Abstract
A multiple filter test for the analysis and detection of rate change points in point processes on the line has been proposed recently. The underlying statistical test investigates the null hypothesis of constant rate. For that purpose, multiple filtered derivative processes are observed simultaneously. Under the null hypothesis, each process G asymptotically takes the form
while L is a zero-mean Gaussian process with unit variance. This result is used to derive a rejection threshold for statistical hypothesis testing. The purpose of this paper is to describe the behavior of G under the alternative hypothesis of rate changes and potential simultaneous variance changes. We derive the approximation
with deterministic functions \(\Delta \) and \(\Lambda \). The function \(\Lambda \) accounts for the systematic deviation of G in the neighborhood of a change point. When only the rate changes, \(\Lambda \) is hat shaped. When also the variance changes, \(\Lambda \) takes the form of a shark’s fin. In addition, the parameter estimates required in practical application are not consistent in the neighborhood of a change point. Therefore, we derive the factor \(\Delta \) termed here the distortion function. It accounts for the lack in consistency and describes the local parameter estimating process relative to the true scaling of the filtered derivative process.
Similar content being viewed by others
References
Antoch J, Hušková M (1994) Procedures for detection of multiple changes in series of independent observations. In: Mandl P, Huskova M (eds) Proceedings of the Fifth Prague Symposium on Asymptotic Statistics. Physica Verlag, Heildelberg, pp 3–20
Antoch J, Hušková M, Prášková Z (1997) Effect of dependence on statistics for determination of change. J Stat Plan Inference 60(2):291–310
Aue A, Horváth L (2013) Structural breaks in time series. J Time Ser Anal 34(1):1–16
Basseville M, Nikiforov IV (1993) Detection of abrupt changes: theory and application. Prentice Hall Information and System Sciences Series. Prentice Hall Inc., Englewood Cliffs
Bertrand P (2000) A local method for estimating change points: the “hat-function”. Statistics 34(3):215–235
Bertrand PR, Fhima M, Guillin A (2011) Off-line detection of multiple change points by the filtered derivative with \(p\)-value method. Seq Anal 30(2):172–207
Billingsley P (1999) Convergence of probability measures, second edn., Wiley series in probability and statistics probability and statisticsWiley, New York A Wiley-Interscience Publication
Brodsky BE, Darkhovsky BS (1993) Nonparametric methods in change-point problems volume 243 of Mathematics and its Applications. Kluwer Academic Publishers Group, Dordrecht
Carlstein E (1988) Nonparametric change-point estimation. Ann Stat 16(1):188–197
Csörgő M, Horváth L (1997) Limit theorems in change-point analysis., Wiley series in probability and statisticsWiley, Chichester With a foreword by David Kendall
Dümbgen L (1991) The asymptotic behavior of some nonparametric change-point estimators. Ann Stat 19(3):1471–1495
Frick K, Munk A, Sieling H (2014) Multiscale change point inference. J R Stat Soc Ser B Stat Methodol 76(3):495–580 With 32 discussions by 47 authors and a rejoinder by the authors
Fryzlewicz P (2014) Wild binary segmentation for multiple change-point detection. Ann Stat 42(6):2243–2281
Gut A, Steinebach J (2002) Truncated sequential change-point detection based on renewal counting processes. Scand J Stat 29(4):693–719
Gut A, Steinebach J (2009) Truncated sequential change-point detection based on renewal counting processes. II. J Stat Plan Inference 139(6):1921–1936
Kirch C, Muhsal B (2014) A MOSUM procedure for the estimation of multiple random change points. Preprint
Matteson DS, James NA (2014) A nonparametric approach for multiple change point analysis of multivariate data. J Am Stat Assoc 109(505):334–345
Messer M (2014) A multiple filter test for the detection of rate changes in renewal processes with varying variance. PhD thesis, Goethe University, Diss., Frankfurt
Messer M, Kirchner M, Schiemann J, Roeper J, Neininger R, Schneider G (2014) A multiple filter test for the detection of rate changes in renewal processes with varying variance. Ann Appl Stat 8(4):2027–2067
Steinebach J, Eastwood VR (1995) On extreme value asymptotics for increments of renewal processes. J Stat Plan Inference 45(1–2):301–312 Extreme value theory and applications (Villeneuve d’Ascq, 1992)
Timmermann HC (2014) Monitoring procedures for detecting gradual changes. Dissertation. urn:nbn:de:hbz:38-56965
Acknowledgments
This work was supported by the German Federal Ministry of Education and Research (BMBF) within the framework of the e:Med research and funding concept (Grant number: 01ZX1404B) and by the Priority Program 1665 of the DFG. We thank Brooks Ferebee for helpful comments on the manuscript.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Unless otherwise specified, we use the following notation (compare Construction 2.3): Let \(T>0\), \(h\in (0,T/2]\), \(t\in \tau _h\) and \(c\in (0,T)\). Further, let \(\{\xi _{1,j}\}_{j \ge 1}\), \(\{\xi _{2,j}\}_{j \ge 1}\) and \(\{\xi _j^{(n)}\}_{j\ge 1}\) denote the sequences of life times that correspond to \(\Phi _1\), \(\Phi _2\) and to the compound process \(\Phi ^{(n)}\), respectively. Analogously, let \((N_{1,t})_{t\ge 0}\), \((N_{2,t})_{t\ge 0}\) and \((N_t^{(n)})_{t\ge 0}\) denote the associated counting processes [see Eq. (1)]. Further, let \((W_{1,t})_{t\ge 0}\) and \((W_{2,t})_{t\ge 0}\) be independent standard Brownian motions.
1.1 Proof of Proposition 2.5
Outline We show the joint convergence in distribution of the rescaled counting processes \((N_{1,t})_{t}\) and \((N_{2,t})_{t}\) to a function of \((W_{1,t})_{t}\) and \((W_{2,t})_{t}\) (compare 24). Then, at time t both processes refer to the information of the entire time interval (0, t]. In a second step, the processes are continuously mapped to the scenario of the two windows \((t-h,t]\) and \((t,t+h]\) which refers to the filtered derivative process \((\Gamma _t)_t\).
Proof of Proposition 2.5
For \(i=1,2\) let the rescaled random walk \((X_{i,t}^{(n)})_{t\ge 0}\) and the rescaled counting process \((Z_{i,t}^{(n)})_{t\ge 0}\) concerning \(\Phi _i\) be given as
for \(t\ge 0\). According to Donsker’s theorem [in the case of RPVVs apply Messer et al. (2014, Proposition A.8.)], we find in \((D[0,\infty ),d_{SK})\) as \(n\rightarrow \infty \) that
implying weak convergence of \((Z_{i,t}^{(n)})_{t\ge 0}\), i.e., it holds in \((D[0,\infty ),d_{SK})\) as \(n\rightarrow \infty \) that \((Z_{i,t}^{(n)})_{t\ge 0}\mathop {\longrightarrow }\limits ^{d} (W_{i,t})_{t\ge 0}\) for \(i=1,2\), as stated in Billingsley (1999, Theorem 14.6.).
We use a different scaling and set
where \(s_{t}^{(n)}, t\in [0,\infty )\) is given in Definition 2.4. Then for \(i=1,2\), we find in \((D[0,\infty ),d_{SK})\) for \(n\rightarrow \infty \)
because \(\left( {\sqrt{n}\sqrt{\sigma _i^2/\mu _i^3}}/{s_{t}^{(n)}}\right) _{t} = \left( {\sqrt{\sigma _i^2/\mu _i^3}}/{s_{t}^{(1)}}\right) _{t} \) is continuous in t and does not depend on n.
Let now \((\widetilde{Z}_{1,t}^{(n)})_{t\ge 0}\) and \((\widetilde{Z}_{2,t}^{(n)})_{t\ge 0}\) denote the processes derived from \(\Phi _1\) and \(\Phi _2\), respectively. Due to independence of \(\Phi _1\) and \(\Phi _2\), we obtain joint convergence in \((D[0,\infty )\times D[0,\infty ), d_{SK} \otimes d_{SK})\) for \(n\rightarrow \infty \)
We consider the continuous map \(\varphi : (D[0,\infty )\times D[0,\infty ), d_{SK} \otimes d_{SK}) \rightarrow (D[h,T-h],d_{SK})\) given by
The continuous mapping theorem applied to (24) with map \(\varphi \) yields in \((D[h,T-h],d_{SK})\) for \(n\rightarrow \infty \)
Thus, it remains to be shown that
where \(\sim \) denotes equality in distribution. In order to show (25) and (26) we differentiate the four cases \(t\in [h,c-h)\), \( t\in [c-h, c)\), \(t \in [c, c+h)\) and \(t\in [c+h,T-h]\).
Derivation of (25)
Case \(t<c-h:\)
For \(t\ge c+h\) we obtain analogous results by exchanging subscripts. For \(t\in [c-h,c)\) we obtain
Analogously, we obtain \(c\le t < c+h\), which proves (25).
Derivation of (26)
For \(t< c-h\) we obtain
The same holds for \(t\ge c+h\) with the subscript exchanged. In the case \(c-h\le t < c\) we obtain
Analogously, we obtain for \(c\le t < c+h\)
Now let \((W_t)_{t\ge 0}\) be a standard Brownian motion, i.e., \((W_t)_{t\ge 0} \sim (W_{1,t})_{t\ge 0} \sim (W_{2,t})_{t\ge 0}\). The process defined in (27), (28) and (29) has continuous sample paths and is given as a function of increments of disjoint intervals of the processes \((W_{1,t})_{t\ge 0}\) and \((W_{2,t})_{t\ge 0}\). Therefore, we can omit the subscripts one and two in (27), (28) and (29) and obtain a process that has continuous sample paths and the same distribution as the former one. By omitting the subscripts, we obtain the limit process L as defined in Eq. (8), which completes the proof of Proposition 2.5. \(\square \)
1.2 Proof of Lemma 3.2
The Proof of Lemma 3.2 works as follows: The uniform convergence \((s_t^{(n)}/\hat{s}_t^{(n)})_{t\in \tau _h}\rightarrow (\Delta )_{t\in \tau _h}\) a.s. is equivalent to the uniform convergence \((\tilde{s}_t^{(n)}/\hat{s}_t^{(n)})_{t\in \tau _h}\rightarrow (1)_{t\in \tau _h}\) a.s. as \(n\rightarrow \infty \). The terms \(\tilde{s}_t^{(n)}\) and \(\hat{s}_t^{(n)}\) are functions of the estimators \(\hat{\mu }_{le},\hat{\mu }_{ri},\hat{\sigma }_{le}^2\) and \(\hat{\sigma }_{ri}^2\) as given in (15). We show the uniform a.s. convergence to their counterparts \(\mu _{le},\mu _{ri},\sigma _{le}^2\) and \(\sigma _{ri}^2\) defined in (19) and (20). More precisely, we show the uniform a.s. convergence of \((\hat{\mu }_{le})_{t\in \tau _h}\) to \((\mu _{le})_{t\in \tau _h}\) and \((\hat{\mu }_{ri})_{t\in \tau _h}\) to \((\mu _{ri})_{t\in \tau _h}\) in Lemma 4.2, and the uniform a.s. convergence of \((\hat{\sigma }^2_{le})_{t\in \tau _h}\) to \((\sigma _{le}^2)_{t\in \tau _h}\) and \((\hat{\sigma }^2_{ri})_{t\in \tau _h}\) to \((\sigma _{ri}^2)_{t\in \tau _h}\) in Lemma 4.3. Thus, the assertion of the Proposition holds true by the structure of the estimator \(\hat{s}^2\) in (15) and the function \(\tilde{s}^2\) in (18) and because convergence of sums and products of càdlàg-valued functions in supremum norm is preserved when the limits are constant. \(\square \)
For completeness of the proof, we show the consistency of the estimators \(\hat{\mu }_{le}\) and \(\hat{\mu }_{ri}\) in Lemma 4.2 and the consistency of \(\hat{\sigma }_{le}^2\) and \(\hat{\sigma }_{ri}^2\) in Lemma 4.3. For that we first show a functional version of the SLLN in the following Lemma.
Lemma 4.1
For the counting process \(N_t^{(n)}\) that corresponds to the process \(\Phi ^{(n)}\), it holds in \((D[h,T-h],d_{\Vert \cdot \Vert })\) as \(n\rightarrow \infty \) almost surely
Proof
Outline: We show the convergence of the right window half as stated in (30). The statement for the left window half follows analogously.
First, we show that for all \(t\ge 0\) and all \(h>0\) it holds almost surely as \(n\rightarrow \infty \)
Then, by a discretization argument this result is extended to hold true in \((D[h,T-h],d_{\Vert \,\cdot \,\Vert })\), as stated in (30).
Derivation of (32)
In order to show the convergence in (32), we distinguish between three cases. First assume \(t\le c-h\). Here, for all \(n=1,2,\ldots \), the corresponding window \((nt,n(t+h)]\) lies left of the change point nc. Thus, the counting process \(N_t^{(n)}\) completely refers to the first RP \(\Phi _1(\mu _1,\sigma _1^2)\), i.e., \(N_{nt}^{(n)}=N_{1,nt}\), while \((N_{1,t})_{t\ge 0}\) denotes the counting process associated with \(\Phi _1\). Then, it can be shown that it holds almost surely for \(n\rightarrow \infty \)
compare e.g., Messer et al. (2014). An analogous statement holds for \(t>c\).
For \(t\in (c-h,c]\), the right window half refers partially to \(\Phi _1\) and \(\Phi _2\). The section (nt, nc] refers to \(\Phi _1\) and the section \((nc,n(t+h)]\) corresponds to \(\Phi _2\). Thus, we decompose \(N_{n(t+h)}^{(n)}-N_{nt}^{(n)} = (N_{2,n(t+h)}-N_{2,nc}) + (N_{1,nc}-N_{1,nt})\). We obtain almost surely for \(n\rightarrow \infty \)
In total, the convergences (33) and (34) yield (32).
Derivation of (30)
In order to show that also convergence in \((D[h,T-h],d_{\Vert \cdot \Vert })\) holds, we even show the convergence in (30) on \([0,T-h]\). It is sufficient to show that almost surely
We show the left inequality of (35). The right one follows analogously. We use a discretization argument. For \(x\in \mathbb {R}\) let \(|\lceil x \rceil |:= \lceil x \rceil + 1\). For \(\varepsilon >0\) with \(T/\varepsilon \in \mathbb {N}\) we decompose the time interval (0, nT] into equidistant sections of length \(n\varepsilon \) (Fig. 5). Then we observe a set \(S_\varepsilon :=\{(kn\varepsilon ,kn\varepsilon + n|\lceil h/\varepsilon \rceil | \varepsilon ] :k=0,1,\ldots ,T/\varepsilon - |\lceil h/\varepsilon \rceil |\}\) of finitely many windows of size \(n|\lceil h/\varepsilon \rceil |\varepsilon \). The windows of \(S_\varepsilon \) are slightly larger than nh and for every \(t\in (0,T-h]\) we find an element of \(S_\varepsilon \) that overlaps the window \((nh,n(t+h)]\) (blue window in Fig. 5). We bound
Now we make use of the fact that the convergence in (30) holds true for a finite number of windows. By letting \(n\rightarrow \infty \) the right hand side of (36) converges to \(1+\delta _\varepsilon \), with
The expression \(\delta _\varepsilon >0\) accounts for the additional portion that results from the enlarged windows. Then, by letting \(\varepsilon \downarrow 0\) the summand \(\delta _\varepsilon \) vanishes, which yields the first inequality in (35). Analogously, for the lower bound of (35) we find finitely many smaller windows of length \(n(\lfloor h/\varepsilon \rfloor - 1)\varepsilon \), such that every window \((nh,n(t+h)])\) contains such a smaller window (red window in Fig. 5). Then, the limit of the infimum can be bounded from below by \(1-\delta _\varepsilon '\) with \(\delta _\varepsilon '>0\) and such that \(\delta _\varepsilon '\rightarrow 0\) as \(\varepsilon \downarrow 0\). Here, \(\delta _\varepsilon '\) refers to the portion that is not covered by choosing the finitely many windows to be slightly smaller than the true window size nh. \(\square \)
Next, we show the uniform a.s. convergences \((\hat{\mu }_{ri}(nh,nt))_{t\in \tau _h}\rightarrow (\mu _{ri}(h,t))_{t\in \tau _h}\) and
\((\hat{\mu }_{le}(nh,nt))_{t\in \tau _h}\rightarrow (\mu _{le}(h,t))_{t\in \tau _h}\) as \(n\rightarrow \infty \). The estimators are given as
and \(\hat{\mu }_{ri}=0\) otherwise and \(\hat{\mu }_{le}\) is given analogously.
Lemma 4.2
For the estimators \(\hat{\mu }_{ri}(nh,nt)\) and \(\hat{\mu }_{le}(nh,nt)\) as given in (15), it holds in \((D[h,T-h],d_{\Vert \cdot \Vert })\) as \(n\rightarrow \infty \) almost surely
Proof
We show the convergence of the right window half as stated in (37). The assertion for the left window half follows analogously. We proceed as in the proof Lemma 4.1. First, we show that for all \(t\ge 0\) and \(h>0\), it holds almost surely as \(n\rightarrow \infty \)
i.e., the sum of the life times in the window half asymptotically equals the window length. Then, this result is extended to \((D[h,T-h],d_{\Vert \,\cdot \,\Vert })\) to conclude the convergence in (37).
Derivation of (39)
Assertion (37) has been shown in Messer et al. (2014) to hold for the individual processes \(\Phi _j(\mu _j,\sigma _j^2)\). Therefore, as we show (37) for the right window half, convergence (39) holds true for \(t\in (0,c-h]\) and \(t\ge c\). For \(t\in (c-h,c]\), the right window half contains parts of \(\Phi _1\) and of \(\Phi _2\). We therefore decompose \((nt,n(t+h)] = (nt,nc] \cup (nc,n(t+h)]\). The section (nt, nc] refers to \(\Phi _1\) and the section \((nc,n(t+h)]\) corresponds to \(\Phi _2\). The life time at the change point c results from \(\Phi _1\) and \(\Phi _2\), and we therefore bound
which allows to use the properties of the individual processes. For the right hand side in (40) it holds as \(n\rightarrow \infty \)
Analogously we obtain the lower bound, such that assertion (39) holds true.
Derivation of (37)
In order to show convergence (39) in \((D[h,T-h],d_{\Vert \,\cdot \,\Vert })\) we even prove it on the interval \([0,T-h]\). For that, we show that almost surely
We show here the left inequality of (41). We use the same discretization argument as in the proof of Lemma 4.1 and decompose the interval (0, nT] into equidistant sections of length \(n \varepsilon \) (Fig. 5). Then we bound
The first summand tends to zero as \(\varepsilon \downarrow 0\) and is independent of n. Further, for every \(\varepsilon >0\), the second summand converges to unity almost surely as \(n \rightarrow \infty \), according to equation (39). Thus, the first inequality in (41) holds. The second inequality in (41) can be shown similarly. Thus, convergence (39) holds in \((D[h,T-h],d_{\Vert \,\cdot \,\Vert })\), and together with Lemma 4.1 the assertion (37) holds true by Slutsky’s Theorem. \(\square \)
To finish the proof of Lemma 3.2, we need to show the uniform a.s. convergences \((\hat{\sigma }_{ri}^2(nh,nt))_{t\in \tau _h}\rightarrow (\sigma _{ri}^2(h,t))_{t\in \tau _h}\) and \((\hat{\sigma }_{ri}^2(nh,nt))_{t\in \tau _h}\rightarrow (\sigma _{ri}^2(h,t))_{t\in \tau _h}\) as \(n\rightarrow \infty \). The estimator \(\hat{\sigma }_{ri}^2\) is given as
and \(\hat{\sigma }^2_{ri}=0\) otherwise. Similarly \(\hat{\sigma }^2_{le}\) is given.
Lemma 4.3
For the estimators \(\hat{\sigma }_{ri}^2(nh,nt)\) and \(\hat{\sigma }_{le}^2(nh,nt)\) as given in (15) it holds in \((D[h,T-h],d_{\Vert \cdot \Vert })\) as \(n\rightarrow \infty \) almost surely
Proof
Again we show the convergence of the right window half as given in (42), while the statement for the left window half follows analogously.
First, we show that for all \(t\ge 0\) and \(h>0\), it holds almost surely as \(n\rightarrow \infty \)
Then, this result is extended to \((D[h,T-h],d_{\Vert \,\cdot \,\Vert })\) which yields (42).
Derivation of (44)
In the following, let \(\hat{\mu }_{j,ri}(nh,nt)\) denote the estimator that corresponds to \(\Phi _j(\mu _j,\sigma _j^2)\). As before, \(\hat{\mu }_{ri}(nh,nt)\) denotes the estimator that refers to the compound process \(\Phi ^{(n)}\).
Note that (44) was shown in Messer et al. (2014) to hold for the individual processes \(\Phi _j(\mu _j,\sigma _j^2)\). Therefore, as we show (42) for the right window, (44) holds for \(t\in (0,c-h]\) and \(t\ge c\). For the remaining case \(t\in (c-h,c]\), we recall that the right window half partially corresponds to \(\Phi _1\) and \(\Phi _2\). Again, we decompose \((nt,n(t+h)] = (nt,nc] \cup (nc,n(t+h)]\), where the sections (nt, nc] and \((nc,n(t+h)]\) refer to \(\Phi _1\) and \(\Phi _2\), respectively. We decompose
The term \(o_{a.s.}(1)\) accounts for the summand that corresponds to the single life time \(\xi _{N_{nc}^{(n)}+1}^{(n)}\) that overlaps the change point and that is not respected in the first two terms of (45). By Borel-Cantelli Lemma, the sequence \(\{(\xi _{N_{nc}^{(n)}+1}^{(n)}-\hat{\mu }_{ri}(nh,nt))^2/nh\}_{n=1,2,\ldots }\) can be shown to vanish almost surely for \(n\rightarrow \infty \) and is therefore abbreviated with \(o_{a.s.}(1)\). For the first summand, we find almost surely as \(n\rightarrow \infty \)
The first summand in (46) shows the a.s. convergence to \(\sigma _1^2\) because it refers only to \(\Phi _1\). The second summand in (46) vanishes a.s. since the left term converges a.s. according to Lemma 4.2 and the right term tends to zero a.s. according to Lemmas 4.1 and 4.2. The third summand in (46) tends to \((\mu _1-\mu _{ri}(h,t))^2\) a.s., since the term in the squared brackets converges to \((\mu _1-\mu _{ri}(h,t))^2\) a.s. according to Lemma 4.2, while the scaled counting process converges to unity a.s. due to Lemma 4.1.
An analogous result can be obtained for the second summand of (45) which yields almost surely for \(n\rightarrow \infty \)
and elementary calculations yield equality to \(\sigma _{ri}^2(h,t)\). The convergence in (44) can be concluded using an analogous discretization argument as in the proofs of Lemmas 4.2 and 4.1, such that the assertion (42) can be concluded. \(\square \)
Rights and permissions
About this article
Cite this article
Messer, M., Schneider, G. The shark fin function: asymptotic behavior of the filtered derivative for point processes in case of change points. Stat Inference Stoch Process 20, 253–272 (2017). https://doi.org/10.1007/s11203-016-9138-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11203-016-9138-0