Close
Register
Close Window

Show Source |    | About   «  21.4. Introduction to Analyzing a Problem   ::   Contents   ::   21.6. Growth Rates Review  »

21.5. Bounds Review

21.5.1. Bounds Review

We define the upper bound for a problem to be the upper bound of the best algorithm we know for that problem, and the lower bound to be the tightest lower bound that we can prove over all algorithms for that problem. While we usually can recognize the upper bound for a given algorithm, finding the tightest lower bound for all possible algorithms is often difficult, especially if that lower bound is more than the "trivial" lower bound determined by measuring the amount of input that must be processed.

The benefits of being able to discover a strong lower bound are significant. In particular, when we can make the upper and lower bounds for a problem meet, this means that we truly understand our problem in a theoretical sense. It also saves us the effort of attempting to discover more (asymptotically) efficient algorithms when no such algorithm can exist.

Often the most effective way to determine the lower bound for a problem is to find a reduction to another problem whose lower bound is already known. However, this approach does not help us when we cannot find a suitable "similar problem". One thing that we will focus on is discovering and proving lower bounds from first principles. Our most significant example of a lower bounds argument so far is the sorting lower bound proof, which shows that the problem of sorting has a lower bound of \(O(n \log n)\) in the worst case.

The lower bound for the problem is the tightest (highest) lower bound that we can prove for all possible algorithms that solve the problem. [1] This can be a difficult bar, given that we cannot possibly know all algorithms for any problem, because there are theoretically an infinite number. However, we can often recognize a simple lower bound based on the amount of input that must be examined. For example, we can argue that the lower bound for any algorithm to find the maximum-valued element in an unsorted list must be \(\Omega(n)\) because any algorithm must examine all of the inputs to be sure that it actually finds the maximum value.

In the case of maximum finding, the fact that we know of a simple algorithm that runs in \(O(n)\) time, combined with the fact that any algorithm needs \(\Omega(n)\) time, is significant. Because our upper and lower bounds meet (within a constant factor), we know that we indeed have a "good" algorithm for solving the problem. It is possible that someone can develop an implementation that is a "little" faster than an existing one, by a constant factor. But we know that its not possible to develop one that is asymptotically better.

We must be careful about how we interpret this last statement, however. The world is certainly better off for the invention of Quicksort, even though Mergesort was available at the time. Quicksort is not asymptotically faster than Mergesort, yet is not merely a "tuning" of Mergesort either. Quicksort is a substantially different approach to sorting. So even when our upper and lower bounds for a problem meet, there are still benefits to be gained from a new, clever algorithm.

21.5.1.1. Analyzing Problems

Our first example for analyzing a problem was Towers of Hanoi. This had the advantage that for a given input size \(n\), there was only one input (the value \(n\)). Of course, this is not always true. For example, there are a number of ways to think about the input to the problem of finding the maximum value in an array of \(n\) records. We could consider an array of \(n\) arbitrary values, in which there are an infinite number of inputs of size \(n\). It might be hard for us to understand the analysis of some problem if we cannot even enumerate all of the possible inputs. So we might prefer a simpler model that we believe will not change the underlying behavor. For example, we could assume that the input is some permutation of the values from 1 to \(n\). While we probably do not want to restrict our find algorithm to such input, if all we care about for analysis purposes is the position of the biggest value within the array, then a permutation of the values from 1 to \(n\) might make it easier for us to think through all of the possibilities.

Another complicating factor that might or might not arise is that different inputs of a given size might have different costs. For example, we probably realize that to find the maximum value in the array, we need to look at all values of the array. So, the order of the values does not really affect the time required. However, consider if our problem is to find where in the array the record with value \(X\) appears, if any such record exists. Now, not only are there many possible inputs of size \(n\), but those inputs have different costs when using, for example, a simple sequential search from the start of the array. This is why we need the concepts of best, average, and worst case inputs.

To make things even worse, the cost to solve the problem for a given input depends on the algorithm that we use! For example, which input of size \(n\) is the worst is different for the algorithm that moves sequentially through the array from the start to the end, as opposed to the algorithm that moves sequentially through the array from the end to the start.

The worst-case cost (for all input of size $n$) is the maximum cost for the algorithm over all problem instances of size \(n\). The best-case cost (for all input of size $n$) is the minimum cost for the algorithm over all problem instances of size \(n\). It is possible that the {best, worst} case cost changes radically with \(n\). That is, even \(n\) might have a very different cost from odd \(n\).

We will use the following notation at varous times this semester. \(\mathcal{A}\) is an algorithm. \(I_n\) is the set of all possible inputs to \(\mathcal{A}\) of size \(n\). \(I\) is an input in \(I_n\). \(f_\mathcal{A}\) is a function that expresses the resource cost of algorithm \(\mathcal{A}\). Using this notation, we can define the worst and best case costs as:

\[\mbox{worst cost}(\mathcal{A}) = \max_{I \in I_n} f_{\mathcal{A}}(I).\]\[\mbox{best cost}(\mathcal{A}) = \min_{I \in I_n} f_{\mathcal{A}}(I).\]

This point that we are considering all of the inputs of size \(n\) is crucial. In other words, we don't pick the \(n\) for which the best (or worst) case occurs. So it would be wrong to say something like "The best case is when \(n=1\)."

If we want the average-case cost, it is even more complicated. We might model this as half way between the best and worst case costs, but this is not often correct. (Think about what circumstances would make it correct, and some situations where it would not be correct.) To account for the true average cost for input of size \(n\), we have to consider the entire collection of such inputs. For each one of these, we need its relative frequency, and its cost. Frequencies for inputs can be hard to determine! For example, the average cost of sequential search is \((n+1)/2\), but only if every position of the array is equally likely to hold the value that we are looking for. And what do we do about the situation where the value is not even in the array?

However, ideally we have all the information that we need to calcluate the average case cost. Then we can calculate the weighted average:

\[\frac{\sum_{I\in I_n} \mathrm{freq}(I) * \mathrm{cost}(I)}{\mathrm{total\ count\ of\ frequencies}}\]

Think about this: Can the average cost be worse than the worst cost? Or better than the best cost?

So now we are ready to give a more precise definition for the lower bound of a problem. As always, we have to define it for some class of inputs. We also have to consider that there are many (infinitely many in theory) algorithms that solve the problem. Recall that to analyze any problem, we have to define a model that includes the definition for problem size and the definition for solution cost. Call such a model \(\mathcal{M}\). Then, \(\mathcal{A_M}\) is the set of all algorithms that solve the problem under model \(\mathcal{M}\). Then, the lower bound of a problem in the worst case is:

\[\min_{{\mathcal A} \in {\mathcal A}_M} \left\{ \max_{I \in I_n} f_{\mathcal A}(I)\right\}\]

21.5.1.2. Modeling the Inputs

Especially when trying to come to grips with what the average case cost of the algorithm will be, it might be easier to think about what is going on if we simplify the model that we use for the class of inputs that we are considering.

Think about the seemingly simple problem of finding the value \(X\) in an (unsorted) array of \(n\) records. What are the inputs to this problem? Of course, an array of \(n\) records! But what does that mean if we try to enumerate all of the inputs of size \(n\)? How many such inputs are there?

Well, if each position in the array can take any value, then there are an infinite number of values for each position. Even if we restrict these values to something like a 64-bit integer, it is still a lot of possibilities to consider!

Given the cognative load involved in thinking about all of those inputs, we might want to consider instead analyzing a simpler set of inputs. For example, we might decide to only consider (for analysis purposes) that the input is a permuation of the numbers 1 to \(n\). The argument here might be that we don't care about the actual values in the array. We only care about whether a given value is \(X\) or not, and so we can simplify the inputs that we consider.

There are two dangers that we have to be aware of when doing such a simplification. First, our simplification has to still reflect reality. If we simplify an array of \(n\) numbers as a permutation of the numbers 1 to \(n\), then we eliminate the inputs that have duplicates. That might make for a wrong analysis. Second, we have to separate the issue of inputs for the purpose of analysis from inputs for the purpose of solving a problem. In the case of sorting, we might want to analyze behavior on a collection of \(n\) records each with a unique key value. Since we don't care about the actual key values, we might simplify this to some permuation of a set of records with key values 1 to \(n\). But, sorting a collection of records whose keys are known to be a permuation of the values 1 to \(n\) is much simpler than sorting a collection of \(n\) arbitrary records! For example, we can sort the permutation in linear time with a simple Binsort.

Going back to the example of finding value \(X\) in an array of \(n\) records, we might want to consider a model that only considers the position of the first occcurance of \(X\) in the array. In other words, we lump all inputs whose first occurance of \(X\) is in the first position into one input. All inputs whose first occurance of \(X\) is in the second position is another input. And so on. Then we can analyze the cost only for those "groups" of inputs that we care about. Of course, we might have difficulty deciding what the proper frequencies are for each of these synthetic input groups. Perhaps it is reasonable to say that each position in the array has equal probability of holding the first occurrance of \(X\). Or perhaps it is not.

[1]Throughout this discussion, it should be understood that any mention of bounds must specify what class of inputs are being considered. Do we mean the bound for the worst case input? The average cost over all inputs? Regardless of which class of inputs we consider, all of the issues raised apply equally.

   «  21.4. Introduction to Analyzing a Problem   ::   Contents   ::   21.6. Growth Rates Review  »

nsf
Close Window