• Constants (like 1, true, or "hello") are free of cost.

• Every arithmetic operation costs \(1\) unit, plus the costs associated with evaluating its subexpressions.

• Every method invocation costs \(1\) unit, plus the costs associated with evaluating the arguments, plus the cost of evaluating the method body.

• Every statement costs \(1\) unit, plus the costs associated with evaluating its subexpressions.

1. Performance of an algorithm is best expressed as a function of the size of the input.

2. Since algorithms are often recursive, the performance of an algorithm at one size often depends the performance of that algorithm at another size.

3. We need the ability to compare one function to another holistically, to express when one function “is no bigger than” or “no smaller than” another function.

4. Comparing two functions ought to behave like comparisons should: it should be reflexive, transitive and antisymmetric.

Do Now!

Figure out this cost. Justify your answer.

1. For each ConsList item, it costs \(1\) unit to have invoked length(), and within that method we have to perform one addition (which costs \(1\) unit) on a constant (\(0\) units) and the result of invoking this.rest.length(). We also need to account for the cost of running that method itself.

2. It costs \(1\) unit to have invoked length, and \(1\) more unit to return 0 in the MtList case.

• When \(n = 0\), the list is empty, and we determined above that the cost is \(1\) to have invoked the method, and \(1\) more for the return 0 statement.

• When \(n > 0\), the list is non-empty. We determined above that the cost is \(1\) to have invoked the method, \(1\) for the addition, \(0\) for the constant, plus whatever the cost of running this.rest.length() is. Above, we ignored this entirely. But now we have a better answer: the rest of this list has size n-1, and we have a function for describing the cost of running the length method — it’s \(T\) itself!

\begin{equation*}T(n) = \begin{cases} 2 &\text{when $n = 0$} \\ 2 + T(n-1) &\text{when $n > 0$} \end{cases}\end{equation*}

Do Now!

Prove that \(T(n) = 2n + 2\) is a closed-form solution to the recurrence above.

\begin{equation*}\begin{aligned} T(n) &= 2 + T(n-1) &\text{when $n > 0$} \\ &= 2 + (2 + T(n-2)) &\text{when $n-1 > 0$} \\ &= 2 + (2 + (2 + T(n-3))) &\text{when $n-2 > 0$} \\ &= \ldots \\ &= 2 + \underbrace{(2 + (2 + \ldots + 2))}_{n\text{ times}} &\text{when we reach the base case} \\ &= 2 + n * 2 = 2n+2 \end{aligned}\end{equation*}

A function \(g(x)\) is said to be an upper bound for a function \(f(x)\) if there exists some constant \(c\) such that for all “sufficiently large” values of \(x\), \(f(x)\) is less than \(c\) times \(g(x)\):

\begin{equation*}\exists c . \exists N . \forall x > N . |f(x)| \leq c|g(x)|\end{equation*}

A function \(f(x)\) is said to be a lower bound for a function \(g(x)\) if there exists some constant \(c\) such that for all “sufficiently large” values of \(x\), \(f(x)\) is less than \(c\) times \(g(x)\):

\begin{equation*}\exists c . \exists N . \forall x > N . f(x) < c g(x)\end{equation*}

• Reflexivity: For every function \(f\), it’s always the case that \(f\) is its own upper bound, i.e. \(f \in O(f)\).

• Transitivity: For all functions \(f\), \(g\) and \(h\), if \(f \in O(g)\) and \(g \in O(h)\), then \(f \in O(h)\).

• Antisymmetry: For all functions \(f\) and \(g\), if \(f \in O(g)\) and \(g \in O(f)\), then \(f\) and \(g\) are essentially “equal” up to constant factors.

Exercise

Prove these properties.

• Suppose we have a constant function \(f(x) = c\). Then we can immediately tell that \(f \in O(1)\) — \(c\) is clearly the upper bound for this function, since it doesn’t grow at all!

• Suppose we have two functions such that \(f \in O(g)\). What can we say about the asymptotic behavior of \(h(x) = k * f(x)\) for some constant \(k\)? We know that \(f \in O(g)\) means that for large enough \(x\), \(|f(x)| \leq c |g(x)|\) for some constant \(c\). With a little algebra, we can easily show that \(|h(x)| \leq (c/k)|g(x)|\) — which means that \(h \in O(g)\) also. In other words, multiplying functions by constants doesn’t affect their asymptotic behavior.

• Suppose again we have two functions such that \(f \in O(g)\). What can we say about \(h(x) = f(x) + k\)? It’s straightforward to show \(h \in O(f)\), and so by transitivity \(h \in O(g)\). In other words, adding constants to functions doesn’t affect their asymptotic behavior.

• When \(n = 0\), we’re inserting an item into an empty list, and this performs one allocation, and one statement. So \(T_{insert}(0) = 1\), and \(M_{insert}(0) = 1\).

• In the best case, the number to be inserted is less than everything else in the list. In that case, we perform one comparison, and construct one new ConsLoInt. In other words, we have

  \begin{equation*}\begin{aligned} T_{insert}^{best}(n) &= 1 \\ M_{insert}^{best}(n) &= 1 \end{aligned}\end{equation*}

• In the worst case, the number to be inserted is greater than everything else in the list, and so must be inserted at the back. In that case, we have

  \begin{equation*}\begin{aligned} T_{insert}^{worst}(n) &= 3 + T_{insert}^{worst}(n-1) \\ &= \underbrace{3 + 3 + \cdots + 3}_{n\text{ times}} + 1 \\ &= 3n + 1 \\ M_{insert}^{worst}(n) &= 1 + M_{insert}^{worst}(n-1) \\ &= \underbrace{1 + 1 + \cdots + 1}_{n\text{ times}} + 1 \\ &= n + 1 \\ \end{aligned}\end{equation*}

  for the number of operations, because we must examine every single item in the list. From the best and worst case results we can conclude that

  \begin{equation*}\begin{alignedat}{2} 1 &\leq T_{insert}(n) &&\leq 3n+1 \\ 1 &\leq M_{insert}(n) &&\leq n+1 \end{alignedat}\end{equation*}

  We can summarize our results as follows:

  Runtime for insert		Best-case		Worst-case
  \(T_{insert}\)		\(\Omega(1)\)		\(O(n)\)
  \(M_{insert}\)		\(\Omega(1)\)		\(O(n)\)

• When \(n = 0\), sort clearly takes constant time, because it does a constant number of operations, and allocates zero objects. So

  \begin{equation*}\begin{aligned} T_{sort}(n) &= 1 \\ M_{sort}(n) &= 0 \end{aligned}\end{equation*}

• In the recursive case, we perform two operations: we sort the rest of the list, and we insert an item into it. This translates neatly to the recurrence relations

  \begin{equation*}\begin{aligned} T_{sort}(n) &= T_{sort}(n-1) + T_{insert}(n-1) \\ M_{sort}(n) &= M_{sort}(n-1) + M_{insert}(n-1) \end{aligned}\end{equation*}

  In the best case, we substitute our best-case formulas for \(T_{insert}^{best}\) and \(M_{insert}^{best}\) to obtain:

  \begin{equation*}\begin{aligned} T_{sort}^{best}(n) &= T_{sort}^{best}(n-1) + T_{insert}^{best}(n-1) \\ &= T_{sort}^{best}(n-1) + 1 \\ &= \underbrace{1 + 1 + \cdots + 1}_{n\text{ times}} + 1 \\ &= n + 1 \\ M_{sort}^{best}(n) &= M_{sort}^{best}(n-1) + M_{insert}^{best}(n-1) \\ &= M_{sort}^{best}(n-1) + 1 \\ &= \underbrace{1 + 1 + \cdots + 1}_{n\text{ times}} + 1 \\ &= n + 1 \end{aligned}\end{equation*}

  In other words, in the best case, insertion sort is linear in the size of the list.
  In the worst case, we substitute our worst-case formulas for \(T_{insert}^{worst}\) and \(M_{insert}^{worst}\) to obtain:

  \begin{equation*}\begin{aligned} T_{sort}^{worst}(n) &= T_{sort}^{worst}(n-1) + T_{insert}^{worst}(n-1) \\ &= T_{sort}^{worst}(n-1) + (3(n-1) + 1) \\ M_{sort}^{worst}(n) &= M_{sort}^{worst}(n-1) + M_{insert}^{worst}(n-1) \\ &= M_{sort}^{wrost}(n-1) + ((n-1) + 1) \\ \end{aligned}\end{equation*}

  Solving these recurrences is a bit trickier, because \(T_{insert}^{worst}\) and \(M_{insert}^{worst}\) are not constant functions. If we unroll the recurrence a few times, we start to see a pattern (here illustrated only for \(T_{sort}^{worst}\), but \(M_{sort}^{worst}\) behaves the same way):

  \begin{equation*}\begin{aligned} T_{sort}^{worst}(n) &= T_{sort}^{worst}(n-1) + T_{insert}^{worst}(n-1) \\[2ex] &\qquad\text{substitute for $T_{insert}^{worst}(n-1)$} \\[2ex] &= T_{sort}^{worst}(n-1) + (3(n-1) + 1) \\[2ex] &\qquad\text{substitute for $T_{sort}^{worst}(n-1)$} \\[2ex] &= (T_{sort}^{worst}(n-2) + (3(n-2) + 1)) + (3(n-1) + 1) \\[2ex] &\qquad\text{substitute for $T_{sort}^{worst}(n-2)$} \\[2ex] &= ((T_{sort}^{worst}(n-3) + (3(n-3) + 1)) + (3(n-2) + 1)) + (3(n-1) + 1) \\[2ex] &\qquad\text{keep unrolling until we reach the base case} \\[2ex] &= T_{sort}^{worst}(0) + (3(1) + 1) + (3(2) + 1) + \cdots + (3(n-2) + 1) + (3(n-1) + 1) \\[2ex] &\qquad\text{rearrange the formula a bit} \\[2ex] &= 1 + \sum_{i=1}^{n-1}(3i + 1) \\ &= 1 + 3\sum_{i=1}^{n-1}i + \sum_{i=1}^{n-1}1 \\ &= 1 + 3(n(n-1)/2) + (n-1) \\ &= 3(n(n-1)/2) + n \\ &\in O(n^2) \end{aligned}\end{equation*}

  In other words, in the worst case, insertion sort is quadratic in the size of the list.

\begin{equation*}\begin{alignedat}{2} 1 &\leq T_{sort}(n) &&\leq 3(n(n-1)/2) + n \\ 1 &\leq M_{sort}(n) &&\leq n(n+1)/2 \end{alignedat}\end{equation*}

Do Now!

When do the best cases and worst cases happen for insertion-sort? Describe the inputs that lead to these cases.

Exercise

Repeat this analysis for selection-sort.

• The swap method performs four operations, in every situation.

  \begin{equation*}\begin{aligned} T_{swap}(n) &= 4 \end{aligned}\end{equation*}

• Analyzing findMinIndex is a bit trickier. Let’s assume that we can execute comp.compare in constant (i.e. \(O(1)\)) time — for concreteness, let’s call that time \(t_{comp}\).

  Do Now!

  What are the best and worst cases for findMinIndex?

  The body of the loop invokes comp.compare once, performs one numeric comparison, and then performs up to two more assignment statements. Based on our assumption, the loop body therefore executes in \(O(1)\) time. But how many times does it execute? Let’s define \(T_{findMinIndex}(n)\) to be the runtime of findMinIndex when the difference arr.size() - startFrom is \(n\): in other words, \(n\) here is the number of times the loop iterates. The performance of findMinIndex is the same in best or worst cases (because the loop always runs the same number of times, and runs the same loop body every time), and is

  \begin{equation*}T_{findMinIndex}(n) = n * (t_{comp} + 4) + 3\end{equation*}

• Analyzing selectionSort is simpler. The body of the loop costs \(T_{findMinIndex}(n) + T_{swap}\). The loop itself executes \(n\) times. Since the performance of \(T_{findMinIndex}\) is the same in the best and worst cases, the runtime of selectionSort is \(T_{selectionSort} \in n * O(n) = O(n^2)\).

  But wait! That analysis is a bit too simplistic: each time we call findMinIndex inside the loop, the length of the ArrayList stays the same, but the starting index increases, so later calls to findMinIndex must be cheaper than earlier ones. A more careful analysis leads us to

  \begin{equation*}\begin{aligned} T_{selectionSort} &= T_{findMinIndex}(n) + T_{findMinIndex}(n-1) + \cdots + T_{findMinIndex}(n - n) \\ &= ((3+t_{comp})(n)+2) + ((3+t_{comp})(n-1) + 2) + \cdots + ((3+t_{comp})(1) + 2) \\ &= \sum_{i=0}^{n} ((3+t_{comp})i+2) \\ &= (3+t_{comp})\sum_{i=0}^{n} i + \sum_{i=0}^{n} 2 \\ &= (3+t_{comp})(n(n+1)/2) + 2n + 2\\ &\in O(n^2) \end{aligned}\end{equation*}

  In this case, the quick analysis leads to the same answer as the more detailed analysis, but for more complicated algorithms, the more detailed analysis may lead to a better upper bound.

Do Now!

When do the best cases and worst cases happen for selection-sort? Describe the inputs that lead to these cases.