A NON-SKOROHOD TOPOLOGY ON THE SKOROHOD SPACE

: A new topology (called S ) is de(cid:12)ned on the space ID of functions x : [0 ; 1] ! IR 1 which are right-continuous and admit limits from the left at each t > 0. Although S cannot be metricized, it is quite natural and shares many useful properties with the traditional Skorohod’s topologies J 1 and M 1 . In particular, on the space P ( ID ) of laws of stochastic processes with trajectories in ID the topology S induces a sequential topology for which both the direct and the converse Prohorov’s theorems are valid, the a.s. Skorohod representation for subsequences exists and (cid:12)nite dimensional convergence outside a countable set holds. Abstract A new topology (called S ) is de(cid:12)ned on the space ID of functions x : [0 ; 1] ! IR 1 which are right-continuous and admit limits from the left at each t > 0. Although S cannot be metricized, it is quite natural and shares many useful properties with the traditional Skorohod’s topologies J 1 and M 1 . In particular, on the space P ( ID ) of laws of stochastic processes with trajectories in ID the topology S induces a sequential topology for which both the direct and the converse Prohorov’s theorems are valid, the a.s. Skorohod representation for subsequences exists and (cid:12)nite dimensional convergence outside a countable set holds.


Introduction
Let ID = ID([0, 1] : IR 1 ) be the space of functions x : [0, 1] → IR 1 which are right-continuous and admit limits from the left at each t > 0. We are going to study a new sequential topology on ID generated by naturally arising criteria of relative compactness. The novelty is that this topology cannot be metricized. Nevertheless we shall show how to build a complete and satisfactory theory of the convergence in distribution with respect to this topology.
Despite the fact that metric topologies are sequential, the process of defining topology through description of the family of convergent sequences is not the common approach, especially in probability theory. We refer to [7] for rather extensive discussion of sequential methods in weak limit theorems of probability theory. Here we shall point out only that every time one proves a limit theorem via the direct Prohorov's theorem, one obtains a result for the sequential topology generated by the original one.
Since Prohorov's theorems are accepted tools of probability theory, the same should happen to sequential methods, as they fit the original ideas of Prohorov and Skorohod much better than the theory based on weak- * convergence of distributions.
In order to show our motivations we begin with a simple, well-known example. Let V + ⊂ ID consists of nonnegative and nondecreasing functions v : [0, 1] → IR + . Suppose that for some subset K ⊂ V + we have: sup Let Q ⊂ [0, 1] be countable dense and let 1 ∈ Q. By (1) we may find a sequence {v n } n∈IN ⊂ K such that for each q ∈ Q v n (q) →ṽ(q), whereṽ(q) are numbers satisfyingṽ(q ) ≤ṽ(q ), q < q , q , q ∈ Q. Hence the function belongs to V + . Moreover, we have as It follows that condition (1) when restricted to V + is a criterion of relative compactness for some, quite natural topology. A very similar procedure may be performed for the space ID. Suppose that sup x∈K sup t∈[0, 1] |x(t)| ≤ C K < +∞, (4) and that for all a < b, a, b ∈ IR sup x∈K N a,b (x) ≤ C a,b K < +∞, (5) where N a,b is the usual number of up-crossing given levels a < b. (Recall that N a,b (x) ≥ k if one can find numbers 0 ≤ t 1 < t 2 < . . . < t 2k−1 < t 2k ≤ 1 such that x(t 2i−1 ) < a and x(t 2i ) > b, i = 1, 2, . . . , k). Let, as previously, Q ⊂ [0, 1], Q 1, be countable dense. By (4) we can extract a sequence {x n } n∈IN ⊂ K such that, as n → ∞ x n (q) →x(q), q ∈ Q.
Now the construction of the limiting function x 0 is not as easy as in the case of V + and one has to use (5) in an essential way to see that is well-defined and belongs to ID. And whether x n converges to x 0 in some topology on ID is not clear at all. Meyer and Zheng [14] considered on ID so-called pseudo-path topology and proved that (4) and (5) imply conditional compactness of K in this topology. However, the pseudo-path topology was shown to be equivalent on ID to the convergence in (Lebesgue) measure. And so neither (4) nor (5) form the proper description of the relative compactness in the pseudo-path topology, for it is easy to find a sequence {x n } of elements of ID which is convergent in measure to an element x 0 ∈ ID and is such that for K = {x n : n = 0, 1, 2, . . .} both (4) and (5) are not satisfied.
In the present paper we construct a topology on ID, say S, which possesses the following properties.
• S is sequential and cannot be metricized.
• There exists a countable family of S-continuous functions which separate points in ID.
• The σ-field B S of Borel subsets for S coincides with the usual σ-field generated by projections (or evaluations) on ID: B S = σ(π t : t ∈ [0, 1]).
• The set P(ID, S) of S-tight probability measures is exactly the set of distributions of stochastic processes with trajectories in ID: P(ID, S) = P(ID).
• S is weaker than Skorohod's M 1 and J 1 topologies. Since J 1 is Polish, S is Lusin in the sense of Fernique. But we do not know whether it is (completely) regular.
In the absence of regularity almost nothing of the theory for Lusin spaces (as described in [3] and [4]) can be used. There exists, however, a different approach, developed in [7] and based on the following strong form of the direct Prohorov's theorem (see [5,Theorem 2]): Theorem 1.1 Let (X , τ) be a topological space on which there exists a countable family of τcontinuous functions separating points in X. Let {µ n } n∈IN be a uniformly tight sequence of laws on X . Then there exists a subsequence n 1 < n 2 < . . . and X -valued random elements It is clear that our space (ID, S) fulfills the only assumption of the above theorem and so in the space P(ID, S) the direct Prohorov's theorem is valid. Notice that we are also very close to the a.s. Skorohod representation.
One can go further in this direction and following [7] define a new sequential topology O( * =⇒) (where • Let X n * −→ D X 0 mean that the laws of processes X n converge in the new sense: L(X n ) * =⇒ L(X 0 ). Suppose X n * −→ D X 0 . Then in each subsequence {X n k } k∈IN one can find a further subsequence {X n k l } l∈IN such that: The last statement improves a corresponding result due to Meyer and Zheng [14], where finite dimensional convergence outside a set of Lebesgue measure null was shown via results on the pseudo-path topology. Results on finite dimensional convergence are of quite different flavour than in the case of Skorohod's metric topologies M 1 and J 1 : for every t = 1 the projection π t : ID → IR 1 , π t (x) = x(t), is nowhere S-continuous and so the standard continuity arguments cannot be applied. There exists, however, a simple procedure, completely paralleling the one described for ID in (4) -(7) and allowing to determine the finite dimensional distributions of the limiting process (hence identifying the limits). Suppose that {X n } is a uniformly S-tight sequence. Choose a dense countable subset Q ⊂ [0, 1], Q 1, and extract a subsequence {X n k } such that for each finite sequence q 1 < q 2 < . . . < q m of elements of Q we have whereν (q 1 ,q 2 ,...,qm) is a probability distribution on IR m . Then one can prove that for each finite sequence t 1 < t 2 < . . . < t m and each approximating sequence q 1,l < q 2,l < . . . < q m,l , q j,l t j , as l → ∞, (q m,l = 1 if t m = 1), probability distributionsν (q 1,l ,q 2,l ,...,q m,l ) weakly converge to some limit ν (t 1 ,t 2 ,...,tm) . Moreover, there is a stochastic process X 0 with trajectories in ID such that for every finite sequence t 1 < t 2 < . . . < t m (X 0 (t 1 ), X 0 (t 2 ), . . . , X 0 (t m )) ∼ ν (t 1 ,t 2 ,...,tm) , and X n k * −→ D X 0 . This simple procedure was known to hold in the Skorohod J 1 -topology ( [20]); it is interesting to see that it preserves validity at much lower level.
The paper is organized as follows.
In Section 2 we define the topology S and give its basic properties. In Section 3 we apply to the space P(ID, S) the machinery developed in [7] which is suitable for spaces with countable continuous separation property (as it is the case). In particular we prove all the results announced above.
In Section 4 we show how the theory can be applied to sets of semimartingales satisfying socalled Condition UT which implies uniform S-tightness and which is known to be important in limit theorems for stochastic integrals and stochastic differential equations ( [8], [17], [12], [18], [6]).

The topology S
We shall define the topology S in several steps, following the strategy for sequential topologies described in Section 3 of [7] (the reader is referred to that paper for definitions and notation). Here is the description of subsequent steps.
3. Define S-topology and construct an L * -convergence using the Kantorovich-Kisyński recipe.
4. Show there exists a countable family of S-continuous functions separating points of ID and conclude the topology S is suitable for needs of probability theory. 5. Investigate in some detail other properties of the topology S (Propositions 2.14 and 2.15).
Let us denote by x ∞ the supremum norm on ID: and by v the total variation of v: For a < b let the number of up-crossings N a,b be defined as above (see the lines after formula (5)) and for η > 0 let the number N η of η-oscillations be defined by the relation: Then the statements (i) and (ii) below are equivalent: Moreover, either set of conditions (12)+(13) and (12)+(14) is equivalent to and sup Proof. Let us observe first that (iii) implies (12): Then the chain of implications (iii)⇒(ii)⇒ (i) follows by inequalities and Now assume (12). First we shall prove (i)⇒(ii). Let C ∞ = sup x∈K x ∞ and suppose that for some η > 0 and every n ∈ IN there is x n ∈ K such that N η (x n ) ≥ n. In particular, for some for each x ∈ I there is a j such that |x − a j | < η/2. In every interval I n,i with ends in points x n (t n,2i−1 ) and x n (t n,2i ) there are at least two points a j−1 and a j which belong to I n,i and are distinct from the interval's ends. Let M j (n) be the number of i's such that both a j−1 and a j belong to I n,i . It follows that for some But N a j 0 −1 ,a j 0 (x n ) is greater or equal to the integer part of (M j 0 (x n ) − 1)/2 and so by (20) sup It remains to prove (ii)⇒(iii). The construction of v x,ε is, in some sense, standard. For ε > 0 let us define (where by convention inf ∅ = +∞) and let Then by the very definition and the lemma follows. 1] is the natural right-continuous filtration on the canonical space ID: Hence v ε (X) is an adapted stochastic process provided X is adapted to a right-continuous filtration.
Before introducing a convergence in ID which generates the S-topology, let us recall some facts on the weak- * topology on V. Any element v ∈ V determines a signed measure ν on ([0, 1], B) given by the formula Since the set of signed measures can be identified with the dual of the Banach space C([0, 1] : IR 1 ), V can be equipped with the weak- * topology. Convergence of elements of V in this topology will be denoted by −→ w . In particular, v n −→ w v 0 means that for every continuous function Definition 2. 3 We shall write x n −→ S x 0 if for every ε > 0 one can find elements v n,ε , n = 0, 1, 2, . . . which are ε-uniformly close to x n 's and weakly- * convergent: for each t outside a countable set D ε ⊂ [0, 1). Taking ε = 1, 1/2, 1/3, . . . we obtain for each t ∈ [0, 1] \ ∞ m=1 D 1/m . Hence the limit for −→ S is determined uniquely. Further, step functions are dense in ID for the uniform topology and so for the constant sequence x n ≡ x 0 , n = 1, 2, . . ., we have x n −→ S x 0 . Since also a subsequence of a sequence convergent in the sense of −→ S is (obviously) −→ S -convergent, we conclude that −→ S is an L-convergence. It follows we have enough information to define a topology.
Definition 2.5 A set F ⊂ ID is closed in S-topology, if it contains all limits of its −→ S -convergent subsequences, i.e. if x n ∈ F , n = 1, 2, . . ., and x n −→ S x 0 then x 0 ∈ F . The convergence of sequences in S-topology will be denoted by * −→ S .

Remark 2.6
Similarly as in many other cases, * −→ S , being an L * -convergence, may be weaker than the original L-convergence −→ S . This is not a real problem in view of the Kantorovich-Kisyński recipe [9], [10]: In particular, relative * −→ S -compactness of K ⊂ ID and relative −→ S -compactness of K coincide, as well as sequential * −→ S -continuity (≡ S-continuity) of g : ID → IR 1 means the same as sequential −→ S -continuity of g.
The reason for our interest in the S-topology is that (4) and (5) provide criteria of relative compactness for topology S. (4) and (5) hold for K ⊂ ID then there exists a sequence {x n } ⊂ K and x 0 ∈ ID such that x n −→ S x 0 .

Lemma 2.7 If
Conversely, if in every sequence {x n } ⊂ K one can find a subsequence {x n k } and x 0 ∈ ID such that x n k −→ S x 0 , then K satisfies both conditions (4) and (5).
Proof. Suppose (4) and (5) are satisfied for K ⊂ ID. Fix for the time being ε > 0 and consider the map v ε (x) defined by (23). By Lemma 2.1 Now let us set ε = 1, 1/2, 1/3, . . . and apply the diagonal procedure in order to find a sequence We have 1 ∈ Q and [0, 1] \ Q is at most countable, hence Q is dense. In particular, for any x ∈ ID, By (34) It follows that v 1/m uniformly converges to some x 0 ∈ ID and The converse part follows immediately from Lemma 2.8 below.
In view of (36) and (41) we have continuity for integral functionals.
An important particular case is the Lebesgue measure on [0, 1] and which gives the S-continuity of mappings Since for u < 1 lim δ 0 and π 1 is S-continuous, we conclude that Borel subsets of (ID, S) coincide with σ(π u : u ∈ [0, 1]). In addition, running u and δ over rational numbers in [0, 1] we get a countable family of S-continuous functions which separate points in ID. In particular, any S-compact subset of ID is metrisable. Another useful statement implied by continuity of (48) is that (ID, S) is a Hausdorff space. Further, it follows from the very definition of the convergence −→ S that x n −→ S x 0 and y n −→ S y 0 implies x n + y n −→ S x 0 + y 0 , i.e. addition is sequentially continuous and so from this point of view S-topology is nicer than the Skorohod's J 1 and M 1 .
It should be, however, emphasized that contrary to Skorohod's topologies, the evaluations π u are for u < 1 nowhere S-continuous. To see this, take x = x 0 ∈ ID and 0 ≤ u < 1 and define Finally, let us notice that S is not metrisable, as the following example shows.   (v) S-Borel subsets B S coincide with the standard σ-algebra generated by evaluations (projections).
We conclude this section with two additional properties of the S-topology.
We have observed that the evaluations at t < 1 are nowhere continuous. It is therefore interesting that the evaluations still can be used for identification of the limit. Proposition 2.14 Let Q ⊂ [0, 1], 1 ∈ Q, be dense. Suppose {x n } is relatively S-compact and x n (q) → x 0 (q), as n → +∞, q ∈ Q. (49) Proof. Suppose along a subsequence {n k } we have x n k −→ S y 0 . Then for q in some dense subset Q ⊂ [0, 1), x n (q ) → y 0 (q ). If y 0 = x 0 , then for some η > 0 and (u, v) ⊂ [0, 1) one has |y 0 (t) − x 0 (t)| > η for t ∈ (u, v). Let u < t 1 < t 2 < . . . < v be such that: Then for each m ∈ IN N η/2 (x n k ) ≥ 2m for k large enough, hence N η/2 (x n k ) → +∞ and {x n k } cannot be S-convergent. Hence y 0 = x 0 . Clearly, not all topologies on ID possess the property investigated in Proposition 2.14. For example, it is easy to find a sequence {x n } ⊂ ID which converges in measure to x 0 ≡ 0 and is such that x n (q) → 1 for each rational q ∈ [0, 1]. Hence Proposition 2.14 is not valid for the "pseudo-path" topology.
Another interesting feature of the S-topology is the continuity of the smoothing operation and µ is an atomless finite measure on [0, 1] (e.g. given by a density p µ (s)).

Remark 2.16
It is easy to see that the topology S can be defined also on the space ID([0, 1] : IR d ) of cádlág functions with values in finite dimensional space IR d . It suffices either to define the convergence by coordinates or to consider in (28) convergence (26) of vector-valued measures and to restrict the attention in Lemma 2.1 to quantity N η only. An infinite dimensional generalization requires, however, more careful approach (e.g. in infinite dimension σ-additive vector valued measures may have infinite variation) and will not be considered here.

Convergence in distribution on (ID, S)
In the previous section we checked the equality B S = σ{π t : t ∈ [0, 1]}. It follows that every probability measure on (ID, B S ) is tight. Let us denote the set of such measures by P(ID). Further, the notions "random element in (ID, B S )" and "stochastic process with trajectories in ID" are synonymous and we see that the theory developed in Section 2 applies to usual objects. It is a nice feature of Lemma 2.7 that we have (i) {X α } is uniformly S-tight.
(ii) { X α ∞ } is a uniformly tight family as well as for each a < b {N a,b (X α )} is uniformly tight.
(iv) For each ε > 0 the family { v ε (X α ) (1)} is uniformly tight (where v ε is defined by (23)). The converse Prohorov's theorem, however, is not clear at all. And we are not going to investigate this question. Instead we shall equip P(ID) with a sequential topology induced by S, which is finer than the weak- * topology, for which both the direct and the converse Prohorov's theorem hold and which is very close to the a.s. Skorohod representation. The advantage of such an approach is evident: we have effective tools (Prohorov's theorems, the a.s. Skorohod representation), more (in general) continuous functionals and if the converse Prohorov's theorem is valid for the weak- * topology, then the convergence of sequences in both topologies coincides (see [7]).

And
The advertized "new" convergence in distribution will be denoted by * −→ D and X n * −→ D X 0 will mean the following: and for each ε > there exists an S-compact subset K ε ⊂ ID such that One can say that (52) and (53) describe "the almost sure convergence in compacts" and that (51), (52) and (53) define the strong a.s. Skorohod representation for subsequences ("strong" because of condition (53)). We shall write µ n * =⇒ µ 0 whenever µ n = L(X n ) and X n * −→ D X 0 . It follows from Lemma 2.7 and Corollary 2.10 that on (ID, S) there exists a countable family of lower semicontinuous functionals {ξ i } i∈II such that K ⊂ ID is conditionally compact iff each ξ i is bounded on K, i ∈ II. Therefore we can apply Theorem 5.2 of [7] to obtain the direct and converse Prohorov's theorems for O S ( * =⇒). We have shown the topology O S ( * =⇒) induced by S is as good as the weak- * topology on P(ID, J 1 ), of course, only from the point of view of proving limit theorems. Theorem 3.2 does not end the list of (formal) similarities between the mentioned topologies. We shall point three other properties of the topology O S ( * =⇒) which belong to standard tools of limit theory. It will be proved first that convergence of finite dimensional distributions can be used for identification of limits in * =⇒-convergence (although the projections are nowhere continuous).
where X 0 is a stochastic process with trajectories in ID.
If {X n } is relatively compact with respect to * −→ D , then X n * −→ D X 0 .
An especially convenient tool for investigations of the topology O S ( * =⇒) is provided by the following decomposition.
Theorem 3.8 A family {X α } of stochastic processes with trajectories in ID is uniformly S-tight if, and only if, for each ε > 0 we can decompose processes X α in the following way: where all processes R α,ε , U α,ε and V α,ε are adapted to the natural right-continuous filtration generated by X α , trajectories of R α,ε are uniformly small: U α,ε and V α,ε are nonnegative and nondecreasing (i.e. have paths in V + ) and both {U α,ε (1)} and {V α,ε (1)} are uniformly tight families of random variables.
Remark 3.9 It follows from (62) that any fact on increasing processes may contribute to the knowledge of S-topology. The very close relation between the weak topology on V + and the Stopology on ID becomes clear, if we realize that, similarly to M 1 , the topology S induces on V + exactly the topology of weak convergence (notice, however, that on V = V + − V + the topology S is coarser than the weak topology, while M 1 is incompatible with the latter!). So we can consider the S-topology as a natural extension on ID of the notion of weak convergence of elements of V + .
Proof. Let µ α be a random measure on ([0, 1], B [0,1] ) given by formula Since M ≤1 × IR + is a Polish space we may assume without loss of generality that almost surely The theorem has been proved. Now we are ready to prove a corresponding result for (ID, S) which (in some sense) improves Theorem 5 of [14], where finite dimensional convergence outside a set of Lebesgue measure 0 was obtained (but only convergence in distribution with respect to the "pseudo-path topology" was assumed). Theorem 3.11 Let {X α } be a uniformly S-tight family of stochastic processes with trajectories in ID. Then there exists a sequence {X n } ⊂ {X α }, a process X 0 with trajectories in ID and a countable subset D ⊂ [0, 1) such that for all finite sets Q 0 = {q 1 < q 1 < . . . < q m } ⊂ [0, 1] \ D (X n (q 1 ), X n (q 2 ), . . . , X n (q m )) −→ D (X 0 (q 1 ), X 0 (q 2 ), . . ., X 0 (q m )). (70) In particular, X n * −→ D X 0 .

Remark 3.12
In Introduction we stated that we do not know whether S is a regular topological space. Except for very special cases (like metrisability or existence of a countable fundamental system of compact sets -what is not the case) regularity is not easy to verify within the class of sequential spaces. One thing requires special attention. We have checked that addition is sequentially continuous, i.e. continuous with respect to the natural sequential topology on the product ("convergence on coordinates"). It does not mean continuity as a function on the product ID × ID with the product topology S × S, i.e. the coarsest topology with respect to which both projections are continuous. It is known that the product topology on the product of linear sequential spaces is again sequential if one of the spaces is e.g. locally compact, what is not true again. (see [2], p. 213, for details). Therefore (ID, S) is not a linear topological space. Some comments are relevant.
• Due to the a.s. Skorohod representation given by Theorem 1.1 only sequential continuity of addition is required by possible applications. Hence the last problem is unimportant.
• If (ID, S) were completely regular, we could repeat the reasoning given after Theorem 3.2.
• It seems to be possible to develop the theory within the standard framework and define a completely regular topology τ , for which S is the sequential topology generated by τconvergence of sequences. But then we encounter all difficulties proper to the theory of weak convergence on nonmetric spaces: lack of the a.s. Skorohod's representation, local character of the direct Prohorov's theorem, etc. In addition, if we use the direct Prohorov's theorem, then we enter again the present framework. Problems of this type are in details described in [7] in the case of infinite dimensional separable Hilbert space equipped with the weak topology.

Uniform S-tightness and semimartingales
Theorem 3.2 provides several sets of conditions equivalent to uniform S-tightness. Inequalities (18) and (19) suggest the easiest way of proving uniform S-tightness: one has to check whether families { X α ∞ } and {N a,b (X α )}, for each a < b, are bounded in probability. For example, if X is a supermartingale, then one can use the classical Doob's inequalities (see [1], Ch. VI): is uniformly S-tight. The most general result of such kind concerns semimartingales and belongs to Stricker [19]. We shall restate it using terminology of the paper [8] and the setting of S-topology. Let {X α } be a family of stochastic processes with trajectories in ID, with X α defined on the stochastic basis (Ω α , F α , {F α t } t∈[0,1] , P α ) and adapted to filtration {F α t } t∈[0,1] . We say that Condition UT holds for {X α }, if the family of elementary stochastic integrals { H α − dX α (1)} with integrands bounded by 1 is uniformly tight. (By an elementary stochastic integral with integrands bounded by 1 we mean random variable of the form where m ∈ IN , 0 = t 0 < t 1 < . . . < t m = T , H α (t i ) ≤ 1 and H α (t i ) is F α (t i )-measurable for i = 0, 1, . . ., m. Condition UT was considered for the first time in [19]. The reader may find conditions which follow the line of (73) and are sufficient for Condition UT in [8]. For equivalent reformulations in terms of predictable characteristics we refer to [13]. Here we stress the fact that the family consisting of a single process X satisfies Condition UT if, and only if, X is a semimartingale (see e.g. [15]). Therefore in what follows we shall deal with semimartingales only.
The essential step in the proof of Theorem 2 of [19] gives us Theorem 4.1 Condition UT implies uniform S-tightness.
The two subsequent results are based on Theorems 3.5 and 3.6 and are not valid for the pseudopath topology. If Condition UT is satisfied for {X n }, then X 0 is a semimartingale and X n * −→ D X 0 .