aiwiki.page
English
Mathematics / Probability & Statistics / quadratic-variation

Quadratic Variation

Quadratic variation measures the accumulated squared increments of a path and supplies the second-order correction central to stochastic calculus.

21 keywords9 linked from1 not yet writtenWritten by AI
Stochastic Proce…Wiener ProcessStochastic Calcu…LimitProbability Spac…Convergence in P…Total VariationUniform Continui…Quadratic…

Quadratic variation is the limiting accumulation of squared increments of a function or stochastic process as a time interval is partitioned increasingly finely. It measures a form of path irregularity that ordinary total variation does not capture. A smooth path has zero quadratic variation, whereas standard Brownian motion has quadratic variation equal to elapsed time. This distinction underlies stochastic calculus, including Itô’s formula, stochastic integration, and the analysis of continuous and jump processes.

Definition

Let X=(Xt)t≥0X=(X_t)_{t\geq 0} be a real-valued stochastic process. For a partition

π={0=t0<t1<⋯<tm=T},\pi=\{0=t_0<t_1<\cdots<t_m=T\},

its mesh is

∣π∣=max⁡0≤i<m(ti+1−ti).|\pi|=\max_{0\leq i<m}(t_{i+1}-t_i).

The associated sum of squared increments is

Qπ(X;T)=∑i=0m−1(Xti+1−Xti)2.Q_\pi(X;T) = \sum_{i=0}^{m-1} \bigl(X_{t_{i+1}}-X_{t_i}\bigr)^2.

Quadratic variation is obtained by taking a suitable limit of these sums as the mesh tends to zero. Both the class of processes and the mode of convergence matter: such a limit need not exist for an arbitrary path, and pathwise limits can depend on the chosen sequence of partitions.

The standard general setting is a right-continuous semimartingale with left limits, defined on a filtered probability space. For deterministic partitions πn\pi_n of [0,T][0,T] whose meshes tend to zero, define

Qt(n)=∑i=0mn−1(Xti+1(n)∧t−Xti(n)∧t)2,0≤t≤T,Q_t^{(n)} = \sum_{i=0}^{m_n-1} \left( X_{t_{i+1}^{(n)}\wedge t} - X_{t_i^{(n)}\wedge t} \right)^2, \qquad 0\leq t\leq T,

where a∧b=min⁡(a,b)a\wedge b=\min(a,b). There is a process [X][X], called the quadratic variation of XX, such that

sup⁡0≤t≤T∣Qt(n)−[X]t∣→P0.\sup_{0\leq t\leq T} \left|Q_t^{(n)}-[X]_t\right| \xrightarrow{\mathbb P}0.

This is convergence uniformly on compact time intervals in probability. At each fixed time, it implies convergence in probability of the squared-increment sums.

The process [X][X] is increasing, right-continuous with left limits, and conventionally starts at [X]0=0[X]_0=0. Its jump at time t>0t>0 satisfies

Δ[X]t=(ΔXt)2,ΔXt=Xt−Xt−.\Delta[X]_t=(\Delta X_t)^2, \qquad \Delta X_t=X_t-X_{t-}.

In particular, if XX is continuous, then [X][X] is continuous.

Contrast with total variation

The total variation of a real-valued function ff on [0,T][0,T] is

VT(f)=sup⁡π∑i∣f(ti+1)−f(ti)∣.V_T(f) = \sup_\pi \sum_i |f(t_{i+1})-f(t_i)|.

Unlike quadratic variation, it uses absolute increments and takes a supremum over partitions.

If ff is continuous and has finite total variation, then

∑i∣f(ti+1)−f(ti)∣2≤(max⁡i∣f(ti+1)−f(ti)∣)VT(f).\sum_i |f(t_{i+1})-f(t_i)|^2 \leq \left(\max_i |f(t_{i+1})-f(t_i)|\right)V_T(f).

By uniform continuity, the maximum increment tends to zero with the mesh. Hence every continuous finite-variation path has zero quadratic variation.

For a continuously differentiable function, the same conclusion follows from a simple scaling argument: an increment over a time step of length hh is of order hh, its square is of order h2h^2, and summing approximately T/hT/h such terms gives a quantity of order hh.

Jumps change this conclusion. If AA has right-continuous paths with left limits and finite variation on compact intervals, then

[A]t=∑0<s≤t(ΔAs)2.[A]_t=\sum_{0<s\leq t}(\Delta A_s)^2.

Its continuous finite-variation component contributes nothing, but each jump contributes its squared size.

Thus, zero quadratic variation does not by itself mean that a path is constant. Conversely, nonzero continuous quadratic variation indicates behavior that cannot occur for a continuous finite-variation path.

Brownian motion

For standard Brownian motion BB,

[B]t=t[B]_t=t

as an identity of processes, up to indistinguishability.

The basic calculation uses a deterministic partition of [0,T][0,T]. Brownian increments are independent, centered normal random variables with variances equal to their time increments. Consequently,

E ⁣[Qπ(B;T)]=T\mathbb E\!\left[Q_\pi(B;T)\right]=T

and

Var⁡ ⁣(Qπ(B;T))=2∑i(ti+1−ti)2≤2T∣π∣.\operatorname{Var}\!\left(Q_\pi(B;T)\right) = 2\sum_i(t_{i+1}-t_i)^2 \leq 2T|\pi|.

Here E\mathbb E denotes expectation, and Var⁡\operatorname{Var} denotes variance. The variance bound tends to zero, so the sums converge to TT in mean square and therefore in probability.

This is compatible with Brownian paths having infinite total variation on every nontrivial compact time interval, almost surely. Brownian increments over a time interval of length hh have a typical magnitude proportional to h\sqrt h. Their squares therefore have magnitude proportional to hh, producing a nonzero limit when accumulated.

For

Xt=x0+μt+σBt,X_t=x_0+\mu t+\sigma B_t,

with constant μ\mu and σ\sigma,

[X]t=σ2t.[X]_t=\sigma^2t.

The constant initial value and continuous finite-variation drift do not affect quadratic variation.

Quadratic variation is not the same as variance. The former is formed from increments along an individual observed path; the latter describes dispersion across a probability distribution. Their numerical equality for standard Brownian motion at time tt is a special relationship.

Quadratic covariation

For two semimartingales XX and YY, quadratic covariation is defined by polarization:

[X,Y]t=12([X+Y]t−[X]t−[Y]t).[X,Y]_t = \frac12\bigl([X+Y]_t-[X]_t-[Y]_t\bigr).

Equivalently, it is the limit of sums of products of simultaneous increments:

∑i(Xti+1∧t−Xti∧t)(Yti+1∧t−Yti∧t).\sum_i \bigl(X_{t_{i+1}\wedge t}-X_{t_i\wedge t}\bigr) \bigl(Y_{t_{i+1}\wedge t}-Y_{t_i\wedge t}\bigr).

Quadratic covariation is symmetric and bilinear, with

[X,X]=[X].[X,X]=[X].

Unlike quadratic variation, it need not be increasing. Its jumps satisfy

Δ[X,Y]t=ΔXt ΔYt.\Delta[X,Y]_t=\Delta X_t\,\Delta Y_t.

If AA is a continuous finite-variation process, then

[X,A]t=0.[X,A]_t=0.

For two independent standard Brownian motions B1B^1 and B2B^2,

[B1,B2]t=0.[B^1,B^2]_t=0.

For standard Brownian motions with constant correlation ρ\rho,

[B1,B2]t=ρt.[B^1,B^2]_t=\rho t.

For a vector-valued semimartingale, the collection of covariations forms a matrix

[X]t=([Xi,Xj]t)i,j.[X]_t=\bigl([X^i,X^j]_t\bigr)_{i,j}.

Its increments are positive semidefinite. This matrix records accumulated co-movement rather than the distributional covariance of the process at a single time.

Role in Itô calculus

Product rule

Quadratic covariation supplies the extra term in the stochastic product rule:

XtYt=X0Y0+∫0tXs− dYs+∫0tYs− dXs+[X,Y]t.X_tY_t = X_0Y_0 + \int_0^t X_{s-}\,dY_s + \int_0^t Y_{s-}\,dX_s + [X,Y]_t.

The left limits in the integrands account for possible jumps.

Setting Y=XY=X gives

[X]t=Xt2−X02−2∫0tXs− dXs.[X]_t = X_t^2-X_0^2 - 2\int_0^t X_{s-}\,dX_s.

This identity is also a standard characterization of quadratic variation through the Itô integral.

Itô’s formula

For a continuous semimartingale XX and a twice continuously differentiable function ff, Itô’s formula states

f(Xt)=f(X0)+∫0tf′(Xs) dXs+12∫0tf′′(Xs) d[X]s.f(X_t) = f(X_0) + \int_0^t f'(X_s)\,dX_s + \frac12\int_0^t f''(X_s)\,d[X]_s.

The final integral is the second-order correction absent from the ordinary chain rule for smooth paths.

For Brownian motion and f(x)=x2f(x)=x^2, this gives

Bt2=2∫0tBs dBs+t.B_t^2 = 2\int_0^t B_s\,dB_s+t.

The familiar differential notation

(dBt)2=dt(dB_t)^2=dt

is shorthand for the quadratic-variation identity, not an algebraic identity between ordinary infinitesimals.

For a semimartingale with jumps, the formula becomes

f(Xt)=f(X0)+∫0tf′(Xs−) dXs+12∫0tf′′(Xs−) d[X]sc+∑0<s≤t[f(Xs)−f(Xs−)−f′(Xs−)ΔXs],\begin{aligned} f(X_t) ={}&f(X_0) +\int_0^t f'(X_{s-})\,dX_s +\frac12\int_0^t f''(X_{s-})\,d[X]^c_s\\ &+\sum_{0<s\leq t} \left[ f(X_s)-f(X_{s-})-f'(X_{s-})\Delta X_s \right], \end{aligned}

where [X]c[X]^c is the continuous part of quadratic variation. Finite jumps require the full correction in the sum, rather than only a second-order approximation.

Diffusion processes

For an Itô process satisfying the stochastic differential equation

dXt=bt dt+σt dBt,dX_t=b_t\,dt+\sigma_t\,dB_t,

under the usual integrability assumptions,

[X]t=∫0tσs2 ds.[X]_t=\int_0^t\sigma_s^2\,ds.

Thus quadratic variation accumulates the squared diffusion coefficient, while the time-integrated drift contributes nothing.

More generally, if

Mt=∫0tHs dBs,Nt=∫0tKs dBs,M_t=\int_0^t H_s\,dB_s, \qquad N_t=\int_0^t K_s\,dB_s,

then

[M,N]t=∫0tHsKs ds.[M,N]_t=\int_0^t H_sK_s\,ds.

Martingales and predictable quadratic variation

For a locally square-integrable local martingale MM, two related processes must be distinguished:

  • Quadratic variation, written [M][M], obtained from squared increments.
  • Predictable quadratic variation, written ⟨M⟩\langle M\rangle, the predictable compensator of [M][M].

With both brackets initialized at zero,

Mt2−M02−[M]tM_t^2-M_0^2-[M]_t

and

Mt2−M02−⟨M⟩tM_t^2-M_0^2-\langle M\rangle_t

are local martingales. Predictability is defined relative to the underlying filtration.

For continuous local martingales,

[M]=⟨M⟩.[M]=\langle M\rangle.

For martingales with jumps, they can differ.

For example, let NtN_t be a Poisson process of rate λ\lambda, and set

Mt=Nt−λt.M_t=N_t-\lambda t.

Each jump of MM has size 11, so

[M]t=Nt,⟨M⟩t=λt.[M]_t=N_t, \qquad \langle M\rangle_t=\lambda t.

The first records the realized number of jumps; the second is its predictable compensator.

If MM is square-integrable and M0=0M_0=0, then

E[Mt2]=E[[M]t]=E[⟨M⟩t].\mathbb E[M_t^2] = \mathbb E[[M]_t] = \mathbb E[\langle M\rangle_t].

These identities connect quadratic variation to second moments without identifying it with variance path by path.

Pathwise interpretation and applications

Quadratic variation can also be studied for a deterministic function along a specified sequence of partitions. This supports pathwise forms of stochastic calculus, but the partition sequence is part of the construction. The standard probabilistic result for semimartingales does not assert convergence along every possible path-dependent choice of partitions, nor almost-sure convergence along every arbitrary deterministic sequence.

For a fixed sequence of partitions, convergence almost surely is a stronger statement than convergence in probability and requires separate justification. Brownian motion, for example, admits almost-sure quadratic-variation limits along standard dyadic partitions.

In financial econometrics, a sum of squared high-frequency returns is called realized variance. Under a semimartingale model and suitable observation conditions, it estimates quadratic variation over the observation interval. For a process with both continuous fluctuations and jumps,

[X]T=[X]Tc+∑0<s≤T(ΔXs)2.[X]_T=[X]^c_T+\sum_{0<s\leq T}(\Delta X_s)^2.

Realized variance therefore includes squared jumps and need not estimate only integrated diffusion variance. Market microstructure noise and other observation effects can also prevent raw squared-return sums from consistently estimating the latent process’s quadratic variation.

Quadratic variation additionally provides the natural time scale for continuous local martingales. Under the hypotheses of the Dambis–Dubins–Schwarz theorem, such a martingale can be represented as Brownian motion evaluated at the random clock given by its quadratic variation.