W1,1(Rn).
Consequently,
Tfℓ⟶Tf,∂j(Tfℓ)⟶∂j(Tf)in L1(Rn).
Uniqueness of the L1 limit now gives Tf=g and ∂j(Tf)=Gj. Hence
(Tf)(y)=h(y)f(φ(y))
and
∂j(Tf)(y)=hj(y)(∂σy(j)f)(φ(y)),1≤j≤n,
for almost every y∈Rn. Thus both formulas remain valid for the extension, with the same map φ, the same permutations σy, and the same weights.
We now use the first-order extension to truncate a preimage under T. A Lipschitz truncation preserves W1,1, whereas for k>1 there is no such general rule on Wk,1.
Lemma 6.3 (The weight is constant). There is a constant c∈R∖{0} such that h=c almost everywhere.
Proof. We work with the first-order extension. Fix a rational ball K⊂Rn and a positive integer a. Choose u∈Cc∞(Rn) equal to a on K. By surjectivity there exists f∈W1,1 with Tf=u. The value formula gives f(φ(y))=a/h(y) on K. Also ∂ju=0 there for each j. Since all hj are nonzero and σy is a permutation, (13) yields (∂if)(φ(y))=0 for all i at almost every y∈K.
We now truncate ∣f∣. The function ψ(t)=min(∣t∣,1) is 1-Lipschitz on R and ψ(0)=0. The Sobolev composition estimate [2 Theorem 4.2] gives ψ(f)∈W1,1 and ∣∂i(ψ(f))∣≤∣∂if∣ almost everywhere. The preceding inequalities hold for all i=1,…,n outside a common Lebesgue null set N⊂Rn. Since φ is nonsingular, the inverse image φ−1(N) is also a Lebesgue null set. Thus, for almost every y∈Rn, we have φ(y)∈/N, and consequently
∣(∂i(ψ(f)))(φ(y))∣≤∣(∂if)(φ(y))∣,1≤i≤n.
Applying (13) to ψ(f) therefore gives ∇T(ψ(f))=0 almost everywhere on K, since (∂if)(φ(y))=0 for all i at almost every y∈K. Its value on K is, by (10),
T(ψ(f))(y)=h(y)min(∣h(y)∣a,1)=:ha(y).
The expression ha is independent of the choice of u and of K. Since ∇T(ψ(f))=0 almost everywhere on K, we have that T(ψ(f)) is almost everywhere constant there. For a fixed a, apply this conclusion on the countable nested balls B(0,R), R∈N. Their constants agree on overlaps, so ha=ca almost everywhere on Rn for one scalar ca.
Finally take a common full-measure set for all integers a≥1 on which 0<∣h∣<∞. At every point of this set ha(y)=h(y) once a≥∣h(y)∣. Thus the same scalar sequence (ca) converges to h(y) at every such y. Its limit gives the required nonzero c. □
Lemma 6.4. The base map has a representative φ(y)=Py+b, where P is a signed permutation matrix. The constant in Lemma 6.3 belongs to {−1,1}.
Proof. Choose θ∈Cc∞(Rn) equal to one on B(0,1). For integers R≥1 and 1≤i≤n, put fR,i(x)=xiθ(x/R). Their derivative is
∂ℓfR,i(x)=δiℓθ(x/R)+(xi/R)(∂ℓθ)(x/R).
With z=x/R, the second term is zi(∂ℓθ)(z), which has a bound independent of R. Put FR,i=c−1TfR,i=fR,i∘φ. Each FR,i is bounded and its weak first derivatives are bounded uniformly in R by (13), since ∣hj/c∣=1. The W1,∞ representative theorem [8 Theorem 4.1] gives globally Lipschitz representatives with a common Lipschitz bound.
Take a full-measure set on which all these representatives agree with fR,i∘φ. For any y in this set, FR,i(y)=φi(y) whenever R>∣φ(y)∣. Passing to the limit R→∞ in the Lipschitz inequality at two points of this set gives the Lipschitz bound for φi. Extension from this dense set gives a globally Lipschitz representative of each coordinate of φ. On {y:∣φ(y)∣<R} the functions FR,i and φi agree. By locality their derivatives agree there almost everywhere, and ∂ℓFR,i=δiℓ on B(0,R). Using (13) and exhausting in R, we obtain
Each column of Dφ has one nonzero entry, and these entries occupy distinct rows, since σy is a permutation for almost every y. Each nonzero entry is a ratio hj(y)/c of real numbers with ∣hj(y)∣=∣c∣, so it is either 1 or −1. Thus Dφ(y) is a signed permutation matrix almost everywhere. Its Euclidean operator norm is one and therefore ∣φ(y)−φ(z)∣≤∣y−z∣.
Apply the same argument to T−1, using U−1 as its ambient extension. This gives a globally 1-Lipschitz representative ψ of the inverse base map. Since both original maps are nonsingular, replacing them by their continuous representatives preserves both inverse identities almost everywhere. By continuity the identities hold everywhere. Applying the bound for ψ at φ(y) and φ(z) gives the opposite inequality, so that
∣y−z∣=∣ψ(φ(y))−ψ(φ(z))∣≤∣φ(y)−φ(z)∣≤∣y−z∣.
Thus φ is a surjective Euclidean isometry. By the Mazur–Ulam theorem [12], φ(y)=Py+b. Its linear part P preserves the Euclidean norm and is therefore orthogonal. Its derivative is the constant matrix P, and (15) therefore makes P a signed permutation matrix. Since h=c and φ(y)=Py+b, identity (14) reads
∣c∣∫Rnv(Py+b)dy=∫Rnv(x)dx.
Let B=B(0,1) and take v=1B. Then
∣c∣∫Rn1B(Py+b)dy=∣B∣.
The change of variables x=Py+b gives dy=∣detP∣−1dx, and hence
The author used OpenAI’s GPT-6 Astra to assist with mathematical exploration, the development and checking of proofs, and the presentation of arguments. The author takes full responsibility for the mathematical content and any remaining errors.
References
References
[1]A. Pełczyński and M. Wojciechowski, Sobolev spaces, in Handbook of the Geometry of Banach Spaces, Vol. 2, W. B. Johnson and J. Lindenstrauss (eds.), North-Holland, Amsterdam, 2003, pp. 1361–1423.
[2]W. Arendt and M. Kreuter, Mapping theorems for Sobolev spaces of vector-valued functions, Studia Math. 240 (2018), 275–299.DOI
[3]M. Biegert and R. Nittka, Isometries between Sobolev spaces, Math. Nachr. 285 (2012), 2059–2077. Preprint version available as arXiv:0807.0360v4.
[4]G. Diestel, Sobolev spaces with only trivial isometries, II, Positivity 13 (2009), 621–630.DOI