Model to Describe Fast Shutoff of CoVID-19 Pandemic Spread

Early CoVID-19 growth obeys: N{t*}=N/I exp[+K/o t*], with K/o = [(ln 2) / (t/dbl)], where t/dbl is the pandemic growth doubling time. Given N{t*}, the daily number of new CoVID-19 cases is {rho}{t*}=dN{t*}/dt*. Implementing society-wide Social Distancing increases the t/dbl doubling time, and a linear function of time for t/dbl was used in our Initial Model: N/o[t] = 1 exp[+K/A t / (1 + {gamma}/o t) ] = exp(+G/o) exp( - Z/o[t] ) , to describe these changes, with G/o = [K/A / {gamma}/o]. However, this equation could not easily model some quickly decreasing {rho}[t] cases, indicating that a second Social Distancing process was involved. This second process is most evident in the initial CoVID-19 data from China, South Korea, and Italy. The Italy data is analyzed here in detail as representative of this second process. Modifying Z/o[t] to allow exponential cutoffs: Z/E[t] = +[G/o / (1 + {gamma}/o t) ] [exp( - {delta}/o t - q/o t^2 ] = Z/o[t] [exp( - {delta}/o t - q/o t^2 ] , provides a new Enhanced Initial Model (EIM), which significantly improves datafits, where N/E[t] = exp(+G/o) exp( - Z/E[t] ). Since large variations are present in {rho}/data[t], these models were generalized into an orthogonal function series, to provide additional data fitting parameters: N(Z) = Sum{m = (0, M/F)} g/m L/m(Z) exp[-Z]. Its first term can give N/o[t] or N/E[t], for Z=Z/o[t] or Z=Z/E[t]. The L/m(Z) are Laguerre Polynomials, with L/0(Z)=1, and {g/m; m= (0, M/F)} are constants derived from each dataset. When {rho}[t]=dN[t]/dt gradually decreases, using Z/o[t] provided good datafits at small M/F values, but was inadequate if {rho}[t] decreased faster. For those cases, Z/E[t] was used in the above N(Z) series to give the most general Enhanced Orthogonal Function [EOF] model developed here. Even with M/F=0 and q/o=0, this EOF model fit the Italy CoVID-19 data for {rho}[t] = dN[t]/dt fairly well. When the {rho}[t] post-peak behavior is not Gaussian, then Z/E[t] with {delta}/o=0, q/o=0; which we call Z/A[t], is also likely to be a sufficient extension of the Z/o[t] model. The EOF model also can model a gradually decreasing {rho}[t] tail using small {{delta}/o, q/o} values [with 6 Figures].


Introduction
Let N f b tg be the total number of CoVID-19 cases in any given locality, with f b tg being the predicted number of daily new CoVID-19 cases, so that: where t dbl is the pandemic doubling time. The start of societywide Social Distancing at b t = 0 can gradually lengthen t dbl for b t > 0. The b t < 0 exponential growth phase is not applicable for estimating Social Distancing e¤ects. For b t > 0, an Initial Model for CoVID-19 pandemic shuto¤ was …rst developed 1 using a linear function of time to describe the t dbl changes: . [1.2] Given measured N data f b tg, the data end-points fN I ; N F g help to set fK o ; S g. An Orthogonal Function Model [OFM] was developed next 2 , with Eq. [1.2] as the …rst term of the orthogonal function series. Each new OFM term provides another …tting parameter, to progressively better match N data f b tg and The OFM improves on the Initial Model, and it works best with gradually decreasing f b tg ["Slow Shuto¤ "]. In contrast, when f b tg decreased quickly ["Fast Shuto¤ "], the Initial Model was not a good data…t, and a few-term OFM series only gave small improvements. This result indicates there is an inherent limit to what the gradually changing t dbl doubling time of Eq. [1.2] can model.
For these cases, typi…ed by CoVID-19 pandemic evolution in Italy, data often showed a stage where data f b tg [exp( o b t)] or data f b tg [exp( q o b t 2 )], which likely represents a second process, independent of the gradually changing t dbl doubling time. An Enhanced Initial Model (EIM) is developed here to include this second process. The prior OFM methods can then be applied, giving an Enhanced Orthogonal Function (EOF) model for this more general case.

Review of Prior Models
The Initial Model of Eq. [1.2] is still needed as the …rst part of the OFM. The Initial Model starts with measured data end-points fN I ; N F g, where b t = (t F t I ) is the largest data time interval so that N f b t = (t F t I )g = N F . Usually S in Eq. [1.2] was chosen …rst, and K o or t dbl adjusted to match the N F data end-point, using an Excel T M _Goal-Seek or its equivalent. The …nal fK o ; S g values were the pair with the minimum root-mean-square (rms) error between the given data and the Eq. [1.2] model. The above fK o ; S g also provides a t = 0 estimate for the pandemic start, and gives fK A ; o g as new data …tting parameters: [1.3c] 2 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The fg m ; m = (0; M F )g constants in Eq. [1.7a] can be arranged in a ! g vector form, with comparable constants for R(Z) from Eq. [1.7b] arranged in a ! C vector form. For M F = 2, it allows Eq. [1.7d] to be written as: The OFM implicitly uses a Linear Y-axis, so its results di¤er from the Initial Model data…t on a Logarithmic Y-axis. As an example, compare the In the Initial Model, G o is …xed so that N o (Z o ) exactly matches fN I ; N F g at the ft I ; t F g boundaries. In the OFM, g 0 = G o is no longer required, so that the OFM N (Z o ) best data…t is not constrained to exactly match fN I ; N F g at ft I ; t F g.
The above R(Z) and Z [t] gives N [t] and [t] as an explicit functions of time:  Fig. 1, the X-axis uses this t = 0 point where N o [t = 0] ! 1, and it shows what Social Distancing e¤ects would have been, if it had been operating throughout the t > 0 period. This o [t] prediction still has a much more gradual drop than the data. This discrepancy indicates that a second Social Distancing process is operating, besides just the gradual t dbl lengthening of the Initial Model.
g , which exhibits the following variety of long-time limits: is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. ; If data does not show evidence of a Gaussian pandemic Fast Shuto¤ , assuming the post-peak [t] data will be Gaussian is likely to provide optimistically inaccurate N [t] predictions for CoVID-19 pandemic evolution. Apparently, this is exactly what was done by the University of Washington IHME (Institute of Health Metrics and Evaluation) in their widely publicized initial preprint 3 of 27 March 2020, with this Gaussian model continuing throughout their subsequent updates 4 6 up through 29 April 2020. IHME changed everything in their 4 May 2020 7 function has the f o ; o g Mitigation Measure operating at the start of Social Distancing, but reverting to the Initial Model in the long-time limit. Combining Eq. [2.6b] and Eq. [3.1a] cases gives this Enhanced Initial Model (EIM) equation: = 0 is a pure exponential, and = 1 has a modi…ed tail that includes its own long-term shuto¤.
1a] prior to any EOF analysis. The = 1 case also has this special symmetry: [ 6 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. ; 2d] which can re-written as: 1a], and the N I (t I ) and N F (t F ) initial and …nal points, give these equations to help set fK A ; o ; o g: where t data is the data start time. Set a preliminary value for b t of f set …rst, to …x the time scale for the N data [t] measured values: [4.8b] The resulting calculated values for both f b N [t I ]; b N [t F ]g can often be much too high or low, compared to the fN I ; N F g measured data, but those values can be renormalized to: value that is needed to obey N [t F ] ! N F can be set by using Excel T M _Goal-Seek or its equivalent, which also sets a particular S I value. Next, the b o value is adjusted to …nd the speci…c fK A ; o g parameter pair that gives S I = 1. This process is needed because these b o 6 = 0 cases do not allow easy determination of t I as in Eq.   Figure 3 has a predicted CoVID-19 pandemic peak of~5; 217 = day at t = 29:134 days on 3=25=2020, with~243; 100 total cases at the pandemic end. This data…t has > 4X error reduction over the o = 0 case, as summarized next: is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. [ where m = f0; M F g sets how many terms are in the Eq. [6.3a] series. Generally M F = 2 is used here. The L m (Z) are the Laguerre Polynomials, with the …rst few L m (Z) being: . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
Since the N (Z) of Eq. [6.3a] has fg m ; m = (0; M F )g, and N (Z) also appears in each g n -equation of Eq. [6.3b], how to determine each g m by itself, can be done as follows. First de…ne: When the N data (Z) is comprised of j = f1; 2; :::Jg discrete values between fZ min A ; Z max A g, with each Z j having an N (j) data (Z j ) value, the Eq. [6.8a] integral needs to be replaced by a sum. Let Z 0 = Z 1 and Z J+1 = Z J , the Q n replacement for Eq. [6.8a] is then: [6.9b] Eq. [6.3b] can then be re-written as a 3 3 matrix M 3 , which relates a data-driven ! Q 3 -vector to a resultant ! g 3 -vector:

10
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. ; +1g, this M 3 becomes the Identity Matrix. The following k m;n (Z) integrals set K m;n : [6.11b] The k m;n (Z) integrals can be determined using Eq. [6.5c], which gives: [6.12f] To extract fg 0 ; g 1 ; g 2 g, the 3 3 symmetric M 3 matrix needs inversion: which determines fg 0 ; g 1 ; g 2 g from the fQ 0 ; Q 1 ; Q 2 g data. A best-…t N (Z) for Z = f0 + ; 1 g results, along with an equivalent …t for R(Z) using Eq. [1.7d].
Instead of having to …nd the best fg 0 ; g 1 ; g 2 g triplet, one could …nd the best fg 0 0 ; g 0 1 g by just using using fQ 0 ; Q 1 g and an M 2 sub-matrix; or one could …nd the best fg + 0 g by itself by just using fQ 0 g and an M 1 sub-matrix: [6.14f] Once the fg m ; m = (0; M F )g constants are found and used in Eq. [1.8], its c 0 value provides the new EOF estimate for the predicted total number of CoVID-19 cases at the pandemic end, re…ning the initial Eq. [5.2]    A , [7.5] determining the constants needed for R(Z A ) in Eq. [1.7b]. Using these fg 0 ; g 1 ; g 2 g values along with Eq. [1.8] gives:  Figure 5, along with the t > t I raw data for the daily new CoVID-19 cases.
The Figure 5 EOF model also gives a t < t I extrapolation, which shows what the combination of processes would look like, if they all had been operating continuously from the CoVID-19 pandemic start. The companion N [t] analytic 12 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. ; https://doi.org/10.1101/2020.08.07.20169904 doi: medRxiv preprint result, along with the t > t I raw data for the total number of CoVID-19 cases is show in Figure 6.
Comparing the size and timing of the [t] pandemic peak, and its Day 200 value, between the EIM (Figs. 3-4) and EOF model (Figs. 5-6 [7.7] showing the EOF model predicts more cases total and more daily new CoVID-19 cases at Day 200, as well as modifying the pandemic peak predictions. While the above analysis used M F = 2, with the Eq. [7.4] ! g 3 setting the best fg 0 ; g 1 ; g 2 g values, this EOF model also provides estimates for the simpler M F = f0; 1g cases, as outlined by Eqs. [6.14a]-[6.14f]. For M F = 1, the best two fg 0 0 ; g 0 1 g values were gotten by only using fQ 0 ; Q 1 g and an M 2 sub-matrix of M 3 . For M F = 0, the best fg + 0 g by itself is derived by using fQ 0 g and the M 1 sub-matrix. These alternative estimates give: (b) The data…ts in Fig. 4 and Fig. 6 show that the extra parameters in the EIM and EOF model …ts the data [t] shape progressively better.
(c) The EOF model shows only relatively small changes of~2:06% in the N [t ! 1] limits (Eq. [7.10]), as an estimate of uncertainty in the EIM.
(d) The Enhanced Initial Model (EIM) function captures much of the progression to a pandemic Fast Shuto¤ , as seen in the Italy data.
The [t] tail may still di¤er from these predictions, due to factors such as: (i) The CoVID-19 dynamics may change in the long-term low [t] regime; (ii) A "second wave" or multiple waves of [t] resurgence may occur, which are beyond the scope of this CoVID-19 pandemic modeling.

13
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Summary and Conclusions
The early stages of the CoVID-19 coronavirus pandemic starts o¤ with a nearly exponential rise in the number of infections with time. De…ning N [t] as the expected total number of CoVID-19 cases vs time, this basic function: 2d] We also examined if the exp( o t) exponential decay could also be subject to a Slow Shuto¤ , giving exp[ o t = (1 + o t)] instead of exp( o t), but that did not match the Italy data. To allow more data …tting parameters beyond just fK A ; o ; o g, an orthogonal function method was developed 2 : 14 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 11, 2020. ; 4b] Methods were developed to derive the fK A ; o ; o g values, and to determine the fg m ; m = (0; +M F )g and fc m ; m = (0; +M F )g constants from data. Whereas our Initial Model and EIM were M F = 0 cases, the M F = 2 case was used here to examine the Italy CoVID-19 data, as an EOF model example.
The bing.com data for Italy up to~6/15/2020 was then analyzed, with Figures 3-6 giving the new Italy results. Both the EIM and the EOF model provided good data…ts, giving similar N [t ! 1] results for the …nal number of CoVID-19 pandemic cases, di¤ering by only~2% at the 1 level.
The [t] post-peak behavior best indicates if a o 6 = 0 model (CoVID-19 pandemic Fast Shuto¤ ) is applicable. The o 6 = 0 case likely is a second Social Distancing process, that operates along with, but is independent of the gradual pandemic doubling time changes. That doubling time change gives rise to a CoVID-19 pandemic Slow Shuto¤ ( o 6 = 0), and that process still operates concurrently with the o 6 = 0 CoVID-19 pandemic Fast Shuto¤ .
This analysis shows a wide variety of CoVID-19 data can be modeled using fK A ; o ; o ; t of f set g as parameters, covering: (I) an exponential rise at CoVID-19 pandemic start; (II) a gradual lengthening of doubling times for a pandemic Slow Shuto¤ ; and (III) an exponential decay for pandemic Fast Shuto¤ s.

15
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted August 11, 2020. ; https://doi.org/10.1101/2020.08.07.20169904 doi: medRxiv preprint