We construct a continuous function, strictly convex in space and strictly decreasing in time, whose composition with a bounded continuous martingale is not a semimartingale. We also construct a continuous, spatially convex, time-increasing function that is not a difference of spatially convex, time-nonincreasing functions on a rectangle. The construction combines parabolic oscillations, diffusion superposition and martingale concatenation to obtain infinite drift variation in the composition.
Introduction
Theorem 1. There are a continuous martingale X with X0=0 and ∣X∣≤2 and a continuous function G:R×[0,∞)→R, strictly convex in x and strictly decreasing in t, such that G(Xt,t) is not a semimartingale on [0,1] in the usual natural filtration of X. The derivative Gx is continuous; G is smooth for t<1 and equals x2/2−t for t≥1.
Theorem 2. There is a continuous f:R2→R, strictly convex in x and strictly increasing in t, whose restriction to Q=(−1,1)2 is not g−h for any g,h:Q→R convex in x and nonincreasing in t. Moreover, f−(x2/2+t) is compactly supported in Q, fx is continuous, and the distributional derivatives satisfy fxx,ft∈Lloc1(R2) but ft∈/L∞(Q).
Theorem 1 answers Lowther’s question [1] by completing the construction from [2]. Parabolic oscillations produce smooth, spatially convex pairs F,P with Ft>0>Pt and arbitrarily large ∫Q∣PxxFt+FxxPt∣. Convex combinations give Theorem 2; diffusion superposition and concatenation turn this divergence into infinite drift variation for Theorem 1.
Oscillation
Put F0(x,t)=x2/2+t, P0(x,t)=x2/2−t and K=(−1/2,1/2)2. Integrals over rectangles are with respect to dxdt.
Lemma 3. There are Fn,Pn∈C∞(R2), equal to F0,P0 off a fixed compact S⊂(−2/3,2/3)2, with ∥Fn−F0∥∞,∥Pn−P0∥∞≤1/16 and positive derivatives
Since a≤21min(A,B,C,D) and the derivatives equal 1 off S, large frequencies preserve positivity. The increments of A−D and B−C tend uniformly to zero.
Write ⟨w⟩=∫Kw and E=⟨A2+B2⟩. Integration by parts in x, with boundary terms O(λ−1), gives
Δ(⟨A⟩,⟨B⟩,⟨AB⟩)=o(1),ΔE=⟨a2⟩+o(1).
Starting from F0,P0, choose the frequency at step n≥0 so that the averaging errors, mismatch increments and function perturbations are each at most δ2−n−1, where δ=1/16. Then
Integrating with (3) gives 4⟨an(An+Bn)⟩≥1−3δ−2δ2=206/256, so Cauchy–Schwarz gives
51<⟨an(An+Bn)⟩≤(2En⟨an2⟩)1/2.
If ∑n⟨an2⟩<∞, (3) bounds En, but the last inequality bounds each summand away from zero, a contradiction. Thus En→∞. On K, ∣CnBn−AnDn∣≥∣Bn2−An2∣−δ(An+Bn)≥(An−Bn)2−δ(An+Bn), so
Jn≥En−2⟨AnBn⟩−δ⟨An+Bn⟩≥En−3⟶∞.
Boundary agreement gives the integral identities in (2); convexity and the slopes ±1 at x=±1 give the slope bounds. Integration by parts gives ∫QAnDn=∫QCnBn; positivity then gives 2In≥Jn.
Proof of Theorem 2. Choose increasing indices nj with Inj≥4j and set
f=F0+j≥1∑2−j(Fnj−F0)=j≥1∑2−jFnj.
The perturbation series converges uniformly and is supported in S. Positivity gives strict convexity and strict time increase. The bounds (2) give continuity of fx and
Suppose f=g−h on Q with g,h convex in x and nonincreasing in t. For u=g,h, spatial continuity and time monotonicity give Borel measurability and local boundedness: on [a,b]×[c,d]⋐Q, u(x,d)≤u(x,t)≤u(x,c). Thus μu=uxx and νu=−ut are positive Radon measures. Choose an open R with S⊂R⋐Q. Since Pn−P0∈C∞(R), integration by parts gives
∫RDndμu+∫RCndνu=μu(R)+νu(R).
Since μg=fxxdxdt+μh, positivity implies ∫RDnfxx≤μg(R)+νg(R)<∞, uniformly in n. Since Dn=fxx=1 almost everywhere on Q∖R, this contradicts (4). For fixed n, integration by parts for the partial sums and L1 convergence give ∫QDnfxx=∫QCnft. As ∫QCn=4, (4) also forces ft∈/L∞(Q).
Martingale blocks
Lemma 4. There are smooth pn on [−1,1]×[0,1], equal to P0 off a fixed compact subset of (−1,1)×(0,1), and continuous martingales Zn such that
satisfies EWn<∞ and EWn→∞, where Wn=Var[0,1]Vn.
Proof. On Q, set
An=An+7−4t,Bn=Bn+1−2x2.
The support condition gives An>0 on Q, Bn>0 in Q and Bn(±1,t)=0. For −1≤t≤1, define the mean-zero probability laws
μtn=48An(x,t)1(−1,1)(x)dx+124+t(δ−1+δ1)
and the coefficient an=2Bn/An on ∣x∣<1, zero elsewhere. The laws are narrowly continuous and the coefficient is bounded for each n. Since ∂tAn=∂xxBn, Bn(±1,t)=0 and ∂xBn(±1,t)=∓4, integration by parts gives, for φ∈C∞(R),
Trevisan’s superposition theorem [3 Theorem 2.5] gives a continuous solution Yn of the martingale problem for 21an∂xx with marginals μtn. The marginal supports imply ∣Yn∣≤1. Smooth compactly supported test functions equal to x and x2 on [−1,1] show that Yn is a martingale with d[Yn]t=an(Ytn,t)dt. If Un is the finite-variation part of Pn(Ytn,t), Itô’s formula gives
Here we discard the nonnegative atomic contribution and use (2).
Set
pn(z,s)=P0(z,s)+41(Pn−P0)(2z,4s−2).
Then (pn)zz=Cn and (pn)s=−Dn at the rescaled point; convexity and (pn)z(±1,s)=±1 give the slope bound. On [1/4,3/4], take Zsn=Y4s−2n/2. Prepend the continuous Brownian martingale E[h(B1)∣FsB], choosing h(B1) with the law of Y−1n/2. Given h(B1), attach Yn using its regular conditional law given Y−1n=2h(B1), independently of the prefix. These conditional laws are martingale laws for almost every initial value, so the join preserves the martingale property. Append the current endpoint plus an independent Brownian motion, stopped at {−1,1} with clock (s−3/4)/(1−s). It exits before s=1 almost surely. Thus Zn is a continuous martingale in its natural filtration.
On the middle interval, pn(Zsn,s)=41Pn(Y4s−2n,4s−2)−21, so (6) gives EWn→∞. Finally, E[Zn]1=1 and smoothness give EWn≤∥(pn)s∥∞+21∥(pn)zz∥∞<∞. □
Proof of Theorem 1. For each m≥1, choose a block with EW>2⋅4m. By the weak law, choose increasing integers rm so that 4rm independent copies have sample average exceeding 4m with probability at least 1−2−m. Take the groups independent, scale group m by ϵ=32−m−rm, and enumerate the copies in increasing group order as (pk,Zk,Vk,Wk,ϵk). The scales satisfy
k∑ϵk2=1,ϵk/ϵk+1∈N,k∑ϵk2Wk=∞a.s.(7)
Each group has total squared scale 3⋅4−m. By Borel–Cantelli, all but finitely many groups succeed, each contributing more than 3.
Following [2], put T0=0, Tk=∑j≤kϵj2 and concatenate:
This is a continuous martingale with EMt2≤1. The endpoint series converges almost surely and in L2, and within-block oscillations are at most 2ϵk→0. Thus M extends continuously to 1, with EM12=1; hold it constant thereafter.
Put Λk=∑j<kϵj+2ϵkZ. By (7), MTk−1∈Λk. On each tile [a−ϵk,a+ϵk]×[Tk−1,Tk], a∈Λk, define
Near tile boundaries, G=P0, so G is smooth for t<1, with Gxx>0 and Gt<0. In block k,
∣G−P0∣≤ϵk2/64,∣Gx−x∣≤2ϵk.
Extend G=P0 for t≥1. Since ϵk→0, G,Gx are continuous at t=1, and strict convexity and time decrease persist: G(x,t)>lims↑1G(x,s)=P0(x,1) for t<1.
Stop M at τ=inf{t:∣Mt∣≥2} and put X=Mτ. The maximal inequality gives P(τ>1)≥3/4. In the usual natural filtration of X, define on [0,1]
At=G(Xt,t)−G(0,0)−∫0tGx(Xs,s)dXs.
The integral is a continuous martingale since Gx is bounded on [−2,2]×[0,1]. Itô’s formula gives, on each [0,Tk],
At=∫0tGt(Xs,s)ds+21∫0tGxx(Xs,s)d[X]s.
Since ∣Zk∣≤1, M stays in its starting tile during block k. On {τ>1}, X=M and d[M]=ϵk2d[Zk] in rescaled time, so (8) and (5) give drift increment ϵk2Vk in block k. Consequently,
Var[0,Tk]A=j≤k∑ϵj2Wj⟶∞a.s. on {τ>1}.
If G(Xt,t) were a semimartingale on [0,1], its continuous finite-variation part, starting at zero, would equal A on each [0,Tk] by uniqueness. Its variation on [0,1] would dominate all these sums, a contradiction. □
[1]G. Lowther, Do convex and decreasing functions preserve the semimartingale property? MathOverflow, Question 43319 (2010).
[2]G. Lowther, Do convex and decreasing functions preserve the semimartingale property? Almost Sure, 26 September 2016. https://almostsuremath.com/2016/09/26/do-convex-and-decreasing-functions-preserve-the-semimartingale-property/.
[3]D. Trevisan, Well-posedness of multidimensional diffusion processes with weakly differentiable coefficients, Electron. J. Probab. 21 (2016), no. 22, 1–41.arxiv.org/abs/1507.01357