betaFeedback

WorksMathematics

Convex, time-decreasing functions need not preserve semimartingales

Abstract

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 XX with X0=0X_0 = 0 and ∣X∣≤2|X| \le2 and a continuous function G:R×[0,∞)→RG : \mathbb{R} \times[0,\infty) \to\mathbb{R}, strictly convex in xx and strictly decreasing in tt, such that G(Xt,t)G(X_t,t) is not a semimartingale on [0,1][0,1] in the usual natural filtration of XX. The derivative GxG_x is continuous; GG is smooth for t<1t < 1 and equals x2/2−tx^2/2 - t for t≥1t \ge1.

Theorem 2. There is a continuous f:R2→Rf : \mathbb{R}^2 \to\mathbb{R}, strictly convex in xx and strictly increasing in tt, whose restriction to Q=(−1,1)2Q = (-1,1)^2 is not g−hg-h for any g,h:Q→Rg,h : Q \to\mathbb{R} convex in xx and nonincreasing in tt. Moreover, f−(x2/2+t)f-(x^2/2+t) is compactly supported in QQ, fxf_x is continuous, and the distributional derivatives satisfy fxx,ft∈Lloc1(R2)f_{xx},f_t \in L^1_{\mathrm{loc}}(\mathbb{R}^2) but ft∉L∞(Q)f_t \notin L^\infty(Q).

Theorem 1 answers Lowther’s question [1] by completing the construction from [2]. Parabolic oscillations produce smooth, spatially convex pairs F,PF,P with Ft>0>PtF_t > 0 > P_t and arbitrarily large ∫Q∣PxxFt+FxxPt∣\int_Q |P_{xx}F_t+F_{xx}P_t|. 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+tF_0(x,t)=x^2/2+t, P0(x,t)=x2/2−tP_0(x,t)=x^2/2-t and K=(−1/2,1/2)2K=(-1/2,1/2)^2. Integrals over rectangles are with respect to dx dtdx\,dt.

Lemma 3. There are Fn,Pn∈C∞(R2)F_n,P_n \in C^\infty(\mathbb{R}^2), equal to F0,P0F_0,P_0 off a fixed compact S⊂(−2/3,2/3)2S \subset(-2/3,2/3)^2, with ∥Fn−F0∥∞,∥Pn−P0∥∞≤1/16\|F_n-F_0\|_\infty,\|P_n-P_0\|_\infty\le1/16 and positive derivatives

(An,Bn,Cn,Dn)=((Fn)xx,(Fn)t,(Pn)xx,−(Pn)t).(A_n,B_n,C_n,D_n)=((F_n)_{xx},(F_n)_t,(P_n)_{xx},-(P_n)_t).

They satisfy

Jn:=∫Q∣CnBn−AnDn∣⟶∞,In:=∫QAnDn=∫QCnBn≥12Jn.(1)J_n:=\int_Q |C_nB_n-A_nD_n| \longrightarrow\infty,\qquad I_n:=\int_Q A_nD_n=\int_Q C_nB_n\ge\frac{1}{2}J_n. \tag*{(1)}

and

∫QAn=∫QBn=∫QCn=∫QDn=4,∥(Fn)x−x∥∞,∥(Pn)x−x∥∞≤2.(2)\int_Q A_n=\int_Q B_n=\int_Q C_n=\int_Q D_n=4,\qquad\|(F_n)_x-x\|_\infty,\|(P_n)_x-x\|_\infty\le2. \tag*{(2)}

Proof. Choose χ∈C∞((−2/3,2/3)2)\chi\in C^\infty((-2/3,2/3)^2) with 0≤χ≤10 \le\chi\le1 and χ=1\chi=1 on KK, and set S=supp⁡χS=\operatorname{supp}\chi. For the current pair F,PF,P, write (A,B,C,D)=(Fxx,Ft,Pxx,−Pt)(A,B,C,D)=(F_{xx},F_t,P_{xx},-P_t) and put

a=χ2(A−1+B−1+C−1+D−1),θ=λx+λ2t.a=\frac{\chi}{2(A^{-1}+B^{-1}+C^{-1}+D^{-1})},\qquad\theta=\lambda x+\lambda^2t.

Replace F,PF,P by F′=F−λ−2acos⁡θF'=F-\lambda^{-2}a\cos\theta and P′=P−λ−2asin⁡θP'=P-\lambda^{-2}a\sin\theta. For fixed F,PF,P, uniformly as λ→∞\lambda\to\infty,

(A′,B′,C′,D′)=(A,B,C,D)+a(cos⁡θ,sin⁡θ,sin⁡θ,cos⁡θ)+O(λ−1).(A',B',C',D')=(A,B,C,D)+a(\cos\theta,\sin\theta,\sin\theta,\cos\theta)+O(\lambda^{-1}).

Since a≤12min⁡(A,B,C,D)a\le\frac{1}{2}\min(A,B,C,D) and the derivatives equal 11 off SS, large frequencies preserve positivity. The increments of A−DA-D and B−CB-C tend uniformly to zero.

Write ⟨w⟩=∫Kw\langle w\rangle=\int_K w and E=⟨A2+B2⟩E=\langle A^2+B^2\rangle. Integration by parts in xx, with boundary terms O(λ−1)O(\lambda^{-1}), gives

Δ(⟨A⟩,⟨B⟩,⟨AB⟩)=o(1),ΔE=⟨a2⟩+o(1).\Delta(\langle A\rangle,\langle B\rangle,\langle AB\rangle)=o(1),\qquad\Delta E=\langle a^2\rangle+o(1).

Starting from F0,P0F_0,P_0, choose the frequency at step n≥0n\ge0 so that the averaging errors, mismatch increments and function perturbations are each at most δ2−n−1\delta2^{-n-1}, where δ=1/16\delta=1/16. Then

∥An−Dn∥∞, ∥Bn−Cn∥∞≤δ,∣⟨An⟩−1∣, ∣⟨Bn⟩−1∣, ∣⟨AnBn⟩−1∣≤δ,En=2+∑j<n⟨aj2⟩+en,∣en∣≤δ.(3)\begin{aligned} \|A_n-D_n\|_\infty,\ \|B_n-C_n\|_\infty&\le\delta,\qquad |\langle A_n\rangle-1|,\ |\langle B_n\rangle-1|,\ |\langle A_nB_n\rangle-1|\le\delta,\\ E_n&=2+\sum_{j<n}\langle a_j^2\rangle+e_n,\qquad|e_n|\le\delta. \tag*{(3)} \end{aligned}

On KK, the bounds Cn≥Bn−δC_n\ge B_n-\delta, Dn≥An−δD_n\ge A_n-\delta imply

4an(An+Bn)≥(An−δ)+(Bn−δ)+≥AnBn−δ(An+Bn).4a_n(A_n+B_n)\ge(A_n-\delta)_+(B_n-\delta)_+\ge A_nB_n-\delta(A_n+B_n).

Integrating with (3) gives 4⟨an(An+Bn)⟩≥1−3δ−2δ2=206/2564\langle a_n(A_n+B_n)\rangle\ge1-3\delta-2\delta^2=206/256, so Cauchy–Schwarz gives

15<⟨an(An+Bn)⟩≤(2En⟨an2⟩)1/2.\frac{1}{5}<\langle a_n(A_n+B_n)\rangle\le(2E_n\langle a_n^2\rangle)^{1/2}.

If ∑n⟨an2⟩<∞\sum_n\langle a_n^2\rangle<\infty, (3) bounds EnE_n, but the last inequality bounds each summand away from zero, a contradiction. Thus En→∞E_n\to\infty. On KK, ∣CnBn−AnDn∣≥∣Bn2−An2∣−δ(An+Bn)≥(An−Bn)2−δ(An+Bn)|C_nB_n-A_nD_n|\ge|B_n^2-A_n^2|-\delta(A_n+B_n)\ge(A_n-B_n)^2-\delta(A_n+B_n), so

Jn≥En−2⟨AnBn⟩−δ⟨An+Bn⟩≥En−3⟶∞.J_n\ge E_n-2\langle A_nB_n\rangle-\delta\langle A_n+B_n\rangle\ge E_n-3\longrightarrow\infty.

Boundary agreement gives the integral identities in (2); convexity and the slopes ±1\pm1 at x=±1x=\pm1 give the slope bounds. Integration by parts gives ∫QAnDn=∫QCnBn\int_Q A_nD_n=\int_Q C_nB_n; positivity then gives 2In≥Jn2I_n\ge J_n.

Proof of Theorem 2. Choose increasing indices njn_j with Inj≥4jI_{n_j}\ge4^j and set

f=F0+∑j≥12−j(Fnj−F0)=∑j≥12−jFnj.f=F_0+\sum_{j\ge1}2^{-j}(F_{n_j}-F_0)=\sum_{j\ge1}2^{-j}F_{n_j}.

The perturbation series converges uniformly and is supported in SS. Positivity gives strict convexity and strict time increase. The bounds (2) give continuity of fxf_x and

fxx=∑j≥12−jAnj,ft=∑j≥12−jBnjin Lloc1(R2).f_{xx}=\sum_{j\ge1}2^{-j}A_{n_j},\qquad f_t=\sum_{j\ge1}2^{-j}B_{n_j}\quad\text{in }L^1_{\mathrm{loc}}(\mathbb{R}^2).

In particular,

∫QDnjfxx≥2−jInj≥2j.(4)\int_Q D_{n_j}f_{xx}\ge2^{-j}I_{n_j}\ge2^j. \tag*{(4)}

Suppose f=g−hf=g-h on QQ with g,hg,h convex in xx and nonincreasing in tt. For u=g,hu=g,h, spatial continuity and time monotonicity give Borel measurability and local boundedness: on [a,b]×[c,d]⋐Q[a,b]\times[c,d]\Subset Q, u(x,d)≤u(x,t)≤u(x,c)u(x,d)\le u(x,t)\le u(x,c). Thus μu=uxx\mu_u=u_{xx} and νu=−ut\nu_u=-u_t are positive Radon measures. Choose an open RR with S⊂R⋐QS\subset R\Subset Q. Since Pn−P0∈C∞(R)P_n-P_0\in C^\infty(R), integration by parts gives

∫RDn dμu+∫RCn dνu=μu(R)+νu(R).\int_R D_n\,d\mu_u+\int_R C_n\,d\nu_u=\mu_u(R)+\nu_u(R).

Since μg=fxx dx dt+μh\mu_g=f_{xx}\,dx\,dt+\mu_h, positivity implies ∫RDnfxx≤μg(R)+νg(R)<∞\int_R D_nf_{xx}\le\mu_g(R)+\nu_g(R)<\infty, uniformly in nn. Since Dn=fxx=1D_n=f_{xx}=1 almost everywhere on Q∖RQ\setminus R, this contradicts (4). For fixed nn, integration by parts for the partial sums and L1L^1 convergence give ∫QDnfxx=∫QCnft\int_QD_nf_{xx}=\int_QC_nf_t. As ∫QCn=4\int_QC_n=4, (4) also forces ft∉L∞(Q)f_t\notin L^\infty(Q).

Martingale blocks

Lemma 4. There are smooth pnp_n on [−1,1]×[0,1][-1,1] \times[0,1], equal to P0P_0 off a fixed compact subset of (−1,1)×(0,1)(-1,1) \times(0,1), and continuous martingales ZnZ^n such that

(pn)zz>0,(pn)s<0,∥(pn)z∥∞≤1,∥pn−P0∥∞≤1/64,(p_n)_{zz} > 0,\qquad(p_n)_s < 0,\qquad\lVert(p_n)_z\rVert_\infty\le1,\qquad\lVert p_n-P_0\rVert_\infty\le1/64,
Z0n=0,∣Zn∣≤1,Z1n∈{−1,1}.Z^n_0=0,\qquad\lvert Z^n\rvert\le1,\qquad Z^n_1 \in\{-1,1\}.

The finite-variation part VnV^n in

pn(Zsn,s)=pn(0,0)+∫0s(pn)z(Zrn,r) dZrn+Vsn(5)p_n(Z^n_s,s)=p_n(0,0)+\int_0^s (p_n)_z(Z^n_r,r)\,\mathrm{d}Z^n_r+V^n_s \tag*{(5)}

satisfies EWn<∞\mathbb{E}W_n<\infty and EWn→∞\mathbb{E}W_n\to\infty, where Wn=Var⁡[0,1]VnW_n=\operatorname{Var}_{[0,1]}V^n.

Proof. On Q‾\overline{Q}, set

A^n=An+7−4t,B^n=Bn+1−2x2.\widehat{A}_n=A_n+7-4t,\qquad\widehat{B}_n=B_n+1-2x^2.

The support condition gives A^n>0\widehat{A}_n>0 on Q‾\overline{Q}, B^n>0\widehat{B}_n>0 in QQ and B^n(±1,t)=0\widehat{B}_n(\pm1,t)=0. For −1≤t≤1-1\le t\le1, define the mean-zero probability laws

μtn=A^n(x,t)481(−1,1)(x) dx+4+t12(δ−1+δ1)\mu^n_t=\frac{\widehat{A}_n(x,t)}{48}1_{(-1,1)}(x)\,\mathrm{d}x+\frac{4+t}{12}(\delta_{-1}+\delta_1)

and the coefficient an=2B^n/A^na_n=2\widehat{B}_n/\widehat{A}_n on ∣x∣<1\lvert x\rvert<1, zero elsewhere. The laws are narrowly continuous and the coefficient is bounded for each nn. Since ∂tA^n=∂xxB^n\partial_t\widehat{A}_n=\partial_{xx}\widehat{B}_n, B^n(±1,t)=0\widehat{B}_n(\pm1,t)=0 and ∂xB^n(±1,t)=∓4\partial_x\widehat{B}_n(\pm1,t)=\mp4, integration by parts gives, for φ∈C∞(R)\varphi\in C^\infty(\mathbb{R}),

ddt∫φ dμtn=148∫−11φ ∂xxB^n dx+φ(−1)+φ(1)12=148∫−11B^nφ′′ dx,\frac{\mathrm{d}}{\mathrm{d}t}\int\varphi\,\mathrm{d}\mu^n_t=\frac{1}{48}\int_{-1}^{1}\varphi\,\partial_{xx}\widehat{B}_n\,\mathrm{d}x+\frac{\varphi(-1)+\varphi(1)}{12}=\frac{1}{48}\int_{-1}^{1}\widehat{B}_n\varphi''\,\mathrm{d}x,

that is,

∂tμtn=12∂xx(anμtn).\partial_t\mu^n_t=\frac{1}{2}\partial_{xx}(a_n\mu^n_t).

Trevisan’s superposition theorem [3 Theorem 2.5] gives a continuous solution YnY^n of the martingale problem for 12an∂xx\frac{1}{2}a_n\partial_{xx} with marginals μtn\mu^n_t. The marginal supports imply ∣Yn∣≤1\lvert Y^n\rvert\le1. Smooth compactly supported test functions equal to xx and x2x^2 on [−1,1][-1,1] show that YnY^n is a martingale with d[Yn]t=an(Ytn,t) dt\mathrm{d}[Y^n]_t=a_n(Y^n_t,t)\,\mathrm{d}t. If UnU^n is the finite-variation part of Pn(Ytn,t)P_n(Y^n_t,t), Itô’s formula gives

EVar⁡[−1,1]Un≥148∫Q∣CnB^n−DnA^n∣≥Jn48−1⟶∞.(6)\mathbb{E}\operatorname{Var}_{[-1,1]}U^n\ge\frac{1}{48}\int_Q\lvert C_n\widehat{B}_n-D_n\widehat{A}_n\rvert\ge\frac{J_n}{48}-1\longrightarrow\infty. \tag*{(6)}

Here we discard the nonnegative atomic contribution and use (2).

Set

pn(z,s)=P0(z,s)+14(Pn−P0)(2z,4s−2).p_n(z,s)=P_0(z,s)+\frac{1}{4}(P_n-P_0)(2z,4s-2).

Then (pn)zz=Cn(p_n)_{zz}=C_n and (pn)s=−Dn(p_n)_s=-D_n at the rescaled point; convexity and (pn)z(±1,s)=±1(p_n)_z(\pm1,s)=\pm1 give the slope bound. On [1/4,3/4][1/4,3/4], take Zsn=Y4s−2n/2Z^n_s=Y^n_{4s-2}/2. Prepend the continuous Brownian martingale E[h(B1)∣FsB]\mathbb{E}[h(B_1)\mid\mathcal{F}^{B}_s], choosing h(B1)h(B_1) with the law of Y−1n/2Y^n_{-1}/2. Given h(B1)h(B_1), attach YnY^n using its regular conditional law given Y−1n=2h(B1)Y^n_{-1}=2h(B_1), 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}\{-1,1\} with clock (s−3/4)/(1−s)(s-3/4)/(1-s). It exits before s=1s=1 almost surely. Thus ZnZ^n is a continuous martingale in its natural filtration.

On the middle interval, pn(Zsn,s)=14Pn(Y4s−2n,4s−2)−12p_n(Z^n_s,s)=\frac{1}{4}P_n(Y^n_{4s-2},4s-2)-\frac{1}{2}, so (6) gives EWn→∞\mathbb{E}W_n\to\infty. Finally, E[Zn]1=1\mathbb{E}[Z^n]_1=1 and smoothness give EWn≤∥(pn)s∥∞+12∥(pn)zz∥∞<∞\mathbb{E}W_n\le\lVert(p_n)_s\rVert_\infty+\frac{1}{2}\lVert(p_n)_{zz}\rVert_\infty<\infty. □

Proof of Theorem 1. For each m≥1m \ge1, choose a block with EW⁡>2⋅4m\operatorname{EW} > 2 \cdot4^m. By the weak law, choose increasing integers rmr_m so that 4rm4^{r_m} independent copies have sample average exceeding 4m4^m with probability at least 1−2−m1 - 2^{-m}. Take the groups independent, scale group mm by ϵ=3 2−m−rm\epsilon= \sqrt{3}\,2^{-m-r_m}, and enumerate the copies in increasing group order as (pk,Zk,Vk,Wk,ϵk)(p^k, Z^k, V^k, W^k, \epsilon_k). The scales satisfy

∑kϵk2=1,ϵk/ϵk+1∈N,∑kϵk2Wk=∞a.s.(7)\sum_k \epsilon_k^2 = 1,\qquad\epsilon_k/\epsilon_{k+1} \in\mathbb{N},\qquad\sum_k \epsilon_k^2 W^k = \infty\quad\text{a.s.} \tag*{(7)}

Each group has total squared scale 3⋅4−m3 \cdot4^{-m}. By Borel–Cantelli, all but finitely many groups succeed, each contributing more than 33.

Following [2], put T0=0T_0 = 0, Tk=∑j≤kϵj2T_k = \sum_{j \le k} \epsilon_j^2 and concatenate:

Mt=∑j<kϵjZ1j+ϵkZ(t−Tk−1)/ϵk2k,Tk−1≤t≤Tk.M_t = \sum_{j<k} \epsilon_j Z_1^j + \epsilon_k Z^k_{(t-T_{k-1})/\epsilon_k^2},\qquad T_{k-1} \le t \le T_k.

This is a continuous martingale with EMt2≤1\mathbb{E}M_t^2 \le1. The endpoint series converges almost surely and in L2L^2, and within-block oscillations are at most 2ϵk→02\epsilon_k \to0. Thus MM extends continuously to 11, with EM12=1\mathbb{E}M_1^2 = 1; hold it constant thereafter.

Put Λk=∑j<kϵj+2ϵkZ\Lambda_k = \sum_{j<k} \epsilon_j + 2\epsilon_k\mathbb{Z}. By (7), MTk−1∈ΛkM_{T_{k-1}} \in\Lambda_k. On each tile [a−ϵk,a+ϵk]×[Tk−1,Tk][a-\epsilon_k,a+\epsilon_k]\times[T_{k-1},T_k], a∈Λka \in\Lambda_k, define

G(x,t)=ax−a22−Tk−1+ϵk2pk(x−aϵk,t−Tk−1ϵk2).(8)G(x,t) = ax - \frac{a^2}{2} - T_{k-1} + \epsilon_k^2p^k\left(\frac{x-a}{\epsilon_k},\frac{t-T_{k-1}}{\epsilon_k^2}\right). \tag*{(8)}

Near tile boundaries, G=P0G=P_0, so GG is smooth for t<1t<1, with Gxx>0G_{xx}>0 and Gt<0G_t<0. In block kk,

∣G−P0∣≤ϵk2/64,∣Gx−x∣≤2ϵk.|G-P_0| \le\epsilon_k^2/64,\qquad|G_x-x| \le2\epsilon_k.

Extend G=P0G=P_0 for t≥1t\ge1. Since ϵk→0\epsilon_k \to0, G,GxG,G_x are continuous at t=1t=1, and strict convexity and time decrease persist: G(x,t)>lim⁡s↑1G(x,s)=P0(x,1)G(x,t) > \lim_{s \uparrow1}G(x,s) = P_0(x,1) for t<1t<1.

Stop MM at τ=inf⁡{t:∣Mt∣≥2}\tau= \inf\{t: |M_t| \ge2\} and put X=MτX=M^\tau. The maximal inequality gives P(τ>1)≥3/4\mathbb{P}(\tau>1) \ge3/4. In the usual natural filtration of XX, define on [0,1][0,1]

At=G(Xt,t)−G(0,0)−∫0tGx(Xs,s) dXs.A_t = G(X_t,t)-G(0,0)-\int_0^t G_x(X_s,s)\,dX_s.

The integral is a continuous martingale since GxG_x is bounded on [−2,2]×[0,1][-2,2]\times[0,1]. Itô’s formula gives, on each [0,Tk][0,T_k],

At=∫0tGt(Xs,s) ds+12∫0tGxx(Xs,s) d[X]s.A_t = \int_0^t G_t(X_s,s)\,ds + \frac{1}{2}\int_0^t G_{xx}(X_s,s)\,d[X]_s.

Since ∣Zk∣≤1|Z^k|\le1, MM stays in its starting tile during block kk. On {τ>1}\{\tau>1\}, X=MX=M and d[M]=ϵk2 d[Zk]d[M]=\epsilon_k^2\,d[Z^k] in rescaled time, so (8) and (5) give drift increment ϵk2Vk\epsilon_k^2V^k in block kk. Consequently,

Var⁡[0,Tk]A=∑j≤kϵj2Wj⟶∞a.s. on {τ>1}.\operatorname{Var}_{[0,T_k]} A = \sum_{j\le k}\epsilon_j^2W^j \longrightarrow\infty\quad\text{a.s. on }\{\tau>1\}.

If G(Xt,t)G(X_t,t) were a semimartingale on [0,1][0,1], its continuous finite-variation part, starting at zero, would equal AA on each [0,Tk][0,T_k] by uniqueness. Its variation on [0,1][0,1] would dominate all these sums, a contradiction. □

References

Email address: [email protected]

References

  1. [1]G. Lowther, Do convex and decreasing functions preserve the semimartingale property? MathOverflow, Question 43319 (2010).
  2. [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. [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

Paper details

vlnc:W_hgme9kwz
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents