Two Different Kinds of Brownian Claims
A Brownian statement may concern a value at one specified time, or it may concern the entire trajectory over an interval. Those are different kinds of claims. The first can often be answered from a normal distribution; the second concerns a sample path and usually needs an argument about several times, or even infinitely many times.
Assume throughout that \(B=(B_t)_{t\geq 0}\) is standard Brownian motion: \(B_0=0\) almost surely, its increments over disjoint time intervals are independent, \(B_t-B_s\sim N(0,t-s)\) for \(0\leq s<t\), and its sample paths are continuous almost surely. The continuity assumption is part of the path-level setup, not a consequence of inspecting one-time marginals.
The distinction matters in pharmacokinetic modeling. Brownian motion may be used as a stochastic driver, but the fact that its value at a particular time has a centered normal distribution does not describe the smoothness, direction, or endpoint behavior of every intervening trajectory.
What an Endpoint Distribution Says
For each fixed \(t>0\), the increment assumptions give \(B_t=B_t-B_0\sim N(0,t)\). Thus a probability about \(B_t\) can be calculated from a single normal distribution. By contrast, \(B_0=0\) almost surely is an explicit starting-value property. The two endpoints of an interval therefore need not have the same kind of distributional statement.
Worked Example: A Fixed Endpoint Is Not Usually Zero
Let the interval end at \(t=4\). Since \(B_4\sim N(0,4)\), its density is
A normal distribution with a density assigns probability zero to any single point, so \(P(B_4=0)=0\). Symmetry about zero gives \(P(B_4>0)=1/2\). In contrast, \(P(B_0=0)=1\) by the defining starting condition.
These calculations concern the values at times 0 and 4. They do not say that the path stays on one side of zero between those times, nor do they describe whether it reaches zero at an intermediate time. Such questions involve the trajectory, not just its endpoint marginal.
A frequent error is to replace “the endpoint is positive with probability one half” by “the path is positive with probability one half.” The latter is incomplete: it needs a time interval and a precise path event, such as positivity at every time in that interval. A single-time probability cannot answer that question.
Monotonicity Is a Pathwise Claim
A path is nondecreasing on \([0,T]\) if \(B_s\leq B_t\) for every pair \(0\leq s<t\leq T\). This requires a simultaneous comparison across the interval. Independent increments provide a direct way to rule it out for standard Brownian motion.
To prove the first claim, divide the interval into \(N\) equal pieces. If a path is nondecreasing, every increment between consecutive grid points must be nonnegative. The \(N\) increments are independent, and each has a centered normal distribution with positive variance \(T/N\). Each therefore has probability \(1/2\) of being nonnegative. The probability that all \(N\) increments are nonnegative is \(2^{-N}\). Since the event of being nondecreasing is contained in this grid event for every \(N\), its probability is at most \(2^{-N}\) for every positive integer \(N\). Letting \(N\) increase shows that its probability is zero. For nonincreasing paths, all \(N\) increments must instead be nonpositive; the same independence and symmetry give the same bound \(2^{-N}\), hence probability zero. \(\square\)
Worked Example: Testing a Monotonicity Claim on a Grid
Take \(T=2\) and use four equal intervals. Their increments are independent and each has distribution \(N(0,1/2)\). If a Brownian path is nondecreasing on \([0,2]\), these four increments must all be nonnegative. Therefore
So the probability of a nondecreasing path is at most \(1/16\). Using \(N\) equal intervals instead gives the bound \(2^{-N}\), which tends to zero. The four-interval calculation alone does not prove probability zero; the argument needs the bounds for arbitrarily large \(N\).
The grid argument is a useful diagnostic. A claim about the path must imply corresponding constraints on every finite grid. If the probability of those necessary finite-grid constraints can be made arbitrarily small, the path claim has probability zero. The argument establishes the result on a specified interval; it does not infer path properties merely from \(P(B_t>0)\) or \(P(B_t<0)\) at one time.
Differentiability at a Fixed Time
Brownian increments become small in absolute size over short intervals, but not on the scale required for a finite derivative. The following result makes that statement precise at any fixed deterministic time. At an endpoint of the time domain, “derivative” here means the available one-sided derivative.
For the right derivative, set \(h_n=2^{-n}\) and define the difference quotient \(Q_n=(B_{t+h_n}-B_t)/h_n\). Each quotient has distribution \(N(0,1/h_n)\), because the increment in its numerator has distribution \(N(0,h_n)\). For any fixed finite \(M>0\), write \(Q_n=Z_n/\sqrt{h_n}\) in distribution, where \(Z_n\sim N(0,1)\). Then
The limit follows because \(M\sqrt{h_n}\to 0\) and the standard normal distribution has a continuous density. If the right derivative were finite, the sequence \(Q_n\) would converge to a finite number and therefore would eventually be bounded by some positive integer \(M\). For fixed \(M\), the event that \(|Q_n|\leq M\) for all sufficiently large \(n\) has probability zero: for each \(N\), the event that this bound holds for every \(n\geq N\) is contained in \(\{|Q_n|\leq M\}\) for each such \(n\), and hence has probability at most \(\inf_{n\geq N}P(|Q_n|\leq M)=0\). Taking the countable union over \(N\) and over positive integers \(M\) still gives probability zero. Thus the right derivative is not finite, almost surely.
For a fixed \(t>0\), use the left quotients \((B_t-B_{t-h_n})/h_n\), restricting to \(n\) large enough that \(h_n\leq t\). Each has distribution \(N(0,1/h_n)\). The same bounded-quotient argument proves that the left derivative is not finite, almost surely. \(\square\)
Worked Example: Why a Finite Slope Fails at Time 1
Consider a proposed right derivative at \(t=1\), and use \(h_n=2^{-n}\). The quotient is
For instance, the probability that this quotient lies between \(-10\) and \(10\) is
As \(n\) increases, the interval \([-10\,2^{-n/2},10\,2^{-n/2}]\) shrinks to the single point zero, so this probability tends to zero. A finite derivative would require the quotients to remain bounded eventually; applying the same calculation to every integer bound proves that this happens with probability zero.
Keep the scope of this proof precise: it concerns a fixed deterministic time. It does not, by itself, prove that one probability-zero event covers all times simultaneously. A claim about differentiability at every time requires an argument that addresses the uncountable collection of possible times; it cannot be obtained just by repeating a fixed-time conclusion informally.
Choosing the Right Argument
There are three useful levels of reasoning. A marginal calculation uses the distribution of \(B_t\) at one specified time. A finite-dimensional calculation uses the joint distribution of values at a finite list of times, including the independent-increment structure. A pathwise argument concerns a property over an interval; the monotonicity proof, for example, used a sequence of increasingly fine finite grids. These approaches answer different questions.
| Claim | Relevant information | What it does not establish by itself |
|---|---|---|
| \(P(B_t=0)\) for fixed \(t>0\) | The one-time normal distribution \(N(0,t)\) | Whether the path reaches zero at some time |
| Monotonicity on \([0,T]\) | Independent increments on arbitrarily fine grids | The direction of the path at one selected time |
| A finite derivative at fixed \(t\) | Difference quotients over shrinking time steps | A simultaneous differentiability claim for all times |
One important qualification concerns finite-dimensional distributions. For the continuous version of Brownian motion, the standard path law is consistent with these finite-dimensional distributions. But a statement about a single marginal is not a substitute for a path argument. Also, if continuity or an equivalent path regularity assumption has not been specified, finite-dimensional distributions alone should not be used to assert that a particular process has continuous paths.
Common Mistakes and Practical Checks
- Confusing “at a fixed time” with “at some time.” \(P(B_t=0)=0\) for a fixed \(t>0\) is not a calculation of the probability that a path hits zero anywhere in an interval.
- Inferring monotonicity from a trend or endpoint sign. A positive value at the end of an interval does not mean all intervening increments were positive.
- Using one grid as a complete proof. A finite grid can give an upper bound for a path event. To prove that the probability is zero, the bound must tend to zero along a sequence of grids.
- Turning a fixed-time result into a simultaneous one. A probability-one statement for each fixed time does not automatically give a single probability-one event on which the statement holds for every time.
- Omitting path assumptions. Normal marginals specify distributions at fixed times; continuity is a separate path-level property in the Brownian motion setup used here.
When checking a Brownian claim, first identify whether its quantifier is about one deterministic time, finitely many times, or every time on an interval. Then use information at that same level: a marginal distribution, a joint increment calculation, or a pathwise argument with explicit regularity assumptions. This prevents a correct distributional fact from being used to support a conclusion it does not imply.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- For fixed \(t=9\), what are \(P(B_9=0)\) and \(P(B_9>0)\)?
- On \([0,1]\), what is the probability that the four increments on the equal four-part grid are all nonnegative?
- Why does ruling out a finite right derivative at one fixed time not, by itself, prove a simultaneous statement about every time?
- Which is the more appropriate tool for a claim about monotonicity on an interval: the distribution of \(B_T\), or independent increments on fine grids? Explain.
Solutions. (1) Since \(B_9\sim N(0,9)\), it has a continuous symmetric density, so \(P(B_9=0)=0\) and \(P(B_9>0)=1/2\). (2) The increments are independent centered normal variables, each nonnegative with probability \(1/2\), giving \((1/2)^4=1/16\). (3) The fixed-time result concerns one specified time; extending it to all times requires controlling an uncountable set of candidate times, which that result does not do. (4) Fine-grid independent increments are appropriate because monotonicity constrains every increment on every grid; an endpoint distribution alone contains no such pathwise information.